Loss value change to NaN

[tvmsouza]

Hello sounakdey,

i tried reproduce your result from paper, and everytime on the end of first epoch of training phase,
the loss value from the model changes to NaN. I think that is some configuration in Keras-JS that needs to be changed, so i tried put the value of epsilon, that's by the default 1e-07, to 1e-05 to avoid this problem. You passed for something like this when you trained your model?

Maybe is my Keras or Tensorflow Version? Whats is the version that you use?

Thank you.

[sounakdey]

No i didnot change the epsilon value.... the code provided is running properly i checked it... i used keras 2.0.8 and tensorflow1.3.0.... even with my previous keras 1 it runned porperly...

[tvmsouza]

Hello again sounakdey,

firsty thanks for answer my question, but the code continued to crack.

I see that you use the example of siamese network from keras.

So i found a solution. I saw that the siamese network from keras had pieces of code that had to be changed.

Like this:

def euclidean_distance(vects):
x, y = vects
return K.sqrt((K.sum(K.square(x - y), axis=1, keepdims=True),)

for this:

def euclidean_distance(vects):
x, y = vects
return K.sqrt(K.maximum(K.sum(K.square(x - y), axis=1, keepdims=True), K.epsilon()))

Now the code is running well.

[sounakdey]

Ahh I got it... It seems it needs a epsilon to stop the squared distance falling beyond certain value... Great to hear that the code is running perfectly...