[Feature] add SFSegNet head

[MengzhangLI]

Implementation of Semantic Flow for Fast and Accurate SceneParsing.
Modified from the official repository.

Done:

• unitest

Todo:

• design the config interface and files
• more experiments especially model with res18 backbone.

Related re-implementation https://github.com/azxj/SFNet.

[codecov]

Codecov Report

Merging #733 (8323699) into master (6526593) will increase coverage by 0.25%.
The diff coverage is 98.00%.

❗ Current head 8323699 differs from pull request most recent head 12def7c. Consider uploading reports for the commit 12def7c to get more accurate results

@@            Coverage Diff             @@
##           master     #733      +/-   ##
==========================================
+ Coverage   85.28%   85.54%   +0.25%     
==========================================
  Files         107      108       +1     
  Lines        5817     5906      +89     
  Branches      952      957       +5     
==========================================
+ Hits         4961     5052      +91     
  Misses        673      673              
+ Partials      183      181       -2     

Flag	Coverage Δ
unittests	85.54% <98.00%> (+0.27%)	⬆️

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

Impacted Files	Coverage Δ
mmseg/models/backbones/resnet.py	99.28% <ø> (ø)
mmseg/models/backbones/swin.py	86.89% <66.66%> (ø)
mmseg/models/necks/fpn.py	54.54% <66.66%> (+0.59%)	⬆️
mmseg/models/backbones/unet.py	94.91% <100.00%> (+0.04%)	⬆️
mmseg/models/backbones/vit.py	84.84% <100.00%> (ø)
mmseg/models/builder.py	92.00% <100.00%> (+0.69%)	⬆️
mmseg/models/decode_heads/__init__.py	100.00% <100.00%> (ø)
mmseg/models/decode_heads/fpn_head.py	100.00% <100.00%> (ø)
mmseg/models/decode_heads/setr_mla_head.py	100.00% <100.00%> (ø)
mmseg/models/decode_heads/setr_up_head.py	100.00% <100.00%> (ø)
... and 6 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 6526593...12def7c. Read the comment docs.

[Junjun2016]

Hi @MengzhangLI
Please resolve the comments.

[Junjun2016]

Hi @MengzhangLI
For every PR, you should checkout to the master first, then pull the upstream master, and finally checkout a new branch.
It can avoid too many unnecessary commit histories.

[MengzhangLI]

Re-implementing SFNet ResNet50 of PaddleSeg.

Inference metric is done.

Note:
stride in model of paddleseg sfnet Res50 is (1, 2, 1, 1), where original paper is (1, 2, 2, 2)
That's why the performance of its sfnet res50 81.49 is better than paper 80.74.

[MengzhangLI]

Re-implementing SFNet ResNet18 of PaddleSeg.

Note: there are some potential risks for paddleseg users:
(1) Actual Output Stride(OS) of Paddleseg SFNet ResNet18 model is not 8 but 32. Although 32 is setting of original paper, It is unintentionally correct settings because paddleseg setting is 8 from here

The reason of OS = 32 is when building backbone models, paddleseg would skip dilation rate from here when layers <= 50.

Thus, stride of SFNet of Paddleseg are (1, 2, 1, 1), os = 8 (Res50) and (1, 2, 2, 2), os = 32 (Res18), respectively.

(2) When using ResNetv1d with 18 layers, paddleseg would add a short downsample operations (conv + BN) at first layer, where official usage (mmsegmentation adopted) would not add. Thus it would have five more variables:
backbone.bb_0_0.short._conv.weight (64, 64, 1, 1)
backbone.bb_0_0.short._batch_norm.weight (64,)
backbone.bb_0_0.short._batch_norm.bias (64,)
backbone.bb_0_0.short._batch_norm._mean (64,)
backbone.bb_0_0.short._batch_norm._variance (64,)

Anyway, inference metric is DONE.

[MengzhangLI]

Related SFNet model: https://github.com/azxj/SFNet

[jyang68sh]

Hi I just runned sfnet_r18-d32_512x1024_80k_cityscapes.py and only get mIoU of 73.92. I am wondering the cause. Did you change the config?

@MengzhangLI

+-------+-------+-------+ | aAcc | mIoU | mAcc | +-------+-------+-------+ | 95.49 | 73.92 | 81.48 | +-------+-------+-------+