Train GAN on array?


Is there a way to train a GAN without images, but rather numpy data/torch tensors? My data would be the equivalent of 5-6+ channel images, and I don’t know if such a format exists and whether torchvision would support the extension. So, is it possible to follow the generic DCGAN tutorial for example, but exchange images for arrays/tensors?


The model itself is expecting tensors as its input. Assuming you can create the input tensors without running into errors from any transforms or image libs, it should work.
I don’t know if the DCGAN architecture is suitable for your format, but you could certainly try to train the model.

1 Like