Trouble with NaN

1

I am trying to train an autoencoder and I using torch.nn.MSELoss() as the loss function. I keep getting this error below:

RuntimeError                              Traceback (most recent call last)
<ipython-input-76-67910671b729> in <module>()
     20 
     21         optimizer.zero_grad()
---> 22         loss.backward()
     23         optimizer.step()
     24 
RuntimeError: Function 'MseLossBackward' returned nan values in its 0th output.

I have tried reading many of the other posts on this topic and I have tried various approaches such as adjusting the learning rate, gradient clipping. The weird thing is the loss doesn’t seem to diverge at all. I am not totally sure why this error keeps occurring for me, any suggestions or help would be appreciated? I am very new to pytorch.

2

Could you check if you have any nan’s in your input? And if there aren’t any in it, could you check if the outputs are producing nan’s?