(this list is no longer maintained, and I am not sure how relevant it is in 2020)
While research in Generative Adversarial Networks (GANs) continues to improve the fundamental stability of these models, we use a bunch of tricks to train them and make them stable day to day.
Here are a summary of some of the tricks.
Here's a link to the authors of this document
If you find a trick that is particularly useful in practice, please open a Pull Request to add it to the document. If we find it to be reasonable and verified, we will merge it in.
- normalize the images between -1 and 1
- Tanh as the last layer of the generator output
In GAN papers, the loss function to optimize G is min (log 1-D), but in practice folks practically use max log D
- because the first formulation has vanishing gradients early on
- Goodfellow et. al (2014)
In practice, works well:
- Flip labels when training generator: real = fake, fake = real
- Dont sample from a Uniform distribution
- Sample from a gaussian distribution
- When doing interpolations, do the interpolation via a great circle, rather than a straight line from point A to point B
- Tom White's Sampling Generative Networks ref code https://github.com/dribnet/plat has more details
- Construct different mini-batches for real and fake, i.e. each mini-batch needs to contain only all real images or all generated images.
- when batchnorm is not an option use instance normalization (for each sample, subtract mean and divide by standard deviation).
- the stability of the GAN game suffers if you have sparse gradients
- LeakyReLU = good (in both G and D)
- For Downsampling, use: Average Pooling, Conv2d + stride
- For Upsampling, use: PixelShuffle, ConvTranspose2d + stride
- PixelShuffle: https://arxiv.org/abs/1609.05158
- Label Smoothing, i.e. if you have two target labels: Real=1 and Fake=0, then for each incoming sample, if it is real, then replace the label with a random number between 0.7 and 1.2, and if it is a fake sample, replace it with 0.0 and 0.3 (for example).
- Salimans et. al. 2016
- make the labels the noisy for the discriminator: occasionally flip the labels when training the discriminator
- Use DCGAN when you can. It works!
- if you cant use DCGANs and no model is stable, use a hybrid model : KL + GAN or VAE + GAN
- Experience Replay
- Keep a replay buffer of past generations and occassionally show them
- Keep checkpoints from the past of G and D and occassionaly swap them out for a few iterations
- All stability tricks that work for deep deterministic policy gradients
- See Pfau & Vinyals (2016)
- optim.Adam rules!
- See Radford et. al. 2015
- Use SGD for discriminator and ADAM for generator
- D loss goes to 0: failure mode
- check norms of gradients: if they are over 100 things are screwing up
- when things are working, D loss has low variance and goes down over time vs having huge variance and spiking
- if loss of generator steadily decreases, then it's fooling D with garbage (says martin)
- Dont try to find a (number of G / number of D) schedule to uncollapse training
- It's hard and we've all tried it.
- If you do try it, have a principled approach to it, rather than intuition
For example
while lossD > A:
train D
while lossG > B:
train G
- if you have labels available, training the discriminator to also classify the samples: auxillary GANs
- Add some artificial noise to inputs to D (Arjovsky et. al., Huszar, 2016)
- adding gaussian noise to every layer of generator (Zhao et. al. EBGAN)
- Improved GANs: OpenAI code also has it (commented out)
- especially when you have noise
- hard to find a schedule of number of D iterations vs G iterations
- Mixed results
- Use an Embedding layer
- Add as additional channels to images
- Keep embedding dimensionality low and upsample to match image channel size
- Provide noise in the form of dropout (50%).
- Apply on several layers of our generator at both training and test time
- https://arxiv.org/pdf/1611.07004v1.pdf
- Soumith Chintala
- Emily Denton
- Martin Arjovsky
- Michael Mathieu