cleaning and reorganizing fuse core code on segmentation #318
Conversation
itaijj
changed the title
itaijj
changed the title
fuse/dl/losses/loss_focalLoss.py
Outdated
|
|
||
|
|
||
| class WeightedFocalLoss(nn.Module): | ||
| "Non weighted version of Focal Loss" |
There was a problem hiding this comment.
non-weighted or weighted?
There was a problem hiding this comment.
waiting to decide if to keep it or not
fuse/dl/losses/loss_focalLoss.py
Outdated
| self.alpha = torch.tensor([alpha, 1 - alpha]) | ||
| self.gamma = gamma | ||
|
|
||
| def forward(self, inputs: Tensor, targets: Tensor) -> Tensor: |
There was a problem hiding this comment.
Why don't you get batch_dict instead of tensors?
There was a problem hiding this comment.
waiting to decide if to keep it or not
fuse/dl/losses/loss_focalLoss.py
Outdated
| self.alpha = torch.tensor([alpha, 1 - alpha]) | ||
| self.gamma = gamma | ||
|
|
||
| def forward(self, inputs: Tensor, targets: Tensor) -> Tensor: |
There was a problem hiding this comment.
Add comments specifying which tensors do you expect, shape and content
There was a problem hiding this comment.
waiting to decide if to keep it or not
fuse/dl/losses/loss_focalLoss.py
Outdated
| return result | ||
|
|
||
|
|
||
| class WeightedFocalLoss(nn.Module): |
There was a problem hiding this comment.
Why did you move it? Are you using it?
If not, let's delete it. It's old code.
There was a problem hiding this comment.
I just think focal loss is not specific for segmentation, and we don't have it anywhere else in fuse so I moved it out of segmentation folder
| @@ -44,7 +42,7 @@ def make_one_hot(input: Tensor, num_classes: int) -> Tensor: | |||
|
|
|||
|
|
|||
| class BinaryDiceLoss(nn.Module): | |||
| def __init__(self, power: int = 1, eps: float = 1.0, reduction: str = "mean"): | |||
| def __init__(self, eps: float = 1e-5, reduction: str = "mean"): | |||
There was a problem hiding this comment.
Why did you change the defaults?
There was a problem hiding this comment.
remove power also from comments
There was a problem hiding this comment.
I think that the epsilon for smoothing should be small number and not 1.
I see in monai the number is 1e-5, do you have any other reference besides @egozi old code to keep it 1??
No description provided.