Lab 5.1 - error - "Negative dimensions are not allowed"

[violetwhitney]

When going through Lab 5.1 without changing any of the content, when I run the cell to build the keras model, I receive an error 'Negative dimensions are not allowed' referring to the hyper-parameter: 'num_hidden_1 = 1000'.

hyper-parameters

build Keras model

error

I attempted adjusting the 'num_hidden_1 = 1000' in the hyper parameters with no success. Any value lower than '479' gives the error 'range exceeds valid bounds', while '480' gives an error of 'float division by zero' and any value between '481-1000' keeps the error of 'negative dimensions are not allowed'.

Is anyone else running into this issue?

[danilnagy]

In your case, is Keras using the Tensorflow or Theano backend? Should be printed in the output of the first cell.

[danilnagy]

I think this has to do with the fact that Keras expects different variable ordering depending on if it is using the Theano or Tensorflow backend:

https://keras.io/backend/

In the 4th code cell, the X array is ordered in the way expected by Tensforflow, assuming you are using this backend. You can also tell Keras explicitly to use the Tensorflow ordering by adding the line:

K.set_image_dim_ordering('tf')

right after the import lines in the first code cell. Let me know if adding this line fixes the issue, I will add it to the assignment.

[violetwhitney]

Perfect! That code bit worked. It was using Theano.

Thanks!

[danilnagy]

Great, I'm glad that fix worked. To make it more robust, I have updated the Lab 5.1 code so that it will work with either variable ordering based on this example:

https://github.com/fchollet/keras/blob/master/examples/mnist_cnn.py

You can check out the changes in the latest commit:

762ef11

I'll keep the issue open for now in case anyone is using the old code and runs into the same problem.

Thanks for bringing this up!