Diary - Ignore value in Cross Entropy function of PyTorch

In PyTorch, there are times when you don’t want to compute the loss for certain labels. Since running a For loop while doing GPU computation is very inefficient, you use a different approach.
If you look at the documentation for CrossEntropy, there’s an argument called ignore_index. The default value is set to -100, which means that if a label is -100, this part is not included in the computation.
Simply putting in a negative value might cause an error, so to be safe let’s put in the magic number -100.

Reference

https://discuss.pytorch.org/t/negative-label-in-crossentropyloss/158873