CODEWITHSUNDEEP
tensorflow
Nipun
Nipun

Reputation: 4319

tensor flow giving nan loss when running with convolutional network

I was trying to create a convolutional NN for house numbers images in tensorflow http://ufldl.stanford.edu/housenumbers/

I am getting cost as nan right in the first step when I am running my code. Here is the link to the github where I have put in my code https://github.com/ibnipun10/TensorFlow/blob/master/convhouseNumbers.ipynb

Please let me know where I am making mistake

Upvotes: 3

Views: 1230

Answers (1)

nessuno
nessuno

Reputation: 27060

The loss is NaN because the gradient is just exploded.

Your code looks ok, but your learning rate is high. Try with a lower learning rate (like 1e-2 or 1e-3) and see if the gradient still explode.

In addiction, I don't know if the images you're using in training have just been converted to float values and scaled in order to have zero mean and unit norm, but usually this is a step needed when working with images that helps to avoid gradient explosions.

Tensorflow have a function for that: tf.image.per_image_whitening

Upvotes: 5

Related Questions