def build_network(conv_dim, z_size):
# define discriminator and generator
D = Discriminator(conv_dim)
G = Generator(z_size=z_size, conv_dim=conv_dim)
# initialize model weights
D.apply(weights_init_normal)
G.apply(weights_init_normal)
return D, G
def train_gan(data_path, save_gen_path, save_disc_path, feedback_fn=None):
data_loader = load_dataset(path=data_path, batch_size=128)
gen = Generator().to('cuda')
disc = Discriminator().to('cuda')
criterion = torch.nn.BCEWithLogitsLoss()
lr = 3e-4
gen_opt = torch.optim.Adam(gen.parameters(), lr=lr, betas=(0.5, 0.999))
disc_opt = torch.optim.Adam(disc.parameters(), lr=lr, betas=(0.5, 0.999))
gen = gen.apply(weights_init)
disc = disc.apply(weights_init)
n_epochs = 5
gen, disc = _train_loop(data_loader=data_loader,
gen=gen,
disc=disc,
criterion=criterion,
gen_opt=gen_opt,
disc_opt=disc_opt,
n_epochs=5,
feedback_fn=feedback_fn)
torch.save(gen.state_dict(), save_gen_path)
torch.save(disc.state_dict(), save_disc_path)