[Feature] Support DeiT3

[okotaku]

Motivation

DeiT III: Revenge of the ViT
github

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.

[CLAassistant]

All committers have signed the CLA.

[okotaku]

@Ezra-Yu Accuracy is a bit lower than the official repo. Do you have any idea?

[Ezra-Yu]

Accuracy is a bit lower than the official repo. Do you have any idea?

Thanks very much! Deit3 is in our planning for the next month.

What is the specific accuracy gap?

[codecov]

Codecov Report

Base: 0.02% // Head: 91.58% // Increases project coverage by +91.55% 🎉

Coverage data is based on head (24f68cd) compared to base (b8b31e9).
Patch coverage: 43.26% of modified lines in pull request are covered.

Additional details and impacted files

@@             Coverage Diff              @@
##           dev-1.x    #1065       +/-   ##
============================================
+ Coverage     0.02%   91.58%   +91.55%     
============================================
  Files          121      128        +7     
  Lines         8217     9601     +1384     
  Branches      1368     1507      +139     
============================================
+ Hits             2     8793     +8791     
+ Misses        8215      622     -7593     
- Partials         0      186      +186     

Flag	Coverage Δ
unittests	91.58% <43.26%> (+91.55%)	⬆️

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

Impacted Files	Coverage Δ
mmcls/apis/inference.py	0.00% <0.00%> (ø)
mmcls/datasets/multi_label.py	100.00% <ø> (+100.00%)	⬆️
mmcls/datasets/dataset_wrappers.py	25.97% <22.53%> (+25.97%)	⬆️
mmcls/datasets/transforms/formatting.py	88.09% <95.65%> (+88.09%)	⬆️
mmcls/__init__.py	100.00% <100.00%> (+84.61%)	⬆️
mmcls/datasets/transforms/__init__.py	100.00% <100.00%> (+100.00%)	⬆️
mmcls/datasets/transforms/auto_augment.py	100.00% <100.00%> (+100.00%)	⬆️
mmcls/datasets/transforms/processing.py	98.81% <100.00%> (+98.81%)	⬆️
mmcls/engine/hooks/class_num_check_hook.py	100.00% <100.00%> (+100.00%)	⬆️
... and 126 more

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[okotaku]

@Ezra-Yu

Backbone	official repo	converted
deit3-b-384	85.0	85.03
deit3-b-384-21k	86.7	86.24
deit3-l-384	85.8	85.94
deit3-l-384-21k	87.7	87.22

[tonysy]

@okotaku Great thanks for your contribution, we would like to build connection with some community developers. If possible, would you like to share your preferred IM software? Maybe we can set a chat group to improve the future collaboration and communication.

[Ezra-Yu]

Have you checked the test_transforms, such crop_ratio and interpolation? If it is all right, it may also be related to the torch version

[okotaku]

@tonysy Hello, I have joined openmmlab slack workspace.
How about it?

[okotaku]

@Ezra-Yu Thanks to your advice I solved!

[Ezra-Yu]

Well done!

[mzr1996]

Seems this PR is denied to be modified by developers and I cannot push modification. Any ideas about it? @okotaku

[okotaku]

@mzr1996 I added you as collabolator. Please check it.