This repository has been archived by the owner on Mar 22, 2021. It is now read-only.
Something's wrong in multiclass dice loss #20
Labels
invalid
This doesn't seem right
Comments
jakubczakon
added a commit
that referenced
this issue
Oct 13, 2018
* Hypercolumn (#16) * fixed lovash loss, added helpers for loss weighing (#14) * updated results exploration, added unet with hypercolumn * updated with lighter hypercolumn setup * Model average (#17) * added prediction average notebook * added simple average notebook * added replication pad instead of zero pad (#18) * changed to heng-like arch, added channel and spatial squeeze and excite, extended hypercolumn (#19) * Update unet_models.py typo in resnet unet fixed * added resnet 18 an50 pretrained options, unified hyper and vanilla in one class (#20) * Update models.py Changed old class import and namings * Loss design (#21) * local * initial * formated results * added focal, added border weighing, added size weighing added focus, added loss desing notebook * fixed wrong focal definition, updated loss api * exp with dropped borders * set best params, not using weighing for now * Dev depth experiments (#23) * add depth layer in input * reduce lr on plateau scheduler * depth channels transformer * fix reduce lr * bugfix * change default config * added adaptive threshold in callbacks (#24) * added adaptive threshold in callbacks * fix * added initial lr selector (#25) * Initial lb selector (#26) * added initial lr selector * small refactor * Auxiliary data small masks (#27) * exping * auxiliary data for border masks generated
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Here https://github.com/neptune-ml/open-solution-salt-detection/blob/master/src/models.py#L270 you write that
targetis a tensor [NxHxW] and later, here: https://github.com/neptune-ml/open-solution-salt-detection/blob/master/src/models.py#L291 you treattargetas if it was a 4D tensor. The docstring is wrong or implementation is wrong.The text was updated successfully, but these errors were encountered: