Simple SN-GAN to generate CryptoPunks. Neural network architecture and training code has been modified from the PyTorch DCGAN example. See Notes for more information.

Punks during training

Linear interpolation between two punks

Generate 64 punks using pretrained model:

import torch
from torchvision.utils import save_image
from train import Generator

model = Generator()
model.load_state_dict(torch.load("models/net_g_epoch_999.pth"))
z = torch.randn(64, 100, 1, 1)
punks = model(z)
save_image(punks, "punks.png", normalize=True)

On CUDA machine:

python train.py

Check out directory for weights and sample images.

Using Spectral Normalization created very nice punks, but had a tendency to create very transparent ones:

Using batch normalization seems to negate this, but causes mode collapse

Instead, penalizing the model for a high mean alpha produces very good punks, with no transparency issues:

# minimize criterion + mean of alpha channel
loss_g = criterion(output, label) + fake[:, -1].mean()

In order to encourage more variety in the outputs, we generate a feature matrix of all the punks (using punks.attributes):

    3D Glasses  Bandana  Beanie  Big Beard  Big Shades  ...  Ape  Human  Zombie  Female  Male
id                                                      ...
0            0        0       0          0           0  ...    0      1       0       1     0
1            0        0       0          0           0  ...    0      1       0       0     1
2            0        0       0          0           0  ...    0      1       0       1     0
3            0        0       0          0           0  ...    0      1       0       0     1
4            0        0       0          0           1  ...    0      1       0       0     1

For each image in the batch, randomly select an attribute (column), and then randomly select a punk that has that attribute. This ensures that the discriminator is exposed to all of the attributes in a more balanced manor than running through all the punks.