from GAN.utils.init import InitNormal
from keras.optimizers import Adam, SGD, RMSprop
if __name__ == '__main__':
nbatch = 128
nmax = nbatch * 100
npxw, npxh = 64, 128
from load import people, load_all
va_data, tr_stream, _ = people(pathfile='protocol/PPPS.txt', size=(npxw, npxh), batch_size=nbatch)
g = Generator(g_size=(8, npxh, npxw), g_nb_filters=128, g_nb_coding=500, g_scales=4, g_init=InitNormal(scale=0.002))#, g_FC=[5000])
d = Discriminator(d_size=g.g_size, d_nb_filters=128, d_scales=4, d_init=InitNormal(scale=0.002))#, d_FC=[5000])
gan = GAN(g, d)
from keras.optimizers import Adam, SGD, RMSprop
gan.fit(tr_stream,
save_dir='./samples/parsing_skeleton/',
k=1,
nbatch=nbatch,
nmax=nmax,
opt=Adam(lr=0.0002, beta_1=0.5, decay=1e-5),
transform=transform_skeleton, #opt=RMSprop(lr=0.01))
inverse_transform=inverse_transform_skeleton)
def main(args):
pl.seed_everything(args.seed)
initial_transforms = [
transforms.CenterCrop(168),
transforms.Resize(64),
]
augmentation_transforms = [
transforms.RandomHorizontalFlip(p=0.5),
]
final_transforms = []
if args.model.lower() == "realnvp":
datainfo = DataInfo("celeba", 3, 64)
prior = Normal(loc=0.0, scale=1.0)
hyperparams = Hyperparameters(base_dim=32, res_blocks=2)
model = RealNVP(datainfo, prior, hyperparams, args.lr)
elif args.model.lower() == "vae":
model = VAE(latent_dim=args.latent_dim,
lr=args.lr,
kl_weight=args.kl_weight)
elif args.model.lower() == "gan":
final_transforms = [
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
]
model = GAN(nz=args.latent_dim,
ngf=args.generator_features,
ndf=args.discriminator_features,
lr=args.lr)
elif args.model.lower() == "glow":
final_transforms = [preprocess]
model = Glow((64, 64, 3),
flow_coupling=args.flow_coupling,
lr=args.lr,
warmup=args.warmup)
else:
raise ValueError(
f"{args.model} is not a valid model name. Use -h flag for help.")
train_transform = transforms.Compose(initial_transforms +
augmentation_transforms +
[transforms.ToTensor()] +
final_transforms)
test_transform = transforms.Compose(initial_transforms +
[transforms.ToTensor()] +
final_transforms)
train_dataset = datasets.CelebA(root=args.data_root,
split="train",
transform=train_transform)
val_dataset = datasets.CelebA(root=args.data_root,
split="test",
transform=test_transform)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers)
val_loader = DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers)
if args.model.lower() == "glow":
model.train()
if args.gpus and args.gpus > 0:
model.to('cuda')
init_batches = []
for x, _ in islice(train_loader, None, args.n_init_batches):
init_batches.append(x)
init_batches = torch.cat(init_batches).to(model.device)
with torch.no_grad():
model(init_batches)
model.cpu()
tt_logger = TestTubeLogger(args.logs_root, name=args.model.lower())
checkpoint_callback = ModelCheckpoint(
verbose=True,
save_last=True,
)
trainer = pl.Trainer.from_argparse_args(
args,
logger=tt_logger,
deterministic=True,
min_epochs=5,
callbacks=[checkpoint_callback],
)
trainer.fit(model, train_loader, val_loader)
if __name__ == '__main__':
nbatch = 128
nmax = nbatch * 100
npxw, npxh = 64, 128
from load import people
va_data, tr_stream, _ = people(pathfile='protocol/cuhk01-train.txt',
size=(npxw, npxh),
batch_size=nbatch)
g = Generator(g_size=(3, npxh, npxw),
g_nb_filters=128,
g_nb_coding=200,
g_scales=4,
g_init=InitNormal(scale=0.002))
d = Discriminator(d_size=g.g_size,
d_nb_filters=128,
d_scales=4,
d_init=InitNormal(scale=0.002))
gan = GAN(g, d)
from keras.optimizers import Adam, SGD, RMSprop
gan.fit(tr_stream,
save_dir='/home/shaofan/Projects/JSTL/transfer/gan/',
k=1,
nbatch=nbatch,
nmax=nmax,
opt=Adam(lr=0.0003, beta_1=0.5, decay=1e-5))
#opt=RMSprop(lr=0.01))
class mnist_stream():
def __init__(self):
(x_train, _), (_, _) = mnist.load_data()
x = np.expand_dims(x_train / 127.5 - 1, 1)
self.x = x
def __call__(self, bs):
return self.x[np.random.choice(self.x.shape[0],
replace=False,
size=(bs, ))]
if __name__ == '__main__':
coding = 200
img_shape = (1, 28, 28)
g = basic_gen((coding, ), img_shape, nf=64, scale=2, FC=[256])
d = basic_dis(img_shape, nf=64, scale=2, FC=[512])
gan = GAN(g, d, init=InitNormal(scale=0.02))
from keras.optimizers import Adam, SGD, RMSprop
gan.fit(mnist_stream(),
niter=5000,
save_dir='./quickshots/mnist',
k=3,
save_iter=100,
nbatch=128,
opt=Adam(lr=0.0002, beta_1=0.5))
return x, y, random_stream
if __name__ == '__main__':
nbatch = 128
x, y, stream = get_mnist(nbatch)
g = Generator(g_size=(1, 28, 28),
g_nb_filters=64,
g_nb_coding=200,
g_scales=2,
g_FC=[1024],
g_init=InitNormal(scale=0.05))
d = Discriminator(d_size=g.g_size,
d_nb_filters=64,
d_scales=2,
d_FC=[1024],
d_init=InitNormal(scale=0.05))
gan = GAN(g, d)
from keras.optimizers import Adam, SGD, RMSprop
gan.fit(stream,
save_dir='./labs/mnist',
k=1,
nbatch=nbatch,
opt=Adam(lr=0.0002, beta_1=0.5, decay=1e-5))
# 10/26/16: if not initialize in a proper way
# all-zero will appears in BN layer and cause nan
# solution: use leaky relu
from GAN.utils.data import transform, inverse_transform
from GAN.utils.init import InitNormal
if __name__ == '__main__':
g = Generator(g_size=(1, 28, 28),
g_nb_filters=64,
g_nb_coding=200,
g_scales=2,
g_FC=[1024],
g_init=InitNormal(scale=0.05))
d = Discriminator(d_size=g.g_size,
d_nb_filters=64,
d_scales=2,
d_FC=[1024],
d_init=InitNormal(scale=0.05))
gan = GAN(g, d)
vis_grid(inverse_transform(g.random_generate(100)), (10, 10), 'sample.png')
g.load_weights('{}/{}_gen_params.h5'.format('samples/mnist_bck/', 600))
vis_grid(inverse_transform(g.random_generate(100)), (10, 10),
'sample2.png')
img = g.random_generate(128)
reverse(g, img, savedir='reverse/mnist/')
def reverse(self, target, min=-1.0, max=+1.0, savedir=None):
import theano
import theano.tensor as T
from ..utils.vis import vis_grid
from ..utils.data import transform, inverse_transform