Gradients: inf or nan in tensor
SiBensberg opened this issue · comments
Hi,
I used the FLIR dataset to try out the training. From time to time the error above happens and all model parameters are nan's. Could ou observe similiar behaviour?
I fixed it through catching nan values and clipping gradients.
Hello,
I have the same problem right now. I followed your solution however still getting nan values.
Could you please share your codes regarding clipping gradients ?