EderSantana / seya

Bringing up some extra Cosmo to Keras.
Other
377 stars 103 forks source link

the loss is nan when i make the memory in NTM bigger? #52

Closed ghost closed 7 years ago

ghost commented 7 years ago

In my task, when i make the memory size (50X50) with clipnorm = 10 the loss gives good numbers but if the memory is (50X512) with clipnorm = 10 or even 5 it gives nan why?

EderSantana commented 7 years ago

I haven't checked that problem. Perhaps the softmax is vanishing. Will check this later. Thanks for the report