Dice Loss Error

[albertnanda]

I have two part question,

1. The example given in the code bugs out i.e. https://github.com/ShannonAI/dice_loss_for_NLP/blob/418d09d285c103176152a97d73f8e7ebcdb1fa49/loss/dice_loss.py#L41 IndexError: Dimension out of range (expected to be in range of [-1, 0], but got 1)
2. The other question is related to the implementation, say the classifier has perfectly predicted the labels, but there would be still some dice loss because of loss = 1 - ((2 * interection + self.smooth) / (torch.sum(torch.square(flat_input, ), -1) + torch.sum(torch.square(flat_target), -1) + self.smooth)) smooth. Is this the expected behavior or am I missing something.

input = torch.FloatTensor([[1., .0, .0, .0],[0., 1, .0, .0]])
target = torch.LongTensor([0, 1])
loss = DiceLoss(with_logits=False,reduction=None,ohem_ratio=0.)
input.requires_grad=True
output = loss(input, target)

Output tensor([1.9998, 1.9998], grad_fn=<AddBackward0>)

[xiaoya-li]

Hey, thanks for asking. Response to quesiton2: As shown in https://github.com/ShannonAI/dice_loss_for_NLP/blob/418d09d285c103176152a97d73f8e7ebcdb1fa49/tasks/tnews/train.py#L139 we recommend using the following setting for multi-class tasks:

$ loss_fct = DiceLoss(square_denominator=True, with_logits=False, index_label_position=True,
                        smooth=1, ohem_ratio=0, alpha=0.01, reduction="none")

• If input_probs = torch.FloatTensor([[0.9, 0.03, 0.03, 0.03],[0.03, 0.9, 0.03, 0.03]]) and target = torch.LongTensor([0, 1]), then the dice loss should be :

$ loss_fct(input_probs, target)
> tensor([0.0079, 0.0079])

• If input_probs = torch.FloatTensor([[0.9, 0.1, .0, .0],[.0, 0.9, 0.1, .0]]) and target = torch.LongTensor([0, 1]), then the dice loss should be :

$ loss_fct(input_probs, target)
> tensor([0.0151, 0.0151])

These are in line with our expectations.

[albertnanda]

@xiaoya-li : Let me try it out, also can you recommend the settings for Multi-Label classification and NER task. NER Task:

inp = torch.FloatTensor([[[.1,.2,.3,.4]]*4,[[.5,.5,0,0]]*4]) #2 sentences, 4 words and 4 tags per word
target = torch.LongTensor([[0,1,1,2],[0,3,2,3]]) #tags 0,3, 2 sentences and 4 words

[kk19990709]

I have also met question 1, do you have any solution?@albertnanda