Fixed the issue referenced, which concerns implementation of Jaccard Loss function.
- How I did it
I changed the axis argument of summation terms from only channels to all except channels, which is the correct way according to what I have studied. Consequently, this returns a per-channel loss vector for each sample.
- How you can verify it
The functionality and correctness demands (other than theoretical reasoning) verification by training via this loss function. I have used this as well as dice loss (also arranged in a similar configuration) for multiple training sessions which yield good performance.
526<!--
Please make sure you've read and understood our contributing guidelines; https://github.com/keras-team/keras-contrib/blob/master/CONTRIBUTING.md -->
- What I did
Fixed the issue referenced, which concerns implementation of Jaccard Loss function.
- How I did it
I changed the axis argument of summation terms from only channels to all except channels, which is the correct way according to what I have studied. Consequently, this returns a per-channel loss vector for each sample.
- How you can verify it
The functionality and correctness demands (other than theoretical reasoning) verification by training via this loss function. I have used this as well as dice loss (also arranged in a similar configuration) for multiple training sessions which yield good performance.
This pull request fixes Issue #526 opened by me