[Feature] Support Wide ResNet

[yasu0001]

Motivation

Based on #678

Modification

Add two wide resnet config files for ImageNet.

[CLAassistant]

All committers have signed the CLA.

[Ezra-Yu]

Thank you for your contribution to mmclassification. Please sign the CLA license so that we can review your code.

We notice there are checkpoints for wide resnet. Could you use the script(as below) to convert torchvison ckpt into mmcls ckpt and make sure that the inference accuracy in Imagnet-1k are the same?

after convert and prepare the magnet-1k val dataset. run

python ./tools/test.py PATH_TO_WIDERESNET CKPT_PATH --metrics accuracy

torchvison_to_mmcls.py

import argparse
from collections import OrderedDict
from pathlib import Path

import torch


def convert(src, dst):
    print('Converting...')
    blobs = torch.load(src, map_location='cpu')
    converted_state_dict = OrderedDict()

    for key, value in blobs.items():
        if not key.startswith('fc'):
            converted_state_dict['backbone.' + key] = value
        else:
            converted_state_dict['head.' + key] = value

    torch.save(converted_state_dict, dst)
    print('Done!')


def main():
    parser = argparse.ArgumentParser(description='Convert model keys')
    parser.add_argument('src', help='src detectron model path')
    parser.add_argument('dst', help='save path')
    args = parser.parse_args()

    dst = Path(args.dst)
    if dst.suffix != '.pth':
        print('The path should contain the name of the pth format file.')
        exit()
    dst.parent.mkdir(parents=True, exist_ok=True)

    convert(args.src, args.dst)


if __name__ == '__main__':
    main()

[yasu0001]

I've confirmed that the accuracy is almost same after model conversion.
The accuracy of torchvision is cited from https://pytorch.org/vision/stable/models.html

Model	Accs@1	Accs@1(torchvision)	Accs@5	Accs@5(torchvision)
Wide ResNet 50	77.7	78.47	93.79	94.09
Wide ResNet 101	78.18	78.85	93.98	94.28
ResNet 50	75.36	76.130	92.52	92.862

I would like to ask you about CLA.
The status of CLA is still pending after agreement the License.
This may be because I used wrong mail address in git.config when I pushed.
What I should do in this case?

[Ezra-Yu]

I've confirmed that the accuracy is almost same after model conversion. The accuracy of torchvision is cited from https://pytorch.org/vision/stable/models.html

Actually, the accuracy should be the same. So there is something different with setting in torchvision. Maybe it causes by Resize backend, You can try:

dict(type='Resize', size=(256, -1), backend='pillow'),

This may be because I used wrong mail address in git.config when I pushed.

Did you mean that you have pushed your commits before config your email address, which caused the username unable to be tracked? Maybe you need to use git rebase to modify your commit information. Or just close this PR and re-create a new one.

[yasu0001]

Actually, the accuracy should be the same. So there is something different with setting in torchvision. Maybe it causes by Resize backend, You can try:
dict(type='Resize', size=(256, -1), backend='pillow'),

I confirmed the accuracy is the same as torchvision by modifying resize backend.

Model	Accs@1	Accs@1(torchvision)	Accs@5	Accs@5(torchvision)
Wide ResNet 50	78.48	78.47	94.08	94.09
Wide ResNet 101	78.84	78.85	94.28	94.28
ResNet 50	76.14	76.13	92.87	92.86

Did you mean that you have pushed your commits before config your email address, which caused the username unable to be tracked? Maybe you need to use git rebase to modify your commit information. Or just close this PR and re-create a new one.

Yes, I modified commit message.

[Ezra-Yu]

Great! Please update the configs.

Later, I will add some docs about wide-ResNet. Could I just commit to your branch directly?

[yasu0001]

Later, I will add some docs about wide-ResNet. Could I just commit to your branch directly?

Sure! please.

[codecov]

Codecov Report

Merging #715 (62bae69) into dev (c708770) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##              dev     #715   +/-   ##
=======================================
  Coverage   84.94%   84.94%           
=======================================
  Files         121      121           
  Lines        7548     7548           
  Branches     1303     1303           
=======================================
  Hits         6412     6412           
  Misses        944      944           
  Partials      192      192           

Flag	Coverage Δ
unittests	84.87% <ø> (ø)

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

---

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 c708770...62bae69. Read the comment docs.

[mzr1996]

LGTM

[Davidgzx]

No support for 3 stage WRN? like WRN 16 to WRN 40?

[Ezra-Yu]

@Davidgzx Can you attach the code and weights link?

[Ezra-Yu]

@Davidgzx In the official repo, the wrn16 or the wrn40 are trained by using cifra10 and cifra100. I think the ImagNet-1k dataset is more general.

[OpenMMLab-Assistant003]

Hi @yasu0001！First of all, we want to express our gratitude for your significant PR in the MMClassification project. Your contribution is highly appreciated, and we are grateful for your efforts in helping improve this open-source project during your personal time. We believe that many developers will benefit from your PR.

We would also like to invite you to join our Special Interest Group (SIG) private channel on Discord, where you can share your experiences, ideas, and build connections with like-minded peers. To join the SIG channel, simply message moderator— OpenMMLab on Discord or briefly share your open-source contributions in the #introductions channel and we will assist you. Look forward to seeing you there! Join us ：https://discord.gg/UjgXkPWNqA

If you have WeChat account，welcome to join our community on WeChat. You can add our assistant ：openmmlabwx. Please add "mmsig + Github ID" as a remark when adding friends：）
Thank you again for your contribution❤ @yasu0001