Is it possible to train it on custom dataset?

[AdarshMJ]

I have implemented a GAN network for predicting future frames of a video. I was wondering how to subject my code to AdaGAN. Any pointers would be really helpful.

[tolstikhin]

Unfortunately, the way this project is currently structured does not easily allow you to take your GAN architecture and run it as a black box with AdaGan scheme.

What you can do instead is to
(a) add your dataset to the DataHandler class of datahandler.py
(b) incorporate your specific GAN application in the gan.py
(c) make AdaGan class of adagan.py use your newly introduced GAN architecture together with your new dataset (look around lines 56-83 of adagan.py

Subject to couple of further minor changes this should work.

[AdarshMJ]

Thank you so much for your insights. I still had a question, The datasets fed to the datahandler.py are they supposed to be in specific format? My GAN network directly takes images as input.

[tolstikhin]

As you can see from datahandler.py, eventually they should end up as tensors in (N, W, H, C) format, first dimension enumerating individual input instances, then width, height, and finally channels. Also you can play around with storing the entire dataset in the memory vs. loading on-demand and memorizing, which could be preferable with larger datasets. For the former, check out for instance load_mnist procedure, for the latter --- load_celeba.