在自定义损失函数中重塑张量

Question

I have a problem similar to this question . I am trying to devise a loss function in keras given as:

def depth_loss_func(lr):
    def loss(actual_depth,pred_depth):
        actual_shape = actual_depth.get_shape().as_list()
        dim = np.prod(actual_shape[1:])
        actual_vec = K.reshape(actual_depth,[-1,dim])
        pred_vec = K.reshape(pred_depth,[-1,dim])
        di = K.log(pred_vec)-K.log(actual_vec)
        di_mean = K.mean(di)
        sq_mean = K.mean(K.square(di))

        return (sq_mean - (lr*di_mean*di_mean))
    return loss

based on the answer given in this question . However, I am getting an error:

 TypeError: unsupported operand type(s) for *: 'NoneType' and 'NoneType'

Specifically this statement gives the following output

(Pdb) actual_depth.get_shape()
TensorShape([Dimension(None), Dimension(None), Dimension(None)])

The backend is TensorFlow. Thanks for your help.

Answer 1

I managed to reproduce you exception with a Tensor of shape (None, None, None, 9) , when calling np.prod() like this:

from keras import backend as K

#create tensor placeholder
z = K.placeholder(shape=(None, None, None, 9))
#obtain its static shape with int_shape from Keras
actual_shape = K.int_shape(z)
#obtain product, error fires here... TypeError between None and None
dim = np.prod(actual_shape[1:])

This happens because you are trying to multiply two elements of type None , even though you sliced your actual_shape (as more than 1 elements where None ). In some cases you can even get TypeError between None and int , if only one none-type element remains after slicing.

Taking a look at the answer you mentioned, they specify what to do in those situations, quoting from it:

For the cases where more than 1 dimension are undefined, we can use tf.shape() with tf.reduce_prod() alternatively.

Based on that, we can translate those operations to the Keras API, by using K.shape() ( docs ) and K.prod() ( docs ), respectively:

z = K.placeholder(shape=(None, None, None, 9))
#obtain Real shape and calculate dim with prod, no TypeError this time
dim = K.prod(K.shape(z)[1:])
#reshape
z2 = K.reshape(z, [-1,dim])

Also, for the case where only one dimension is undefined remember to use K.int_shape(z) or its wrapper K.get_variable_shape(z) instead of just get_shape() , as also defined in the backend ( docs ). Hope this solves your problem.