I'm adapting a GAN described here used for generating binary output. It's trained on binarized MNIST data, with a size of 28x28 so 784 values. I want to adapt it to train on and generate 1D vectors with a length of ~4000. Since the original GAN uses MLP and not convolution the 1D part should not matter, but I'm wondering if the 5x increased input size will cause problems. Should I increase the size of the hidden layers in a case like this, and if yes, by how much? Are there good rules of thumb here?
Additionally, I will also only have ~2500 training samples compared to the 60k in MNIST. Should this have influence on the size of the hidden layers as well? And would it help to 'recycle' or reuse training data?