Support for changing the size of the training data and thus the output images on an already trained model?

[ProGamerGov]

In an effort to speed up training times, I would like to train a model first on 32x32px images, and then on 64x64px, 128x128px, 256x256px, etc...

Currently when running:

hypergan train faceGAN/data/images -s 64x64x3 -f png -b 32 --config rmsprop-lsgan-64x64 --sampler static_batch --sample_every 5

Or:

hypergan train faceGAN/data/images -s 64x64x3 -f png -b 32 --config rmsprop-lsgan --sampler static_batch --sample_every 5

The models saved in .hypergan can only continue their training if you use the same size value. This sort of scaling up technique that I want to try seems to be successful for style transfer system, and it seems that logically the model can exploit it's already learned knowledge to speed up training on larger images.

[martyn]

I think you need to use the API for this with something like stack gan. https://arxiv.org/abs/1612.03242

Switching sizes actually changes the architecture of the network, which invalidates all the weights.

[ProGamerGov]

@martyn Would it be possible to make creating stacked sets of different size models (for creating images larger than 256), possible with the command line input, in addition to the API?

It would nice if it were easier to replicate results like these, with HyperGAN: https://mtyka.github.io/machine/learning/2017/06/06/highres-gan-faces.html

https://www.reddit.com/r/MachineLearning/comments/6frfb5/p_portraits_of_imaginary_people_gans_at_4000x4000/