[Enchance] support infererence with padding

[FreyWang]

• Before padding
  
• After padding
  

Motivation

I use mmseg to train my model and found the mask has a left offset to the image. Finally I found that it is caused by inference input without padding. For example, A input with origin shape (1920, 1080) and img_scale=(2048, 512) will be resized to (910 , 512). When infer it with resnet50-d8，the final feature map is 114 * 64 and then resize it to 910*512. Attention that 114x8=912. So it is same with pad the input to (912, 512) and resize feature to (910, 512). With 2-pixel left offset

BC-breaking (Optional)

Attention that when dict(type='Pad', size_divisor=32), is add, the model performance will intend to be improved, I have test several mIoU rise

• configs/deeplabv3plus/deeplabv3plus_r50-d8_512x512_80k_ade20k.py
  from 42.72 to 42.78
• configs/ocrnet/ocrnet_hr18s_512x512_80k_ade20k.py
  from 35.06 to 35.53
• configs/knet/knet_s3_upernet_swin-l_8x2_512x512_adamw_80k_ade20k.py
  from 52.05 to 52.61

Use cases (Optional)

Add Pad when inference to make sure that input can be divide to model stride.
[Update]
Attention that Pad must behinde RandomFlip when flip is True

test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(2048, 512),
        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip'),
            dict(type='Pad', size_divisor=32),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]

[codecov]

Codecov Report

Merging #1607 (3e7426d) into master (104429b) will decrease coverage by 0.14%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #1607      +/-   ##
==========================================
- Coverage   90.40%   90.26%   -0.15%     
==========================================
  Files         141      142       +1     
  Lines        8443     8485      +42     
  Branches     1423     1431       +8     
==========================================
+ Hits         7633     7659      +26     
- Misses        572      586      +14     
- Partials      238      240       +2     

Flag	Coverage Δ
unittests	90.26% <100.00%> (-0.15%)	⬇️

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

Impacted Files	Coverage Δ
mmseg/datasets/pipelines/test_time_aug.py	96.55% <100.00%> (+0.25%)	⬆️
mmseg/models/segmentors/encoder_decoder.py	88.66% <100.00%> (+0.31%)	⬆️
mmseg/apis/train.py	19.51% <0.00%> (ø)
mmseg/utils/__init__.py	100.00% <0.00%> (ø)
mmseg/utils/util_distribution.py	48.38% <0.00%> (ø)
mmseg/datasets/samplers/distributed_sampler.py	56.52% <0.00%> (+4.14%)	⬆️

---

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 104429b...3e7426d. Read the comment docs.

[FreyWang]

It may effect pytorch2onnx and i will refactor the code later

[MeowZheng]

@FreyWang have you compared the model performance before and after this modification? Would you like to share the results with us? Thank you

[FreyWang]

@FreyWang have you compared the model performance before and after this modification? Would you like to share the results with us? Thank you
I only test several model such as deeplabv3+，ocrnet, knet and setr, all gain 0~0.5 mIoU improve. 🥲 Test all config is difficult, some automated scripts is needed
Attention that Pad must behinde RandomFlip when flip is True

[FreyWang]

a

the performance is same after this commit if pad is not add to config

[MeowZheng]

@FreyWang have you compared the model performance before and after this modification? Would you like to share the results with us? Thank you
I only test several model such as deeplabv3+，ocrnet, knet and setr, all gain 0~0.5 mIoU improve. smiling_face_with_tear Test all config is difficult, some automated scripts is needed
Attention that Pad must behinde RandomFlip when flip is True

would you like to present the TTA result with pad and without pad of deeplabv3+?

[FreyWang]

@FreyWang have you compared the model performance before and after this modification? Would you like to share the results with us? Thank you
I only test several model such as deeplabv3+，ocrnet, knet and setr, all gain 0~0.5 mIoU improve. smiling_face_with_tear Test all config is difficult, some automated scripts is needed
Attention that Pad must behinde RandomFlip when flip is True

would you like to present the TTA result with pad and without pad of deeplabv3+?

configs/deeplabv3plus/deeplabv3plus_r50-d8_512x512_80k_ade20k.py
from 42.72 to 42.78 after pad is added.
emmm, what is the meaning of TTA. I have only ade20k dataset now

[FreyWang]

@FreyWang have you compared the model performance before and after this modification? Would you like to share the results with us? Thank you
I only test several model such as deeplabv3+，ocrnet, knet and setr, all gain 0~0.5 mIoU improve. smiling_face_with_tear Test all config is difficult, some automated scripts is needed
Attention that Pad must behinde RandomFlip when flip is True

would you like to present the TTA result with pad and without pad of deeplabv3+?

I got what you said and I will find time to commit the test code😂

[FreyWang]

@MeowZheng after pad
configs/deeplabv3plus/deeplabv3plus_r50-d8_512x512_80k_ade20k.py
default: 42.72 -> 42.78
ms+flip: 43.75->43.77
configs/setr/setr_mla_512x512_160k_b8_ade20k.py
default: 47.34->47.58
ms+flip: 49.05->49.06