Loss sometimes goes to nan even with the gradient clipping

[carpedm20]

Didn't figure out why yet and any advice for this is welcome!

[jli05]

not sure if it’s related but do softmax(xxxx + eps)

On 31 Dec 2015, at 23:13, Taehoon Kim notifications@github.com wrote:

Didn't figure out why yet and any advice for this is welcome!

—
Reply to this email directly or view it on GitHub #2.

[carpedm20]

@jli05 Thanks! I'll try it. I could only learn NTM with max_length=10 since now without nan loss. If it becomes more than 10, I think we need more than 100000 epochs which is different from referenced code.

[EderSantana]

@carpedm20 in my NTM implementation (and in a couple of others I saw out there) nans were usually caused by one of the following:

• Initializing the memory to zero. The memory appears in the denominator of the cosine distance and that makes it nan. Check if that is not your case and possibly add a small constant in the denominator and avoid initializing the memory to all zeros (make it a small constant).
• negative sharpening value. that creates a complex number and also makes the cost function go nan

I think there was a third case but I don't remember right now. Good luck debugging! :D

[lixiangnlp]

@EderSantana could you explain what is the meaning of negative sharpening value? Thanks

[EderSantana]

the sharpening value is is uses as pow(input, sharpening). So it can't be negative. Use a nonlinearity like softplus to avoid getting negative values: sharpening = tf.nn.softplus(sharpening).

[therealjtgill]

Having a negative sharpening value wouldn't make a real become imaginary. But in the paper Graves explicitly states that the sharpening value is >= 1, so softplus(gamma) + 1 would work fine.

a^(-b) = 1/(a^b)