The question about KLD loss

[haoopan]

Excuse me again!
When I use the KLDivLoss as the criterion in ./block/models/criterions/kl_divergence.py rather than crossentropy, I get the loss is 0 and the acc is worse , as following picture shows:

But in paper MFH ,KLD loss is better than CrossEntropy, So, how could this be? Thanks!

[Cadene]

@haoopan sorry we don't provide support for KLD loss.
We didn't manage to make it work.
If I remember correctly, even with this low loss it can learn, but not as well as standard cross entropy.
There might be a bug somewhere.
Did you fix i maybe???

Best