Add pretrained Wide ResNet

[szagoruyko]

I trained WRN-50-2 and WRN-101-2 with master torchvision, which now allows making WRN models with simple width_per_group argument. I did not use the standard training procedure for ResNet though, here are the differences:

• SGD with cosine learning rate and warm restarts for 256 epochs (~0.2% to top1 accuracy)
• FP16 training with batchnorm in FP32 with apex O2

so the checkpoints are in torch.float16 to save space.

model	top1, top5 error
WRN-50-2	21.49, 5.91
WRN-101-2	21.16, 5.72

idk do we want these in torchvision? I could put them in wide-residual-networks instead.

[codecov-io]

Codecov Report

Merging #912 into master will increase coverage by 0.09%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master     #912      +/-   ##
==========================================
+ Coverage   63.87%   63.97%   +0.09%     
==========================================
  Files          66       66              
  Lines        5273     5279       +6     
  Branches      793      793              
==========================================
+ Hits         3368     3377       +9     
+ Misses       1673     1671       -2     
+ Partials      232      231       -1

Impacted Files	Coverage Δ
torchvision/models/resnet.py	88.27% <100%> (+0.45%)	⬆️
torchvision/transforms/transforms.py	82.73% <0%> (+0.6%)	⬆️

---

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 8616227...fdcb388. Read the comment docs.

[fmassa]

Thanks a lot for the PR Sergey!

This is almost good to merge.
Can you add an entry for it in https://github.com/pytorch/vision/blob/master/docs/source/models.rst, with the accuracies as well in the table?

Then I'll copy the model weights to the pytorch website and let you know

[szagoruyko]

Thanks for looking at it @fmassa , I added:

• docstrings for wide_resnet50_2 and wide_resnet101_2 functions (with explanation of diff vs ResNet)
• docs at docs/source/models.rst
• wide_resnet50_2 and wide_resnet101_2 are now properly imported

[fmassa]

Sorry for the delay in replying.

I've uploaded the pre-trained weights to

https://download.pytorch.org/models/wide_resnet50_2-95faca4d.pth
https://download.pytorch.org/models/wide_resnet101_2-32ee1156.pth

Can you update the URLs and fix the conflicts? Then it's good to go!

[szagoruyko]

@fmassa done, updated links and rebased on master

[fmassa]

LGTM, thanks!

Just waiting for CI to finish before merging the PR

[fmassa]

Thanks Sergey!