MaximumMarginGANs

Code for paper: Support Vector Machines, Wasserstein's distance and gradient-penalty GANs maximize a margin

Discussion at https://ajolicoeur.wordpress.com/MaximumMarginGANs.

This basically the same code as https://github.com/AlexiaJM/relativistic-f-divergences, but with more options.

Citation

If you use our novel gradient penalties or would like to mention that gradient penalties correspond to having a maximum-margin discriminator, please cite us in your work:

@article{jolicoeur2019connections}
  title={Connections between Support Vector Machines, Wasserstein distance and gradient-penalty GANs},
  author={Jolicoeur-Martineau, Alexia},
  journal={arXiv preprint arXiv:1910.06922},
  year={2019}
}

Sample PyTorch code to use L1, L2, Linfinity gradient penalties with hinge or LS:

# Best setting (novel Hinge Linfinity gradient penalty)
grad_penalty_Lp_norm = 'Linf'
penalty_type = 'hinge'

# Default setting from WGAN-GP and most cases (L2 gradient penalty)
grad_penalty_Lp_norm = 'L2'
penalty_type = 'LS'

# Calculate gradient
penalty = 20 # 10 is the more usual choice
u.resize_(batch_size, 1, 1, 1)
u.uniform_(0, 1)
x_both = x.data*u + x_fake.data*(1-u) # interpolation between real and fake samples
x_both = x_both.cuda()
x_both = Variable(x_both, requires_grad=True)
y0 = D(x_both)
grad = torch.autograd.grad(outputs=y0, inputs=x_both, grad_outputs=grad_outputs, retain_graph=True, 
create_graph=True, only_inputs=True)[0]
x_both.requires_grad_(False)
grad = grad.view(current_batch_size,-1)
			
if grad_penalty_Lp_norm = 'Linf': # Linfinity gradient norm penalty (Corresponds to L1 margin, BEST results)
  grad_abs = torch.abs(grad) # Absolute value of gradient
  grad_norm , _ = torch.max(grad_abs,1)
elif grad_penalty_Lp_norm = 'L1': # L1 gradient norm penalty (Corresponds to Linfinity margin, WORST results)
  grad_norm = grad.norm(1,1) 
else: # L2 gradient norm penalty (Corresponds to L2 margin, this is what people generally use)
  grad_norm = grad.norm(2,1)

if penalty_type == 'LS': # The usual choice, penalize values below 1 and above 1 (too constraining to properly estimate the Wasserstein distance)
  constraint = (grad_norm-1).pow(2)
elif penalty_type == 'hinge': # Penalize values above 1 only (best choice)
  constraint = torch.nn.ReLU()(grad_norm - 1)

constraint = constraint.mean()
grad_penalty = penalty*constraint
grad_penalty.backward(retain_graph=True)

Needed

• Python 3.6
• Pytorch (Latest from source)
• Tensorflow (Latest from source, needed to get FID)
• Cat Dataset (http://academictorrents.com/details/c501571c29d16d7f41d159d699d0e7fb37092cbd)

To do beforehand

• Change all folders locations in GAN.py (and startup_tmp.sh, fid_script.sh, experiments.sh if you want FID and replication of the paper)
• Make sure that there are existing folders at the locations you used
• To get the CAT dataset: open and run each necessary lines of setting_up_script.sh in same folder as preprocess_cat_dataset.py (It will automatically download the cat datasets, if this doesn't work well download it from http://academictorrents.com/details/c501571c29d16d7f41d159d699d0e7fb37092cbd)

To run models

• HingeGAN Linfinity grad norm penalty with max(0, ||grad||-1):
  • python GAN.py --loss_D 3 --image_size 32 --CIFAR10 True --grad_penalty True --l1_margin --penalty-type 'hinge'
• WGAN Linfinity grad norm penalty with max(0, ||grad||-1):
  • python GAN.py --loss_D 4 --image_size 32 --CIFAR10 True --grad_penalty True --l1_margin --penalty-type 'hinge'
• WGAN L2 grad norm penalty with (||grad||-1)^2 (i.e., WGAN-GP):
  • python GAN.py --loss_D 4 --image_size 32 --CIFAR10 True --grad_penalty True

To replicate the paper

• Open experiments.sh and run the lines you want