MagmaDNN / magmadnn

MagmaDNN: a simple deep learning framework in c++
MIT License
45 stars 7 forks source link

[BUG] NaN loss #18

Closed Dando18 closed 5 years ago

Dando18 commented 5 years ago

Describe the bug Overtraining a model produces NaN loss with the cross entropy loss function.

To Reproduce Training the simple_network example several times (i.e. call fit 3 times) and the loss will become NaN and the model won't train.

Expected behavior no NaN.

Likely fixed by using numerically stable log in cross entropy.