Training is not working. Loss gets worse

[kalle264]

Hi,
did you use your implementation successfully for training?

It's working for me when I just calculate the loss for a prediction of a pretrained model, but when I want to train with it, the loss increases from batch to batch and then just stays at the diagonal distance for every following epoch.

My code works just fine with a different loss.
Do you have any idea what could cause that problem?

[N0vel]

Hi,
Try to tune learning rate and the parameters of the loss itself (it's a weighted sum of two components)