I had been trying to train my autoencoder with a GAN component on and off for a couple of months and it just didn't seem to be working very well. I thought that maybe the autoencoder and the discriminator errors were somehow cancelling each other out or something. Just for the hell of it I decided to try to use the discriminator to optimize a reconstructed image to look real, just to see what the result would be. Instead of optimizing the weights, I created a Variable of the input and optimized that instead. To my surprise I ended up with weird splotches of primary colors against a white background, it actually made the image look less and less real rather than more. After seeing that I decided that there must be some major problem with my code so I went through it in greater detail.
I decided to train all three networks from scratch (the three being the encoder, the decoder and the discriminator) to see what would happen. I was surprised to see that the generator did not seem to be learning ANYTHING and neither did the discriminator. I found a tutorial on creating a GAN in PyTorch and I went through the training code to see how it differed from mine.
I had written my code to optimize it for speed, training the autoencoder without the GAN already took about 4 hours per epoch on a (free) K80 on Colab so I didn't want to slow that down much more, so I tried to minimize the numebr of times data had to be passed through the networks. The tutorial did not do that. First it ran a batch of real data through the discriminator, computed the gradients but did NOT back propagate them. Then it used the generator to generate a batch of faked data, passed that through the discriminator, computed the gradients, added them to the gradients from the first batch and THEN did the back prop. Then it ran the same batch of faked data through the discriminator again, and used that to update the generator. This was different from my code in several major ways:
- I was using a single batch containing half real images and half reconstructed images to train my discriminator.
- I was training passing data through each network one single time per batch.
- I wasn't detaching the reconstructed data before passing them through the discriminator.
After updating my code to bring it more in line with the tutorial both networks began to learn, I think that major change was detaching the reconstructed images before putting them through the discriminator. However I noticed a few strange things regarding the discriminator batches:
- If I used a single batch containing both real and constructed images to train the discriminator it learned very quickly, it's loss approached 0 very quickly, and the discriminator loss component of the generator overwhelmed the autoencoder loss, which sort of fluctuated but didn't decrease very much.
- If I trained using two batches, each containing only images for a single label, it's accuracy hovered around 50% and the autoencoder loss decreased rapidly.
I read in a couple of places that using separate batches was a trick to make GANs train better, but no one really had an explanation for why this worked. What I am currently doing it using separate batches most of the time, before every n batches I use a single batch to encourage the discriminator to learn a bit more. I've tested values for n of 8, 16, 32 and 64. Most of those seemed to result in the worst of both worlds, nothing really seemed to improve, but with n = 64 the autoencoder loss is again decreasing, although slowly, and the discriminator accuracy is hovering around 52% rather than the 49-50% it was at using all separate batches.
To me using separate batches doesn't intuitively make sense, I don't see how the network can really learn to differentiate between classes when it only sees one class at a time. Of course the gradients are then added, and the differences should cancel out, with what's left indicating how to differentiate the classes; but to me it seems much more efficient to learn from mixed batches. One would never consider training a network on, say, the CIFAR dataset with each batch consisting exclusively of a single class. Maybe that's the point, to slow down the discriminator's learning enough for the generator to keep up? Anyway I will continue to experiment and see what works and what doesn't work.