[Improvement] Support soft label for CrossEntropyLoss

[irvingzhang0512]

Motivation

Soft labels are required for data augmentation methods like mixup, cutmix, etc. For now, official CrossEntropyLoss doesn't support soft labels.

Details

• If number of classes is 4.
  • Hard labels example: [0, 1, 2]
  • Corresponding soft labels are [[1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0]]
  • Mixup/Cutmix may create soft labels like [[0.8, 0.2, 0, 0], [0.1, 0.9, 0, 0], [0, 0, 0.6, 0.4]]
• It's difficult to get a proper name for this pr and related args.
  • Sparse is borrowed from TensorFlow and MXNet, TensorFlow has CategoricalCrossentropy(accept non-sparse, one-hot like labels) and SparseCategorialCrossentropy(just like CrossEntropyLoss in pytorch).
  • BTW, one-hot like labels is more clear and meaningful for me.
• For now, this pr detects loss type by the shape of label.

TODO

• Modify CrossEntropyLoss in mmaction/models/losses/cross_entropy_loss.py.
• unittest
• changelog.

[codecov]

Codecov Report

Merging #625 (ed7dda2) into master (80a9ae1) will increase coverage by 0.01%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #625      +/-   ##
==========================================
+ Coverage   84.62%   84.63%   +0.01%     
==========================================
  Files         127      127              
  Lines        8948     8958      +10     
  Branches     1503     1506       +3     
==========================================
+ Hits         7572     7582      +10     
  Misses       1020     1020              
  Partials      356      356              

Flag	Coverage Δ
unittests	84.62% <100.00%> (+0.01%)	⬆️

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

Impacted Files	Coverage Δ
mmaction/models/losses/cross_entropy_loss.py	100.00% <100.00%> (ø)

---

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 80a9ae1...ca2bb88. Read the comment docs.

[innerlee]

Some minor comments.

[kennymckormick]

Thanks for the PR~ Besides have u find some settings in which soft labels work in action recognition? I have tried to find one but failed. If you find such a setting, you can add the config file which uses soft label for training and achieves better results.

[irvingzhang0512]

Thanks for the PR~ Besides have u find some settings in which soft labels work in action recognition? I have tried to find one but failed. If you find such a setting, you can add the config file which uses soft label for training and achieves better results.

I plan to implement mixup/cutmix in two weeks, which use soft label.