I try to implement a keras backend function f, to be used in a custom loss function. When I compute f(x, y) or even K.eval(f(x, y)) it is fine, yet printing K.eval(f(x, y)) crashes the kernel.

Pycharm writes

Process finished with exit code -1073741819 (0xC0000005)

Jupyter notebook outputs

"The kernel crashed, it will restart automatically".

Runing this in a terminal also crashes python.

I've tried uninstalling and reinstalling keras. I have python 3.6, keras 2.2.4, tensorflow 1.13.1, and numpy 1.15.2. OS is windows 10.

A minimal example :

from keras import backend as K
import numpy as np    

def f(x, y):
    return x + y

b = K.constant(np.array([1, 0, 0, 0]))
c = K.constant(np.array([0, 1, 0, 1]))
# this works if you comment the rest out :
evaluation = K.eval(f(b, c))
# this KILLS the freaking KERNEL !!!
print(evaluation)

A solution (?)

I updated numpy to 1.16.3 and it seems to work again.

It could come from your tensorflow version. I tried your code with the same versions except for the Tensorflow version which is 1.10 (instead of 1.13.1) and it works.