loss is nan

[eeric]

your example was right, but when batch_size was 128, loss was nan, actually it occur error on 36,37th line in circleloss.py.

[qianjinhao]

your example was right, but when batch_size was 128, loss was nan, actually it occur error on 36,37th line in circleloss.py.

One example in the code is a tensor with a batch size of 128, and the result is not nan, so did you use the 'cos' method to calculate the similarity?

[eeric]

yes.

[TinyZeaMays]

@qianjinhao

You may use soft_plus and logsumexp instead.

https://github.com/TinyZeaMays/CircleLoss/blob/d002ecd0c7e395f6e39bf8a2a96fd05b83afa93f/circle_loss.py#L37

[qianjinhao]

@qianjinhao

You may use soft_plus and logsumexp instead.

https://github.com/TinyZeaMays/CircleLoss/blob/d002ecd0c7e395f6e39bf8a2a96fd05b83afa93f/circle_loss.py#L37

@qianjinhao

You may use soft_plus and logsumexp instead.

https://github.com/TinyZeaMays/CircleLoss/blob/d002ecd0c7e395f6e39bf8a2a96fd05b83afa93f/circle_loss.py#L37

Thanks for the reminder. Have you tested the difference between the two implementations?

[vodanhbk95]

@qianjinhao

You may use soft_plus and logsumexp instead.

https://github.com/TinyZeaMays/CircleLoss/blob/d002ecd0c7e395f6e39bf8a2a96fd05b83afa93f/circle_loss.py#L37

@TinyZeaMays Can you tell me what is the difference between your implement and @qianjinhao 's implement. Cause I think it's true formula for circle loss