[Enhancement] Support ignore_index for sigmoid BCE

[xvjiarui]

Fixed #162
Fixed #201

[codecov]

Codecov Report

Merging #210 (02ac5e6) into master (3d18775) will increase coverage by 0.02%.
The diff coverage is 73.33%.

@@            Coverage Diff             @@
##           master     #210      +/-   ##
==========================================
+ Coverage   84.26%   84.28%   +0.02%     
==========================================
  Files          90       90              
  Lines        4315     4322       +7     
  Branches      684      685       +1     
==========================================
+ Hits         3636     3643       +7     
- Misses        537      538       +1     
+ Partials      142      141       -1     

Flag	Coverage Δ
unittests	84.28% <73.33%> (+0.02%)	⬆️

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

Impacted Files	Coverage Δ
mmseg/models/decode_heads/decode_head.py	91.66% <ø> (ø)
mmseg/models/losses/cross_entropy_loss.py	78.12% <73.33%> (+2.68%)	⬆️

---

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 3d18775...f560849. Read the comment docs.

[JinxiaoWang]

Hi, I changed files as https://github.com/open-mmlab/mmsegmentation/pull/210/commits/0f316127623a32c65585b3d10e80bb770ea17789 did.
When I set use_sigmoid=True , AssertError : BCE loss does not support ignore_index.
Then I add 'ignore_index = None ' to binary_cross_entropy function.
I meet a new error : IndexError: shape mismatch: indexing tensors could not be broadcast together with shapes [591, 3], [591, 3, 256, 256]
I think a more detailed tutorial for add custom loss contained by pytorch should added to mmsegmentation.
And, mmseg/models/losses should contain more custom losses.

[hellock]

Task linked: CU-f6t2qg fix BCE loss

[hellock]

Hi, I changed files as https://github.com/open-mmlab/mmsegmentation/pull/210/commits/0f316127623a32c65585b3d10e80bb770ea17789 did.
When I set use_sigmoid=True , AssertError : BCE loss does not support ignore_index.
Then I add 'ignore_index = None ' to binary_cross_entropy function.
I meet a new error : IndexError: shape mismatch: indexing tensors could not be broadcast together with shapes [591, 3], [591, 3, 256, 256]
I think a more detailed tutorial for add custom loss contained by pytorch should added to mmsegmentation.
And, mmseg/models/losses should contain more custom losses.

@xvjiarui Could you have a look?

[xvjiarui]

Hi @DreamerWangDr
You may try it again.