[Feature] Support ViTPose

[LareinaM]

Motivation

Add implementation of ViTPose on MMPose 1.0

Modification

• Add six config files
• Add layer decay optimizer
• In HeatmapHead, add parameter upsample, default to zero (no effect on previous codes), resize in BaseHead

BC-breaking (Optional)

Use cases (Optional)

Checklist

**Before

• I have read and followed the workflow indicated in the CONTRIBUTING.md to create this PR.
• Pre-commit or linting tools indicated in CONTRIBUTING.md are used to fix the potential lint issues.
• Bug fixes are covered by unit tests, the case that causes the bug should be added in the unit tests.
• New functionalities are covered by complete unit tests. If not, please add more unit tests to ensure correctness.
• The documentation has been modified accordingly, including docstring or example tutorials.

After PR:

• CLA has been signed and all committers have signed the CLA in this PR.

[jin-s13]

Have you tried training a base ViT model to check the accuracy?

[LareinaM]

Result of the current implementation

With classic decoder

Arch	Input Size	AP	AR
ViTPose-S	256x192	0.739	0.792
ViTPose-B	256x192	0.757	0.810
ViTPose-L	256x192	0.782	0.834
ViTPose-H	256x192	0.788	0.839

With simple decoder

Arch	Input Size	AP	AR
ViTPose-S	256x192	0.736	0.790
ViTPose-B	256x192	0.756	0.809
ViTPose-L	256x192	0.781	0.833
ViTPose-H	256x192	0.789	0.839

Result of original ViTPose implementation

With classic decoder

Model	Input Size	AP	AR
ViTPose-S	256x192	0.738	0.792
ViTPose-B	256x192	0.758	0.811
ViTPose-L	256x192	0.783	0.835
ViTPose-H	256x192	0.791	0.841

With simple decoder

Model	Input Size	AP	AR
ViTPose-S	256x192	0.735	0.789
ViTPose-B	256x192	0.755	0.809
ViTPose-L	256x192	0.782	0.834
ViTPose-H	256x192	0.789	0.840

[codecov]

Codecov Report

Patch coverage: 6.34% and project coverage change: -0.46 ⚠️

Comparison is base (d341f11) 82.22% compared to head (e22d6c8) 81.77%.

❗ Current head e22d6c8 differs from pull request most recent head 23b7838. Consider uploading reports for the commit 23b7838 to get more accurate results

Additional details and impacted files

@@             Coverage Diff             @@
##           dev-1.x    #1876      +/-   ##
===========================================
- Coverage    82.22%   81.77%   -0.46%     
===========================================
  Files          225      227       +2     
  Lines        13375    13438      +63     
  Branches      2269     2285      +16     
===========================================
- Hits         10998    10989       -9     
- Misses        1864     1933      +69     
- Partials       513      516       +3     

Flag	Coverage Δ
unittests	81.77% <6.34%> (-0.46%)	⬇️

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

Impacted Files	Coverage Δ
mmpose/engine/optim_wrappers/__init__.py	0.00% <0.00%> (ø)
...engine/optim_wrappers/layer_decay_optim_wrapper.py	0.00% <0.00%> (ø)
mmpose/models/heads/heatmap_heads/heatmap_head.py	82.19% <23.07%> (-5.78%)	⬇️
mmpose/models/heads/base_head.py	81.81% <33.33%> (-2.31%)	⬇️

... and 2 files with indirect coverage changes

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 in Codecov by Sentry.
📢 Do you have feedback about the report comment? Let us know in this issue.

[ly015]

The argument extra is convenient for extending the head class but may confuse users. We should keep a well-defined interface where every argument has a clear meaning and a detailed usage description in the docstring.

Here I think maybe we can use existing arguments conv_out_channels, conv_kernel_sizes, and has_final_layer to configure the final conv layers.

[LareinaM]

To keep the code clear and simple, would it be better to split the dictionary into two parameters, e.g. input_upsample (defaults to 0) and final_kernel_size (defaults to 1)?

[ly015]

I suggest adding a new argument, e.g. input_upsample or input_rescale.