Convert kernel_mask into a constant tensor

[Larst0]

When I use the given implementation for training, I always get NaN values at the output. Sometimes this happens after a few training steps and sometimes after a few epochs (depending on the training data used).

While debugging, I noticed that the kernel_mask was updated. I think this is because K.ones(shape=...) returns a trainable variable if all entries in the passed shape are >0. In the original PyTorch implementation the kernel_mask is initialized using weight_maskUpdater = torch.ones(...), which by default creates a non-trainable tensor (since requires_grad=False).

After replacing K.ones(...) with K.constant(...) the NaN values no longer occur.