[Feature] Support Resnet strikes back

[MengzhangLI]

Motivation

Transfer learning of paper 'Resnet strikes back: An improved training procedure in timm'.

paper URL https://arxiv.org/pdf/2110.00476

Related PR: open-mmlab/mmdetection#7001.

Notification:

(1) Current MMSegmentation default ResNet backbone is ResNetV1c, which performs better than standard ResNet backbone.
(2) This PR only implements on PSPNet.
(3) With more RSB (Resnet strikes back) experiments on different decoder heads (like DeepLabV3, FCN and so on) and RSB strategy on ResNetV1c backbone, we would consider open an independent resnet_strikes_back folder in ./configs/ like MMDetection.

Results

Method	Backbone	optimizer	learning rate	weight decay	Detail config	Dataset	mIoU	mIoU (ms + flip)
PSPNet	ResNet RSB R50-D8	AdamW	0.0005	0.05	KNET	Cityscapes	78.47	79.45
PSPNet	ResNet V1c R50-D8	SGD	0.01	0.0005	None	Cityscapes	78.55	79.49
PSPNet	ResNet R50-D8	SGD	0.01	0.0005	None	Cityscapes	78.22	79.46
PSPNet	ResNet RSB R50-D32	AdamW	0.0005	0.05	KNET	Cityscapes	74.09	77.18
PSPNet	ResNet V1c R50-D32	SGD	0.01	0.0005	None	Cityscapes	73.88	76.85
PSPNet	ResNet R50-D32	SGD	0.01	0.0005	None	Cityscapes	72.61	75.51

Where KNET means learning policy referred from K-Net semantic segmentation configs, which works well with AdamW optimizer:

optimizer_config = dict(grad_clip=dict(max_norm=1, norm_type=2))
# learning policy
lr_config = dict(
    _delete_=True,
    policy='step',
    warmup='linear',
    warmup_iters=1000,
    warmup_ratio=0.001,
    step=[60000, 72000],
    by_epoch=False)

[codecov]

Codecov Report

Merging #1390 (4f02398) into master (3d0c2eb) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #1390   +/-   ##
=======================================
  Coverage   90.39%   90.39%           
=======================================
  Files         133      133           
  Lines        7906     7906           
  Branches     1318     1318           
=======================================
  Hits         7147     7147           
  Misses        536      536           
  Partials      223      223           

Flag	Coverage Δ
unittests	90.39% <ø> (ø)

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

Impacted Files	Coverage Δ
mmseg/core/evaluation/class_names.py	88.33% <ø> (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 3d0c2eb...4f02398. Read the comment docs.