How to replace torch.nn.CrossEntropyLoss by monai.losses.FocalLoss?

[dzenanz]

I am trying to check whether focal loss improves the training of my network. Just replacing CrossEntropyLoss by FocalLoss leads to an error.

When I set use_focal_loss = True

if use_focal_loss:
    loss_function = monai.losses.FocalLoss(weight=class_weights)
else:
    loss_function = torch.nn.CrossEntropyLoss(weight=class_weights)

I get: ValueError: logits and target ndim must match, got logits=2 target=1.

If I also add at the end of forward:

if use_focal_loss:
    outputs = torch.nn.functional.log_softmax(outputs, dim=1)

the error message is unchanged.

[dzenanz]

Anyone?

[wyli]

perhaps set to_one_hot_y=True?
example:

>>> import torch
>>> from monai.losses import FocalLoss
>>> pred = torch.tensor([[1, 0], [0, 1], [1, 0]], dtype=torch.float32)
>>> grnd = torch.tensor([0, 1, 0], dtype=torch.int64)
>>> fl = FocalLoss(to_onehot_y=True)
>>> fl(pred, grnd)
tensor(0.0113)
>>> pred.shape
torch.Size([3, 2])
>>> grnd.shape
torch.Size([3])
>>> 

[dzenanz]

With MONAI 0.4:

loss_function = monai.losses.FocalLoss(weight=class_weights, to_onehot_y=True)

TypeError: init() got an unexpected keyword argument 'to_onehot_y'

With MONAI 0.5 it works. Thanks a lot Wenqi!