Experiencing model collapse

[demiahmed]

Hi first of all thanks for this notebook it works very well as expected. However, in my dataset size of around 5k diverse images I experience a model collapse. i.e, the variability of the generator is too low (it makes images that are almost the same - cycles around 2 or 3 images only).

I would like to increase the size of the latent vector. Could you please point me in the right direction?

This is what I'm currently doing

old

self.G.add(Reshape(target_shape = [1, 1, 4096], input_shape = [4096]))

new

self.G.add(Reshape(target_shape = [1, 1, 4096], input_shape = [100]))

But I am getting an error while compiling the model. Any help or pointers would be much appreciated.

[TTK95]

You want to reshape a Tensor with shape 1x100 into a Tensor with shape 1x1x4096..
try: self.G.add(Reshape(target_shape = [1, 1, 100], input_shape = [100]))

but if you want to increase the input vector, try:
self.G.add(Reshape(target_shape = [1, 1, 4096*X], input_shape = [4096*X])) where X = 1,2,3,...