sparse_categorical_crossentropy vs categorical_crossentropy Loss Function
Closed this issue · 2 comments
AmbiTyga commented
Why did you used sparse_categorical_crossentropy?
I don't find true labels one-hot encoded while processing the datasets.
Please elaborate why did you used sparse_categorical_crossentropy instead of categorical_crossentropy.
filipefborba commented
If I remember correctly, sparse_categorical_crossentropy is used for non-one-hot-encoded datasets. Also, the classes are mutually exclusive, so we don't need probabilities...
atomar2000 commented
yes, thats right....in this the labels are mutually exclusive and there is no need to find a multi-hot vector.....we just need to find a one-hot vector, that is, only one label as the output.....
therefore, categorical_crossentropy should've been used in my opinion.