To learn about generative adversarial models.
For this lab, you will need to implement a generative adversarial network (GAN). Specifically, we will be using the technique outlined in the paper Improved Training of Wasserstein GANs.
You should turn in an iPython notebook that shows two plots. The first plot should be random samples from the final generator. The second should show interpolation between two faces by interpolating in
You must also turn in your code, but your code does not need to be in a notebook if it's easier to turn it in separately (but please zip your code and notebook together in a single zip file).
NOTE: this lab is complex. Please read through the entire spec before diving in.
Also, note that training on this dataset will likely take some time. Please make sure you start early enough to run the training long enough!
Your code/image will be graded on the following:
The dataset you will be using is the "celebA" dataset, a set of 202,599 face images of celebrities. Each image is 178×218. You should download the “aligned and cropped” version of the dataset. Here is a direct download link (1.4G), and here is additional information about the dataset.
This lab will help you develop several new pytorch skills, as well as understand some best practices needed for building large models. In addition, we'll be able to create networks that generate neat images!
One of the advantages of the “Improved WGAN Training” algorithm is that many different kinds of topologies can be used. For this lab, I recommend one of three options:
Again, you are encouraged to use either a DCGAN-like architecture, or a ResNet.
Our reference implementation used 4 convolution layers, each followed by a batch norm layer and leaky relu (leak 0.2) No batch norm on the first layer.
Note that the discriminator simply outputs a single scalar value. This value should unconstrained (ie, can be positive or negative), so you should not use a relu/sigmoid on the output of your network.
The implementation of the improved Wasserstein GAN training algorithm (hereafter called “WGAN-GP”) is fairly straightforward, but involves a few new details:
requires_grad = Truefor the parameters of the discriminator. An easier way to do this would be to iterate through the discriminator model parameters and set
param.requires_grad = True
#initialize your generator and discriminator models #initialize separate optimizer for both gen and disc #initialize your dataset and dataloader for e in epochs: for true_img in trainloader: #train discriminator# #because you want to be able to backprop through the params in discriminator for p in disc_model.parameters(): p.requires_grad = True for p in gen_model.parameters(): p.requires_grad = False for n in range(critic_iters): disc_optim.zero_grad() # generate noise tensor z # calculate disc loss: you will need autograd.grad # call dloss.backward() and disc_optim.step() #train generator# for p in disc_model.parameters(): p.requires_grad = False for p in gen_model.parameters(): p.requires_grad = True gen_optim.zero_grad() # generate noise tensor z # calculate loss for gen # call gloss.backward() and gen_optim.step()
Your final deliverable is two images. The first should be a set of randomly generated faces. This is as simple as generating random
z variables, and then running them through your generator.
For the second image, you must pick two random
z values, then linearly interpolate between them (using about 8-10 steps). Plot the face corresponding to each interpolated
See the beginning of this lab spec for examples of both images.
We have recently tried turning off the batchnorms in both the generator and discriminator, and have gotten good results – you may want to start without them, and only add them if you need them. Plus, it's faster without the batchnorms.
The reference implementation was trained for 8 hours on a GTX 1070. It ran for 25 epochs (ie, scan through all 200,000 images), with batches of size 64 (3125 batches / epoch).
However, we were able to get reasonable (if blurry) faces after training for 2-3 hours.
I didn't try to optimize the hyperparameters; these are the values that I used:
beta1 = 0.5 # 0 beta2 = 0.999 # 0.9 lambda = 10 ncritic = 1 # 5 learning_rate = 0.0002 # 0.0001 batch_size = 200 batch_norm_decay=0.9 batch_norm_epsilon=1e-5
Changing to number of critic steps from 5 to 1 didn't seem to matter; changing the alpha parameters to 0.0001 didn't seem to matter; but changing beta1 and beta2 to the values suggested in the paper (0.0 and 0.9, respectively) seemed to make things a lot worse. Different set of numbers might works well for different people. So play around with the numbers that work well for you.
This code should be helpful to get the data:
!wget --load-cookies cookies.txt 'https://docs.google.com/uc?export=download&confirm='"$(wget --save-cookies cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=0B7EVK8r0v71pZjFTYXZWM3FlRnM' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')"'&id=0B7EVK8r0v71pZjFTYXZWM3FlRnM' -O img_align_celeba.zip !unzip -q img_align_celeba !mkdir test !mv img_align_celeba test
And using the data in a dataset class:
class CelebaDataset(Dataset): def __init__(self, root, size=128, train=True): super(CelebaDataset, self).__init__() self.dataset_folder = torchvision.datasets.ImageFolder(os.path.join(root) ,transform = transforms.Compose([transforms.Resize((size,size)),transforms.ToTensor()])) def __getitem__(self,index): img = self.dataset_folder[index] return img def __len__(self): return len(self.dataset_folder)