Hashed cross-product transformation in PyTorch

I want to implement a hashed cross product transformation like the one Keras uses:

No gradient provided for any variable when using tf.numpy_function loss

I want to code a custom loss function that uses the Wasserstein distance from scipy: