adapt loss to full mask #1426
adapt loss to full mask #1426
Conversation
|
|
Codecov ReportPatch coverage has no change and project coverage change:
Additional details and impacted files@@ Coverage Diff @@
## dev-1.x #1426 +/- ##
============================================
+ Coverage 0.02% 86.77% +86.75%
============================================
Files 121 170 +49
Lines 8217 14049 +5832
Branches 1368 2237 +869
============================================
+ Hits 2 12191 +12189
+ Misses 8215 1471 -6744
- Partials 0 387 +387
Flags with carried forward coverage won't be shown. Click here to find out more. see 171 files with indirect coverage changes Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here. ☔ View full report in Codecov by Sentry. |
| @@ -23,7 +23,8 @@ def wrapped(inputs, data_samples, **kwargs): | |||
| task_data_samples.append(data_sample.get(task_name)) | |||
|
|
|||
| if len(task_data_samples) == 0: | |||
| return {'loss': torch.tensor(0.), 'mask_size': torch.tensor(0.)} | |||
| loss = (inputs[0] * 0).sum() | |||
There was a problem hiding this comment.
@marouaneamz i think you should explain with a comment here
|
closed since #1530 merged. |
in the pull request #1229 , for the extreme case that we have a full-mask we do for the loss = 0 which will cause problems to calculate the gradient of the loss. here i use the torch vector of the image to create a loss =0