[Feature] barlowtwins

[xcnick]

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily get feedback. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

Please describe the motivation of this PR and the goal you want to achieve through this PR.

Modification

This PR implements barlow twins algorithm.

BC-breaking (Optional)

Does the modification introduce changes that break the backward compatibility of the downstream repositories?
If so, please describe how it breaks the compatibility and how the downstream projects should modify their code to keep compatibility with this PR.

Use cases (Optional)

If this PR introduces a new feature, it is better to list some use cases here and update the documentation.

Checklist

Before PR:

• Pre-commit or other linting tools are used to fix the potential lint issues.
• Bug fixes are fully covered by unit tests, the case that causes the bug should be added in the unit tests.
• The modification is covered by complete unit tests. If not, please add more unit test to ensure the correctness.
• The documentation has been modified accordingly, like docstring or example tutorials.

After PR:

• If the modification has potential influence on downstream or other related projects, this PR should be tested with those projects, like MMDet or MMSeg.
• CLA has been signed and all committers have signed the CLA in this PR.

[codecov]

Codecov Report

Merging #207 (86319d7) into dev_v0.9.0 (ed31de7) will increase coverage by 0.25%.
The diff coverage is 96.15%.

@@              Coverage Diff               @@
##           dev_v0.9.0     #207      +/-   ##
==============================================
+ Coverage       70.05%   70.30%   +0.25%     
==============================================
  Files             113      114       +1     
  Lines            4051     4099      +48     
  Branches          655      658       +3     
==============================================
+ Hits             2838     2882      +44     
- Misses           1089     1092       +3     
- Partials          124      125       +1     

Flag	Coverage Δ
unittests	70.30% <96.15%> (+0.25%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmselfsup/models/heads/latent_pred_head.py	95.83% <92.00%> (-4.17%)	⬇️
mmselfsup/models/algorithms/__init__.py	100.00% <100.00%> (ø)
mmselfsup/models/algorithms/barlowtwins.py	100.00% <100.00%> (ø)
mmselfsup/models/heads/__init__.py	100.00% <100.00%> (ø)
mmselfsup/datasets/utils.py	36.84% <0.00%> (-0.57%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update ed31de7...86319d7. Read the comment docs.

[fangyixiao18]

Thank you for your contribution. Do you have any results or models of barlowtwins based on this PR?

[xcnick]

Thank you for your contribution. Do you have any results or models of barlowtwins based on this PR?

Sorry, due to lack of computing resources, I only performed the pretraining steps on RTX2080Ti * 2, and observed loss decreasing. Therefore, the config parameters may need to be further adjusted.

I will use 4 * V100 to verify this PR later if possible.

[fangyixiao18]

Sorry, due to lack of computing resources, I only performed the pretraining steps on RTX2080Ti * 2, and observed loss decreasing. Therefore, the config parameters may need to be further adjusted.

I will use 4 * V100 to verify this PR later if possible.

OK, can you checkout a new branch from dev_v0.7.0 and re-create a PR to the dev_v0.7.0 branch? And if you have any updated results, you can push new commits to the new PR.
We look forward that your PR can be merged and the BarlowTwins can be released in version 0.7.0 at the begining of the next month.

[xcnick]

Sorry, due to lack of computing resources, I only performed the pretraining steps on RTX2080Ti * 2, and observed loss decreasing. Therefore, the config parameters may need to be further adjusted.
I will use 4 * V100 to verify this PR later if possible.

OK, can you checkout a new branch from dev_v0.7.0 and re-create a PR to the dev_v0.7.0 branch? And if you have any updated results, you can push new commits to the new PR. We look forward that your PR can be merged and the BarlowTwins can be released in version 0.7.0 at the begining of the next month.

OK.

[xcnick]

Using the parameters in the code for pretraining in 4*V100, the loss value is about 2000. The top1 acc is 40% when finetune in imagenet1k. So the model did not converge.
Could you help me for the model training?

[YuanLiuuuuuu]

Using the parameters in the code for pretraining in 4*V100, the loss value is about 2000. The top1 acc is 40% when finetune in imagenet1k. So the model did not converge. Could you help me for the model training?

Using the parameters in the code for pretraining in 4*V100, the loss value is about 2000. The top1 acc is 40% when finetune in imagenet1k. So the model did not converge. Could you help me for the model training?

If you do not have enough computation resources, you can choose a setting, which has lower computation overhead. Under this setting, you can run the official code and get the log. Finally, you can reproduce the results, according to the log.

[YuanLiuuuuuu]

We have already created a new branch, dev_v0.8.0, for next released version. Please re-submit a PR to merge all your modifications to that branch as soon as possible. Thanks a lot !

[YuanLiuuuuuu]

Using the parameters in the code for pretraining in 4*V100, the loss value is about 2000. The top1 acc is 40% when finetune in imagenet1k. So the model did not converge. Could you help me for the model training?

Please add the same decorator to your algorithm to pass CI.

[fangyixiao18]

Do you have any updated training results?

[xcnick]

Do you have any updated training results?

Sorry, the learning rate parameter is adjusted according to the official code, but the loss is around 2000. In contrast, the loss is around 500 in official log when training epoch is 100.

[Jiahao000]

Better to reorganize the algorithms in model zoo in their chronological order. Others LGTM.

[YuanLiuuuuuu]

@fangyixiao18 Please update index.rst in docs