if args.use_morph_network:
Gz = EncoderMorphNet28(args.l_size,
args.h_size,
args.use_mish,
n_channels=1,
cap_variance=True,
block_Gz_morph_grads=args.no_morph_loss_on_Gz)
else:
Gz = Encoder28(args.l_size,
args.h_size,
args.use_mish,
n_channels=1,
cap_variance=True)
Gx = Generator28(args.l_size,
args.h_size,
args.use_mish,
n_channels=1,
sigmoid_out=True)
D = ALIDiscriminator28(args.l_size,
args.h_size,
use_bn=args.use_batchnorm_in_D,
use_mish=args.use_mish,
n_channels=1,
dropout=args.dropout_rate,
fc_h_size=args.fc_h_size)
if args.use_morph_network:
G_no_mn_optimizer = torch.optim.Adam(list(Gz.Gz_params()) +
list(Gx.parameters()),
lr=args.lr,
betas=(0.5, 0.999))
dataset = CelebaCropped(split="train",
download=True,
transform=transforms.Compose([
transforms.Resize(28),
transforms.ToTensor(),
transforms.Lambda(lambda img: img * 2 - 1)
]))
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=12)
G = Generator28(args.l_size,
args.h_size,
args.use_mish,
not args.no_bias_in_G,
n_channels=3)
D = Discriminator28(args.h_size,
use_bn=args.use_batchnorm_in_D,
use_mish=args.use_mish,
n_channels=3,
dropout=args.dropout_rate)
G_optimizer = torch.optim.Adam(G.parameters(), lr=args.lr, betas=(0.5, 0.999))
D_optimizer = torch.optim.Adam(D.parameters(), lr=args.lr, betas=(0.5, 0.999))
if args.cuda:
G = G.cuda()
D = D.cuda()
D.init_weights()
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4)
print("Dataset length: ", len(dataset))
Gz = Encoder28(args.l_size,
args.h_size,
args.use_mish,
n_channels=3,
cap_variance=True,
use_lr_norm=args.use_lr_norm)
Gx = Generator28(args.l_size,
args.h_size,
args.use_mish,
n_channels=3,
sigmoid_out=True,
use_lr_norm=args.use_lr_norm)
D = ALIDiscriminator28(args.l_size,
args.h_size,
use_bn=args.use_batchnorm_in_D,
use_mish=args.use_mish,
n_channels=3,
dropout=args.dropout_rate,
fc_h_size=args.fc_h_size)
G_optimizer = torch.optim.Adam(list(Gz.parameters()) + list(Gx.parameters()),
lr=args.lr,
betas=(0.5, 0.999))
D_optimizer = torch.optim.Adam(D.parameters(), lr=args.lr, betas=(0.5, 0.999))
# Code for loading frs model when frs based reconstruction loss is used
]))
valid_pairs_dataset = CelebaCroppedPairs(split="valid",
download=True,
transform=transforms.Compose([
transforms.Resize(28),
transforms.ToTensor(),
]))
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4)
Gz = Encoder28(args.l_size, args.h_size, n_channels=3)
Gx = Generator28(args.l_size, args.h_size, n_channels=3, sigmoid_out=True)
D = VAEGANDiscriminator28(args.h_size,
use_bn=args.use_batchnorm_in_D,
n_channels=3,
dropout=args.dropout_rate)
Gz_optimizer = torch.optim.Adam(Gz.parameters(),
lr=args.lr,
betas=(0.5, 0.999))
Gx_optimizer = torch.optim.Adam(Gx.parameters(),
lr=args.lr,
betas=(0.5, 0.999))
D_optimizer = torch.optim.Adam(D.parameters(), lr=args.lr, betas=(0.5, 0.999))
if args.cuda:
Gz = Gz.cuda()
Gx = Gx.cuda()
args)
dataset = CelebaCropped(split="train",
download=True,
transform=transforms.Compose([
transforms.Resize(28),
transforms.ToTensor(),
]))
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=12)
G = Generator28(args.l_size,
args.h_size,
args.use_mish,
bias=False,
n_channels=3,
sigmoid_out=True)
D = Discriminator28(args.h_size,
use_bn=False,
use_mish=args.use_mish,
n_channels=3,
dropout=args.dropout_rate,
use_logits=True)
G_optimizer = torch.optim.Adam(G.parameters(), lr=args.lr, betas=(0.5, 0.999))
D_optimizer = torch.optim.Adam(D.parameters(), lr=args.lr, betas=(0.5, 0.999))
if args.cuda:
G = G.cuda()
D = D.cuda()
output_path = util.output.init_experiment_output_dir("mnist", "gan", args)
dataset = data.MNIST("data/downloads/mnist",
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Lambda(lambda img: img * 2 - 1)
]))
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=12)
G = Generator28(args.l_size, args.h_size, args.use_mish)
D = Discriminator28(args.h_size,
use_bn=args.use_batchnorm_in_D,
use_mish=args.use_mish)
if args.cuda:
G = G.cuda()
D = D.cuda()
G.init_weights()
D.init_weights()
G_optimizer = torch.optim.Adam(G.parameters(), lr=args.lr, betas=(0.5, 0.999))
D_optimizer = torch.optim.Adam(D.parameters(), lr=args.lr, betas=(0.5, 0.999))
listeners = [
valid_dataset = MNIST("data/downloads/mnist",
train=False,
download=True,
transform=transforms.Compose([
transforms.Resize(28),
transforms.ToTensor(),
transforms.Lambda(lambda img: img * 2 - 1)
]))
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=12)
Gz = Encoder28(args.l_size, args.h_size, args.use_mish, n_channels=1)
Gx = Generator28(args.l_size, args.h_size, args.use_mish, False, n_channels=1)
D = ALIDiscriminator28(args.l_size,
args.h_size,
use_bn=args.use_batchnorm_in_D,
use_mish=args.use_mish,
n_channels=1,
dropout=args.dropout_rate,
fc_h_size=args.fc_h_size)
G_optimizer = torch.optim.Adam(list(Gz.parameters()) + list(Gx.parameters()),
lr=args.lr,
betas=(0.5, 0.999))
D_optimizer = torch.optim.Adam(D.parameters(), lr=args.lr, betas=(0.5, 0.999))
if args.cuda:
Gz = Gz.cuda()
Gx = Gx.cuda()
dataset = CelebaCroppedTriplets(split="train", download=True, transform=transforms.Compose([
transforms.Resize(28),
transforms.ToTensor(),
]))
valid_dataset = CelebaCroppedTriplets(split="valid", download=True, transform=transforms.Compose([
transforms.Resize(28),
transforms.ToTensor(),
]))
dataloader = torch.utils.data.DataLoader(dataset, batch_size=args.batch_size, shuffle=True, num_workers=4)
print("Dataset length: ", len(dataset))
Gz = Encoder28(args.l_size, args.h_size, args.use_mish, n_channels=3, cap_variance=True, add_dense_layer=False, deterministic=True)
Gx = Generator28(args.l_size, args.h_size, args.use_mish, n_channels=3, sigmoid_out=True, add_dense_layer=False)
D = ALIDiscriminator28(args.l_size, args.h_size, use_bn=args.use_batchnorm_in_D, use_mish=args.use_mish, n_channels=3, dropout=args.dropout_rate, fc_h_size=args.fc_h_size)
G_optimizer = torch.optim.Adam(list(Gz.parameters()) + list(Gx.parameters()), lr=args.lr, betas=(0.5, 0.999))
D_optimizer = torch.optim.Adam(D.parameters(), lr=args.lr, betas=(0.5, 0.999))
if args.cuda:
Gz = Gz.cuda()
Gx = Gx.cuda()
D = D.cuda()
Gz.init_weights()
Gx.init_weights()
D.init_weights()
listeners = [
LossReporter(),