Discriminator loss of WGAN-gp drop sharply to zero

[AboMad]

I am trying to train WGAN-GP with a grayscale images dataset of about 13000 samples. I am using ResNet architecture to generate 64x64 images based on the implementation here (https://github.com/jalola/improved-wgan-pytorch) I faced the problem of the sharp dropping of the discriminator loss. I tried to solve that by:

reduce lr
randomly switch between real and fack images
apply L2 regularization (for both G & D models)
apply many types of data Augmentation But still stuck with the problem. figures show the D (multiply by -1) and G losses of 10K iterations.

Can I have an idea of the reason for that sharp drop?