Occasion 'nan' in training w/. semihard triplet loss about triplet-loss-pytorch HOT 3 CLOSED

Hi 😄

How many categories do you have in the training set? It is easier to have NAN values in the validation set rather than in the training set due to missing similar pairs.

The easiest way to solve this is to increase the batch size but, of course, this will increase the computational cost. This might be a solution if you are using a very small batch size or if your number of categories is not very large. A batch size of 128 and 256 usually works with ~1,000 categories or more.

from triplet-loss-pytorch.

Understood, thanks. I compared your code with TensorFlow addons as well and your implementation should be good.

from triplet-loss-pytorch.

Happy to help

from triplet-loss-pytorch.