[Bug fixed]Fix dice_loss errors #417

xiexinch · 2021-03-13T08:39:42Z

Fix training errors #411

codecov · 2021-03-13T10:42:16Z

Codecov Report

Merging #417 (ab5ebc8) into master (7e1b24d) will decrease coverage by 0.06%.
The diff coverage is 76.92%.

❗ Current head ab5ebc8 differs from pull request most recent head 22f9fb4. Consider uploading reports for the commit 22f9fb4 to get more accurate results

@@            Coverage Diff             @@
##           master     #417      +/-   ##
==========================================
- Coverage   86.44%   86.38%   -0.07%     
==========================================
  Files          97       97              
  Lines        4958     4957       -1     
  Branches      805      804       -1     
==========================================
- Hits         4286     4282       -4     
- Misses        517      519       +2     
- Partials      155      156       +1

Flag	Coverage Δ
unittests	`86.38% <76.92%> (-0.07%)`	⬇️

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

Impacted Files	Coverage Δ
mmseg/apis/test.py	`22.68% <ø> (ø)`
mmseg/models/segmentors/base.py	`44.95% <0.00%> (-0.42%)`	⬇️
mmseg/apis/inference.py	`66.07% <50.00%> (ø)`
mmseg/models/decode_heads/fcn_head.py	`100.00% <100.00%> (ø)`
mmseg/models/losses/dice_loss.py	`100.00% <100.00%> (+1.96%)`	⬆️
mmseg/datasets/pipelines/transforms.py	`97.12% <0.00%> (-0.87%)`	⬇️

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 7e1b24d...22f9fb4. Read the comment docs.

zhongqiu1245 · 2021-03-13T14:07:40Z

Hi
I used your modified dice_loss to test, the original error disappeared, but a new error appeared.
My data is only two types (background and people), in the VOC format.

No error is reported when loss is the Cross_Entropy or LovaszLoss
It will appear when loss is DiceLoss:
"num_classes=(1 if num_classes == 2 else num_classes))
RuntimeError: Class values must be smaller than num_classes."

In the config file, I call DiceLoss like this：
loss_decode=dict(type='DiceLoss')
Is this method right?
Thank you!

xiexinch · 2021-03-13T15:50:28Z

Hi
I used your modified dice_loss to test, the original error disappeared, but a new error appeared.
My data is only two types (background and people), in the VOC format.

No error is reported when loss is the Cross_Entropy or LovaszLoss
It will appear when loss is DiceLoss:
"num_classes=(1 if num_classes == 2 else num_classes))
RuntimeError: Class values must be smaller than num_classes."

In the config file, I call DiceLoss like this：
loss_decode=dict(type='DiceLoss')
Is this method right?
Thank you!

Hi @zhongqiu1245
This error occurs if num_classes == 2, I am trying to fix it, sorry..

zhongqiu1245 · 2021-03-19T05:54:53Z

Thanks, it works perfectly
Sorry to interrupt you again, I want to try to change dice loss to tversky loss, but I am having trouble calculating the false positives and false negatives on the denominator.
How do you calculate false positives and false negatives?
Why is the denominator of your dice loss the sum of squares? Is there any benefit to doing this?
Thank you

xvjiarui · 2021-03-19T22:24:38Z

mmseg/models/losses/dice_loss.py

-        assert loss_type in ['multi_class', 'binary']
-        if loss_type == 'multi_class':
-            self.cls_criterion = dice_loss
-        else:
-            self.cls_criterion = binary_dice_loss
+        # assert loss_type in ['multi_class', 'binary']
+        # if loss_type == 'multi_class':
+        #     # self.cls_criterion = dice_loss
+        #     self.binary_cls = False
+        # else:
+        #     # self.cls_criterion = binary_dice_loss
+        #     self.binary_cls = True
+        # self.cls_criterion = dice_loss


Why comment these lines?

Cause dice_loss() method has handled these two cases, I think it's no need to add extra codes.

In this case, we could delete this code.

xvjiarui · 2021-03-28T04:22:16Z

Kindly ping @xiexinch

xiexinch · 2021-03-29T01:46:30Z

Kindly ping @xiexinch

I will update it soon, sorry..

ChenLiufeng · 2021-12-16T05:08:47Z

Hi
I used your modified dice_loss to test, the original error disappeared, but a new error appeared.
My data is only two types (background and people), in the VOC format.
No error is reported when loss is the Cross_Entropy or LovaszLoss
It will appear when loss is DiceLoss:
"num_classes=(1 if num_classes == 2 else num_classes))
RuntimeError: Class values must be smaller than num_classes."

In the config file, I call DiceLoss like this：
loss_decode=dict(type='DiceLoss')
Is this method right?
Thank you!

Hi @zhongqiu1245 This error occurs if num_classes == 2, I am trying to fix it, sorry..
@xiexinch I changed loss_decode=dict(type='FocalLoss'). And my num_classes=2. I met the same problem.

Could you fix it? Thanks!

* fix training bugs * fix unitest error * fix error in num_classes==2 case * delete comments

* Update conditional_image_generation.mdx * Update unconditional_image_generation.mdx

* update template * update emoji

fix training bugs

c29ab07

xiexinch closed this Mar 13, 2021

fix unitest error

36fd0ca

xiexinch reopened this Mar 13, 2021

fix error in num_classes==2 case

b8a76ab

xiexinch mentioned this pull request Mar 19, 2021

Dice loss error #432

Closed

xvjiarui reviewed Mar 19, 2021

View reviewed changes

delete comments

22f9fb4

xvjiarui approved these changes Mar 29, 2021

View reviewed changes

xvjiarui merged commit 1fc3e37 into open-mmlab:master Mar 29, 2021

xiexinch mentioned this pull request Mar 30, 2021

refactor test organization #440

Merged

MengzhangLI mentioned this pull request Dec 6, 2021

[Fix] Remove unexist arg in docstring #1106

Merged

bowenroom pushed a commit to bowenroom/mmsegmentation that referenced this pull request Feb 25, 2022

[Bug fixed]Fix dice_loss errors (open-mmlab#417)

71be1c2

* fix training bugs * fix unitest error * fix error in num_classes==2 case * delete comments

aravind-h-v pushed a commit to aravind-h-v/mmsegmentation that referenced this pull request Mar 27, 2023

[Docs] Pipelines for inference (open-mmlab#417)

6b9906f

* Update conditional_image_generation.mdx * Update unconditional_image_generation.mdx

sibozhang pushed a commit to sibozhang/mmsegmentation that referenced this pull request Mar 22, 2024

Update issue template (open-mmlab#417)

cae8ae6

* update template * update emoji

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Bug fixed]Fix dice_loss errors #417

[Bug fixed]Fix dice_loss errors #417

xiexinch commented Mar 13, 2021

codecov bot commented Mar 13, 2021 •

edited

Loading

zhongqiu1245 commented Mar 13, 2021 •

edited

Loading

xiexinch commented Mar 13, 2021

zhongqiu1245 commented Mar 19, 2021 •

edited

Loading

xvjiarui Mar 19, 2021

xiexinch Mar 20, 2021

xvjiarui Mar 20, 2021

xvjiarui commented Mar 28, 2021

xiexinch commented Mar 29, 2021

ChenLiufeng commented Dec 16, 2021

[Bug fixed]Fix dice_loss errors #417

[Bug fixed]Fix dice_loss errors #417

Conversation

xiexinch commented Mar 13, 2021

codecov bot commented Mar 13, 2021 • edited Loading

Codecov Report

zhongqiu1245 commented Mar 13, 2021 • edited Loading

xiexinch commented Mar 13, 2021

zhongqiu1245 commented Mar 19, 2021 • edited Loading

xvjiarui Mar 19, 2021

Choose a reason for hiding this comment

xiexinch Mar 20, 2021

Choose a reason for hiding this comment

xvjiarui Mar 20, 2021

Choose a reason for hiding this comment

xvjiarui commented Mar 28, 2021

xiexinch commented Mar 29, 2021

ChenLiufeng commented Dec 16, 2021

codecov bot commented Mar 13, 2021 •

edited

Loading

zhongqiu1245 commented Mar 13, 2021 •

edited

Loading

zhongqiu1245 commented Mar 19, 2021 •

edited

Loading