[Fix] Add avg_non_ignore in cross entropy loss

[MengzhangLI]

Motivation

Right now, MMSegmentation do not support ignore_index in loss average over non-ignored elements in CE loss.

fix #578.

old pr: #1388.

In old PR: #578, the author fixed it while it left a bug if all labels in one image is ignore_index, avg_factor would be zero and ZeroDivisionError would be rasied in weight_reduce_loss function.

Usage

Just take fcn_unet_s5-d16_128x128_40k_chase_db1.py, its base config for example:

original config:

# model settings
...
        loss_decode=dict(
            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
...

Just add ignore_index in decoder head or auxiliary head and add avg_non_ignore=True:

# model settings
...
        loss_decode=dict(
            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0, avg_non_ignore=True),
...

cfg	original	avg_non_ignore=True	avg_non_ignore=False
deeplabv3_r18-d8_512x1024_80k_cityscapes	76.70 78.27(ms+flip)	77.07 78.82(ms+flip)	-
deeplabv3_r50-d8_512x512_80k_ade20k	42.42 43.28(ms+flip)	43.08 44.31(ms+flip)	42.40 43.65(ms+flip)

[MeowZheng]

ping @ZwwWayne and @RangiLyu take a look

[codecov]

Codecov Report

Merging #1409 (a4e4b0a) into master (98984bf) will decrease coverage by 0.13%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #1409      +/-   ##
==========================================
- Coverage   90.39%   90.26%   -0.14%     
==========================================
  Files         133      136       +3     
  Lines        7906     7961      +55     
  Branches     1318     1326       +8     
==========================================
+ Hits         7147     7186      +39     
- Misses        536      550      +14     
- Partials      223      225       +2     

Flag	Coverage Δ
unittests	90.26% <100.00%> (-0.14%)	⬇️

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

Impacted Files	Coverage Δ
mmseg/models/losses/cross_entropy_loss.py	82.71% <100.00%> (+3.30%)	⬆️
mmseg/models/losses/utils.py	87.50% <100.00%> (+5.92%)	⬆️
mmseg/datasets/hrf.py	77.77% <0.00%> (-2.23%)	⬇️
mmseg/datasets/drive.py	77.77% <0.00%> (-2.23%)	⬇️
mmseg/datasets/chase_db1.py	77.77% <0.00%> (-2.23%)	⬇️
mmseg/datasets/pascal_context.py	75.00% <0.00%> (-1.48%)	⬇️
mmseg/datasets/isaid.py	93.93% <0.00%> (-0.18%)	⬇️
mmseg/core/utils/__init__.py	100.00% <0.00%> (ø)
mmseg/core/utils/dist_util.py	55.55% <0.00%> (ø)
mmseg/datasets/samplers/distributed_sampler.py	52.38% <0.00%> (ø)
... and 4 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 0d51115...a4e4b0a. Read the comment docs.