Upload 259 files

.gitignore

@@ -0,0 +1,145 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+.vscode
+.vscode/*
+.DS_Store
+
+output/*
+work_dirs
+work_dirs/*
+work_dirs/
+
+data/temp/*
+slurm_tools/
+slurm_run.sh
+
+core*
+dist_url_*

LICENSE

@@ -0,0 +1,224 @@
+Copyright (c) 2022 - present, SenseTime. All Rights Reserved.
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2022 - present, SenseTime
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+
+
+X-modaler
+
+Copyright 2021 Jingdong Technology Information Technology Co., Ltd
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+  http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+
+
+
+

README.md

@@ -1,3 +1,201 @@
----
-license: apache-2.0
----
+# Uni-Perceiver
+
+This repository contains training (pre-training, fine-tuning, prompt-tuning), evaluation code and pretrained models for the following papers:
+
+
+> [Uni-Perceiver](https://arxiv.org/abs/2112.01522): Pre-training Unified Architecture for Generic Perception for Zero-shot and Few-shot Tasks, CVPR 2022.
+
+> [Uni-Perceiver-MoE](https://arxiv.org/abs/2206.04674): Learning Sparse Generalist Models with Conditional MoEs, NeurIPS 2022.
+
+
+
+## Introduction
+
+__Uni-Perceiver__ is a generalist model (generic perception model) that can process a variety of modalities and tasks with unified modeling and shared
+parameters. Different perception tasks are modeled as the same formulation, that is, finding the maximum likelihood target for each input through the similarity of their representations. Meanwhile,  Uni-Perceiver is pre-trained on several uni-modal and multi-modal tasks, and evaluated on a variety of downstream tasks, including novel tasks that did not appear in the pre-training stage.
+Thanks to the unified formulation, it shows the ability of zero-shot inference on novel tasks, and shows promising performance close to or on par with SOTA results by prompt tuning or finetuning.
+
+![UnPerceiver-intro](./figs/overview.png)
+
+In __Uni-Perceiver-MoE__, we found that the interference among different tasks and modalities can lead to performance degradation of generalist models on some tasks compared with task-specialized models. We introduce the Conditional Mixture-of-Experts (Conditional MoEs) to mitigate such interference. By incorporating the proposed Conditional MoEs, Uni-Perceiver-MoE can effectively mitigate the interference across tasks and modalities, and achieves state-of-the-art results on a series of downstream tasks via prompt tuning on 1% of downstream data. Moreover, the introduction of Conditional MoEs still holds the generalization ability of generalist models to conduct zero-shot inference on new tasks,
+
+![UnPerceiver-moe-intro](./figs/overview_moe.png)
+
+
+## Main Results and Pretrained Models
+
+### Base Models
+
+<table border="1" width="100%">
+  <tr align="center">
+    <th>Task</th>
+    <th>Image Classification</th>
+    <th colspan="2">Image Caption</th>
+    <th colspan="4">Image Retrieval</th>
+    <th>Video Classification</th><th>Video Caption</th><th colspan="2">Video Retrieval</th>
+  </tr>
+  <tr align="center">
+    <td>Dataset</td><td>ImageNet-1k</td><td>MSCOCO</td><td>Flickr30k</td><td colspan="2">MSCOCO</td><td colspan="2">Flickr30k</td><td>Kinetics-400</td><td>MSVD</td><td colspan="2">MSVD</td>
+  </tr>
+  <tr align="center">
+    <td>Split</td><td>ILSVRC 2012 val</td><td>Karpathy test</td><td>test</td><td colspan="2">Karpathy test</td><td colspan="2">test</td><td>test-dev</td><td>val</td><td>val</td><td colspan="2">val</td>
+  </tr>
+  <tr align="center">
+    <td>Metric</td><td>Acc@1</td><td>BLEU-4</td><td>BLEU-4</td><td>R@1 i2t</td><td>R@1 t2i</td><td>R@1 i2t</td><td>R@1 t2i</td><td>Acc@1</td><td>BLEU-4</td><td>R@1 v2t</td><td>R@1 t2v</td>
+  </tr>
+  </tr>
+  <tr align="center">
+    <td>Uni-Perceiver<sub>BASE</sub> w/o Tuning</td><td>79.2 </td><td>32.0</td><td>14.7 </td><td>64.9 </td><td>50.7 </td><td>82.3 </td><td>71.1</td> <td>74.5 </td><td>22.6 </td><td>50.3</td><td>38.7 </td>
+  </tr>
+  <tr align="center">
+    <td>Uni-Perceiver<sub>BASE</sub> PT (1%)</td><td>80.9 </td><td>35.5</td><td>30.2</td><td>68.4 </td><td>51.9 </td><td>91.0 </td><td>76.0 </td><td>74.8 </td><td>59.5 </td><td>62.7 </td><td>43.8 </td>
+  </tr>
+   <tr align="center">
+    <td>Uni-Perceiver<sub>BASE</sub> FT (100%)</td><td>84.0</td><td>36.4 </td><td>31.2 </td><td>69.8</td><td>53.9 </td><td>92.7</td><td>77.5</td><td>77.7 </td><td>63.3 </td><td>62.8</td><td>45.8 </td>
+  </tr>
+
+  <tr align="center">
+    <td>Uni-Perceiver-MoE<sub>BASE</sub> w/o Tuning</td><td>80.3 </td><td>33.2</td><td>15.9 </td><td>64.6 </td><td>51.6 </td><td>82.1 </td><td>75.8</td> <td>76.8 </td><td>23.4 </td><td>52.8</td><td>40.0 </td>
+  </tr>
+  <tr align="center">
+    <td>Uni-Perceiver-MoE<sub>BASE</sub> PT (1%)</td><td>82.0 </td><td>36.8</td><td>30.7</td><td>68.9 </td><td>52.6 </td><td>91.3 </td><td>78.5 </td><td>77.2 </td><td>60.0 </td><td>65.6 </td><td>45.3 </td>
+  </tr>
+   <tr align="center">
+    <td>Uni-Perceiver-MoE<sub>BASE</sub> FT (100%)</td><td>84.5</td><td>37.3 </td><td>32.4 </td><td>70.5</td><td>54.1 </td><td>93.6</td><td>79.8</td><td>79.3 </td><td>65.4 </td><td>65.0</td><td>47.8 </td>
+  </tr>
+</table>
+
+
+### Large Models
+
+<table border="1" width="100%">
+  <tr align="center">
+    <th>Task</th>
+    <th>Image Classification</th>
+    <th colspan="2">Image Caption</th>
+    <th colspan="4">Image Retrieval</th>
+    <th>Video Classification</th><th>Video Caption</th><th colspan="2">Video Retrieval</th>
+  </tr>
+  <tr align="center">
+    <td>Dataset</td><td>ImageNet-1k</td><td>MSCOCO</td><td>Flickr30k</td><td colspan="2">MSCOCO</td><td colspan="2">Flickr30k</td><td>Kinetics-400</td><td>MSVD</td><td colspan="2">MSVD</td>
+  </tr>
+  <tr align="center">
+    <td>Split</td><td>ILSVRC 2012 val</td><td>Karpathy test</td><td>test</td><td colspan="2">Karpathy test</td><td colspan="2">test</td><td>test-dev</td><td>val</td><td>val</td><td colspan="2">val</td>
+  </tr>
+  <tr align="center">
+    <td>Metric</td><td>Acc@1</td><td>BLEU-4</td><td>BLEU-4</td><td>R@1 i2t</td><td>R@1 t2i</td><td>R@1 i2t</td><td>R@1 t2i</td><td>Acc@1</td><td>BLEU-4</td><td>R@1 v2t</td><td>R@1 t2v</td>
+  </tr>
+  <tr align="center">
+    <td>Uni-Perceiver<sub>LARGE</sub> w/o Tuning</td><td>82.7 </td><td> 35.3 </td><td> 15.1 </td><td>67.8 </td><td>54.1 </td><td> 83.7</td><td> 74.2 </td><td> 79.5</td><td>24.7 </td><td> 45.4 </td><td>34.2 </td>
+  </tr>
+  <tr align="center">
+    <td>Uni-Perceiver<sub>LARGE</sub> PT (1%)</td><td>84.2 </td><td>38.6 </td><td> 32.9</td><td> 73.3 </td><td>56.2 </td><td>92.1 </td><td> 80.0</td><td> 80.0</td><td> 67.2</td><td> 65.5 </td><td>48.6 </td>
+  </tr>
+   <tr align="center">
+    <td>Uni-Perceiver<sub>LARGE</sub> FT (100%)</td><td>86.2 </td><td> 39.2 </td><td> 35.5 </td><td>74.4 </td><td>57.9 </td><td>94.7 </td><td> 82.1</td><td>81.9 </td><td>68.3 </td><td> 65.2 </td><td>50.8 </td>
+  </tr>
+
+  <tr align="center">
+    <td>Uni-Perceiver-MoE<sub>LARGE</sub> w/o Tuning</td><td>83.4 </td><td> 35.5 </td><td> 15.8 </td><td>67.9 </td><td>55.3 </td><td> 83.6</td><td> 75.9 </td><td> 82.1</td><td>24.6 </td><td> 45.7 </td><td>41.9 </td>
+  </tr>
+  <tr align="center">
+    <td>Uni-Perceiver-MoE<sub>LARGE</sub> PT (1%)</td><td>84.9 </td><td>39.3 </td><td> 33.7</td><td> 73.3 </td><td>57.1 </td><td>92.4 </td><td> 80.6</td><td> 83.0</td><td> 67.6</td><td> 66.4 </td><td>50.3 </td>
+  </tr>
+   <tr align="center">
+    <td>Uni-Perceiver-MoE<sub>LARGE</sub> FT (100%)</td><td>86.4 </td><td> 40.5 </td><td> 36.2 </td><td>74.7 </td><td>58.3 </td><td>94.1 </td><td> 83.7</td><td>84.2 </td><td>68.9 </td><td> 67.6 </td><td>52.3 </td>
+  </tr>
+</table>
+
+  * The numbers are slightly better than the original paper of Uni-Perceiver, which are from the reproduced version of Uni-Perceiver used as the baseline of [Uni-Perceiver-MoE](https://arxiv.org/abs/2206.04674).
+  * The image resolution for all tasks is `224x224`.
+  * See [OtherResults.md](data/other_results.md) for results on more tasks and datasets.
+
+
+
+## Usage
+### Requirements
+* Linux, CUDA>=10.1, GCC>=5.4
+
+* Python >=3.7
+
+* pytorch >= 1.8.0 
+
+* JAVA >= 1.8 (for caption task evaluation)
+
+
+### Installation
+```bash
+git clone https://github.com/fundamentalvision/Uni-Perceiver
+cd Uni-Perceiver
+pip install -r requirements.txt
+```
+
+
+### Data
+See [prepare_data.md](data/prepare_data.md).
+
+### Pre-trained Model Weights
+See [checkpoints.md](data/checkpoints.md).
+
+
+### Pre-training 
+See [pretraining.md](data/pretraining.md).
+
+### Fine-tuning 
+See [finetuning.md](data/finetuning.md).
+
+###  Prompt-tuning
+
+See [prompt_tuning.md](data/prompt_tuning.md).
+
+
+###  Inference
+
+See [inference.md](data/inference.md).
+
+### TODO 
+
+* release more pretrained models
+  - [ ]  Uni-Perceiver Tiny model
+  - [ ]  Uni-Perceiver Small model
+  - [ ]  Uni-Perceiver Huge model
+
+* support more datasets and tasks
+
+
+
+## License
+Uni-Perceiver is licensed under the [Apache-2.0 License](./LICENSE).
+
+
+<br></br>
+
+## Citing Uni-Perceiver
+If you find Uni-Perceiver useful in your research, please consider giving a star ⭐ and citing:
+```bibtex
+ @article{zhu2021uni,
+  title={Uni-Perceiver: Pre-training Unified Architecture for Generic Perception for Zero-shot and Few-shot Tasks},
+  author={Zhu, Xizhou and Zhu, Jinguo and Li, Hao and Wu, Xiaoshi and Wang, Xiaogang and Li, Hongsheng and Wang, Xiaohua and Dai, Jifeng},
+  journal={arXiv preprint arXiv:2112.01522},
+  year={2021}
+
+}
+```
+
+```bibtex
+@article{zhu2022uni,
+  title={Uni-Perceiver-MoE: Learning Sparse Generalist Models with Conditional MoEs},
+  author={Zhu, Jinguo and Zhu, Xizhou and Wang, Wenhai and Wang, Xiaohua and Li, Hongsheng and Wang, Xiaogang and Dai, Jifeng},
+  journal={arXiv preprint arXiv:2206.04674},
+  year={2022}
+}
+```
+
+### Acknowledgements
+Many thanks to following codes that help us a lot in building this codebase:
+* [Detectron2](https://github.com/facebookresearch/detectron2)
+* [X-modaler](https://github.com/YehLi/xmodaler)
+* [deit](https://github.com/facebookresearch/deit)
+* [VL-BERT](https://github.com/jackroos/VL-BERT)
+* [TimeSformer](https://github.com/facebookresearch/TimeSformer)
+* [CLIP](https://github.com/openai/CLIP)

configs/BERT_L12_H192_experiments/4tasks_training.yaml

@@ -0,0 +1,729 @@
+_BASE_: "base_model_bert_l12_h192.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet1k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+  
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+
+
+TASKS:
+
+  - 
+    NAME: imagenet
+    DATASETS:
+      TRAIN: 'ImageNetDataset'
+      # VAL: 'ImageNetDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet1k'
+      TARGET_SET: ['ImageNet1k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 4
+      # TEST_BATCH_SIZE: 2
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/imagenet'
+      S3_PATH: 'cluster2:s3://imagenet'
+      ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
+      SAMPLING_WEIGHT: 1.0
+      CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
+      MIXUP: 0.8
+      CUTMIX: 1.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 1000
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      # LOSS_FP32: True
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 1.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  -
+    NAME: yfcc_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'YFCC'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: True
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC12M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC3M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: sbu_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'SBU'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: vg_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'VG'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: True
+
+  -
+    NAME: mscoco_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'MSCOCO'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 100
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      NAME: 'RetrievalEvaler'
+      VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
+      TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
+      GENERATION_MODE: False
+
+  -
+    NAME: yfcc_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'YFCC'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: True
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC12M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC3M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+
+  -
+    NAME: vg_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'VG'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: sbu_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'SBU'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: flickr30k_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      TEST: 'ImageTextPairDataset'
+      DATASET_NAME: 'FLICKR'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 128
+      TEST_BATCH_SIZE: 128
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/flickr30k_images/flickr30k_images/flickr30k_images'
+      ANNO_FOLDER:  'open_source_dataset/flickr30k'
+      S3_PATH: "s3://open_dataset/flickr30k/flickr30k_images"
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False 
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 77
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      NAME: 'RetrievalEvaler'
+      VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
+      TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
+      GENERATION_MODE: False
+
+  -
+    NAME: flickr30k_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      TEST: 'ImageTextPairDataset'
+      DATASET_NAME: 'FLICKR'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 32 
+      TEST_BATCH_SIZE: 8
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/flickr30k_images/flickr30k_images/flickr30k_images'
+      ANNO_FOLDER:  'open_source_dataset/flickr30k'
+      S3_PATH: "s3://open_dataset/flickr30k/flickr30k_images"
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False 
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+      TASK_TYPE: caption
+      # DATA_PERCENTAGE: 0.01
+    MODEL:
+      MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/flickr30k/captions_val.json'
+      TEST_ANNFILE: 'open_source_dataset/flickr30k/captions_test.json'
+      GENERATION_MODE: True
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  IN_TUNING: True # use IN1k instead of 22k
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    NUM_HIDDEN_LAYERS: 1
+    DROP_PATH_PROB_FIXED: True
+
+    UNIFY_QKV: True
+  
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 150000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 500000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 5000
+  MIN_LR: 0.000001
+
+
+
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+

configs/BERT_L12_H192_experiments/4tasks_training_small_datasets.yaml

@@ -0,0 +1,292 @@
+_BASE_: "base_model_bert_l12_h192.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet1k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'small_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'small_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+TASKS:
+
+  - 
+    NAME: imagenet
+    DATASETS:
+      TRAIN: 'ImageNetDataset'
+      VAL: 'ImageNetDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet1k'
+      TARGET_SET: ['ImageNet1k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 4
+      # TEST_BATCH_SIZE: 2
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'small_source_dataset/imagenet'
+      ANNO_FOLDER:  'small_source_dataset/imagenet/meta'
+      SAMPLING_WEIGHT: 1.0
+      MIXUP: 0.8
+      CUTMIX: 1.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 1000
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      # LOSS_FP32: True
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'small_source_dataset/imagenet/meta/val.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+
+  -
+    NAME: bookswiki_pretrain
+    DATASETS:
+      TRAIN: 'GeneralCorpusDataset'
+      TASK_TYPE: 'text_mlm'
+      DATASET_NAME: 'BooksWiki'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 128
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  'small_source_dataset/bert_pretrain_data/bookswiki'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      SEQ_PER_SAMPLE: 128
+      MIN_SEQ_PER_SAMPLE: 128
+      APPEND_EOS: True
+      ONE_STREAM: False
+      SAMPLING_WEIGHT: 1.0
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 128
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'small_source_dataset/mscoco_caption/coco_origin'
+      ANNO_FOLDER:  'small_source_dataset/mscoco_caption/annotations'
+      SEQ_PER_SAMPLE:  1
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 1.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'small_source_dataset/mscoco_caption/annotations/captions_val5k.json'
+      TEST_ANNFILE: 'small_source_dataset/mscoco_caption/annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  -
+    NAME: mscoco_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'MSCOCO'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 100
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'small_source_dataset/mscoco_caption/coco_origin'
+      ANNO_FOLDER:  'small_source_dataset/mscoco_caption/annotations'
+      SEQ_PER_SAMPLE:  1
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      NAME: 'RetrievalEvaler'
+      GENERATION_MODE: False
+
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  IN_TUNING: True # use IN1k instead of 22k
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.0
+    NUM_HIDDEN_LAYERS: 1
+    DROP_PATH_PROB_FIXED: True
+
+    UNIFY_QKV: True
+  
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 150000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 500000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 5000
+  MIN_LR: 0.000001
+
+
+
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+

configs/BERT_L12_H192_experiments/7tasks_berttiny_training.yaml

@@ -0,0 +1,416 @@
+_BASE_: "base_model_bert_l12_h192.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet1k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+  
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+
+
+TASKS:
+
+  - 
+    NAME: imagenet
+    DATASETS:
+      TRAIN: 'ImageNetDataset'
+      VAL: 'ImageNetDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet1k'
+      TARGET_SET: ['ImageNet1k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 720
+      TEST_BATCH_SIZE: 256
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/imagenet'
+      S3_PATH: 'cluster2:s3://imagenet'
+      ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
+      SAMPLING_WEIGHT: 2.5
+      CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
+      MIXUP: 0.8
+      CUTMIX: 1.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 1000
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      # LOSS_FP32: True
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+  
+  -
+    NAME: bookswiki_pretrain
+    DATASETS:
+      TRAIN: 'GeneralCorpusDataset'
+      TASK_TYPE: 'text_mlm'
+      DATASET_NAME: 'BooksWiki'
+      TARGET_SET: ['Vocab_Word']
+      VERSION: 'v2'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  'open_source_dataset/text_corpus' # 'open_source_dataset/bert_pretrain_data/bookswiki'
+      # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      SEQ_PER_SAMPLE: 128
+      MIN_SEQ_PER_SAMPLE: 128
+      APPEND_EOS: True
+      ONE_STREAM: False
+      SAMPLING_WEIGHT: 3.5
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 128
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+ ########## Image Captioning ########### 
+
+
+  -
+    NAME: cc12m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC12M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER: 'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.6889
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC3M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  's3://cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.8780
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: vg_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'VG'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5895
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: True
+  
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.3817
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 1.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  -
+    NAME: sbu_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'SBU'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      ANNO_FOLDER:  'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.4618
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  IN_TUNING: True # use IN1k instead of 22k
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.0
+    DROP_PATH_PROB_FIXED: True
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 32
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 150000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 500000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 5000
+  MIN_LR: 0.000001
+
+
+
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+

configs/BERT_L12_H192_experiments/7tasks_berttiny_training_apex_o2.yaml

@@ -0,0 +1,9 @@
+_BASE_: "7tasks_berttiny_training.yaml"
+
+####################################### Optimizer #######################################
+SOLVER:
+  
+  AMP_FP16: False
+  APEX_FP16: True # dangerous
+  APEX_OPT_LEVEL: 'O2'
+  CHECKPOINT_PERIOD: 100000

configs/BERT_L12_H192_experiments/7tasks_berttiny_training_lamb.yaml

@@ -0,0 +1,418 @@
+_BASE_: "base_model_bert_l12_h192.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet1k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+  
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+
+
+TASKS:
+
+  - 
+    NAME: imagenet
+    DATASETS:
+      TRAIN: 'ImageNetDataset'
+      VAL: 'ImageNetDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet1k'
+      TARGET_SET: ['ImageNet1k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 720
+      TEST_BATCH_SIZE: 256
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/imagenet'
+      S3_PATH: 'cluster2:s3://imagenet'
+      ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
+      SAMPLING_WEIGHT: 2.5
+      CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
+      MIXUP: 0.8
+      CUTMIX: 1.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 1000
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      # LOSS_FP32: True
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+  
+  -
+    NAME: bookswiki_pretrain
+    DATASETS:
+      TRAIN: 'GeneralCorpusDataset'
+      TASK_TYPE: 'text_mlm'
+      DATASET_NAME: 'BooksWiki'
+      TARGET_SET: ['Vocab_Word']
+      VERSION: 'v2'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  'open_source_dataset/text_corpus' # 'open_source_dataset/bert_pretrain_data/bookswiki'
+      # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      SEQ_PER_SAMPLE: 128
+      MIN_SEQ_PER_SAMPLE: 128
+      APPEND_EOS: True
+      ONE_STREAM: False
+      SAMPLING_WEIGHT: 3.5
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 128
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+ ########## Image Captioning ########### 
+
+
+  -
+    NAME: cc12m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC12M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.6889
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC3M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  's3://cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.8780
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: vg_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'VG'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5895
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: True
+  
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.3817
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 1.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  -
+    NAME: sbu_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'SBU'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.4618
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  IN_TUNING: True # use IN1k instead of 22k
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.0
+    DROP_PATH_PROB_FIXED: True
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 32
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'LAMB'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 150000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 500000
+  BASE_LR: 0.01
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 5000
+  MIN_LR: 0.000001
+
+
+
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+

configs/BERT_L12_H192_experiments/7tasks_berttiny_training_moe.yaml

@@ -0,0 +1,25 @@
+_BASE_: "7tasks_berttiny_training.yaml"
+
+MOE: 
+  MOE: True 
+  MOE_TYPE: 'attribute'
+  TAG_Transform: True
+  ATTRIBUTE_LENGTH: 8
+  EP_WORLD_SIZE: 1 # tag moe only 
+  NUM_EXPERTS: 8
+  TOP_K: 2
+  CAPACITY_FACTOR: 3.0 
+  EVAL_MIN_CAPACITY: 4.0
+  MIN_CAPACITY: 4
+  NOISY_GATE_POLICY: 'vmoe'
+  MOE_PARAM_GROUP: True 
+  MOE_EXPERT_TYPE: 'FFN,SA'
+  SA_LINEAR_OUT_MOE: True
+  MOE_EXPERT_LOCATION: 'all' # 'odd'
+  # MOE_LAYER_START_IDX: 3
+  # MOE_LAYER_END_IDX: 21
+  # MOE_LAYER_START_IDX: 18
+  # MOE_LAYER_END_IDX: 12 
+  BATCH_PRIO: True 
+  USE_TUTEL: True
+  FFN_SHARE_GATE_DECISION: True

configs/BERT_L12_H192_experiments/7tasks_berttiny_training_moe_lsfp32_gate_softmax_layernorm_fp16.yaml

@@ -0,0 +1,42 @@
+_BASE_: "7tasks_berttiny_training.yaml"
+
+MOE: 
+  MOE: True 
+  MOE_TYPE: 'attribute'
+  TAG_Transform: True
+  ATTRIBUTE_LENGTH: 8
+  EP_WORLD_SIZE: 1 # tag moe only 
+  NUM_EXPERTS: 8
+  TOP_K: 2
+  CAPACITY_FACTOR: 3.0 
+  EVAL_MIN_CAPACITY: 4.0
+  MIN_CAPACITY: 4
+  NOISY_GATE_POLICY: 'vmoe'
+  MOE_PARAM_GROUP: True 
+  MOE_EXPERT_TYPE: 'FFN,SA'
+  SA_LINEAR_OUT_MOE: True
+  MOE_EXPERT_LOCATION: 'all' # 'odd'
+  # MOE_LAYER_START_IDX: 3
+  # MOE_LAYER_END_IDX: 21
+  # MOE_LAYER_START_IDX: 18
+  # MOE_LAYER_END_IDX: 12 
+  BATCH_PRIO: True 
+  USE_TUTEL: True
+  FFN_SHARE_GATE_DECISION: True
+
+MODEL:
+  LAYER_SCALE_FP32: True
+  GATE_FP32: False
+  TAG_TRANSFORM_FP32: False
+
+
+SOLVER:
+ 
+
+  FORCE_SOFTMAX_FP16: True
+  FORCE_LN_FP16: True
+  FORCE_NORM_FP16: True
+  # FORCE_TEMP_FP16: True
+  FORCE_EMBED_FP16: True
+
+  # FORCE_EXPERT_ADDING_FP16: True

configs/BERT_L12_H192_experiments/7tasks_berttiny_training_moe_scale_before.yaml

@@ -0,0 +1,444 @@
+_BASE_: "base_model_bert_l12_h192.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet1k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+  
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+
+
+TASKS:
+
+  - 
+    NAME: imagenet
+    DATASETS:
+      TRAIN: 'ImageNetDataset'
+      VAL: 'ImageNetDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet1k'
+      TARGET_SET: ['ImageNet1k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 720
+      # TEST_BATCH_SIZE: 2
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/imagenet'
+      S3_PATH: 'cluster2:s3://imagenet'
+      ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
+      SAMPLING_WEIGHT: 2.5
+      CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
+      MIXUP: 0.8
+      CUTMIX: 1.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 1000
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      # LOSS_FP32: True
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+  
+  -
+    NAME: bookswiki_pretrain
+    DATASETS:
+      TRAIN: 'GeneralCorpusDataset'
+      TASK_TYPE: 'text_mlm'
+      DATASET_NAME: 'BooksWiki'
+      TARGET_SET: ['Vocab_Word']
+      VERSION: 'v2'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  'open_source_dataset/text_corpus' # 'open_source_dataset/bert_pretrain_data/bookswiki'
+      # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      SEQ_PER_SAMPLE: 128
+      MIN_SEQ_PER_SAMPLE: 128
+      APPEND_EOS: True
+      ONE_STREAM: False
+      SAMPLING_WEIGHT: 3.5
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 128
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+ ########## Image Captioning ########### 
+
+
+  -
+    NAME: cc12m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC12M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.6889
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC3M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  's3://cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.8780
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: vg_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'VG'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5895
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: True
+  
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.3817
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 1.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  -
+    NAME: sbu_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'SBU'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.4618
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  IN_TUNING: True # use IN1k instead of 22k
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.0
+    DROP_PATH_PROB_FIXED: True
+    SCALE_MULTI_BEFORE: True 
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 32
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 150000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 500000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 5000
+  MIN_LR: 0.000001
+
+
+
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+
+
+
+MOE: 
+  MOE: True 
+  MOE_TYPE: 'attribute'
+  TAG_Transform: True
+  ATTRIBUTE_LENGTH: 8
+  EP_WORLD_SIZE: 1 # tag moe only 
+  NUM_EXPERTS: 8
+  TOP_K: 2
+  CAPACITY_FACTOR: 3.0 
+  EVAL_MIN_CAPACITY: 4.0
+  MIN_CAPACITY: 4
+  NOISY_GATE_POLICY: 'vmoe'
+  MOE_PARAM_GROUP: True 
+  MOE_EXPERT_TYPE: 'FFN,SA'
+  SA_LINEAR_OUT_MOE: True
+  MOE_EXPERT_LOCATION: 'all' # 'odd'
+  # MOE_LAYER_START_IDX: 3
+  # MOE_LAYER_END_IDX: 21
+  # MOE_LAYER_START_IDX: 18
+  # MOE_LAYER_END_IDX: 12 
+  BATCH_PRIO: True 
+  USE_TUTEL: True
+  FFN_SHARE_GATE_DECISION: True

configs/BERT_L12_H192_experiments/base_model_bert_l12_h192.yaml

@@ -0,0 +1,73 @@
+
+######################################### MODEL #########################################
+MODEL:
+  VOCAB_SIZE: 49411 # include <BOS>/<EOS>
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder_v3'
+  ENCODER_DIM: 192
+  DECODER: ''
+  DECODER_DIM: 192
+
+  PREDICTOR: 'EmbedClsAsRetrievalPredictor'
+  FEATURE_GATHER: True
+  LEARN_TEMP: True
+  PRED_USE_NORM: True
+  PRED_TEMPERATURE: 0.07
+
+  BertParamsInit: True
+
+  CLS_TOKEN: False
+
+  QUEUE_LEN: 1024
+  MAX_LABEL_LEN: 12
+
+  OUTPUT_PROJ: True # output projection
+ 
+
+# #################################### Token embedding ####################################
+  TOKEN_EMBED:
+    NAME: 'TokenBaseEmbedding'
+    DIM: 192
+    ACTIVATION: 'none'
+    USE_NORM: True
+    DROPOUT: 0.0
+    POSITION: 'NNEmbeddingEncoding'
+    POSITION_MAX_LEN: 512
+    TYPE_VOCAB_SIZE: 2
+
+# #################################### Visual embedding ####################################
+  VISUAL_EMBED:
+    NAME: 'none'
+
+# #################################### video embedding ####################################
+  VIDEO_EMBED:
+    NAME: 'VideoBaseEmbedding'
+    IN_DIM: 768
+    OUT_DIM: 192
+    ACTIVATION: 'none'
+    USE_NORM: True
+    DROPOUT: 0.0
+    TYPE_SIZE: 1 # video to encoder
+    POSITION: 'NNEmbeddingEncoding'
+    MAX_LENGTH: 1600
+    PATCH_SIZE_S: 16
+    PATCH_SIZE_T: 1
+    DIVIDE_ST_POS: True
+    USE_VISUAL_TOKENIZER: True
+    USE_VISUAL_POS: True
+    MAX_FRAMES: 8
+
+####################################### BERT ############################################
+  BERT:
+    DROP_PATH_PROB: 0.0
+    HIDDEN_SIZE: 192
+    HIDDEN_DROPOUT_PROB: 0.
+    HIDDEN_ACT: "gelu"
+    NUM_ATTENTION_HEADS: 3
+    INTERMEDIATE_SIZE: 768
+    INTERMEDIATE_DROP: 0.
+    FFN_DROPOUT_PROB: 0.
+    ATTENTION_PROBS_DROPOUT_PROB: 0.
+    NUM_HIDDEN_LAYERS: 12
+    NUM_GENERATION_LAYERS: 0
+  

configs/BERT_L12_H192_experiments/in1k_training.yaml

@@ -0,0 +1,197 @@
+_BASE_: "base_model_bert_l12_h192.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet1k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+  
+  # -
+  #   NAME: 'Vocab_Word'
+  #   SHARED_TARGETS_CFG:
+  #     FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+  #     DISTRIBUTED: True
+
+
+
+TASKS:
+
+  - 
+    NAME: imagenet
+    DATASETS:
+      TRAIN: 'ImageNetDataset'
+      VAL: 'ImageNetDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet1k'
+      TARGET_SET: ['ImageNet1k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 4
+      TEST_BATCH_SIZE: 4
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/imagenet'
+      S3_PATH: 'cluster2:s3://imagenet'
+      ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
+      SAMPLING_WEIGHT: 1.0
+      CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
+      MIXUP: 0.8
+      CUTMIX: 1.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 1000
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      # LOSS_FP32: True
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  IN_TUNING: True # use IN1k instead of 22k
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    NUM_HIDDEN_LAYERS: 1
+    DROP_PATH_PROB_FIXED: True
+
+    UNIFY_QKV: True
+  
+
+  OLD_CHECKPONT: True
+  
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  # POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  # CHECKPOINT_FILETER: False 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 150000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 500000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 5000
+  MIN_LR: 0.000001
+
+
+
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+

configs/BERT_L12_H192_experiments/in1k_training_moe.yaml

@@ -0,0 +1,219 @@
+_BASE_: "base_model_bert_l12_h192.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet1k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+  
+  # -
+  #   NAME: 'Vocab_Word'
+  #   SHARED_TARGETS_CFG:
+  #     FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+  #     DISTRIBUTED: True
+
+
+
+TASKS:
+
+  - 
+    NAME: imagenet
+    DATASETS:
+      TRAIN: 'ImageNetDataset'
+      VAL: 'ImageNetDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet1k'
+      TARGET_SET: ['ImageNet1k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 4
+      TEST_BATCH_SIZE: 4
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/imagenet'
+      S3_PATH: 'cluster2:s3://imagenet'
+      ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
+      SAMPLING_WEIGHT: 1.0
+      CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
+      MIXUP: 0.8
+      CUTMIX: 1.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 1000
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      # LOSS_FP32: True
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  IN_TUNING: True # use IN1k instead of 22k
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.0
+    DROP_PATH_PROB_FIXED: True
+
+    UNIFY_QKV: True
+  
+
+  OLD_CHECKPONT: True
+  
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  # POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  # CHECKPOINT_FILETER: False 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 150000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 500000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 5000
+  MIN_LR: 0.000001
+
+
+
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+
+MOE: 
+  MOE: True 
+  MOE_TYPE: 'attribute'
+  TAG_Transform: True
+  ATTRIBUTE_LENGTH: 8
+  EP_WORLD_SIZE: 1 # tag moe only 
+  NUM_EXPERTS: 8
+  TOP_K: 2
+  CAPACITY_FACTOR: 3.0 
+  EVAL_MIN_CAPACITY: 4.0
+  MIN_CAPACITY: 4
+  NOISY_GATE_POLICY: 'vmoe'
+  MOE_PARAM_GROUP: True 
+  MOE_EXPERT_TYPE: 'FFN,SA'
+  SA_LINEAR_OUT_MOE: True
+  MOE_EXPERT_LOCATION: 'all' # 'odd'
+  # MOE_LAYER_START_IDX: 3
+  # MOE_LAYER_END_IDX: 21
+  # MOE_LAYER_START_IDX: 18
+  # MOE_LAYER_END_IDX: 12 
+  BATCH_PRIO: True 
+  USE_TUTEL: True
+  FFN_SHARE_GATE_DECISION: True

configs/BERT_L12_H192_experiments/moe_debug.yaml

@@ -0,0 +1,536 @@
+_BASE_: "base_model_bert_l12_h192.yaml"
+
+SHARED_TARGETS:
+
+  # - 
+  #   NAME: 'ImageNet1k'
+  #   SHARED_TARGETS_CFG:
+  #     FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+  #     DISTRIBUTED: False
+  
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  # - 
+  #   NAME: 'Kinetics400'
+  #   SHARED_TARGETS_CFG:
+  #     FILE_PATH: 'open_source_dataset/k400_class_name_CLIP_with_endoftext.pkl'
+  #     DISTRIBUTED: False
+
+
+
+TASKS:
+
+  # - 
+  #   NAME: imagenet
+  #   DATASETS:
+  #     TRAIN: 'ImageNetDataset'
+  #     VAL: 'ImageNetDataset'
+  #     TASK_TYPE: 'image_classification'
+  #     DATASET_NAME: 'ImageNet1k'
+  #     TARGET_SET: ['ImageNet1k']
+      
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 720
+  #     # TEST_BATCH_SIZE: 2
+  #     NUM_WORKERS: 4
+  #     FEATS_FOLDER: 'cluster2:s3://imagenet'
+  #     ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
+  #     SAMPLING_WEIGHT: 2.5
+  #     CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
+  #     MIXUP: 0.8
+  #     CUTMIX: 1.0
+  #     MIXUP_PROB: 1.0
+  #     MIXUP_SWITCH_PROB: 0.5
+  #     MIXUP_MODE: 'batch'
+  #     MIXUP_LABEL_SMOOTHING: 0.1
+  #   MODEL:
+  #     MAX_SEQ_LEN: -1
+  #     LABELS_NUM: 1000
+  #     TEMP_NAME: logit_scale_img_cls
+  #   LOSSES:
+  #     NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 1.0
+  #     REDUCTION: 'mean'
+  #     # LOSS_FP32: True
+  #   INFERENCE:
+  #     NAME: 'ImageNetEvaler'
+  #     ID_KEY: 'image_id'
+  #     VALUE: 'cls_logits'
+  #     VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+  #     TEST_ANNFILE: ''
+  #     GENERATION_MODE: False
+  
+  # -
+  #   NAME: K400_retrieve
+  #   DATASETS:
+  #     TRAIN: 'VideoDataSet'
+  #     VAL: 'VideoDataSet'
+  #     TASK_TYPE: 'video_classification'
+  #     DATASET_NAME: 'K400'
+  #     TARGET_SET: ['Kinetics400']
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 12 # 256
+  #     TEST_BATCH_SIZE: 4 # debug
+  #     NUM_WORKERS: 4 # debug 4
+  #     FEATS_FOLDER: 'open_source_dataset/K400_official'
+  #     ANNO_FOLDER:  'open_source_dataset/K400_official'
+  #     S3_PATH: 's3://K400/'
+  #     FRAMES_PER_CLIP: 8
+  #     STRIDE: 32
+  #     FILE_EXTENSION: ''
+  #     ANNO_FILE: 'annotation.json'
+  #     TIMESFORMER_AUG: True
+  #     SAMPLING_WEIGHT: 1.0
+  #   MODEL:
+  #     MAX_SEQ_LEN: -1
+  #     TEMP_NAME: logit_scale_video_cls
+  #   LOSSES:
+  #     NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+  #     LABELSMOOTHING: 0.1
+  #     LOSS_WEIGHT: 1.0
+  #   INFERENCE:
+  #     NAME: 'MiTEvaler'
+  #     ID_KEY: 'video_name'
+  #     VALUE: 'label'
+  #     VAL_ANNFILE: 'open_source_dataset/K400_official/annotation.json'
+  #     TEST_ANNFILE: ''
+  #     GENERATION_MODE: False
+  #     NUM_VIEWS: 1
+  
+  # -
+  #   NAME: bookswiki_pretrain
+  #   DATASETS:
+  #     TRAIN: 'GeneralCorpusDataset'
+  #     TASK_TYPE: 'text_mlm'
+  #     DATASET_NAME: 'BooksWiki'
+  #     TARGET_SET: ['Vocab_Word']
+  #     VERSION: 'v2'
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 512
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 2
+  #     ANNO_FOLDER:  'open_source_dataset/text_corpus' # 'open_source_dataset/bert_pretrain_data/bookswiki'
+  #     # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+  #     SEQ_PER_SAMPLE:  1
+  #     SAMPLER: NodeDistributed
+  #     CACHE_MODE: True
+  #     SEQ_PER_SAMPLE: 128
+  #     MIN_SEQ_PER_SAMPLE: 128
+  #     APPEND_EOS: True
+  #     ONE_STREAM: False
+  #     SAMPLING_WEIGHT: 3.5
+  #     RANDOM_MASK: True
+  #   MODEL:
+  #     MAX_SEQ_LEN: 128
+  #     TEMP_NAME: logit_scale_text_mlm
+  #   LOSSES:
+  #     NAMES: ['CrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 0.33333
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     GENERATION_MODE: False
+  # -
+  #   NAME: mscoco_retrieve
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     TEST: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_retrieval'
+  #     DATASET_NAME: 'MSCOCO'
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 100
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 1
+  #     FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+  #     ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+  #     S3_PATH: 's3://coco/'
+  #     SEQ_PER_SAMPLE:  1
+  #     CACHE_MODE: True
+  #     CIRCULAR_CACHE_MODE: False
+  #     ZIP_MODE: False
+  #     CACHE_ORIGIN_IMAGE: False
+  #     RANDOM_CAPTION: False
+  #     AS_NUMPY_AS_POSSIBLE: False
+  #     SAMPLING_WEIGHT: 1.0
+  #     TRANSFORM: 'clip_transforms'
+  #   MODEL:
+  #     MAX_SEQ_LEN: 50
+  #     TEMP_NAME: logit_scale_retrieve
+  #   LOSSES:
+  #     NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+  #     LABELSMOOTHING: 0.1
+  #     LOSS_WEIGHT: 1.0
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     ID_KEY: 'image_id'
+  #     VALUE: 'caption'
+  #     NAME: 'RetrievalEvaler'
+  #     VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
+  #     TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
+  #     GENERATION_MODE: False
+
+ ########## Image Captioning ########### 
+
+
+  # -
+  #   NAME: cc12m_caption
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_caption'
+  #     DATASET_NAME: 'CC12M'
+  #     TARGET_SET: ['Vocab_Word']
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 300
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 2
+  #     S3_ANNO_FOLDER:  's3://cc12m/'
+      # ANNO_FOLDER:  'open_source_dataset/c12m/'
+  #     ANNO_FILENAME: 'train_available.json'
+  #     FEATS_FOLDER: 'open_source_dataset/c12m/'
+  #     S3_PATH: 's3://cc12m/'
+  #     SEQ_PER_SAMPLE:  1
+  #     SAMPLER: NodeDistributed
+  #     CACHE_MODE: True
+  #     CIRCULAR_CACHE_MODE: False
+  #     ZIP_MODE: False
+  #     CACHE_ORIGIN_IMAGE: False
+  #     RANDOM_CAPTION: False
+  #     AS_NUMPY_AS_POSSIBLE: False
+  #     SAMPLING_WEIGHT: 1.6889
+  #     TRANSFORM: 'clip_transforms'
+  #   MODEL:
+  #     MAX_SEQ_LEN: 50
+  #     TEMP_NAME: logit_scale_caption
+  #   LOSSES:
+  #     NAMES: ['CrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 0.33333
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     GENERATION_MODE: False
+
+  # -
+  #   NAME: cc3m_caption
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_caption'
+  #     DATASET_NAME: 'CC3M'
+  #     TARGET_SET: ['Vocab_Word']
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 300
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 2
+  #     ANNO_FOLDER:  's3://cc3m/'
+  #     ANNO_FILENAME: 'train_spacy.json'
+  #     FEATS_FOLDER: 'open_source_dataset/cc3m/'
+  #     S3_PATH: 's3://cc3m/'
+  #     SEQ_PER_SAMPLE:  1
+  #     SAMPLER: NodeDistributed
+  #     CACHE_MODE: True
+  #     CIRCULAR_CACHE_MODE: False
+  #     ZIP_MODE: False
+  #     CACHE_ORIGIN_IMAGE: False
+  #     RANDOM_CAPTION: False
+  #     AS_NUMPY_AS_POSSIBLE: False
+  #     SAMPLING_WEIGHT: 0.8780
+  #     TRANSFORM: 'clip_transforms'
+  #   MODEL:
+  #     MAX_SEQ_LEN: 50
+  #     TEMP_NAME: logit_scale_caption
+  #   LOSSES:
+  #     NAMES: ['CrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 0.33333
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     GENERATION_MODE: False
+
+  # -
+  #   NAME: vg_caption
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_caption'
+  #     DATASET_NAME: 'VG'
+  #     TARGET_SET: ['Vocab_Word']
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 300
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 2
+  #     FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+  #     ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+  #     S3_PATH: 's3://visual_genome/images'
+  #     ANNO_FILENAME: 'vg_captions_128filter.json'
+  #     SEQ_PER_SAMPLE:  1
+  #     CACHE_MODE: True
+  #     CIRCULAR_CACHE_MODE: False
+  #     ZIP_MODE: False
+  #     CACHE_ORIGIN_IMAGE: False
+  #     RANDOM_CAPTION: False
+  #     AS_NUMPY_AS_POSSIBLE: False
+  #     SAMPLING_WEIGHT: 0.5895
+  #     TRANSFORM: 'clip_transforms'
+  #   MODEL:
+  #     MAX_SEQ_LEN: 30
+  #     TEMP_NAME: logit_scale_caption
+  #   LOSSES:
+  #     NAMES: ['CrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 0.33333
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     GENERATION_MODE: True
+  
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 32
+      TEST_BATCH_SIZE: 2
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.3817
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 2.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  # -
+  #   NAME: sbu_caption
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_caption'
+  #     DATASET_NAME: 'SBU'
+  #     TARGET_SET: ['Vocab_Word']
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 300
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 1
+  #     S3_ANNO_FOLDER: 's3://SBU/annotations'
+      # ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+  #     ANNO_FILENAME: 'subcaption.json'
+  #     FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+  #     S3_PATH: 's3://SBU/images'
+  #     SEQ_PER_SAMPLE:  1
+  #     SAMPLER: NodeDistributed
+  #     CACHE_MODE: True
+  #     CIRCULAR_CACHE_MODE: False
+  #     ZIP_MODE: False
+  #     CACHE_ORIGIN_IMAGE: False
+  #     RANDOM_CAPTION: False
+  #     AS_NUMPY_AS_POSSIBLE: False
+  #     SAMPLING_WEIGHT: 0.4618
+  #     TRANSFORM: 'clip_transforms'
+  #   MODEL:
+  #     MAX_SEQ_LEN: 50
+  #     TEMP_NAME: logit_scale_caption
+  #   LOSSES:
+  #     NAMES: ['CrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 0.33333
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  IN_TUNING: True # use IN1k instead of 22k
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.0
+    DROP_PATH_PROB_FIXED: True
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+  LAYER_SCALE_FP32: True
+  GATE_FP32: False
+  TAG_TRANSFORM_FP32: False
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 32
+  STRATEGY: 'turn'
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 150000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 500000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+  FORCE_SOFTMAX_FP16: True
+  FORCE_LN_FP16: True
+  FORCE_NORM_FP16: True
+  # FORCE_TEMP_FP16: True
+  FORCE_EMBED_FP16: True
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 5000
+  MIN_LR: 0.000001
+
+
+
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+
+MOE: 
+  MOE: True 
+  MOE_TYPE: 'attribute'
+  TAG_Transform: True
+  ATTRIBUTE_LENGTH: 8
+  EP_WORLD_SIZE: 1 # tag moe only 
+  NUM_EXPERTS: 8
+  TOP_K: 2
+  CAPACITY_FACTOR: 3.0 
+  EVAL_MIN_CAPACITY: 4.0
+  MIN_CAPACITY: 4
+  NOISY_GATE_POLICY: 'vmoe'
+  MOE_PARAM_GROUP: True 
+  MOE_EXPERT_TYPE: 'FFN,SA'
+  SA_LINEAR_OUT_MOE: True
+  MOE_EXPERT_LOCATION: 'all' # 'odd'
+  # MOE_LAYER_START_IDX: 3
+  # MOE_LAYER_END_IDX: 21
+  # MOE_LAYER_START_IDX: 18
+  # MOE_LAYER_END_IDX: 12 
+  BATCH_PRIO: True 
+  USE_TUTEL: True
+  FFN_SHARE_GATE_DECISION: True

configs/BERT_L12_H192_experiments/moe_debug_load_ds_checkpoint.yaml

@@ -0,0 +1,541 @@
+_BASE_: "base_model_bert_l12_h192.yaml"
+
+SHARED_TARGETS:
+
+  # - 
+  #   NAME: 'ImageNet1k'
+  #   SHARED_TARGETS_CFG:
+  #     FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+  #     DISTRIBUTED: False
+  
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  # - 
+  #   NAME: 'Kinetics400'
+  #   SHARED_TARGETS_CFG:
+  #     FILE_PATH: 'open_source_dataset/k400_class_name_CLIP_with_endoftext.pkl'
+  #     DISTRIBUTED: False
+
+
+
+TASKS:
+
+  # - 
+  #   NAME: imagenet
+  #   DATASETS:
+  #     TRAIN: 'ImageNetDataset'
+  #     VAL: 'ImageNetDataset'
+  #     TASK_TYPE: 'image_classification'
+  #     DATASET_NAME: 'ImageNet1k'
+  #     TARGET_SET: ['ImageNet1k']
+      
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 720
+  #     # TEST_BATCH_SIZE: 2
+  #     NUM_WORKERS: 4
+  #     FEATS_FOLDER: 'cluster2:s3://imagenet'
+  #     ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
+  #     SAMPLING_WEIGHT: 2.5
+  #     CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
+  #     MIXUP: 0.8
+  #     CUTMIX: 1.0
+  #     MIXUP_PROB: 1.0
+  #     MIXUP_SWITCH_PROB: 0.5
+  #     MIXUP_MODE: 'batch'
+  #     MIXUP_LABEL_SMOOTHING: 0.1
+  #   MODEL:
+  #     MAX_SEQ_LEN: -1
+  #     LABELS_NUM: 1000
+  #     TEMP_NAME: logit_scale_img_cls
+  #   LOSSES:
+  #     NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 1.0
+  #     REDUCTION: 'mean'
+  #     # LOSS_FP32: True
+  #   INFERENCE:
+  #     NAME: 'ImageNetEvaler'
+  #     ID_KEY: 'image_id'
+  #     VALUE: 'cls_logits'
+  #     VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+  #     TEST_ANNFILE: ''
+  #     GENERATION_MODE: False
+  
+  # -
+  #   NAME: K400_retrieve
+  #   DATASETS:
+  #     TRAIN: 'VideoDataSet'
+  #     VAL: 'VideoDataSet'
+  #     TASK_TYPE: 'video_classification'
+  #     DATASET_NAME: 'K400'
+  #     TARGET_SET: ['Kinetics400']
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 12 # 256
+  #     TEST_BATCH_SIZE: 4 # debug
+  #     NUM_WORKERS: 4 # debug 4
+  #     FEATS_FOLDER: 'open_source_dataset/K400_official'
+  #     ANNO_FOLDER:  'open_source_dataset/K400_official'
+  #     S3_PATH: 's3://K400/'
+  #     FRAMES_PER_CLIP: 8
+  #     STRIDE: 32
+  #     FILE_EXTENSION: ''
+  #     ANNO_FILE: 'annotation.json'
+  #     TIMESFORMER_AUG: True
+  #     SAMPLING_WEIGHT: 1.0
+  #   MODEL:
+  #     MAX_SEQ_LEN: -1
+  #     TEMP_NAME: logit_scale_video_cls
+  #   LOSSES:
+  #     NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+  #     LABELSMOOTHING: 0.1
+  #     LOSS_WEIGHT: 1.0
+  #   INFERENCE:
+  #     NAME: 'MiTEvaler'
+  #     ID_KEY: 'video_name'
+  #     VALUE: 'label'
+  #     VAL_ANNFILE: 'open_source_dataset/K400_official/annotation.json'
+  #     TEST_ANNFILE: ''
+  #     GENERATION_MODE: False
+  #     NUM_VIEWS: 1
+  
+  # -
+  #   NAME: bookswiki_pretrain
+  #   DATASETS:
+  #     TRAIN: 'GeneralCorpusDataset'
+  #     TASK_TYPE: 'text_mlm'
+  #     DATASET_NAME: 'BooksWiki'
+  #     TARGET_SET: ['Vocab_Word']
+  #     VERSION: 'v2'
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 512
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 2
+  #     ANNO_FOLDER:  'open_source_dataset/text_corpus' # 'open_source_dataset/bert_pretrain_data/bookswiki'
+  #     # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+  #     SEQ_PER_SAMPLE:  1
+  #     SAMPLER: NodeDistributed
+  #     CACHE_MODE: True
+  #     SEQ_PER_SAMPLE: 128
+  #     MIN_SEQ_PER_SAMPLE: 128
+  #     APPEND_EOS: True
+  #     ONE_STREAM: False
+  #     SAMPLING_WEIGHT: 3.5
+  #     RANDOM_MASK: True
+  #   MODEL:
+  #     MAX_SEQ_LEN: 128
+  #     TEMP_NAME: logit_scale_text_mlm
+  #   LOSSES:
+  #     NAMES: ['CrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 0.33333
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     GENERATION_MODE: False
+  # -
+  #   NAME: mscoco_retrieve
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     TEST: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_retrieval'
+  #     DATASET_NAME: 'MSCOCO'
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 100
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 1
+  #     FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+  #     ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+  #     S3_PATH: 's3://coco/'
+  #     SEQ_PER_SAMPLE:  1
+  #     CACHE_MODE: True
+  #     CIRCULAR_CACHE_MODE: False
+  #     ZIP_MODE: False
+  #     CACHE_ORIGIN_IMAGE: False
+  #     RANDOM_CAPTION: False
+  #     AS_NUMPY_AS_POSSIBLE: False
+  #     SAMPLING_WEIGHT: 1.0
+  #     TRANSFORM: 'clip_transforms'
+  #   MODEL:
+  #     MAX_SEQ_LEN: 50
+  #     TEMP_NAME: logit_scale_retrieve
+  #   LOSSES:
+  #     NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+  #     LABELSMOOTHING: 0.1
+  #     LOSS_WEIGHT: 1.0
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     ID_KEY: 'image_id'
+  #     VALUE: 'caption'
+  #     NAME: 'RetrievalEvaler'
+  #     VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
+  #     TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
+  #     GENERATION_MODE: False
+
+ ########## Image Captioning ########### 
+
+
+  # -
+  #   NAME: cc12m_caption
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_caption'
+  #     DATASET_NAME: 'CC12M'
+  #     TARGET_SET: ['Vocab_Word']
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 300
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 2
+  #     S3_ANNO_FOLDER:  's3://cc12m/'
+      # ANNO_FOLDER:  'open_source_dataset/c12m/'
+  #     ANNO_FILENAME: 'train_available.json'
+  #     FEATS_FOLDER: 'open_source_dataset/c12m/'
+  #     S3_PATH: 's3://cc12m/'
+  #     SEQ_PER_SAMPLE:  1
+  #     SAMPLER: NodeDistributed
+  #     CACHE_MODE: True
+  #     CIRCULAR_CACHE_MODE: False
+  #     ZIP_MODE: False
+  #     CACHE_ORIGIN_IMAGE: False
+  #     RANDOM_CAPTION: False
+  #     AS_NUMPY_AS_POSSIBLE: False
+  #     SAMPLING_WEIGHT: 1.6889
+  #     TRANSFORM: 'clip_transforms'
+  #   MODEL:
+  #     MAX_SEQ_LEN: 50
+  #     TEMP_NAME: logit_scale_caption
+  #   LOSSES:
+  #     NAMES: ['CrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 0.33333
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     GENERATION_MODE: False
+
+  # -
+  #   NAME: cc3m_caption
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_caption'
+  #     DATASET_NAME: 'CC3M'
+  #     TARGET_SET: ['Vocab_Word']
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 300
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 2
+  #     S3_ANNO_FOLDER: 's3://cc3m/'
+      # ANNO_FOLDER:  'open_source_dataset/cc3m/'
+  #     ANNO_FILENAME: 'train_spacy.json'
+  #     FEATS_FOLDER: 'open_source_dataset/cc3m/'
+  #     S3_PATH: 's3://cc3m/'
+  #     SEQ_PER_SAMPLE:  1
+  #     SAMPLER: NodeDistributed
+  #     CACHE_MODE: True
+  #     CIRCULAR_CACHE_MODE: False
+  #     ZIP_MODE: False
+  #     CACHE_ORIGIN_IMAGE: False
+  #     RANDOM_CAPTION: False
+  #     AS_NUMPY_AS_POSSIBLE: False
+  #     SAMPLING_WEIGHT: 0.8780
+  #     TRANSFORM: 'clip_transforms'
+  #   MODEL:
+  #     MAX_SEQ_LEN: 50
+  #     TEMP_NAME: logit_scale_caption
+  #   LOSSES:
+  #     NAMES: ['CrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 0.33333
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     GENERATION_MODE: False
+
+  # -
+  #   NAME: vg_caption
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_caption'
+  #     DATASET_NAME: 'VG'
+  #     TARGET_SET: ['Vocab_Word']
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 300
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 2
+  #     FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+  #     ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+  #     S3_PATH: 's3://visual_genome/images'
+  #     ANNO_FILENAME: 'vg_captions_128filter.json'
+  #     SEQ_PER_SAMPLE:  1
+  #     CACHE_MODE: True
+  #     CIRCULAR_CACHE_MODE: False
+  #     ZIP_MODE: False
+  #     CACHE_ORIGIN_IMAGE: False
+  #     RANDOM_CAPTION: False
+  #     AS_NUMPY_AS_POSSIBLE: False
+  #     SAMPLING_WEIGHT: 0.5895
+  #     TRANSFORM: 'clip_transforms'
+  #   MODEL:
+  #     MAX_SEQ_LEN: 30
+  #     TEMP_NAME: logit_scale_caption
+  #   LOSSES:
+  #     NAMES: ['CrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 0.33333
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     GENERATION_MODE: True
+  
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 32
+      TEST_BATCH_SIZE: 2
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.3817
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 2.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  # -
+  #   NAME: sbu_caption
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_caption'
+  #     DATASET_NAME: 'SBU'
+  #     TARGET_SET: ['Vocab_Word']
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 300
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 1
+  #     S3_ANNO_FOLDER: 's3://SBU/annotations'
+      # ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+  #     ANNO_FILENAME: 'subcaption.json'
+  #     FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+  #     S3_PATH: 's3://SBU/images'
+  #     SEQ_PER_SAMPLE:  1
+  #     SAMPLER: NodeDistributed
+  #     CACHE_MODE: True
+  #     CIRCULAR_CACHE_MODE: False
+  #     ZIP_MODE: False
+  #     CACHE_ORIGIN_IMAGE: False
+  #     RANDOM_CAPTION: False
+  #     AS_NUMPY_AS_POSSIBLE: False
+  #     SAMPLING_WEIGHT: 0.4618
+  #     TRANSFORM: 'clip_transforms'
+  #   MODEL:
+  #     MAX_SEQ_LEN: 50
+  #     TEMP_NAME: logit_scale_caption
+  #   LOSSES:
+  #     NAMES: ['CrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 0.33333
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  IN_TUNING: True # use IN1k instead of 22k
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.0
+    DROP_PATH_PROB_FIXED: True
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+  VIDEO_EMBED:
+    ADD_TYPE_EMBED: True
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 32
+  STRATEGY: 'turn'
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 150000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 500000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+  FORCE_SOFTMAX_FP16: True
+  FORCE_LN_FP16: True
+  FORCE_NORM_FP16: True
+  # FORCE_TEMP_FP16: True
+  FORCE_EMBED_FP16: True
+
+# # used for debug only 
+  FORCE_WG_RECAST: True
+  FORCE_EXPERT_ADDING_FP16: True
+  
+  # ！！！ note that the VIDEO_EMBED.ADD_TYPE_EMBED=True is current config
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 5000
+  MIN_LR: 0.000001
+
+
+
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+
+
+
+MOE: 
+  MOE: True 
+  MOE_TYPE: 'attribute'
+  TAG_Transform: True
+  ATTRIBUTE_LENGTH: 8
+  EP_WORLD_SIZE: 1 # tag moe only 
+  NUM_EXPERTS: 8
+  TOP_K: 2
+  CAPACITY_FACTOR: 3.0 
+  EVAL_MIN_CAPACITY: 4.0
+  MIN_CAPACITY: 4
+  NOISY_GATE_POLICY: 'vmoe'
+  MOE_PARAM_GROUP: True 
+  MOE_EXPERT_TYPE: 'FFN,SA'
+  SA_LINEAR_OUT_MOE: True
+  MOE_EXPERT_LOCATION: 'all' # 'odd'
+  # MOE_LAYER_START_IDX: 3
+  # MOE_LAYER_END_IDX: 21
+  # MOE_LAYER_START_IDX: 18
+  # MOE_LAYER_END_IDX: 12 
+  BATCH_PRIO: True 
+  USE_TUTEL: True
+  FFN_SHARE_GATE_DECISION: True

configs/BERT_L12_H192_experiments/mscoco_caption_debug.yaml

@@ -0,0 +1,234 @@
+_BASE_: "base_model_bert_l12_h192.yaml"
+
+SHARED_TARGETS:
+
+  
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+
+
+TASKS:
+
+  -
+    NAME: mscoco_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'MSCOCO'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 100
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      NAME: 'RetrievalEvaler'
+      VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
+      TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
+      GENERATION_MODE: False
+
+  
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.3817
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.33333
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 1.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  
+ 
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  IN_TUNING: True # use IN1k instead of 22k
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.0
+    DROP_PATH_PROB_FIXED: True
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 32
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 150000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 500000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 5000
+  MIN_LR: 0.000001
+
+
+
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+

configs/BERT_L12_H192_experiments/vqa_debug.yaml

@@ -0,0 +1,189 @@
+_BASE_: "base_model_bert_l12_h192.yaml"
+
+SHARED_TARGETS:
+
+  
+  -
+    NAME: 'VQA_Answer'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/VQA_Answers_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+TASKS:
+  -
+    NAME: vqa
+    DATASETS:
+      TRAIN: 'VQADataset'
+      VAL: 'VQADataset'
+      DATASET_NAME: 'VQA'
+      TASK_TYPE: 'vqa'
+      TARGET_SET: ['VQA_Answer']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 256
+      TEST_BATCH_SIZE: 128
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER: 'open_source_dataset/VQA'
+      SEQ_PER_SAMPLE:  1
+      MAX_FEAT_NUM: 51
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+      DO_AS_GEN: True
+      SINGLE_CLASS: True
+    MODEL:
+      # VOCAB_SIZE: 49409 # include <BOS>/<EOS>
+      PREDICTOR: 'MLPClassifer'
+      # MM_PREDICTOR:
+      #   LABELS_NUM: 3129
+      #   PREDICT: 'first_one'
+      # PRED_DROPOUT: 0.5
+      MAX_SEQ_LEN: 23
+      # QUERY_EMBED:
+      #   NAME: QueryBaseEmbedding
+      #   DIM: 512
+      #   QUERY_SIZE: 10 # more than 1 is ok
+      #   ACTIVATION: 'none'
+      #   USE_NORM: True
+      #   DROPOUT: 0.1
+      #   POSITION: 'none' # must be none now
+      #   TYPE_VOCAB_SIZE: -1 # must < 0
+    LOSSES:
+      # not single class 
+      # NAMES: ['BCEWithLogits']
+      # LOSS_WEIGHT: 0.05
+      # for single class
+      NAMES: ['CrossEntropy']
+      LOSS_WEIGHT: 0.1
+    INFERENCE:
+      VOCAB: 'CLIP'
+      NAME: 'VQAEvaler'
+      ID_KEY: 'question_id'
+      VALUE: 'answer'
+      VAL_ANNFILE: 'open_source_dataset/VQA/val_target.pkl'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+
+
+######################################### Engine #########################################
+ENGINE:
+  NAME: 'UnifiedTrainer'
+
+######################################### Scheduled sampling #########################################
+SCHEDULED_SAMPLING:
+  START_EPOCH: 0
+  INC_EVERY_EPOCH: 5
+  INC_PROB: 0.05
+  MAX_PROB: 0.25
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True
+
+######################################### MODEL #########################################
+MODEL:
+  TEMP_NAME: logit_scale_downstream
+  # VOCAB_SIZE: 49409 # include <BOS>/<EOS>
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+  # ENCODER_DIM: 512
+  # DECODER: 'UnifiedTransformerDecoder'
+  # DECODER_DIM: 512
+  
+  BertParamsInit: True
+  # WEIGHTS: open_source_dataset/our_model/cc3m_encoder_decoder_warm1w_150k_retrivetask_gatherfeature_caption_mlm/model_Epoch_90000_Iter_0089999.pth
+
+  CLS_TOKEN: True
+  # PREDICTOR: 'BasePredictor'
+  # PRED_DROPOUT: 0.5
+  # MAX_SEQ_LEN: 20
+
+# #################################### Token embedding ####################################
+  # TOKEN_EMBED:
+  #   NAME: 'TokenBaseEmbedding'
+  #   DIM: 512
+  #   ACTIVATION: 'none'
+  #   USE_NORM: True
+  #   DROPOUT: 0.1
+  #   POSITION: 'NNEmbeddingEncoding'
+  #   POSITION_MAX_LEN: 512
+  #   TYPE_VOCAB_SIZE: 2
+
+# #################################### Visual embedding ####################################
+  # VISUAL_EMBED:
+  #   NAME: 'VisualPatchEmbedding'
+  #   IN_DIM: 3
+  #   OUT_DIM: 512
+  #   ACTIVATION: 'none'
+  #   USE_NORM: True
+  #   DROPOUT: 0.0
+  #   PATCH_SIZE: 16
+
+####################################### BERT ############################################
+  BERT:
+    DROP_PATH_PROB: 0.05
+    # HIDDEN_SIZE: 512
+    HIDDEN_SIZE: 192
+    HIDDEN_DROPOUT_PROB: 0.
+    HIDDEN_ACT: "gelu"
+    NUM_ATTENTION_HEADS: 8
+    INTERMEDIATE_SIZE: 2048
+    INTERMEDIATE_DROP: 0.
+    FFN_DROPOUT_PROB: 0.
+    ATTENTION_PROBS_DROPOUT_PROB: 0.
+    NUM_HIDDEN_LAYERS: 6
+    NUM_GENERATION_LAYERS: 6
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'AdamW'
+  # EPOCH: 1
+  MAX_ITER: 30000
+  CHECKPOINT_PERIOD: 5000
+  CHECKPOINT_MAX_SAVE: 5
+  EVAL_PERIOD: 1000
+  BASE_LR: 0.00005
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.01
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.999]
+  EPS: 1e-8
+  GRAD_CLIP: 5.0
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  CHECKPOINT_MAPPING:
+    # - 
+    #  ORIGIN: cc3m_caption 
+    #  DEST: mscoco 
+    - 
+     ORIGIN: cc3m_retrieve 
+     DEST: flickr30k
+
+  CHECKPOINT_MAP: True
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 1000
+  MIN_LR: 0.00000001
+
+# ####################################### losses ####################################### 
+# LOSSES:
+#   NAMES: ['LabelSmoothing']
+#   LABELSMOOTHING: 0.1
+
+####################################### decode strategy ####################################### 
+# DECODE_STRATEGY:
+#   NAME: 'BeamSearcher'
+#   BEAM_SIZE: 2
+
+####################################### evaluation ####################################### 
+INFERENCE:
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+find_unused_parameters: true

configs/BERT_L12_H384_experiments/base_model_bert_l12_h384.yaml

@@ -0,0 +1,80 @@
+
+######################################### MODEL #########################################
+MODEL:
+  VOCAB_SIZE: 49411 # include <BOS>/<EOS>
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: ''
+  ENCODER_DIM: 384
+  DECODER: ''
+  DECODER_DIM: 384
+
+  PREDICTOR: 'EmbedClsAsRetrievalPredictor'
+  FEATURE_GATHER: True
+  LEARN_TEMP: True
+  PRED_USE_NORM: True
+  PRED_TEMPERATURE: 0.07
+
+  BertParamsInit: True
+
+  CLS_TOKEN: False
+
+  QUEUE_LEN: 1024
+  MAX_LABEL_LEN: 12
+
+  OUTPUT_PROJ: True # output projection
+ 
+
+# #################################### Token embedding ####################################
+  TOKEN_EMBED:
+    NAME: 'TokenBaseEmbedding'
+    DIM: 384
+    ACTIVATION: 'none'
+    USE_NORM: True
+    DROPOUT: 0.0
+    POSITION: 'NNEmbeddingEncoding'
+    POSITION_MAX_LEN: 512
+    TYPE_VOCAB_SIZE: 2
+
+# #################################### Visual embedding ####################################
+  VISUAL_EMBED:
+    NAME: 'VisualPatchEmbedding'
+    IN_DIM: 3
+    OUT_DIM: 384
+    ACTIVATION: 'none'
+    USE_NORM: True
+    DROPOUT: 0.0
+    PATCH_SIZE: 16
+    TYPE_SIZE: 1 # image to encoder
+
+# #################################### video embedding ####################################
+  VIDEO_EMBED:
+    NAME: 'VideoBaseEmbedding'
+    IN_DIM: 768
+    OUT_DIM: 384
+    ACTIVATION: 'none'
+    USE_NORM: True
+    DROPOUT: 0.0
+    TYPE_SIZE: 1 # video to encoder
+    POSITION: 'NNEmbeddingEncoding'
+    MAX_LENGTH: 1600
+    PATCH_SIZE_S: 16
+    PATCH_SIZE_T: 1
+    DIVIDE_ST_POS: True
+    USE_VISUAL_TOKENIZER: True
+    USE_VISUAL_POS: True
+    MAX_FRAMES: 8
+
+####################################### BERT ############################################
+  BERT:
+    DROP_PATH_PROB: 0.1
+    HIDDEN_SIZE: 384
+    HIDDEN_DROPOUT_PROB: 0.
+    HIDDEN_ACT: "gelu"
+    NUM_ATTENTION_HEADS: 6
+    INTERMEDIATE_SIZE: 1536
+    INTERMEDIATE_DROP: 0.
+    FFN_DROPOUT_PROB: 0.
+    ATTENTION_PROBS_DROPOUT_PROB: 0.
+    NUM_HIDDEN_LAYERS: 12
+    NUM_GENERATION_LAYERS: 0
+  

configs/BERT_L12_H384_experiments/in1k_training.yaml

@@ -0,0 +1,189 @@
+_BASE_: "base_model_bert_l12_h384.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet1k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+  # -
+  #   NAME: 'Vocab_Word'
+  #   SHARED_TARGETS_CFG:
+  #     FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+  #     DISTRIBUTED: True
+
+TASKS:
+
+  - 
+    NAME: imagenet
+    DATASETS:
+      TRAIN: 'ImageNetDataset'
+      VAL: 'ImageNetDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet1k'
+      TARGET_SET: ['ImageNet1k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 32
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 4 # will be used as numworker for testing loader 
+      FEATS_FOLDER: 'open_source_dataset/imagenet'
+      S3_PATH: 'cluster2:s3://imagenet'
+      ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
+      SAMPLING_WEIGHT: 1.0
+      CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
+      MIXUP: 0.8
+      CUTMIX: 1.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 1000
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      # LOSS_FP32: True
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  IN_TUNING: True # use IN1k instead of 22k
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+    UNIFY_QKV: True
+  
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 200000
+  CHECKPOINT_PERIOD: 10
+  EVAL_PERIOD: 500000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.3
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 20000
+  MIN_LR: 0.000001
+
+
+
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+

configs/BERT_L12_H768_experiments/16tasks_training.yaml

@@ -0,0 +1,738 @@
+_BASE_: "base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet22k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_22k_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  - 
+    NAME: 'MomentsInTime'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/MiT_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+  - 
+    NAME: 'Kinetics700'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/k700_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+TASKS:
+
+  - 
+    NAME: imagenet22k
+    DATASETS:
+      TRAIN: 'ImageNet22KDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet22k'
+      TARGET_SET: ['ImageNet22k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 720
+      # TEST_BATCH_SIZE: 2
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/imagenet22k'
+      S3_PATH: 'cluster2:s3://imagenet22k'
+      ANNO_FOLDER:  'open_source_dataset/'
+      SAMPLING_WEIGHT: 2.486
+      MIXUP: 0.8
+      CUTMIX: 1.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 21842
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+
+  -
+    NAME: K700_retrieve
+    DATASETS:
+      TRAIN: 'VideoDataSet'
+      TASK_TYPE: 'video_classification'
+      DATASET_NAME: 'K700'
+      TARGET_SET: ['Kinetics700']
+    DATALOADER: 
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 24 
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/K700'
+      ANNO_FOLDER:  'open_source_dataset/K700'
+      S3_PATH: 's3://K700/'
+      FRAMES_PER_CLIP: 4
+      STRIDE: 32
+      FILE_EXTENSION: ''
+      ANNO_FILE: 'annotation.json'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 0.76
+    
+    MODEL:
+      MAX_SEQ_LEN: -1
+      TEMP_NAME: logit_scale_video_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.1
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+  
+  -
+    NAME: MomentsInTime
+    DATASETS:
+      TRAIN: 'VideoDataSet'
+      TASK_TYPE: 'video_classification'
+      DATASET_NAME: 'MiT'
+      TARGET_SET: ['MomentsInTime']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 112
+      TEST_BATCH_SIZE: 8
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/MomentsInTime'
+      ANNO_FOLDER:  'open_source_dataset/MomentsInTime'
+      S3_PATH:  's3://MomentsInTime/'
+      FRAMES_PER_CLIP: 3
+      STRIDE: 32
+      FILE_EXTENSION: ''
+      ANNO_FILE: 'annotation.json'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 0.44
+    
+    MODEL:
+      MAX_SEQ_LEN: -1
+      TEMP_NAME: logit_scale_video_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.1
+    INFERENCE:
+      NAME: 'MiTEvaler'
+      ID_KEY: 'video_name'
+      VALUE: 'label'
+      VAL_ANNFILE: 'open_source_dataset/MomentsInTime/annotation.json'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+      NUM_VIEWS: 1
+
+  -
+    NAME: bookswiki_pretrain
+    DATASETS:
+      TRAIN: 'GeneralCorpusDataset'
+      TASK_TYPE: 'text_mlm'
+      DATASET_NAME: 'BooksWiki'
+      TARGET_SET: ['Vocab_Word']
+      VERSION: 'v2'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  'open_source_dataset/text_corpus' # 'open_source_dataset/bert_pretrain_data/bookswiki'
+      # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      SEQ_PER_SAMPLE: 128
+      MIN_SEQ_PER_SAMPLE: 128
+      APPEND_EOS: True
+      ONE_STREAM: False
+      SAMPLING_WEIGHT: 2.75
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 128
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+  -
+    NAME: yfcc_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'YFCC'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: True
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5840
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC12M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5057
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC3M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.26295
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+  
+  -
+    NAME: vg_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'VG'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1766
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: True
+
+
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1144
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 1.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  -
+    NAME: sbu_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'SBU'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1383
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: yfcc_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'YFCC'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: True
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5840
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC12M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5057
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC3M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.26295
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: vg_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'VG'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1766
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: mscoco_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'MSCOCO'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1144
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      NAME: 'RetrievalEvaler'
+      VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
+      TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
+      GENERATION_MODE: False
+  
+  -
+    NAME: sbu_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'SBU'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1383
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+    UNIFY_QKV: True
+  
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 160
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 32
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 400000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 10000000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.2
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 10000
+  MIN_LR: 0.000001
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+

configs/BERT_L12_H768_experiments/16tasks_training_apex_o2.yaml

@@ -0,0 +1,11 @@
+_BASE_: "16tasks_training.yaml"
+
+####################################### Optimizer #######################################
+SOLVER:
+  
+  AMP_FP16: False
+  APEX_FP16: True # dangerous
+  APEX_OPT_LEVEL: 'O2'
+  MIN_LOSS_SCLE: 128.0
+  CHECKPOINT_PERIOD: 10000
+

configs/BERT_L12_H768_experiments/16tasks_training_basedense_stage1_64gpu.yaml

@@ -0,0 +1,739 @@
+_BASE_: "base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet22k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_22k_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  - 
+    NAME: 'MomentsInTime'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/MiT_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+  - 
+    NAME: 'Kinetics700'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/k700_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+TASKS:
+
+  - 
+    NAME: imagenet22k
+    DATASETS:
+      TRAIN: 'ImageNet22KDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet22k'
+      TARGET_SET: ['ImageNet22k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 720
+      # TEST_BATCH_SIZE: 2
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/imagenet22k'
+      S3_PATH: 'cluster2:s3://imagenet22k'
+      ANNO_FOLDER:  'open_source_dataset/'
+      SAMPLING_WEIGHT: 2.486
+      MIXUP: 0.8
+      CUTMIX: 1.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 21842
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+
+  -
+    NAME: K700_retrieve
+    DATASETS:
+      TRAIN: 'VideoDataSet'
+      TASK_TYPE: 'video_classification'
+      DATASET_NAME: 'K700'
+      TARGET_SET: ['Kinetics700']
+    DATALOADER: 
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 24 
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/K700'
+      ANNO_FOLDER:  'open_source_dataset/K700'
+      S3_PATH: 's3://K700/'
+      FRAMES_PER_CLIP: 4
+      STRIDE: 32
+      FILE_EXTENSION: ''
+      ANNO_FILE: 'annotation.json'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 0.76
+    
+    MODEL:
+      MAX_SEQ_LEN: -1
+      TEMP_NAME: logit_scale_video_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.1
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+  
+  -
+    NAME: MomentsInTime
+    DATASETS:
+      TRAIN: 'VideoDataSet'
+      TASK_TYPE: 'video_classification'
+      DATASET_NAME: 'MiT'
+      TARGET_SET: ['MomentsInTime']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 112
+      TEST_BATCH_SIZE: 8
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/MomentsInTime'
+      ANNO_FOLDER:  'open_source_dataset/MomentsInTime'
+      S3_PATH:  's3://MomentsInTime/'
+      FRAMES_PER_CLIP: 3
+      STRIDE: 32
+      FILE_EXTENSION: ''
+      ANNO_FILE: 'annotation.json'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 0.44
+    
+    MODEL:
+      MAX_SEQ_LEN: -1
+      TEMP_NAME: logit_scale_video_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.1
+    INFERENCE:
+      NAME: 'MiTEvaler'
+      ID_KEY: 'video_name'
+      VALUE: 'label'
+      VAL_ANNFILE: 'open_source_dataset/MomentsInTime/annotation.json'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+      NUM_VIEWS: 1
+
+  -
+    NAME: bookswiki_pretrain
+    DATASETS:
+      TRAIN: 'GeneralCorpusDataset'
+      TASK_TYPE: 'text_mlm'
+      DATASET_NAME: 'BooksWiki'
+      TARGET_SET: ['Vocab_Word']
+      VERSION: 'v2'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  'open_source_dataset/text_corpus' # 'open_source_dataset/bert_pretrain_data/bookswiki'
+      # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      SEQ_PER_SAMPLE: 128
+      MIN_SEQ_PER_SAMPLE: 128
+      APPEND_EOS: True
+      ONE_STREAM: False
+      SAMPLING_WEIGHT: 2.75
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 128
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+  -
+    NAME: yfcc_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'YFCC'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: True
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5840
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC12M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5057
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC3M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.26295
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+  
+  -
+    NAME: vg_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'VG'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1766
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: True
+
+
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1144
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 1.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  -
+    NAME: sbu_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'SBU'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1383
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: yfcc_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'YFCC'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: True
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5840
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC12M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5057
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC3M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.26295
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: vg_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'VG'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1766
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: mscoco_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'MSCOCO'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1144
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      NAME: 'RetrievalEvaler'
+      VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
+      TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
+      GENERATION_MODE: False
+  
+  -
+    NAME: sbu_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'SBU'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1383
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+    UNIFY_QKV: True
+  
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 160
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+  OLD_CHECKPONT: True 
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 32
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 200000
+  CHECKPOINT_PERIOD: 10000
+  EVAL_PERIOD: 10000000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.2
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 10000
+  MIN_LR: 0.000001
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+

configs/BERT_L12_H768_experiments/16tasks_training_basedense_stage2_64gpu.yaml

@@ -0,0 +1,750 @@
+_BASE_: "base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet22k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_22k_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  - 
+    NAME: 'MomentsInTime'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/MiT_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+  - 
+    NAME: 'Kinetics700'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/k700_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+TASKS:
+
+  - 
+    NAME: imagenet22k
+    DATASETS:
+      TRAIN: 'ImageNet22KDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet22k'
+      TARGET_SET: ['ImageNet22k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 440
+      # TEST_BATCH_SIZE: 2
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/imagenet22k'
+      S3_PATH: 'cluster2:s3://imagenet22k'
+      ANNO_FOLDER:  'open_source_dataset/'
+      SAMPLING_WEIGHT: 2.486
+      MIXUP: 0.0
+      CUTMIX: 0.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 21842
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      LABELSMOOTHING: 0.1
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+      # VAL_ANNFILE: '/mnt/lustrenew/lihao2/projects/xmodaler_2/val_debug.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+
+  -
+    NAME: K700_retrieve
+    DATASETS:
+      TRAIN: 'VideoDataSet'
+      TASK_TYPE: 'video_classification'
+      DATASET_NAME: 'K700'
+      TARGET_SET: ['Kinetics700']
+    DATALOADER: 
+      TRAIN_BATCH_SIZE: 12
+      TEST_BATCH_SIZE: 24 
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/K700'
+      ANNO_FOLDER:  'open_source_dataset/K700'
+      S3_PATH: 's3://K700/'
+      FRAMES_PER_CLIP: 8
+      STRIDE: 32
+      FILE_EXTENSION: ''
+      ANNO_FILE: 'annotation.json'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 1.0
+    
+    MODEL:
+      MAX_SEQ_LEN: -1
+      TEMP_NAME: logit_scale_video_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.05
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+  
+  -
+    NAME: MomentsInTime
+    DATASETS:
+      TRAIN: 'VideoDataSet'
+      TASK_TYPE: 'video_classification'
+      DATASET_NAME: 'MiT'
+      TARGET_SET: ['MomentsInTime']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 68
+      TEST_BATCH_SIZE: 8
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/MomentsInTime'
+      ANNO_FOLDER:  'open_source_dataset/MomentsInTime'
+      S3_PATH:  's3://MomentsInTime/'
+      FRAMES_PER_CLIP: 3
+      STRIDE: 32
+      FILE_EXTENSION: ''
+      ANNO_FILE: 'annotation.json'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 0.2
+    
+    MODEL:
+      MAX_SEQ_LEN: -1
+      TEMP_NAME: logit_scale_video_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.05
+    INFERENCE:
+      NAME: 'MiTEvaler'
+      ID_KEY: 'video_name'
+      VALUE: 'label'
+      VAL_ANNFILE: 'open_source_dataset/MomentsInTime/annotation.json'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+      NUM_VIEWS: 1
+
+  -
+    NAME: bookswiki_pretrain
+    DATASETS:
+      TRAIN: 'GeneralCorpusDataset'
+      TASK_TYPE: 'text_mlm'
+      DATASET_NAME: 'BooksWiki'
+      TARGET_SET: ['Vocab_Word']
+      VERSION: 'v2'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  'open_source_dataset/text_corpus' # 'open_source_dataset/bert_pretrain_data/bookswiki'
+      # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      SEQ_PER_SAMPLE: 128
+      MIN_SEQ_PER_SAMPLE: 128
+      APPEND_EOS: True
+      ONE_STREAM: False
+      SAMPLING_WEIGHT: 2.75
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 128
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.25
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+  -
+    NAME: yfcc_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'YFCC'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5840
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC12M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5057
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC3M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.26295
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+  
+  -
+    NAME: vg_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'VG'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1766
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: True
+
+
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1144
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 1.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  -
+    NAME: sbu_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'SBU'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1383
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: yfcc_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'YFCC'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5840
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.25
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC12M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5057
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.25
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC3M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.26295
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.25
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: vg_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'VG'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1766
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.25
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: mscoco_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'MSCOCO'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1144
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.25
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      NAME: 'RetrievalEvaler'
+      VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
+      TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
+      GENERATION_MODE: False
+  
+  -
+    NAME: sbu_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'SBU'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1383
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.25
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+    UNIFY_QKV: True
+  
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  CHECKPOINT_FILETER: False 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+  OLD_CHECKPONT: True 
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 32
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 45000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 10000000
+  BASE_LR: 0.00002
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 5000
+  MIN_LR: 0.000001
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+

configs/BERT_L12_H768_experiments/16tasks_training_basemoe_stage1_56gpu.yaml

@@ -0,0 +1,733 @@
+_BASE_: "base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet22k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_22k_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  - 
+    NAME: 'MomentsInTime'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/MiT_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+  - 
+    NAME: 'Kinetics700'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/k700_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+TASKS:
+
+  - 
+    NAME: imagenet22k
+    DATASETS:
+      TRAIN: 'ImageNet22KDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet22k'
+      TARGET_SET: ['ImageNet22k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 720
+      # TEST_BATCH_SIZE: 2
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/imagenet22k'
+      S3_PATH: 'cluster2:s3://imagenet22k'
+      ANNO_FOLDER:  'open_source_dataset/'
+      SAMPLING_WEIGHT: 2.486
+      MIXUP: 0.8
+      CUTMIX: 1.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 21842
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+
+  -
+    NAME: K700_retrieve
+    DATASETS:
+      TRAIN: 'VideoDataSet'
+      TASK_TYPE: 'video_classification'
+      DATASET_NAME: 'K700'
+      TARGET_SET: ['Kinetics700']
+    DATALOADER: 
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 24 
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/K700'
+      ANNO_FOLDER:  'open_source_dataset/K700'
+      S3_PATH: 's3://K700/'
+      FRAMES_PER_CLIP: 4
+      STRIDE: 32
+      FILE_EXTENSION: ''
+      ANNO_FILE: 'annotation.json'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 0.76
+    
+    MODEL:
+      MAX_SEQ_LEN: -1
+      TEMP_NAME: logit_scale_video_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.1
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+  
+  -
+    NAME: MomentsInTime
+    DATASETS:
+      TRAIN: 'VideoDataSet'
+      TASK_TYPE: 'video_classification'
+      DATASET_NAME: 'MiT'
+      TARGET_SET: ['MomentsInTime']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 112
+      TEST_BATCH_SIZE: 8
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/MomentsInTime'
+      ANNO_FOLDER:  'open_source_dataset/MomentsInTime'
+      S3_PATH:  's3://MomentsInTime/'
+      FRAMES_PER_CLIP: 3
+      STRIDE: 32
+      FILE_EXTENSION: ''
+      ANNO_FILE: 'annotation.json'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 0.44
+    
+    MODEL:
+      MAX_SEQ_LEN: -1
+      TEMP_NAME: logit_scale_video_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.1
+    INFERENCE:
+      NAME: 'MiTEvaler'
+      ID_KEY: 'video_name'
+      VALUE: 'label'
+      VAL_ANNFILE: 'open_source_dataset/MomentsInTime/annotation.json'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+      NUM_VIEWS: 1
+
+  -
+    NAME: bookswiki_pretrain
+    DATASETS:
+      TRAIN: 'GeneralCorpusDataset'
+      TASK_TYPE: 'text_mlm'
+      DATASET_NAME: 'BooksWiki'
+      TARGET_SET: ['Vocab_Word']
+      VERSION: 'v2'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  'open_source_dataset/text_corpus' # 'open_source_dataset/bert_pretrain_data/bookswiki'
+      # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      SEQ_PER_SAMPLE: 128
+      MIN_SEQ_PER_SAMPLE: 128
+      APPEND_EOS: True
+      ONE_STREAM: False
+      SAMPLING_WEIGHT: 2.75
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 128
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+  -
+    NAME: yfcc_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'YFCC'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: True
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5840
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC12M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5057
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC3M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.26295
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+  
+  -
+    NAME: vg_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'VG'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1766
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: True
+
+
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1144
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 1.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  -
+    NAME: sbu_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'SBU'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 300
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1383
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: yfcc_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'YFCC'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: True
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5840
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC12M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5057
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC3M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.26295
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: vg_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'VG'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1766
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: mscoco_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'MSCOCO'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1144
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      NAME: 'RetrievalEvaler'
+      VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
+      TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
+      GENERATION_MODE: False
+  
+  -
+    NAME: sbu_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'SBU'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1383
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+    UNIFY_QKV: True
+  
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 160
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+  OLD_CHECKPONT: True 
+
+  LAYER_SCALE_FP32: True
+  GATE_FP32: False
+  TAG_TRANSFORM_FP32: False
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 32
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 230000
+  CHECKPOINT_PERIOD: 10000
+  EVAL_PERIOD: 10000000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.2
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 10000
+  MIN_LR: 0.000001
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+MOE: 
+  MOE: True 
+  MOE_TYPE: 'attribute'
+  TAG_Transform: True
+  ATTRIBUTE_LENGTH: 8
+  EP_WORLD_SIZE: 1 # tag moe only 
+  NUM_EXPERTS: 8
+  TOP_K: 2
+  CAPACITY_FACTOR: 3.0 
+  EVAL_MIN_CAPACITY: 4.0
+  MIN_CAPACITY: 4
+  NOISY_GATE_POLICY: 'vmoe'
+  MOE_PARAM_GROUP: True 
+  MOE_EXPERT_TYPE: 'FFN,SA'
+  SA_LINEAR_OUT_MOE: True
+  MOE_EXPERT_LOCATION: 'odd' # 'odd'
+  # MOE_LAYER_START_IDX: 3
+  # MOE_LAYER_END_IDX: 21
+  # MOE_LAYER_START_IDX: 18
+  # MOE_LAYER_END_IDX: 12 
+  BATCH_PRIO: True 
+  USE_TUTEL: True
+  FFN_SHARE_GATE_DECISION: True

configs/BERT_L12_H768_experiments/16tasks_training_basemoe_stage2_56gpu.yaml

@@ -0,0 +1,744 @@
+_BASE_: "base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet22k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_22k_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  - 
+    NAME: 'MomentsInTime'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/MiT_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+  - 
+    NAME: 'Kinetics700'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/k700_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+TASKS:
+
+  - 
+    NAME: imagenet22k
+    DATASETS:
+      TRAIN: 'ImageNet22KDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet22k'
+      TARGET_SET: ['ImageNet22k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 440
+      # TEST_BATCH_SIZE: 2
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/imagenet22k'
+      S3_PATH: 'cluster2:s3://imagenet22k'
+      ANNO_FOLDER:  'open_source_dataset/'
+      SAMPLING_WEIGHT: 2.486
+      MIXUP: 0.0
+      CUTMIX: 0.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 21842
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      LABELSMOOTHING: 0.1
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+      # VAL_ANNFILE: '/mnt/lustrenew/lihao2/projects/xmodaler_2/val_debug.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+
+  -
+    NAME: K700_retrieve
+    DATASETS:
+      TRAIN: 'VideoDataSet'
+      TASK_TYPE: 'video_classification'
+      DATASET_NAME: 'K700'
+      TARGET_SET: ['Kinetics700']
+    DATALOADER: 
+      TRAIN_BATCH_SIZE: 12
+      TEST_BATCH_SIZE: 24 
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/K700'
+      ANNO_FOLDER:  'open_source_dataset/K700'
+      S3_PATH: 's3://K700/'
+      FRAMES_PER_CLIP: 8
+      STRIDE: 32
+      FILE_EXTENSION: ''
+      ANNO_FILE: 'annotation.json'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 1.0
+    
+    MODEL:
+      MAX_SEQ_LEN: -1
+      TEMP_NAME: logit_scale_video_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.1
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+  
+  -
+    NAME: MomentsInTime
+    DATASETS:
+      TRAIN: 'VideoDataSet'
+      TASK_TYPE: 'video_classification'
+      DATASET_NAME: 'MiT'
+      TARGET_SET: ['MomentsInTime']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 68
+      TEST_BATCH_SIZE: 8
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/MomentsInTime'
+      ANNO_FOLDER:  'open_source_dataset/MomentsInTime'
+      S3_PATH:  's3://MomentsInTime/'
+      FRAMES_PER_CLIP: 3
+      STRIDE: 32
+      FILE_EXTENSION: ''
+      ANNO_FILE: 'annotation.json'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 0.2
+    
+    MODEL:
+      MAX_SEQ_LEN: -1
+      TEMP_NAME: logit_scale_video_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.1
+    INFERENCE:
+      NAME: 'MiTEvaler'
+      ID_KEY: 'video_name'
+      VALUE: 'label'
+      VAL_ANNFILE: 'open_source_dataset/MomentsInTime/annotation.json'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+      NUM_VIEWS: 1
+
+  -
+    NAME: bookswiki_pretrain
+    DATASETS:
+      TRAIN: 'GeneralCorpusDataset'
+      TASK_TYPE: 'text_mlm'
+      DATASET_NAME: 'BooksWiki'
+      TARGET_SET: ['Vocab_Word']
+      VERSION: 'v2'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  'open_source_dataset/text_corpus' # 'open_source_dataset/bert_pretrain_data/bookswiki'
+      # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      SEQ_PER_SAMPLE: 128
+      MIN_SEQ_PER_SAMPLE: 128
+      APPEND_EOS: True
+      ONE_STREAM: False
+      SAMPLING_WEIGHT: 2.75
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 128
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+  -
+    NAME: yfcc_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'YFCC'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5840
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC12M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5057
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC3M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.26295
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+  
+  -
+    NAME: vg_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'VG'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1766
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: True
+
+
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1144
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 1.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  -
+    NAME: sbu_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'SBU'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1383
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: yfcc_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'YFCC'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5840
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC12M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5057
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC3M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.26295
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: vg_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'VG'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1766
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: mscoco_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'MSCOCO'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1144
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      NAME: 'RetrievalEvaler'
+      VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
+      TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
+      GENERATION_MODE: False
+  
+  -
+    NAME: sbu_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'SBU'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1383
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+    UNIFY_QKV: True
+  
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  CHECKPOINT_FILETER: False 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+  OLD_CHECKPONT: True 
+
+  LAYER_SCALE_FP32: True
+  GATE_FP32: False
+  TAG_TRANSFORM_FP32: False
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 32
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 50000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 10000000
+  BASE_LR: 0.00002
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 5000
+  MIN_LR: 0.000001
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+MOE: 
+  MOE: True 
+  MOE_TYPE: 'attribute'
+  TAG_Transform: True
+  ATTRIBUTE_LENGTH: 8
+  EP_WORLD_SIZE: 1 # tag moe only 
+  NUM_EXPERTS: 8
+  TOP_K: 2
+  CAPACITY_FACTOR: 3.0 
+  EVAL_MIN_CAPACITY: 4.0
+  MIN_CAPACITY: 4
+  NOISY_GATE_POLICY: 'vmoe'
+  MOE_PARAM_GROUP: True 
+  MOE_EXPERT_TYPE: 'FFN,SA'
+  SA_LINEAR_OUT_MOE: True
+  MOE_EXPERT_LOCATION: 'odd' # 'odd'
+  # MOE_LAYER_START_IDX: 3
+  # MOE_LAYER_END_IDX: 21
+  # MOE_LAYER_START_IDX: 18
+  # MOE_LAYER_END_IDX: 12 
+  BATCH_PRIO: True 
+  USE_TUTEL: True
+  FFN_SHARE_GATE_DECISION: True

configs/BERT_L12_H768_experiments/16tasks_training_stage2_64gpu_v1.yaml

@@ -0,0 +1,750 @@
+_BASE_: "base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet22k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_22k_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+  - 
+    NAME: 'MomentsInTime'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/MiT_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+  - 
+    NAME: 'Kinetics700'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/k700_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+TASKS:
+
+  - 
+    NAME: imagenet22k
+    DATASETS:
+      TRAIN: 'ImageNet22KDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet22k'
+      TARGET_SET: ['ImageNet22k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 440
+      # TEST_BATCH_SIZE: 2
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/imagenet22k'
+      S3_PATH: 'cluster2:s3://imagenet22k'
+      ANNO_FOLDER:  'open_source_dataset/'
+      SAMPLING_WEIGHT: 2.486
+      MIXUP: 0.0
+      CUTMIX: 0.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 21842
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      LABELSMOOTHING: 0.1
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+      # VAL_ANNFILE: '/mnt/lustrenew/lihao2/projects/xmodaler_2/val_debug.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+
+  -
+    NAME: K700_retrieve
+    DATASETS:
+      TRAIN: 'VideoDataSet'
+      TASK_TYPE: 'video_classification'
+      DATASET_NAME: 'K700'
+      TARGET_SET: ['Kinetics700']
+    DATALOADER: 
+      TRAIN_BATCH_SIZE: 12
+      TEST_BATCH_SIZE: 24 
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/K700'
+      ANNO_FOLDER:  'open_source_dataset/K700'
+      S3_PATH: 's3://K700/'
+      FRAMES_PER_CLIP: 8
+      STRIDE: 32
+      FILE_EXTENSION: ''
+      ANNO_FILE: 'annotation.json'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 1.0
+    
+    MODEL:
+      MAX_SEQ_LEN: -1
+      TEMP_NAME: logit_scale_video_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.1
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+  
+  -
+    NAME: MomentsInTime
+    DATASETS:
+      TRAIN: 'VideoDataSet'
+      TASK_TYPE: 'video_classification'
+      DATASET_NAME: 'MiT'
+      TARGET_SET: ['MomentsInTime']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 68
+      TEST_BATCH_SIZE: 8
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/MomentsInTime'
+      ANNO_FOLDER:  'open_source_dataset/MomentsInTime'
+      S3_PATH:  's3://MomentsInTime/'
+      FRAMES_PER_CLIP: 3
+      STRIDE: 32
+      FILE_EXTENSION: ''
+      ANNO_FILE: 'annotation.json'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 0.2
+    
+    MODEL:
+      MAX_SEQ_LEN: -1
+      TEMP_NAME: logit_scale_video_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.1
+    INFERENCE:
+      NAME: 'MiTEvaler'
+      ID_KEY: 'video_name'
+      VALUE: 'label'
+      VAL_ANNFILE: 'open_source_dataset/MomentsInTime/annotation.json'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+      NUM_VIEWS: 1
+
+  -
+    NAME: bookswiki_pretrain
+    DATASETS:
+      TRAIN: 'GeneralCorpusDataset'
+      TASK_TYPE: 'text_mlm'
+      DATASET_NAME: 'BooksWiki'
+      TARGET_SET: ['Vocab_Word']
+      VERSION: 'v2'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 512
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  'open_source_dataset/text_corpus' # 'open_source_dataset/bert_pretrain_data/bookswiki'
+      # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      SEQ_PER_SAMPLE: 128
+      MIN_SEQ_PER_SAMPLE: 128
+      APPEND_EOS: True
+      ONE_STREAM: False
+      SAMPLING_WEIGHT: 2.75
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 128
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+  -
+    NAME: yfcc_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'YFCC'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5840
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC12M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5057
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'CC3M'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.26295
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+  
+  -
+    NAME: vg_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'VG'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1766
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: True
+
+
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1144
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 50
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 1.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+  -
+    NAME: sbu_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'SBU'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 200
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1383
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: yfcc_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'YFCC'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  'cluster2:s3://yfcc'
+      ANNO_FOLDER:  'open_source_dataset/yfcc'
+      ANNO_FILENAME: 'yfcc100m_subset_available_untokenized.json'
+      FEATS_FOLDER: 'open_source_dataset/yfcc/'
+      S3_PATH: 'cluster2:s3://yfcc/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5840
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc12m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC12M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER:  's3://cc12m/'
+      ANNO_FOLDER:  'open_source_dataset/c12m/'
+      ANNO_FILENAME: 'train_available.json'
+      FEATS_FOLDER: 'open_source_dataset/c12m/'
+      S3_PATH: 's3://cc12m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5057
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: cc3m_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'CC3M'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      S3_ANNO_FOLDER: 's3://cc3m/'
+      ANNO_FOLDER:  'open_source_dataset/cc3m/'
+      ANNO_FILENAME: 'train_spacy.json'
+      FEATS_FOLDER: 'open_source_dataset/cc3m/'
+      S3_PATH: 's3://cc3m/'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.26295
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: vg_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'VG'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
+      ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
+      S3_PATH: 's3://visual_genome/images'
+      ANNO_FILENAME: 'vg_captions_128filter.json'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1766
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  -
+    NAME: mscoco_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'MSCOCO'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1144
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      NAME: 'RetrievalEvaler'
+      VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
+      TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
+      GENERATION_MODE: False
+  
+  -
+    NAME: sbu_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'SBU'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 320
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 1
+      S3_ANNO_FOLDER: 's3://SBU/annotations'
+      ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
+      ANNO_FILENAME: 'subcaption.json'
+      FEATS_FOLDER: 'open_source_dataset/sbucaption/'
+      S3_PATH: 's3://SBU/images'
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.1383
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 50
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+    UNIFY_QKV: True
+  
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  CHECKPOINT_FILETER: False 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+  OLD_CHECKPONT: True 
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 32
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 45000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 10000000
+  BASE_LR: 0.00002
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 5000
+  MIN_LR: 0.000001
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+

configs/BERT_L12_H768_experiments/base_model_bert_l12_h768.yaml

@@ -0,0 +1,73 @@
+
+######################################### MODEL #########################################
+MODEL:
+  VOCAB_SIZE: 49411 # include <BOS>/<EOS>
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder_v3'
+  ENCODER_DIM: 768
+  DECODER: ''
+  DECODER_DIM: 768
+
+  PREDICTOR: 'EmbedClsAsRetrievalPredictor'
+  FEATURE_GATHER: True
+  LEARN_TEMP: True
+  PRED_USE_NORM: True
+  PRED_TEMPERATURE: 0.07
+
+  BertParamsInit: True
+
+  CLS_TOKEN: False
+
+  QUEUE_LEN: 1024
+  MAX_LABEL_LEN: 12
+
+  OUTPUT_PROJ: True # output projection
+ 
+
+# #################################### Token embedding ####################################
+  TOKEN_EMBED:
+    NAME: 'TokenBaseEmbedding'
+    DIM: 768
+    ACTIVATION: 'none'
+    USE_NORM: True
+    DROPOUT: 0.0
+    POSITION: 'NNEmbeddingEncoding'
+    POSITION_MAX_LEN: 512
+    TYPE_VOCAB_SIZE: 2
+
+# #################################### Visual embedding ####################################
+  VISUAL_EMBED:
+    NAME: 'none'
+
+# #################################### video embedding ####################################
+  VIDEO_EMBED:
+    NAME: 'VideoBaseEmbedding'
+    IN_DIM: 768
+    OUT_DIM: 768
+    ACTIVATION: 'none'
+    USE_NORM: True
+    DROPOUT: 0.0
+    TYPE_SIZE: 1 # video to encoder
+    POSITION: 'NNEmbeddingEncoding'
+    MAX_LENGTH: 1600
+    PATCH_SIZE_S: 16
+    PATCH_SIZE_T: 1
+    DIVIDE_ST_POS: True
+    USE_VISUAL_TOKENIZER: True
+    USE_VISUAL_POS: True
+    MAX_FRAMES: 8
+
+####################################### BERT ############################################
+  BERT:
+    DROP_PATH_PROB: 0.1
+    HIDDEN_SIZE: 768
+    HIDDEN_DROPOUT_PROB: 0.
+    HIDDEN_ACT: "gelu"
+    NUM_ATTENTION_HEADS: 12
+    INTERMEDIATE_SIZE: 3072
+    INTERMEDIATE_DROP: 0.
+    FFN_DROPOUT_PROB: 0.
+    ATTENTION_PROBS_DROPOUT_PROB: 0.
+    NUM_HIDDEN_LAYERS: 12
+    NUM_GENERATION_LAYERS: 0
+  

configs/BERT_L12_H768_experiments/bw_mlm_training.yaml

@@ -0,0 +1,309 @@
+_BASE_: "base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  # - 
+  #   NAME: 'ImageNet1k'
+  #   SHARED_TARGETS_CFG:
+  #     FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+  #     DISTRIBUTED: False
+
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+TASKS:
+
+  # - 
+  #   NAME: imagenet
+  #   DATASETS:
+  #     TRAIN: 'ImageNetDataset'
+  #     # VAL: 'ImageNetDataset'
+  #     TASK_TYPE: 'image_classification'
+  #     DATASET_NAME: 'ImageNet1k'
+  #     TARGET_SET: ['ImageNet1k']
+      
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 256
+  #     # TEST_BATCH_SIZE: 2
+  #     NUM_WORKERS: 4
+  #     FEATS_FOLDER: 'cluster2:s3://imagenet'
+  #     ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
+  #     SAMPLING_WEIGHT: 1.0
+  #     CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
+  #     MIXUP: 0.8
+  #     CUTMIX: 1.0
+  #     MIXUP_PROB: 1.0
+  #     MIXUP_SWITCH_PROB: 0.5
+  #     MIXUP_MODE: 'batch'
+  #     MIXUP_LABEL_SMOOTHING: 0.1
+  #   MODEL:
+  #     MAX_SEQ_LEN: -1
+  #     LABELS_NUM: 1000
+  #     TEMP_NAME: logit_scale_img_cls
+  #   LOSSES:
+  #     NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 1.0
+  #     REDUCTION: 'mean'
+  #     # LOSS_FP32: True
+  #   INFERENCE:
+  #     NAME: 'ImageNetEvaler'
+  #     ID_KEY: 'image_id'
+  #     VALUE: 'cls_logits'
+  #     VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+  #     TEST_ANNFILE: ''
+  #     GENERATION_MODE: False
+
+  -
+    NAME: bookswiki_pretrain
+    DATASETS:
+      TRAIN: 'GeneralCorpusDataset'
+      TASK_TYPE: 'text_mlm'
+      DATASET_NAME: 'BooksWiki'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 256
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 2
+      ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+      # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+      SEQ_PER_SAMPLE:  1
+      SAMPLER: NodeDistributed
+      CACHE_MODE: True
+      SEQ_PER_SAMPLE: 128
+      MIN_SEQ_PER_SAMPLE: 128
+      APPEND_EOS: True
+      ONE_STREAM: False
+      SAMPLING_WEIGHT: 1.0
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 128
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.5
+      REDUCTION: 'mean'
+    INFERENCE:
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+  # -
+  #   NAME: mscoco_caption
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     # VAL: 'ImageTextPairDataset'
+  #     # TEST: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_caption'
+  #     DATASET_NAME: 'MSCOCO'
+  #     TARGET_SET: ['Vocab_Word']
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 200
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 4
+  #     FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+  #     ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+  #     S3_PATH: 's3://coco/'
+  #     SEQ_PER_SAMPLE:  1
+      # CACHE_MODE: True
+      # CIRCULAR_CACHE_MODE: False
+      # ZIP_MODE: False
+      # CACHE_ORIGIN_IMAGE: False
+      # RANDOM_CAPTION: False
+      # AS_NUMPY_AS_POSSIBLE: False
+  #     SAMPLING_WEIGHT: 0.5
+  #     TRANSFORM: 'clip_transforms'
+  #     RANDOM_MASK: True
+  #   MODEL:
+  #     MAX_SEQ_LEN: 30
+  #     EVAL_MAX_SEQ_LEN: 21
+  #     TEMP_NAME: logit_scale_caption
+  #   LOSSES:
+  #     NAMES: ['CrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 0.5
+  #     REDUCTION: 'mean'
+  #   DECODE_STRATEGY:
+  #     NAME: 'CaptionBeamSearcherV3'
+  #     BEAM_SIZE: 2
+  #     # LEN_PENALTY: 1.0
+  #   INFERENCE:
+  #     NAME: 'COCOEvaler'
+  #     VOCAB: 'CLIP'
+  #     ID_KEY: 'image_id'
+  #     VALUE: 'caption'
+  #     VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+  #     TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+  #     GENERATION_MODE: True
+
+  # -
+  #   NAME: mscoco_retrieve
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     # TEST: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_retrieval'
+  #     DATASET_NAME: 'MSCOCO'
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 256
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 1
+  #     FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+  #     ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+  #     S3_PATH: 's3://coco/'
+      # CACHE_MODE: True
+      # CIRCULAR_CACHE_MODE: False
+      # ZIP_MODE: False
+      # CACHE_ORIGIN_IMAGE: False
+      # RANDOM_CAPTION: False
+      # AS_NUMPY_AS_POSSIBLE: False
+  #     SEQ_PER_SAMPLE:  1
+  #     SAMPLING_WEIGHT: 0.5
+  #     TRANSFORM: 'clip_transforms'
+  #   MODEL:
+  #     MAX_SEQ_LEN: 30
+  #     TEMP_NAME: logit_scale_retrieve
+  #   LOSSES:
+  #     NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+  #     LABELSMOOTHING: 0.1
+  #     LOSS_WEIGHT: 0.5
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     ID_KEY: 'image_id'
+  #     VALUE: 'caption'
+  #     NAME: 'RetrievalEvaler'
+  #     VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
+  #     TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
+  #     GENERATION_MODE: False
+
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  IN_TUNING: True # use IN1k instead of 22k
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+    UNIFY_QKV: True
+  
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 450000
+  CHECKPOINT_PERIOD: 50000
+  EVAL_PERIOD: 500000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 20000
+  MIN_LR: 0.000001
+
+
+
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+

configs/BERT_L12_H768_experiments/finetuning/GLUE_finetuning_experiments/GLUE_CoLA_mlm_finetune.yaml

@@ -0,0 +1,89 @@
+_BASE_: "base.yaml"
+
+SHARED_TARGETS:
+  - 
+    NAME: 'CoLA'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/GLUE_classnames/CoLA_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+TASKS:
+  -
+    NAME: CoLA
+    DATASETS:
+      TRAIN: 'GLUEDataset'
+      # TEST: 'GLUEDataset'
+      VAL: 'GLUEDataset'
+      TASK_TYPE: 'text_classification'
+      DATASET_NAME: 'CoLA'
+      TARGET_SET: ['CoLA']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 16
+      TEST_BATCH_SIZE: 64
+      NUM_WORKERS: 4
+      ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/glue_data/'
+
+
+    MODEL:
+      MAX_SEQ_LEN: 256
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      # LOSS_WEIGHT: 1
+      REDUCTION: 'mean'
+      LOSS_FP32: False
+    INFERENCE:
+      NAME: 'GLUEEvaler'
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+  
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+######################################### MODEL #########################################
+MODEL:
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  # EPOCH: 1
+  MAX_ITER: 5600
+  CHECKPOINT_PERIOD: 1000000
+  EVAL_PERIOD: 200
+  CHECKPOINT_MAX_SAVE: 1
+  BASE_LR: 0.00001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.1
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.98]
+  EPS: 1e-8
+  GRAD_CLIP: 0.5
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 20
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 400
+  MIN_LR: 0.00000001
+
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/finetuning/GLUE_finetuning_experiments/GLUE_MNLI_mlm_finetune.yaml

@@ -0,0 +1,89 @@
+_BASE_: "base.yaml"
+
+SHARED_TARGETS:
+  - 
+    NAME: 'MNLI'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/GLUE_classnames/MNLI_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+TASKS:
+  -
+    NAME: MNLI
+    DATASETS:
+      TRAIN: 'GLUEDataset'
+      # TEST: 'GLUEDataset'
+      VAL: 'GLUEDataset'
+      TASK_TYPE: 'text_classification'
+      DATASET_NAME: 'MNLI_Match'
+      TARGET_SET: ['MNLI']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 32
+      TEST_BATCH_SIZE: 32
+      NUM_WORKERS: 4
+      ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/glue_data/'
+
+    MODEL:
+      MAX_SEQ_LEN: 256
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      # LOSS_WEIGHT: 1
+      REDUCTION: 'mean'
+      LOSS_FP32: False
+    INFERENCE:
+      NAME: 'GLUEEvaler'
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+  
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+######################################### MODEL #########################################
+MODEL:
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  MAX_ITER: 125000
+  CHECKPOINT_PERIOD: 125000
+  EVAL_PERIOD: 5000
+  CHECKPOINT_MAX_SAVE: 1
+  BASE_LR: 0.00001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.1
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.98]
+  EPS: 1e-8
+  GRAD_CLIP: 0.5
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 20
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 7500
+  MIN_LR: 0.00000001
+
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/finetuning/GLUE_finetuning_experiments/GLUE_MRPC_mlm_finetune.yaml

@@ -0,0 +1,88 @@
+_BASE_: "base.yaml"
+
+SHARED_TARGETS:
+  - 
+    NAME: 'MRPC'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/GLUE_classnames/MRPC_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+TASKS:
+  -
+    NAME: MRPC
+    DATASETS:
+      TRAIN: 'GLUEDataset'
+      # TEST: 'GLUEDataset'
+      VAL: 'GLUEDataset'
+      TASK_TYPE: 'text_classification'
+      DATASET_NAME: 'MRPC'
+      TARGET_SET: ['MRPC']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 16
+      TEST_BATCH_SIZE: 64
+      NUM_WORKERS: 4
+      ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/glue_data/'
+
+    MODEL:
+      MAX_SEQ_LEN: 256
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      # LOSS_WEIGHT: 1
+      REDUCTION: 'mean'
+      LOSS_FP32: False
+    INFERENCE:
+      NAME: 'GLUEEvaler'
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+  
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+######################################### MODEL #########################################
+MODEL:
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  MAX_ITER: 2500
+  CHECKPOINT_PERIOD: 10000
+  EVAL_PERIOD: 100
+  CHECKPOINT_MAX_SAVE: 2
+  BASE_LR: 0.00001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.1
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.98]
+  EPS: 1e-8
+  GRAD_CLIP: 0.5
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 20
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 150
+  MIN_LR: 0.00000001
+
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/finetuning/GLUE_finetuning_experiments/GLUE_QNLI_mlm_finetune.yaml

@@ -0,0 +1,85 @@
+_BASE_: "base.yaml"
+
+SHARED_TARGETS:
+  - 
+    NAME: 'QNLI'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/GLUE_classnames/QNLI_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+TASKS:
+  -
+    NAME: QNLI
+    DATASETS:
+      TRAIN: 'GLUEDataset'
+      # TEST: 'GLUEDataset'
+      VAL: 'GLUEDataset'
+      TASK_TYPE: 'text_classification'
+      DATASET_NAME: 'QNLI'
+      TARGET_SET: ['QNLI']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 16
+      TEST_BATCH_SIZE: 64
+      NUM_WORKERS: 4
+      ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/glue_data/'
+
+    MODEL:
+      MAX_SEQ_LEN: 256
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      # LOSS_WEIGHT: 1
+      REDUCTION: 'mean'
+      LOSS_FP32: False
+    INFERENCE:
+      NAME: 'GLUEEvaler'
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+######################################### MODEL #########################################
+MODEL:
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  MAX_ITER: 34000
+  CHECKPOINT_PERIOD: 200000
+  EVAL_PERIOD: 2000
+  CHECKPOINT_MAX_SAVE: 2
+  BASE_LR: 0.00001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.1
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.98]
+  EPS: 1e-8
+  GRAD_CLIP: 0.5
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 20
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 2000
+  MIN_LR: 0.00000001
+
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/finetuning/GLUE_finetuning_experiments/GLUE_QQP_mlm_finetune.yaml

@@ -0,0 +1,84 @@
+_BASE_: "base.yaml"
+
+SHARED_TARGETS:
+  - 
+    NAME: 'QQP'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/GLUE_classnames/QQP_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+TASKS:
+  -
+    NAME: QQP
+    DATASETS:
+      TRAIN: 'GLUEDataset'
+      # TEST: 'GLUEDataset'
+      VAL: 'GLUEDataset'
+      TASK_TYPE: 'text_classification'
+      DATASET_NAME: 'QQP'
+      TARGET_SET: ['QQP']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 32
+      TEST_BATCH_SIZE: 64
+      NUM_WORKERS: 4
+      ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/glue_data/'
+
+    MODEL:
+      MAX_SEQ_LEN: 256
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      # LOSS_WEIGHT: 1
+      REDUCTION: 'mean'
+      LOSS_FP32: False
+    INFERENCE:
+      NAME: 'GLUEEvaler'
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+######################################### MODEL #########################################
+MODEL:
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  MAX_ITER: 115000
+  CHECKPOINT_PERIOD: 200000
+  EVAL_PERIOD: 5000
+  CHECKPOINT_MAX_SAVE: 2
+  BASE_LR: 0.00001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.1
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.98]
+  EPS: 1e-8
+  GRAD_CLIP: 0.5
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 20
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 28000
+  MIN_LR: 0.00000001
+
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/finetuning/GLUE_finetuning_experiments/GLUE_RTE_mlm_finetune.yaml

@@ -0,0 +1,92 @@
+_BASE_: "base.yaml"
+
+SHARED_TARGETS:
+  - 
+    NAME: 'RTE'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/GLUE_classnames/RTE_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+
+TASKS:
+  -
+    NAME: RTE
+    DATASETS:
+      TRAIN: 'GLUEDataset'
+      # TEST: 'GLUEDataset'
+      VAL: 'GLUEDataset'
+      TASK_TYPE: 'text_classification'
+      DATASET_NAME: 'RTE'
+      TARGET_SET: ['RTE']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 16
+      TEST_BATCH_SIZE: 64
+      NUM_WORKERS: 4
+      ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/glue_data/'
+
+
+    MODEL:
+      MAX_SEQ_LEN: 256
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      # LOSS_WEIGHT: 1
+      REDUCTION: 'mean'
+      LOSS_FP32: False
+    INFERENCE:
+      NAME: 'GLUEEvaler'
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+  
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+######################################### MODEL #########################################
+MODEL:
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  MAX_ITER: 2500
+  CHECKPOINT_PERIOD: 10000
+  EVAL_PERIOD: 100
+  CHECKPOINT_MAX_SAVE: 2
+  BASE_LR: 0.00002
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.1
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.98]
+  EPS: 1e-8
+  GRAD_CLIP: 0.5
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 20
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 150
+  MIN_LR: 0.00000001
+
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/finetuning/GLUE_finetuning_experiments/GLUE_SST2_mlm_finetune.yaml

@@ -0,0 +1,89 @@
+_BASE_: "base.yaml"
+
+SHARED_TARGETS:
+  - 
+    NAME: 'SST-2'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/GLUE_classnames/SST-2_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+TASKS:
+  -
+    NAME: SST-2
+    DATASETS:
+      TRAIN: 'GLUEDataset'
+      # TEST: 'GLUEDataset'
+      VAL: 'GLUEDataset'
+      TASK_TYPE: 'text_classification'
+      DATASET_NAME: 'SST-2'
+      TARGET_SET: ['SST-2']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 32
+      TEST_BATCH_SIZE: 64
+      NUM_WORKERS: 4
+      ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/glue_data/'
+
+    MODEL:
+      MAX_SEQ_LEN: 256
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      # LOSS_WEIGHT: 1
+      REDUCTION: 'mean'
+      LOSS_FP32: False
+    INFERENCE:
+      NAME: 'GLUEEvaler'
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+  
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+######################################### MODEL #########################################
+MODEL:
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  MAX_ITER: 22000
+  CHECKPOINT_PERIOD: 100000
+  EVAL_PERIOD: 1000
+  CHECKPOINT_MAX_SAVE: 2
+  BASE_LR: 0.00001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.1
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.98]
+  EPS: 1e-8
+  GRAD_CLIP: 0.5
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 20
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 1500
+  MIN_LR: 0.00000001
+
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/finetuning/GLUE_finetuning_experiments/base.yaml

@@ -0,0 +1,22 @@
+_BASE_: "../../base_model_bert_l12_h768.yaml"
+
+
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3

configs/BERT_L12_H768_experiments/finetuning/flickr30k_caption_finetuning.yaml

@@ -0,0 +1,151 @@
+_BASE_: "../base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+TASKS:
+  -
+    NAME: flickr30k_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'FLICKR'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 32
+      TEST_BATCH_SIZE: 8
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/flickr30k_images/flickr30k_images/flickr30k_images'
+      ANNO_FOLDER:  'open_source_dataset/flickr30k'
+      S3_PATH: "s3://open_dataset/flickr30k/flickr30k_images"
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 30
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      # NAMES: ['CrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 2.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/flickr30k/captions_val.json'
+      TEST_ANNFILE: 'open_source_dataset/flickr30k/captions_test.json'
+      GENERATION_MODE: True
+
+
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  
+  POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  CHECKPOINT_FILETER: False 
+  OLD_CHECKPONT: True 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 4000
+  CHECKPOINT_PERIOD: 50000
+  EVAL_PERIOD: 500
+  BASE_LR: 0.000002
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.0001
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 500
+  MIN_LR: 0.000001
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/finetuning/flickr30k_retrieval_finetuning.yaml

@@ -0,0 +1,132 @@
+_BASE_: "../base_model_bert_l12_h768.yaml"
+
+
+TASKS:
+
+  
+  
+  -
+    NAME: flickr30k_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'FLICKR'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 256
+      TEST_BATCH_SIZE: 128
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/flickr30k_images/flickr30k_images/flickr30k_images'
+      ANNO_FOLDER:  'open_source_dataset/flickr30k'
+      S3_PATH: 's3://open_dataset/flickr30k/flickr30k_images'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      NAME: 'RetrievalEvaler'
+      GENERATION_MODE: False
+
+  
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  
+  POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  CHECKPOINT_FILETER: False 
+  OLD_CHECKPONT: True 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 5000
+  CHECKPOINT_PERIOD: 50000
+  EVAL_PERIOD: 500
+  BASE_LR: 0.000005
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.0001
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 200
+  MIN_LR: 0.000001
+
+find_unused_parameters: true
+

configs/BERT_L12_H768_experiments/finetuning/in1k_training.yaml

@@ -0,0 +1,135 @@
+_BASE_: "../base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet1k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+
+
+TASKS:
+
+  - 
+    NAME: imagenet
+    DATASETS:
+      TRAIN: 'ImageNetDataset'
+      VAL: 'ImageNetDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet1k'
+      TARGET_SET: ['ImageNet1k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 128
+      TEST_BATCH_SIZE: 256
+      NUM_WORKERS: 4 # will be used as numworker for testing loader 
+      FEATS_FOLDER: 'open_source_dataset/imagenet'
+      S3_PATH: 'cluster2:s3://imagenet'
+      ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
+      SAMPLING_WEIGHT: 1.0
+      CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
+      MIXUP: 0.0
+      CUTMIX: 0.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 1000
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      LABELSMOOTHING: 0.1
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  
+  POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  CHECKPOINT_FILETER: False 
+  OLD_CHECKPONT: True 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 20000
+  CHECKPOINT_PERIOD: 20000
+  EVAL_PERIOD: 2000
+  BASE_LR: 0.00002
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.00000001
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.999]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  LOSS_SCALE_WINDOW: 200
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 2000
+  MIN_LR: 0.00000001
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/finetuning/in1k_training_384inputsize.yaml

@@ -0,0 +1,134 @@
+_BASE_: "../base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet1k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+
+
+TASKS:
+
+  - 
+    NAME: imagenet
+    DATASETS:
+      TRAIN: 'ImageNetDataset'
+      VAL: 'ImageNetDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet1k'
+      TARGET_SET: ['ImageNet1k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 256
+      NUM_WORKERS: 4 # will be used as numworker for testing loader 
+      FEATS_FOLDER: 'open_source_dataset/imagenet'
+      S3_PATH: 'cluster2:s3://imagenet'
+      ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
+      SAMPLING_WEIGHT: 1.0
+      CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
+      MIXUP: 0.0
+      CUTMIX: 0.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 1000
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      LABELSMOOTHING: 0.1
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 384
+  PATCH_SIZE: 16
+  POSEMBED_SCALE: !!python/object/apply:eval ["160/384"]
+  CHECKPOINT_FILETER: False 
+  OLD_CHECKPONT: True 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 40000
+  CHECKPOINT_PERIOD: 40000
+  EVAL_PERIOD: 2000
+  BASE_LR: 0.00002
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.000001
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.999]
+  EPS: 1e-6
+  GRAD_CLIP: 0.0
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  LOSS_SCALE_WINDOW: 200
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 4000
+  MIN_LR: 0.000001
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/finetuning/k400_training.yaml

@@ -0,0 +1,133 @@
+_BASE_: "../base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'Kinetics400'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/k400_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+
+
+TASKS:
+
+  -
+    NAME: K400_retrieve
+    DATASETS:
+      TRAIN: 'VideoDataSet'
+      VAL: 'VideoDataSet'
+      TASK_TYPE: 'video_classification'
+      DATASET_NAME: 'K400'
+      TARGET_SET: ['Kinetics400']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 8 # 256
+      TEST_BATCH_SIZE: 4 # debug
+      NUM_WORKERS: 4 # debug 4
+      FEATS_FOLDER: 'open_source_dataset/K400_official'
+      ANNO_FOLDER:  'open_source_dataset/K400_official'
+      S3_PATH: 's3://K400/'
+      FRAMES_PER_CLIP: 8
+      STRIDE: 32
+      FILE_EXTENSION: ''
+      ANNO_FILE: 'annotation.json'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 1.0
+      MULTI_VEIW_NUM: 4
+      MULTI_VEIW: 'v2'
+    MODEL:
+      MAX_SEQ_LEN: -1
+      TEMP_NAME: logit_scale_video_cls
+    LOSSES:
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+    INFERENCE:
+      NAME: 'MiTEvaler'
+      ID_KEY: 'video_name'
+      VALUE: 'label'
+      VAL_ANNFILE: 'open_source_dataset/K400_official/annotation.json'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+      NUM_VIEWS: 1
+      
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  
+  POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  CHECKPOINT_FILETER: False 
+  OLD_CHECKPONT: True 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 40000
+  CHECKPOINT_PERIOD: 50000
+  EVAL_PERIOD: 2000
+  BASE_LR: 0.000005
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.0001
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  LOSS_SCALE_WINDOW: 200
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 2000
+  MIN_LR: 0.000001
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/finetuning/mscoco_caption_finetuning.yaml

@@ -0,0 +1,150 @@
+_BASE_: "../base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+TASKS:
+  -
+    NAME: mscoco_caption
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      # VAL: 'ImageTextPairDataset'
+      TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_caption'
+      DATASET_NAME: 'MSCOCO'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 32
+      TEST_BATCH_SIZE: 8
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+      RANDOM_MASK: True
+    MODEL:
+      MAX_SEQ_LEN: 30
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 2.0
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 2.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+      TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+      GENERATION_MODE: True
+
+
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.2
+    DROP_PATH_PROB_FIXED: True
+
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  
+  POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  CHECKPOINT_FILETER: False 
+  OLD_CHECKPONT: True 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 10000
+  CHECKPOINT_PERIOD: 50000
+  EVAL_PERIOD: 500
+  BASE_LR: 0.00002
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.0001
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 500
+  MIN_LR: 0.000001
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/finetuning/mscoco_retrieval_finetuning.yaml

@@ -0,0 +1,132 @@
+_BASE_: "../base_model_bert_l12_h768.yaml"
+
+
+TASKS:
+
+  
+  
+  -
+    NAME: mscoco_retrieve
+    DATASETS:
+      TRAIN: 'ImageTextPairDataset'
+      TEST: 'ImageTextPairDataset'
+      TASK_TYPE: 'image_retrieval'
+      DATASET_NAME: 'MSCOCO'
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 256
+      TEST_BATCH_SIZE: 128
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+      S3_PATH: 's3://coco/'
+      SEQ_PER_SAMPLE:  1
+      CACHE_MODE: True
+      CIRCULAR_CACHE_MODE: False
+      ZIP_MODE: False
+      CACHE_ORIGIN_IMAGE: False
+      RANDOM_CAPTION: False
+      AS_NUMPY_AS_POSSIBLE: False
+      SAMPLING_WEIGHT: 0.5
+      TRANSFORM: 'clip_transforms'
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      NAME: 'RetrievalEvaler'
+      GENERATION_MODE: False
+
+  
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.2
+    DROP_PATH_PROB_FIXED: True
+
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  
+  POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  CHECKPOINT_FILETER: False 
+  OLD_CHECKPONT: True 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 10000
+  CHECKPOINT_PERIOD: 50000
+  EVAL_PERIOD: 500
+  BASE_LR: 0.000005
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.0001
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 500
+  MIN_LR: 0.000001
+
+find_unused_parameters: true
+

configs/BERT_L12_H768_experiments/finetuning/msvd_caption_finetuning.yaml

@@ -0,0 +1,144 @@
+_BASE_: "../base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  -
+    NAME: 'Vocab_Word'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+TASKS:
+  -
+    NAME: msvd_caption
+    DATASETS:
+      TRAIN: 'MSVDDataset'
+      TEST: 'MSVDDataset'
+      TASK_TYPE: 'video_caption'
+      DATASET_NAME: 'MSVDDataset'
+      TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 2 #6
+      TEST_BATCH_SIZE: 4
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/msvd_dataset/YouTubeClips'
+      ANNO_FOLDER:  'open_source_dataset/msvd_dataset/new_annotations'
+      STRIDE: 32
+      FRAMES_PER_CLIP: 4
+      S3_PATH: 's3://msvd/YouTubeClips/'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 1.0
+    MODEL:
+      MAX_SEQ_LEN: 30
+      EVAL_MAX_SEQ_LEN: 21
+      TEMP_NAME: logit_scale_caption
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      # NAMES: ['CrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    DECODE_STRATEGY:
+      NAME: 'CaptionBeamSearcherV3'
+      BEAM_SIZE: 2
+      # LEN_PENALTY: 2.0
+    INFERENCE:
+      NAME: 'COCOEvaler'
+      VOCAB: 'CLIP'
+      ID_KEY: 'image_id'
+      VALUE: 'caption'
+      VAL_ANNFILE: 'open_source_dataset/msvd_dataset/new_annotations/caption_msvd_val_cocostyle.json'
+      TEST_ANNFILE: 'open_source_dataset/msvd_dataset/new_annotations/caption_msvd_test_cocostyle.json'
+      GENERATION_MODE: True
+
+
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  
+  POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  CHECKPOINT_FILETER: False 
+  OLD_CHECKPONT: True 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 1000
+  CHECKPOINT_PERIOD: 500
+  EVAL_PERIOD: 200
+  BASE_LR: 0.00002
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.0001
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 100
+  MIN_LR: 0.000001
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/finetuning/msvd_retrieval_finetuning.yaml

@@ -0,0 +1,129 @@
+_BASE_: "../base_model_bert_l12_h768.yaml"
+
+
+TASKS:
+
+  
+  
+  -
+    NAME: msvd_retrieval
+    DATASETS:
+      TRAIN: 'MSVDDataset'
+      TEST: 'MSVDDataset'
+      TASK_TYPE: 'video_retrieval'
+      DATASET_NAME: 'MSVDDataset'
+      # TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 16
+      TEST_BATCH_SIZE: 8
+      NUM_WORKERS: 8
+      FEATS_FOLDER: 'open_source_dataset/msvd_dataset/YouTubeClips'
+      ANNO_FOLDER:  'open_source_dataset/msvd_dataset/new_annotations'
+      STRIDE: 32
+      FRAMES_PER_CLIP: 4
+      S3_PATH: 's3://msvd/YouTubeClips/'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 1.0
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      NAME: 'RetrievalEvaler'
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+  
+  VIDEO_EMBED:
+    MAX_FRAMES: 8
+
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  
+  # POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  CHECKPOINT_FILETER: False 
+  OLD_CHECKPONT: True 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 8
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 2000
+  CHECKPOINT_PERIOD: 5000
+  EVAL_PERIOD: 200
+  BASE_LR: 0.000005
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.0001
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 200
+  MIN_LR: 0.000001
+
+find_unused_parameters: true
+

configs/BERT_L12_H768_experiments/finetuning/msvd_retrieval_finetuning_frames8.yaml

@@ -0,0 +1,125 @@
+_BASE_: "../base_model_bert_l12_h768.yaml"
+
+
+TASKS:
+
+  
+  
+  -
+    NAME: msvd_retrieval
+    DATASETS:
+      TRAIN: 'MSVDDataset'
+      TEST: 'MSVDDataset'
+      TASK_TYPE: 'video_retrieval'
+      DATASET_NAME: 'MSVDDataset'
+      # TARGET_SET: ['Vocab_Word']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 8
+      TEST_BATCH_SIZE: 8
+      NUM_WORKERS: 4
+      FEATS_FOLDER: 'open_source_dataset/msvd_dataset/YouTubeClips'
+      ANNO_FOLDER:  'open_source_dataset/msvd_dataset/new_annotations'
+      STRIDE: 32
+      FRAMES_PER_CLIP: 8
+      S3_PATH: 's3://msvd/YouTubeClips/'
+      TIMESFORMER_AUG: True
+      SAMPLING_WEIGHT: 1.0
+    MODEL:
+      MAX_SEQ_LEN: 30
+      TEMP_NAME: logit_scale_retrieve
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+    INFERENCE:
+      NAME: 'RetrievalEvaler'
+      GENERATION_MODE: False
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  
+  POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  CHECKPOINT_FILETER: False 
+  OLD_CHECKPONT: True 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 5000
+  CHECKPOINT_PERIOD: 50000
+  EVAL_PERIOD: 500
+  BASE_LR: 0.000005
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.0001
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 200
+  MIN_LR: 0.000001
+
+find_unused_parameters: true
+

configs/BERT_L12_H768_experiments/finetuning/vqa_finetuning_debug.yaml

@@ -0,0 +1,127 @@
+_BASE_: "../base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  
+  -
+    NAME: 'VQA_Answer'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/VQA_Answers_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: True
+
+TASKS:
+  -
+    NAME: vqa
+    DATASETS:
+      TRAIN: 'VQADataset'
+      VAL: 'VQADataset'
+      # TEST: 'VQADataset'
+      DATASET_NAME: 'VQA'
+      TASK_TYPE: 'vqa'
+      TARGET_SET: ['VQA_Answer']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 64
+      TEST_BATCH_SIZE: 128
+      NUM_WORKERS: 2
+      FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+      ANNO_FOLDER: 'open_source_dataset/VQA'
+      SEQ_PER_SAMPLE:  1
+      MAX_FEAT_NUM: 51
+      SAMPLING_WEIGHT: 1.0
+      TRANSFORM: 'clip_transforms'
+      DO_AS_GEN: True
+      SINGLE_CLASS: True
+    MODEL:
+      MAX_SEQ_LEN: 23
+      TEMP_NAME: logit_scale_downstream
+    LOSSES:
+      # not single class 
+      # NAMES: ['BCEWithLogits']
+      # LOSS_WEIGHT: 0.05
+      # for single class
+      NAMES: ['CrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 0.1
+    INFERENCE:
+      VOCAB: 'CLIP'
+      NAME: 'VQAEvaler'
+      ID_KEY: 'question_id'
+      VALUE: 'answer'
+      VAL_ANNFILE: 'open_source_dataset/VQA/val_target.pkl'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+
+
+######################################### Engine #########################################
+ENGINE:
+  NAME: 'UnifiedTrainer'
+
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  BERT:
+    DROP_PATH_PROB: 0.1
+
+
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+  TEMP_NAME: logit_scale_downstream
+  PRED_TEMPERATURE: 0.03
+  LEARN_TEMP: False
+  CLS_TOKEN: True
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+  
+  POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
+  CHECKPOINT_FILETER: False 
+  OLD_CHECKPONT: True 
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  # EPOCH: 1
+  MAX_ITER: 20000
+  CHECKPOINT_PERIOD: 1000
+  EVAL_PERIOD: 1000
+  CHECKPOINT_MAX_SAVE: 2
+  BASE_LR: 0.00004
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.999]
+  EPS: 1e-8
+  GRAD_CLIP: 0.0
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  LOSS_SCALE_WINDOW: 500
+
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 1000
+  MIN_LR: 0.00000001
+
+
+find_unused_parameters: true

configs/BERT_L12_H768_experiments/in1k_training.yaml

@@ -0,0 +1,310 @@
+_BASE_: "base_model_bert_l12_h768.yaml"
+
+SHARED_TARGETS:
+
+  - 
+    NAME: 'ImageNet1k'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+
+  # -
+  #   NAME: 'Vocab_Word'
+  #   SHARED_TARGETS_CFG:
+  #     FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
+  #     DISTRIBUTED: True
+
+TASKS:
+
+  - 
+    NAME: imagenet
+    DATASETS:
+      TRAIN: 'ImageNetDataset'
+      VAL: 'ImageNetDataset'
+      TASK_TYPE: 'image_classification'
+      DATASET_NAME: 'ImageNet1k'
+      TARGET_SET: ['ImageNet1k']
+      
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 128
+      TEST_BATCH_SIZE: 128
+      NUM_WORKERS: 4 # will be used as numworker for testing loader 
+      FEATS_FOLDER: 'open_source_dataset/imagenet'
+      S3_PATH: 'cluster2:s3://imagenet'
+      ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
+      SAMPLING_WEIGHT: 1.0
+      CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
+      MIXUP: 0.8
+      CUTMIX: 1.0
+      MIXUP_PROB: 1.0
+      MIXUP_SWITCH_PROB: 0.5
+      MIXUP_MODE: 'batch'
+      MIXUP_LABEL_SMOOTHING: 0.1
+    MODEL:
+      MAX_SEQ_LEN: -1
+      LABELS_NUM: 1000
+      TEMP_NAME: logit_scale_img_cls
+    LOSSES:
+      NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
+      LOSS_WEIGHT: 1.0
+      REDUCTION: 'mean'
+      # LOSS_FP32: True
+    INFERENCE:
+      NAME: 'ImageNetEvaler'
+      ID_KEY: 'image_id'
+      VALUE: 'cls_logits'
+      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
+      TEST_ANNFILE: ''
+      GENERATION_MODE: False
+
+  # -
+  #   NAME: bookswiki_pretrain
+  #   DATASETS:
+  #     TRAIN: 'GeneralCorpusDataset'
+  #     TASK_TYPE: 'text_mlm'
+  #     DATASET_NAME: 'BooksWiki'
+  #     TARGET_SET: ['Vocab_Word']
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 128
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 2
+  #     ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+  #     # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
+  #     SEQ_PER_SAMPLE:  1
+  #     SAMPLER: NodeDistributed
+  #     CACHE_MODE: True
+  #     SEQ_PER_SAMPLE: 128
+  #     MIN_SEQ_PER_SAMPLE: 128
+  #     APPEND_EOS: True
+  #     ONE_STREAM: False
+  #     SAMPLING_WEIGHT: 1.0
+  #     RANDOM_MASK: True
+  #   MODEL:
+  #     MAX_SEQ_LEN: 128
+  #     TEMP_NAME: logit_scale_text_mlm
+  #   LOSSES:
+  #     NAMES: ['CrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 0.33333
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     GENERATION_MODE: False
+
+  # -
+  #   NAME: mscoco_caption
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     # VAL: 'ImageTextPairDataset'
+  #     # TEST: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_caption'
+  #     DATASET_NAME: 'MSCOCO'
+  #     TARGET_SET: ['Vocab_Word']
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 64
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 4
+  #     FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+  #     ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+  #     S3_PATH: 's3://coco/'
+  #     SEQ_PER_SAMPLE:  1
+      # CACHE_MODE: True
+      # CIRCULAR_CACHE_MODE: False
+      # ZIP_MODE: False
+      # CACHE_ORIGIN_IMAGE: False
+      # RANDOM_CAPTION: False
+      # AS_NUMPY_AS_POSSIBLE: False
+  #     SAMPLING_WEIGHT: 1.0
+  #     TRANSFORM: 'clip_transforms'
+  #     RANDOM_MASK: True
+  #   MODEL:
+  #     MAX_SEQ_LEN: 50
+  #     EVAL_MAX_SEQ_LEN: 21
+  #     TEMP_NAME: logit_scale_caption
+  #   LOSSES:
+  #     NAMES: ['CrossEntropy', 'Accuracy']
+  #     LOSS_WEIGHT: 0.33333
+  #     REDUCTION: 'mean'
+  #   DECODE_STRATEGY:
+  #     NAME: 'CaptionBeamSearcherV3'
+  #     BEAM_SIZE: 2
+  #     # LEN_PENALTY: 1.0
+  #   INFERENCE:
+  #     NAME: 'COCOEvaler'
+  #     VOCAB: 'CLIP'
+  #     ID_KEY: 'image_id'
+  #     VALUE: 'caption'
+  #     VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
+  #     TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
+  #     GENERATION_MODE: True
+
+  # -
+  #   NAME: mscoco_retrieve
+  #   DATASETS:
+  #     TRAIN: 'ImageTextPairDataset'
+  #     # TEST: 'ImageTextPairDataset'
+  #     TASK_TYPE: 'image_retrieval'
+  #     DATASET_NAME: 'MSCOCO'
+  #   DATALOADER:
+  #     TRAIN_BATCH_SIZE: 100
+  #     TEST_BATCH_SIZE: 32
+  #     NUM_WORKERS: 1
+  #     FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
+  #     ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
+  #     S3_PATH: 's3://coco/'
+  #     SEQ_PER_SAMPLE:  1
+      # CACHE_MODE: True
+      # CIRCULAR_CACHE_MODE: False
+      # ZIP_MODE: False
+      # CACHE_ORIGIN_IMAGE: False
+      # RANDOM_CAPTION: False
+      # AS_NUMPY_AS_POSSIBLE: False
+  #     SAMPLING_WEIGHT: 1.0
+  #     TRANSFORM: 'clip_transforms'
+  #   MODEL:
+  #     MAX_SEQ_LEN: 50
+  #     TEMP_NAME: logit_scale_retrieve
+  #   LOSSES:
+  #     NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+  #     LABELSMOOTHING: 0.1
+  #     LOSS_WEIGHT: 1.0
+  #     REDUCTION: 'mean'
+  #   INFERENCE:
+  #     VOCAB: 'CLIP'
+  #     ID_KEY: 'image_id'
+  #     VALUE: 'caption'
+  #     NAME: 'RetrievalEvaler'
+  #     VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
+  #     TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
+  #     GENERATION_MODE: False
+
+
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+ 
+MODEL:
+  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
+  ENCODER: 'UnifiedBertEncoder'
+
+  IN_TUNING: True # use IN1k instead of 22k
+  SHARE_LAYERNORM: True
+  BERT:
+    NORMALIZE_DECISION: "BERTPre" 
+    DROP_PATH_PROB: 0.1
+    DROP_PATH_PROB_FIXED: True
+
+    UNIFY_QKV: True
+  
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+
+  MAEParamsInit: True
+  POSEMBEDFIX: True
+
+
+  IMG_INPUT_SIZE: 224
+  PATCH_SIZE: 16
+
+  LAYER_SCALE: True 
+  LAYER_SCALE_INIT: 1e-3
+
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 
+
+
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  TORCH_OPTIMIZER: True
+  PARAMS_SEPERATE: True
+  # PARAMS_GROUP: True
+  # EPOCH: 1
+  MAX_ITER: 200000
+  CHECKPOINT_PERIOD: 50000
+  EVAL_PERIOD: 500000
+  BASE_LR: 0.001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.05
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  WEIGHT_DECAY_EMBEDDING: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.95]
+  EPS: 1e-6
+  GRAD_CLIP: 0.1
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+
+  WRITE_PERIOD: 50
+  MIN_LOSS_SCLE: 2048.0
+  # BF16: False # True
+  # ZEROSTAGE: 2
+
+  LOSS_SCALE_WINDOW: 200
+
+
+
+
+
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 20000
+  MIN_LR: 0.000001
+
+
+
+
+####################################### evaluation ####################################### 
+INFERENCE:
+
+  VOCAB: 'CLIP'
+  ITER_BASED: True
+
+
+find_unused_parameters: true
+
+# ENCODERS:
+#   -
+#     NAME: VisualEncoder
+#     TYPE: VisualEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True
+
+#   -
+#     NAME: TextEncoder
+#     TYPE: TextEncoder
+#     DROP_PATH_PROB: 0.0
+#     HIDDEN_SIZE: 192
+#     HIDDEN_DROPOUT_PROB: 0.
+#     HIDDEN_ACT: "gelu"
+#     NUM_ATTENTION_HEADS: 3
+#     INTERMEDIATE_SIZE: 768
+#     INTERMEDIATE_DROP: 0.
+#     FFN_DROPOUT_PROB: 0.
+#     ATTENTION_PROBS_DROPOUT_PROB: 0.
+#     NUM_HIDDEN_LAYERS: 6
+#     NUM_GENERATION_LAYERS: 0
+#     DROP_PATH_PROB_FIXED: True 
+

configs/BERT_L12_H768_experiments/moe_finetuning/GLUE_finetuning_experiments/GLUE_CoLA_mlm_finetune.yaml

@@ -0,0 +1,89 @@
+_BASE_: "base.yaml"
+
+SHARED_TARGETS:
+  - 
+    NAME: 'CoLA-target'
+    SHARED_TARGETS_CFG:
+      FILE_PATH: 'open_source_dataset/GLUE_classnames/CoLA_class_name_CLIP_with_endoftext.pkl'
+      DISTRIBUTED: False
+TASKS:
+  -
+    NAME: CoLA
+    DATASETS:
+      TRAIN: 'GLUEDataset'
+      # TEST: 'GLUEDataset'
+      VAL: 'GLUEDataset'
+      TASK_TYPE: 'text_classification'
+      DATASET_NAME: 'CoLA'
+      TARGET_SET: ['CoLA-target']
+    DATALOADER:
+      TRAIN_BATCH_SIZE: 16
+      TEST_BATCH_SIZE: 64
+      NUM_WORKERS: 4
+      ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/glue_data/'
+
+
+    MODEL:
+      MAX_SEQ_LEN: 256
+      TEMP_NAME: logit_scale_text_mlm
+    LOSSES:
+      NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
+      LABELSMOOTHING: 0.1
+      # LOSS_WEIGHT: 1
+      REDUCTION: 'mean'
+      LOSS_FP32: False
+    INFERENCE:
+      NAME: 'GLUEEvaler'
+      VOCAB: 'CLIP'
+      GENERATION_MODE: False
+
+
+  
+
+ENGINE:
+  NAME: 'UnifiedTrainer'
+
+DATALOADER:
+  USE_WEIGHTED_SAMPLER: True
+  UNIFIED_DATASET: True 
+  NUM_WORKERS: 16
+
+######################################### MODEL #########################################
+MODEL:
+  MODEL_EMA: False
+  MODEL_EMA_DECAY: 0.9999
+  
+####################################### Optimizer #######################################
+SOLVER:
+  NAME: 'Adam'
+  # EPOCH: 1
+  MAX_ITER: 5600
+  CHECKPOINT_PERIOD: 1000000
+  EVAL_PERIOD: 200
+  CHECKPOINT_MAX_SAVE: 1
+  BASE_LR: 0.00001
+  BIAS_LR_FACTOR: 1.0
+  WEIGHT_DECAY: 0.1
+  WEIGHT_DECAY_NORM: 0.0
+  WEIGHT_DECAY_BIAS: 0.0
+  MOMENTUM: 0.9
+  DAMPENING: 0.0
+  NESTEROV: 0.0
+  BETAS: [0.9, 0.98]
+  EPS: 1e-8
+  GRAD_CLIP: 0.5
+  GRAD_CLIP_TYPE: 'norm'
+  ACCUM_ITER: 0
+  AMP_FP16: True
+  APEX_FP16: False # dangerous
+  WRITE_PERIOD: 20
+  
+####################################### lr scheduler ####################################### 
+LR_SCHEDULER:
+  NAME: 'WarmupCosine'
+  WARMUP: 400
+  MIN_LR: 0.00000001
+
+
+
+find_unused_parameters: true