[Feature] Support two new loss function#600
[Feature] Support two new loss function#600MengzhangLI wants to merge 2 commits intoopen-mmlab:masterfrom
Conversation
|
|
|
Hi @MengzhangLI |
Codecov Report
@@ Coverage Diff @@
## master #600 +/- ##
==========================================
- Coverage 86.26% 81.83% -4.43%
==========================================
Files 101 105 +4
Lines 5278 5626 +348
Branches 854 896 +42
==========================================
+ Hits 4553 4604 +51
- Misses 561 858 +297
Partials 164 164
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
|
Hi @MengzhangLI |
|
Can we implement topk in other ways, e.g. OHEM. |
Of course, but it needs some time. BTW, the way I adopted in DiceTopK is just follow nnUNet default settings: https://github.com/JunMa11/SegLoss/blob/master/test/nnUNetV1/loss_functions/TopK_loss.py |
|
I will update these loss functions in several ways: |
* Fix typo in docstring. * Allow dtype to be overridden on model load. This may be a temporary solution until open-mmlab#567 is addressed. * Create latents in float32 The denoising loop always computes the next step in float32, so this would fail when using `bfloat16`.
* Fix 599 * fix unittest & docs & default value * update docs
Hi, jiarui, junjun, zesen and xincheng:
These are two new loss functions, which are tversky loss and DiceTopK loss.
I tested these two new functions on several 2D medical image datasets, i.e., DRIVE, STARE and CHASE_DB1.
Here is the result:
Best,