def get_network(LEARNING_RATE: float, device: str):
G = Generator().to(device)
D = Discriminator().to(device)
criterion = nn.BCELoss()
d_optimizer = optim.Adam(D.parameters(), lr=LEARNING_RATE)
g_optimizer = optim.Adam(G.parameters(), lr=LEARNING_RATE)
return G, D, criterion, d_optimizer, g_optimizer
def train(opt, dataloader_m=None):
# Loss function
adversarial_loss = torch.nn.BCELoss()
# Initialize generator and discriminator
generator = TinyNet()
discriminator = Discriminator()
cuda = True if torch.cuda.is_available() else False
if cuda:
generator.cuda()
discriminator.cuda()
adversarial_loss.cuda()
Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
# Configure data loader
os.makedirs('./data/mnist', exist_ok=True)
dataloader = torch.utils.data.DataLoader(datasets.MNIST(
'./data/mnist',
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])),
batch_size=opt.batch_size,
shuffle=True)
# Optimizers
optimizer_G = torch.optim.Adam(generator.parameters(),
lr=opt.lr,
betas=(0.9, 0.99))
optimizer_D = torch.optim.Adam(discriminator.parameters(),
lr=opt.lr,
betas=(0.9, 0.99))
# ----------
# Training
# ----------
for epoch in range(opt.n_epochs):
for i, (imgs, imgs_ns, labels) in enumerate(dataloader):
# Configure input
real_imgs = Variable(labels.type(Tensor))
# Adversarial ground truths
valid = Variable(Tensor(imgs.size(0), 1).fill_(1.0),
requires_grad=False)
fake = Variable(Tensor(imgs.size(0), 1).fill_(0.0),
requires_grad=False)
# -----------------
# Train Generator
# -----------------
optimizer_G.zero_grad()
# Sample noise as generator input
#z = Variable(Tensor(np.random.normal(0, 1, (imgs.shape[0], opt.latent_dim)))) # [64, 256]
# Generate a batch of images
gen_imgs = generator(imgs.cuda()) + imgs_ns.cuda() #[64, 64, 64]
# Loss measures generator's ability to fool the discriminator
g_loss = adversarial_loss(discriminator(gen_imgs), valid)
g_loss.backward()
optimizer_G.step()
# ---------------------
# Train Discriminator
# ---------------------
optimizer_D.zero_grad()
# Measure discriminator's ability to classify real from generated samples
real_loss = adversarial_loss(discriminator(real_imgs), valid)
fake_loss = adversarial_loss(discriminator(gen_imgs.detach()),
fake)
d_loss = (real_loss + fake_loss) / 2
d_loss.backward()
optimizer_D.step()
print("[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f]" %
(epoch, opt.n_epochs, i, len(dataloader), d_loss.item(),
g_loss.item()))
batches_done = epoch * len(dataloader) + i
if batches_done % opt.sample_interval == 0:
save_image(gen_imgs.data[:25],
'images/%d.png' % batches_done,
nrow=5,
normalize=True)
weights_path = checkpoint_path.format(epoch=batches_done,
loss=g_loss)
print(
'saving generator weights file to {}'.format(weights_path))
torch.save(generator.state_dict(), weights_path)
optimizer_encorder = optim.RMSprop(params=NetE.parameters(),
lr=lr,
alpha=0.9,
eps=1e-8,
weight_decay=0,
momentum=0,
centered=False)
optimizer_decoder = optim.RMSprop(params=NetG.parameters(),
lr=lr,
alpha=0.9,
eps=1e-8,
weight_decay=0,
momentum=0,
centered=False)
optimizer_discriminator = optim.RMSprop(params=NetD.parameters(),
lr=lr,
alpha=0.9,
eps=1e-8,
weight_decay=0,
momentum=0,
centered=False)
data, _ = next(iter(dataloader))
fixed_batch = Variable(data).to(device)
vutils.save_image(fixed_batch,
'%s/real_samples.png' % opt.outf,
normalize=True)
margin = 0.6
equilibrium = 0.68
Dc = Discriminator(3+1).to(device)
if args.dummy_input: # debug purpose
BATCH_SIZE = 16
s1 = torch.randn(BATCH_SIZE, 3, args.resolution, args.resolution).to(device)
contour = torch.randn(BATCH_SIZE, 1, args.resolution, args.resolution).to(device)
fake = G(s1, contour)
print('fake.shape', fake.shape)
ds_out = Ds(torch.cat([fake, s1], dim=1))
dc_out = Dc(torch.cat([fake, contour], dim=1))
print('Ds_out.shape', ds_out.shape)
print('Dc_out.shape', dc_out.shape)
sys.exit(0)
optimG = optim.Adam(G.parameters(), lr=args.lrG, betas=(0, 0.999))
optimDc = optim.Adam(Dc.parameters(), lr=args.lrD, betas=(0, 0.999))
optimDs = optim.Adam(Ds.parameters(), lr=args.lrD, betas=(0, 0.999))
# prepare dataset
dataset = IconDataset(root=args.dataset, resolution=args.resolution, pad_ratio=args.pad_ratio)
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=True,
drop_last=True,
num_workers=4,
pin_memory=True,
)
# sample fixed inputs
for s1, s2, s3, contour in dataloader:
break
batch_size=args.batchsize,
num_workers=4 * args.ngpu,
worker_init_fn=worker_init_fn)
print("Training Datas:", len(dataset))
if not os.path.exists(args.weight_folder):
os.mkdir(args.weight_folder)
if not os.path.exists(args.result_folder):
os.mkdir(args.result_folder)
G = Generator(config)
D = Discriminator(config)
G_optim = optim.Adam(G.parameters(), lr=args.lr, betas=(0.5, 0.9))
D_optim = optim.Adam(D.parameters(), lr=args.lr, betas=(0.5, 0.9))
writer = SummaryWriter()
print("Running On:", device)
train_handler = Train_Handler(args.dataset_name,
args.epochs,
G_optim,
D_optim,
dataloader,
device=device,
writer=writer,
save_per_epoch=args.save_per_epoch,
print_per_iteration=args.print_per_iteration,
plot_per_iteration=args.plot_per_iteration,
weight_folder=args.weight_folder,
flatten_test_index]
else:
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = y[train_index], y[test_index]
torch_dataset = torch.utils.data.TensorDataset(torch.FloatTensor(X_train),
torch.LongTensor(y_train))
data_loader = torch.utils.data.DataLoader(torch_dataset,
batch_size=batch_size,
shuffle=True,
drop_last=True)
loss_classifier_list = []
for epoch in range(1, num_epochs + 1):
learning_rate = base_lr * math.pow(0.9, epoch / lr_step)
optimizer = torch.optim.Adam([
{
'params': discriminator.parameters()
},
],
lr=learning_rate,
weight_decay=l2_decay)
discriminator.train()
iter_data = iter(data_loader)
num_iter = len(data_loader)
total_clas_loss = 0
num_batches = 0
for it in range(0, num_iter):
data, label = iter_data.next()
if it % len(data_loader) == 0:
iter_data = iter(data_loader)
data = Variable(torch.FloatTensor(data))
shuffle=True,
num_workers=0,
pin_memory=True)
# G = torch.nn.DataParallel( Generator(zdim = config.G['zdim'], use_batchnorm = config.G['use_batchnorm'] , use_residual_block = config.G['use_residual_block'] , num_classes = config.G['num_classes'])).cuda()
# D = torch.nn.DataParallel( Discriminator(use_batchnorm = config.D['use_batchnorm'])).cuda()
G = Generator(zdim=config.G['zdim'],
use_batchnorm=config.G['use_batchnorm'],
use_residual_block=config.G['use_residual_block'],
num_classes=config.G['num_classes']).cuda()
D = Discriminator(use_batchnorm=config.D['use_batchnorm']).cuda()
optimizer_G = torch.optim.Adam(filter(lambda p: p.requires_grad,
G.parameters()),
lr=1e-4)
optimizer_D = torch.optim.Adam(filter(lambda p: p.requires_grad,
D.parameters()),
lr=1e-5)
last_epoch = -1
if config.train['resume_model'] is not None:
e1 = resume_model(G, config.train['resume_model'])
e2 = resume_model(D, config.train['resume_model'])
assert e1 == e2
last_epoch = e1
if config.train['resume_optimizer'] is not None:
e3 = resume_optimizer(optimizer_G, G, config.train['resume_optimizer'])
e4 = resume_optimizer(optimizer_D, D, config.train['resume_optimizer'])
assert e1 == e2 and e2 == e3 and e3 == e4
last_epoch = e1
dataloader = torch.utils.data.DataLoader(
datasets.MNIST('./mnist', train=True, download=True, transform=transform),
batch_size=batch_size,
shuffle=True,
drop_last=True
)
G = Generator().to(device)
D = Discriminator().to(device)
criterion = nn.BCELoss()
optimizerG = torch.optim.Adam(
G.parameters(), lr=learning_rate, betas=(0.5, 0.999))
optimizerD = torch.optim.SGD(
D.parameters(), lr=1e-3, weight_decay=1e-7, momentum=0)
label = torch.FloatTensor(batch_size)
real_label, fake_label = 1, 0
for epoch in range(train_epoch):
for i, (imgs, _) in enumerate(dataloader):
# fix G, train D
optimizerD.zero_grad()
# load imgs to cuda
imgs = imgs.to(device)
# Let D output to 1
output = D(imgs)
label.data.fill_(real_label)
label = label.to(device)
output = torch.max(output, dim=1).values
G_B2A.load_state_dict(torch.load('pretrained/G_B2A.pth'))
D_A.load_state_dict(torch.load('pretrained/D_A.pth'))
D_B.load_state_dict(torch.load('pretrained/D_B.pth'))
else:
G_A2B.apply(weights_init)
G_B2A.apply(weights_init)
D_A.apply(weights_init)
D_B.apply(weights_init)
criterion_GAN = torch.nn.MSELoss()
criterion_cycle = torch.nn.L1Loss()
criterion_identity = torch.nn.L1Loss()
G_params = list(G_A2B.parameters()) + list(G_B2A.parameters())
D_params = list(D_A.parameters()) + list(D_B.parameters())
optimizer_G = optim.Adam(G_params, lr=0.0002, betas=[0.5, 0.999])
optimizer_D = optim.Adam(D_params, lr=0.0002, betas=[0.5, 0.999])
fake_A_buffer = utils.ReplayBuffer()
fake_B_buffer = utils.ReplayBuffer()
logger = utils.Logger(opt.n_epochs, len(loader_A), vis)
step = 0
losses = {'loss_G': 0, 'loss_G_identity': 0, 'loss_G_GAN': 0, 'loss_G_cycle': 0, 'loss_D': 0}
for epoch in range(1, opt.n_epochs+1):
for (A, B) in zip(enumerate(loader_A), enumerate(loader_B)):
step = step + 1
discriminator.cuda()
def weights_init_normal(m):
classname = m.__class__.__name__
if classname.find("Conv") != -1:
torch.nn.init.normal_(m.weight.data, 0.0, 0.02)
elif classname.find("BatchNorm2d") != -1:
torch.nn.init.normal_(m.weight.data, 1.0, 0.02)
torch.nn.init.constant_(m.bias.data, 0.0)
generator.apply(weights_init_normal)
discriminator.apply(weights_init_normal)
criterion = nn.BCELoss()
criterion.cuda()
optimizer_disc = optim.Adam(discriminator.parameters(),lr=learning_rate, betas=(0.5, 0.999))
optimizer_gen = optim.Adam(generator.parameters(), lr=learning_rate, betas=(0.5, 0.999))
Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
print("Training...")
#TRAINING
for epoch in range(n_epoch):
for i, imgs in enumerate(training_data_loader):
real = Variable(Tensor(imgs.shape[0], 1).fill_(0.9), requires_grad=False)
fake = Variable(Tensor(imgs.shape[0], 1).fill_(0.1), requires_grad=False)
real_imgs = Variable(imgs.type(Tensor))
z_vec = torch.randn(imgs.shape[0], 100, 1, 1, device='cuda')
optimizer_gen.zero_grad()
generator_images = generator(z_vec)
class MUNIT(nn.Module):
def __init__(self, opt):
super(MUNIT, self).__init__()
# generators and discriminators
self.gen_a = Generator(opt.ngf, opt.style_dim, opt.mlp_dim)
self.gen_b = Generator(opt.ngf, opt.style_dim, opt.mlp_dim)
self.dis_a = Discriminator(opt.ndf)
self.dis_b = Discriminator(opt.ndf)
#random style code
self.s_a = torch.randn(opt.display_size,
opt.style_dim,
1,
1,
requires_grad=True).cuda()
self.s_b = torch.randn(opt.display_size,
opt.style_dim,
1,
1,
requires_grad=True).cuda()
#optimizers
dis_params = list(self.dis_a.parameters()) + list(
self.dis_b.parameters())
gen_params = list(self.gen_a.parameters()) + list(
self.gen_b.parameters())
self.dis_opt = torch.optim.Adam(dis_params,
lr=opt.lr,
beta=opt.beta1,
weight_delay=opt.weight_delay)
self.gen_opt = torch.optim.Adam(gen_params,
lr=opt.lr,
beta=opt.beta1,
weight_delay=opt.weight_delay)
# nerwork weight initialization
self.apply(weight_init('kaiming'))
self.dis_a.apply(weight_init('gaussian'))
self.dis_b.apply(weight_init('gaussian'))
def forward(self, x_a, x_b):
c_a, s_a_prime = self.gen_a.encode(x_a)
c_b, s_b_prime = self.gen_b.encode(x_b)
x_a2b = self.gen_b.decode(c_a, self.s_b)
x_b2a = self.gen_a.decode(c_b, self.s_a)
return x_a2b, x_b2a
def backward_G(self, x_a, x_b):
#encoding and decoding
s_a = torch.randn(x_a.size(0), opt.style_dim, 1, 1,
requires_grad=True).cuda()
s_b = torch.randn(x_b.size(0), opt.style_dim, 1, 1,
requires_grad=True).cuda()
c_a, s_a_prime = self.gen_a.encode(x_a)
c_b, s_b_prime = self.gen_b.encode(x_b)
x_a_rec = self.gen_a.decode(c_a, s_a_prime)
x_b_rec = self.gen_b.decode(c_b, s_b_prime)
x_a2b = self.gen_b.decode(c_a, s_b)
x_b2a = self.gen_a.decode(c_b, s_a)
c_a_rec, s_b_rec = self.gen_b.encode(x_a2b)
c_b_rec, s_a_rec = self.gen_a.encode(x_b2a)
x_a_fake = self.dis_a(x_b2a)
x_b_fake = self.dis_b(x_a2b)
#loss function
self.loss_xa = recon_loss(x_a_rec - x_a)
self.loss_xb = recon_loss(x_b_rec - x_b)
self.loss_ca = recon_loss(c_a_rec - c_a)
self.loss_cb = recon_loss(c_b_rec - c_b)
self.loss_sa = recon_loss(s_a_rec - s_a)
self.loss_sb = recon_loss(s_b_rec - s_b)
if opt.x_cyc_w > 0:
x_a2b2a = self.gen_b.decode(c_a_rec, s_a_prime)
x_b2a2b = self.gen_a.decode(c_b_rec, s_b_prime)
self.loss_cyc_a2b = recon_loss(x_a2b2a, x_a)
self.loss_cyc_b2a = recon_loss(x_b2a2b, x_b)
else:
self.loss_cyc_a2b = 0
self.loss_cyc_b2a = 0
self.loss_gen_a = gen_loss(x_a_fake, opt.gan_type)
self.loss_gen_b = gen_loss(x_b_fake, opt.gan_type)
self.gen_total_loss = opt.gan_w*(self.loss_gen_a+self.loss_gen_b)+\
opt.x_w*(self.loss_xa+self.loss_xb)+\
opt.c_w*(self.loss_ca+self.loss_cb)+\
opt.s_w*(self.loss_sa+self.loss_sb)+\
opt.x_cyc_w*(self.loss_cyc_a2b+self.loss_cyc_b2a)
self.gen_total_loss.backward()
return self.gen_total_loss
def barkward_D(self, x_a, x_b):
s_a = torch.randn(x_a.size(0), opt.style_dim, 1, 1,
requires_grad=True).cuda()
s_b = torch.randn(x_b.size(0), opt.style_dim, 1, 1,
requires_grad=True).cuda()
c_a, _ = self.gen_a.encode(x_a.detach())
c_b, _ = self.gen_b.encode(x_b.detach())
x_a2b = self.gen_b.decode(c_a, s_b)
x_b2a = self.gen_a.decode(c_b, s_a)
x_a_real = self.dis_a(x_a.detach())
x_a_fake = self.dis_a(x_b2a)
x_b_real = self.dis_b(x_b.detach())
x_b_fake = self.dis_b(x_a2b)
self.loss_dis_a = dis_loss(x_a.detach(), x_b2a, opt.gan_type)
self.loss_dis_b = dis_loss(x_b.detach(), x_a2b, opt.gan_type)
self.dis_total_loss = opt.gan_w * (self.loss_dis_a + self.loss.dis_b)
self.dis_total_loss.backward()
return self.dis_total_loss
def optimize_parameters(self, x_a, x_b):
self.gen_opt.zero_grad()
self.dis_opt.zero_grad()
self.backward_G(x_a, x_b)
self.barkward_D(x_a, x_b)
self.gen_opt.step()
self.dis_opt.step()
def sample(self, x_a, x_b, num_style):
x_a2b = []
x_b2a = []
for i in range(x_a.size(0)):
s_a = torch.randn(num_style, opt.style_dim, 1, 1).cuda()
s_b = torch.randn(num_style, opt.style_dim, 1, 1).cuda()
c_a_i, _ = self.gen_a.encode(x_a[i].unsqueeze(0))
c_b_i, _ = self.gen_b.encode(x_b[i].unsqueeze(0))
x_i_a2b = []
x_i_b2a = []
for j in range(num_style):
#[1,opt.style_dim,1,1]
s_a_j = s_a[j].unsqueeze(0)
s_b_j = s_b[j].unsqueeze(0)
x_i_a2b.append(self.gen_b.decode(c_a_i, s_b_j))
x_i_b2a.append(self.gen_a.decode(c_b_i, s_a_j))
#[num_style,c,h,w]
x_i_a2b = torch.cat(x_i_a2b, dim=0)
x_i_b2a = torch.cat(x_i_b2a, dim=0)
x_a2b.append(x_i_a2b)
x_b2a.append(x_i_b2a)
#[batch_size,num_style,c,h,w]
x_a2b = torch.stack(x_a2b)
x_b2a = torch.stack(x_b2a)
#[batch_size*num_style,c,h,w]
x_a2b = x_a2b.view(-1, x_a2b.size()[2:])
x_b2a = x_b2a.view(-1, x_b2a.size()[2:])
return x_a2b, x_b2a
def test(self, input, direction, style):
output = []
if direction == 'a2b':
encoder = self.gen_a.encode()
decoder = self.gen_b.decode()
else:
encoder = self.gen_b.encode()
decoder = self.gen_a.decode()
content, _ = encode(input.unsqueeze(0))
for i in range(style.size()):
output.append(decoder(content, style[i].unsqueeze(0)))
output = torch.cat(output, dim=0)
return output
def save(self, save_dir, iterations):
save_gen = os.path.join(save_dir, 'gen_%8d.pt' % (iteration + 1))
save_dis = os.path.join(save_dir, 'dis_%8d.pt' % (iteration + 1))
save_opt = os.path.join(save_dir, 'optimizer.pt')
torch.save({
'a': self.gen_a.state_dict(),
'b': self.gen_b.state_dict()
}, save_gen)
torch.save({
'a': self.dis_a.state_dict(),
'b': self.dis_b.state_dict()
}, save_dis)
torch.save(
{
'gen': self.gen_opt.state_dict(),
'dis': self.dis_opt.state_dict()
}, save_opt)
def train(args):
# set the logger
logger = Logger('./logs')
# GPU enabling
if (args.gpu != None):
use_cuda = True
dtype = torch.cuda.FloatTensor
torch.cuda.set_device(args.gpu)
print("Current device: %s" % torch.cuda.get_device_name(args.gpu))
# define networks
g_AtoB = Generator().type(dtype)
g_BtoA = Generator().type(dtype)
d_A = Discriminator().type(dtype)
d_B = Discriminator().type(dtype)
# optimizers
optimizer_generators = Adam(
list(g_AtoB.parameters()) + list(g_BtoA.parameters()), INITIAL_LR)
optimizer_d_A = Adam(d_A.parameters(), INITIAL_LR)
optimizer_d_B = Adam(d_B.parameters(), INITIAL_LR)
# loss criterion
criterion_mse = torch.nn.MSELoss()
criterion_l1 = torch.nn.L1Loss()
# get training data
dataset_transform = transforms.Compose([
transforms.Resize(int(IMAGE_SIZE * 1),
Image.BICUBIC), # scale shortest side to image_size
transforms.RandomCrop(
(IMAGE_SIZE, IMAGE_SIZE)), # random center image_size out
transforms.ToTensor(), # turn image from [0-255] to [0-1]
transforms.Normalize(mean=(0.5, 0.5, 0.5),
std=(0.5, 0.5, 0.5)) # normalize
])
dataloader = DataLoader(ImgPairDataset(args.dataroot, dataset_transform,
'train'),
batch_size=BATCH_SIZE,
shuffle=True)
# get some test data to display periodically
test_data_A = torch.tensor([]).type(dtype)
test_data_B = torch.tensor([]).type(dtype)
for i in range(NUM_TEST_SAMPLES):
imgA = ImgPairDataset(args.dataroot, dataset_transform,
'test')[i]['A'].type(dtype).unsqueeze(0)
imgB = ImgPairDataset(args.dataroot, dataset_transform,
'test')[i]['B'].type(dtype).unsqueeze(0)
test_data_A = torch.cat((test_data_A, imgA), dim=0)
test_data_B = torch.cat((test_data_B, imgB), dim=0)
fileStrA = 'visualization/test_%d/%s/' % (i, 'B_inStyleofA')
fileStrB = 'visualization/test_%d/%s/' % (i, 'A_inStyleofB')
if not os.path.exists(fileStrA):
os.makedirs(fileStrA)
if not os.path.exists(fileStrB):
os.makedirs(fileStrB)
fileStrA = 'visualization/test_original_%s_%04d.png' % ('A', i)
fileStrB = 'visualization/test_original_%s_%04d.png' % ('B', i)
utils.save_image(
fileStrA,
ImgPairDataset(args.dataroot, dataset_transform,
'test')[i]['A'].data)
utils.save_image(
fileStrB,
ImgPairDataset(args.dataroot, dataset_transform,
'test')[i]['B'].data)
# replay buffers
replayBufferA = utils.ReplayBuffer(50)
replayBufferB = utils.ReplayBuffer(50)
# training loop
step = 0
for e in range(EPOCHS):
startTime = time.time()
for idx, batch in enumerate(dataloader):
real_A = batch['A'].type(dtype)
real_B = batch['B'].type(dtype)
# some examples seem to have only 1 color channel instead of 3
if (real_A.shape[1] != 3):
continue
if (real_B.shape[1] != 3):
continue
# -----------------
# train generators
# -----------------
optimizer_generators.zero_grad()
utils.learning_rate_decay(INITIAL_LR, e, EPOCHS,
optimizer_generators)
# GAN loss
fake_A = g_BtoA(real_B)
disc_fake_A = d_A(fake_A)
fake_B = g_AtoB(real_A)
disc_fake_B = d_B(fake_B)
replayBufferA.push(torch.tensor(fake_A.data))
replayBufferB.push(torch.tensor(fake_B.data))
target_real = Variable(torch.ones_like(disc_fake_A)).type(dtype)
target_fake = Variable(torch.zeros_like(disc_fake_A)).type(dtype)
loss_gan_AtoB = criterion_mse(disc_fake_B, target_real)
loss_gan_BtoA = criterion_mse(disc_fake_A, target_real)
loss_gan = loss_gan_AtoB + loss_gan_BtoA
# cyclic reconstruction loss
cyclic_A = g_BtoA(fake_B)
cyclic_B = g_AtoB(fake_A)
loss_cyclic_AtoBtoA = criterion_l1(cyclic_A,
real_A) * CYCLIC_WEIGHT
loss_cyclic_BtoAtoB = criterion_l1(cyclic_B,
real_B) * CYCLIC_WEIGHT
loss_cyclic = loss_cyclic_AtoBtoA + loss_cyclic_BtoAtoB
# identity loss
loss_identity = 0
loss_identity_A = 0
loss_identity_B = 0
if (args.use_identity == True):
identity_A = g_BtoA(real_A)
identity_B = g_AtoB(real_B)
loss_identity_A = criterion_l1(identity_A,
real_A) * 0.5 * CYCLIC_WEIGHT
loss_identity_B = criterion_l1(identity_B,
real_B) * 0.5 * CYCLIC_WEIGHT
loss_identity = loss_identity_A + loss_identity_B
loss_generators = loss_gan + loss_cyclic + loss_identity
loss_generators.backward()
optimizer_generators.step()
# -----------------
# train discriminators
# -----------------
optimizer_d_A.zero_grad()
utils.learning_rate_decay(INITIAL_LR, e, EPOCHS, optimizer_d_A)
fake_A = replayBufferA.sample(1).detach()
disc_fake_A = d_A(fake_A)
disc_real_A = d_A(real_A)
loss_d_A = 0.5 * (criterion_mse(disc_real_A, target_real) +
criterion_mse(disc_fake_A, target_fake))
loss_d_A.backward()
optimizer_d_A.step()
optimizer_d_B.zero_grad()
utils.learning_rate_decay(INITIAL_LR, e, EPOCHS, optimizer_d_B)
fake_B = replayBufferB.sample(1).detach()
disc_fake_B = d_B(fake_B)
disc_real_B = d_B(real_B)
loss_d_B = 0.5 * (criterion_mse(disc_real_B, target_real) +
criterion_mse(disc_fake_B, target_fake))
loss_d_B.backward()
optimizer_d_B.step()
#log info and save sample images
if ((idx % 250) == 0):
# eval on some sample images
g_AtoB.eval()
g_BtoA.eval()
test_B_hat = g_AtoB(test_data_A).cpu()
test_A_hat = g_BtoA(test_data_B).cpu()
fileBaseStr = 'test_%d_%d' % (e, idx)
for i in range(NUM_TEST_SAMPLES):
fileStrA = 'visualization/test_%d/%s/%03d_%04d.png' % (
i, 'B_inStyleofA', e, idx)
fileStrB = 'visualization/test_%d/%s/%03d_%04d.png' % (
i, 'A_inStyleofB', e, idx)
utils.save_image(fileStrA, test_A_hat[i].data)
utils.save_image(fileStrB, test_B_hat[i].data)
g_AtoB.train()
g_BtoA.train()
endTime = time.time()
timeForIntervalIterations = endTime - startTime
startTime = endTime
print(
'Epoch [{:3d}/{:3d}], Training [{:4d}/{:4d}], Time Spent (s): [{:4.4f}], Losses: [G_GAN: {:4.4f}][G_CYC: {:4.4f}][G_IDT: {:4.4f}][D_A: {:4.4f}][D_B: {:4.4f}]'
.format(e, EPOCHS, idx, len(dataloader),
timeForIntervalIterations, loss_gan, loss_cyclic,
loss_identity, loss_d_A, loss_d_B))
# tensorboard logging
info = {
'loss_generators':
loss_generators.item(),
'loss_gan_AtoB':
loss_gan_AtoB.item(),
'loss_gan_BtoA':
loss_gan_BtoA.item(),
'loss_cyclic_AtoBtoA':
loss_cyclic_AtoBtoA.item(),
'loss_cyclic_BtoAtoB':
loss_cyclic_BtoAtoB.item(),
'loss_cyclic':
loss_cyclic.item(),
'loss_d_A':
loss_d_A.item(),
'loss_d_B':
loss_d_B.item(),
'lr_optimizer_generators':
optimizer_generators.param_groups[0]['lr'],
'lr_optimizer_d_A':
optimizer_d_A.param_groups[0]['lr'],
'lr_optimizer_d_B':
optimizer_d_B.param_groups[0]['lr'],
}
if (args.use_identity):
info['loss_identity_A'] = loss_identity_A.item()
info['loss_identity_B'] = loss_identity_B.item()
for tag, value in info.items():
logger.scalar_summary(tag, value, step)
info = {
'test_A_hat':
test_A_hat.data.numpy().transpose(0, 2, 3, 1),
'test_B_hat':
test_B_hat.data.numpy().transpose(0, 2, 3, 1),
}
for tag, images in info.items():
logger.image_summary(tag, images, step)
step += 1
# save after every epoch
g_AtoB.eval()
g_BtoA.eval()
d_A.eval()
d_B.eval()
if use_cuda:
g_AtoB.cpu()
g_BtoA.cpu()
d_A.cpu()
d_B.cpu()
if not os.path.exists("models"):
os.makedirs("models")
filename_gAtoB = "models/" + str('g_AtoB') + "_epoch_" + str(
e) + ".model"
filename_gBtoA = "models/" + str('g_BtoA') + "_epoch_" + str(
e) + ".model"
filename_dA = "models/" + str('d_A') + "_epoch_" + str(e) + ".model"
filename_dB = "models/" + str('d_B') + "_epoch_" + str(e) + ".model"
torch.save(g_AtoB.state_dict(), filename_gAtoB)
torch.save(g_BtoA.state_dict(), filename_gBtoA)
torch.save(d_A.state_dict(), filename_dA)
torch.save(d_B.state_dict(), filename_dB)
if use_cuda:
g_AtoB.cuda()
g_BtoA.cuda()
d_A.cuda()
d_B.cuda()
def main():
os.environ["CUDA_VISIBLE_DEVICES"] = opt.gpus
start_epoch = 0
train_image_dataset = image_preprocessing(opt.dataset, 'train')
data_loader = DataLoader(train_image_dataset, batch_size=opt.batch_size,
shuffle=True, num_workers=opt.num_workers)
criterion = least_squares
euclidean_l1 = nn.L1Loss()
G = Generator(ResidualBlock, layer_count=9)
F = Generator(ResidualBlock, layer_count=9)
Dx = Discriminator()
Dy = Discriminator()
G_optimizer = optim.Adam(G.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2))
F_optimizer = optim.Adam(F.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2))
Dx_optimizer = optim.Adam(Dx.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2))
Dy_optimizer = optim.Adam(Dy.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2))
if torch.cuda.is_available():
G = nn.DataParallel(G)
F = nn.DataParallel(F)
Dx = nn.DataParallel(Dx)
Dy = nn.DataParallel(Dy)
G = G.cuda()
F = F.cuda()
Dx = Dx.cuda()
Dy = Dy.cuda()
if opt.checkpoint is not None:
G, F, Dx, Dy, G_optimizer, F_optimizer, Dx_optimizer, Dy_optimizer, start_epoch = load_ckp(opt.checkpoint, G, F, Dx, Dy, G_optimizer, F_optimizer, Dx_optimizer, Dy_optimizer)
print('[Start] : Cycle GAN Training')
logger = Logger(opt.epochs, len(data_loader), image_step=10)
for epoch in range(opt.epochs):
epoch = epoch + start_epoch + 1
print("Epoch[{epoch}] : Start".format(epoch=epoch))
for step, data in enumerate(data_loader):
real_A = to_variable(data['A'])
real_B = to_variable(data['B'])
fake_B = G(real_A)
fake_A = F(real_B)
# Train Dx
Dx_optimizer.zero_grad()
Dx_real = Dx(real_A)
Dx_fake = Dx(fake_A)
Dx_loss = patch_loss(criterion, Dx_real, True) + patch_loss(criterion, Dx_fake, 0)
Dx_loss.backward(retain_graph=True)
Dx_optimizer.step()
# Train Dy
Dy_optimizer.zero_grad()
Dy_real = Dy(real_B)
Dy_fake = Dy(fake_B)
Dy_loss = patch_loss(criterion, Dy_real, True) + patch_loss(criterion, Dy_fake, 0)
Dy_loss.backward(retain_graph=True)
Dy_optimizer.step()
# Train G
G_optimizer.zero_grad()
Dy_fake = Dy(fake_B)
G_loss = patch_loss(criterion, Dy_fake, True)
# Train F
F_optimizer.zero_grad()
Dx_fake = Dx(fake_A)
F_loss = patch_loss(criterion, Dx_fake, True)
# identity loss
loss_identity = euclidean_l1(real_A, fake_A) + euclidean_l1(real_B, fake_B)
# cycle consistency
loss_cycle = euclidean_l1(F(fake_B), real_A) + euclidean_l1(G(fake_A), real_B)
# Optimize G & F
loss = G_loss + F_loss + opt.lamda * loss_cycle + opt.lamda * loss_identity * (0.5)
loss.backward()
G_optimizer.step()
F_optimizer.step()
if (step + 1 ) % opt.save_step == 0:
print("Epoch[{epoch}]| Step [{now}/{total}]| Dx Loss: {Dx_loss}, Dy_Loss: {Dy_loss}, G_Loss: {G_loss}, F_Loss: {F_loss}".format(
epoch=epoch, now=step + 1, total=len(data_loader), Dx_loss=Dx_loss.item(), Dy_loss=Dy_loss,
G_loss=G_loss.item(), F_loss=F_loss.item()))
batch_image = torch.cat((torch.cat((real_A, real_B), 3), torch.cat((fake_A, fake_B), 3)), 2)
torchvision.utils.save_image(denorm(batch_image[0]), opt.training_result + 'result_{result_name}_ep{epoch}_{step}.jpg'.format(result_name=opt.result_name,epoch=epoch, step=(step + 1) * opt.batch_size))
# http://localhost:8097
logger.log(
losses={
'loss_G': G_loss,
'loss_F': F_loss,
'loss_identity': loss_identity,
'loss_cycle': loss_cycle,
'total_G_loss': loss,
'loss_Dx': Dx_loss,
'loss_Dy': Dy_loss,
'total_D_loss': (Dx_loss + Dy_loss),
},
images={
'real_A': real_A,
'real_B': real_B,
'fake_A': fake_A,
'fake_ B': fake_B,
},
)
torch.save({
'epoch': epoch,
'G_model': G.state_dict(),
'G_optimizer': G_optimizer.state_dict(),
'F_model': F.state_dict(),
'F_optimizer': F_optimizer.state_dict(),
'Dx_model': Dx.state_dict(),
'Dx_optimizer': Dx_optimizer.state_dict(),
'Dy_model': Dy.state_dict(),
'Dy_optimizer': Dy_optimizer.state_dict(),
}, opt.save_model + 'model_{result_name}_CycleGAN_ep{epoch}.ckp'.format(result_name=opt.result_name, epoch=epoch))
def train_network(args):
device = torch.device('cuda' if args.gpu_no >= 0 else 'cpu')
generator = Generator(unet_flag=args.unet_flag).to(device).train()
discriminator = Discriminator().to(device).train()
optimizer_g = torch.optim.Adam(generator.parameters(),
lr=args.lr,
betas=(args.beta1, args.beta2))
optimizer_d = torch.optim.Adam(discriminator.parameters(),
lr=args.lr,
betas=(args.beta1, args.beta2))
bce_criterion = nn.BCELoss()
l1_criterion = nn.L1Loss()
dataset = FacadeFolder(args.data_A, args.data_B, args.imsize,
args.cropsize, args.cencrop)
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers)
loss_seq = {'d_real': [], 'd_fake': [], 'g_gan': [], 'g_l1': []}
for epoch in range(args.epoch):
for iteration, (item_A, item_B) in enumerate(dataloader, 1):
item_A, item_B = item_A.to(device), item_B.to(device)
"""
Train Discriminator
"""
fake_B = generator(item_A)
fake_discrimination = discriminator(fake_B.detach(), item_A)
fake_loss = bce_criterion(
fake_discrimination,
torch.zeros_like(fake_discrimination).to(device))
real_discrimination = discriminator(item_B, item_A)
real_loss = bce_criterion(
real_discrimination,
torch.ones_like(real_discrimination).to(device))
discriminator_loss = (real_loss +
fake_loss) * args.discriminator_weight
loss_seq['d_real'].append(real_loss.item())
loss_seq['d_fake'].append(fake_loss.item())
optimizer_d.zero_grad()
discriminator_loss.backward()
optimizer_d.step()
"""
Train Generator
"""
fake_B = generator(item_A)
fake_discrimination = discriminator(fake_B, item_A)
fake_loss = bce_criterion(
fake_discrimination,
torch.ones_like(fake_discrimination).to(device))
l1_loss = l1_criterion(fake_B, item_B)
generator_loss = fake_loss + l1_loss * args.l1_weight
loss_seq['g_gan'].append(fake_loss.item())
loss_seq['g_l1'].append(l1_loss.item())
optimizer_g.zero_grad()
generator_loss.backward()
optimizer_g.step()
"""
Check training loss
"""
if iteration % args.check_iter == 0:
check_str = "%s: epoch:[%d/%d]\titeration:[%d/%d]" % (
time.ctime(), epoch, args.epoch, iteration,
len(dataloader))
for key, value in loss_seq.items():
check_str += "\t%s: %2.2f" % (
key, lastest_arverage_value(value))
print(check_str)
imsave(torch.cat([item_A, item_B, fake_B], dim=0),
args.save_path + 'training_image.jpg')
# save check point
torch.save(
{
'iteration': iteration,
'g_state_dict': generator.state_dict(),
'd_state_dict': discriminator.state_dict(),
'loss_seq': loss_seq
}, args.save_path + 'check_point.pth')
return discriminator, generator
# Initialize BCELoss function
criterion = nn.BCELoss()
# Create batch of latent vectors that we will use to visualize
# the progression of the generator
if args.truncnorm:
fixed_noise = torch.from_numpy(truncated_z_sample(64, args.nz, args.tc_th, args.manualSeed)).float().to(device)
else:
fixed_noise = torch.randn(64, args.nz, 1, 1, device=device)
# Establish convention for real and fake labels during training
real_label = 1
fake_label = 0
# Setup Adam optimizers for both G and D
optimizerD = optim.Adam(netD.parameters(), lr=args.lr, betas=(args.beta1, 0.999))
optimizerG = optim.Adam(netG.parameters(), lr=args.lr, betas=(args.beta1, 0.999))
# Training Loop
# Lists to keep track of progress
img_list = []
G_losses = []
D_losses = []
iters = 0
writer = SummaryWriter()
print("Starting Training Loop...")
# For each epoch
for epoch in range(args.num_epochs):
gamma, batch_size, device, epochs, load_checkpoint, mutil_gpu,
device_ids))
if __name__ == "__main__":
img_size = (32, 32, 3)
latent_dim = 128
writer = SummaryWriter()
discriminator = Discriminator(img_size=img_size)
generator = Generator(latent_dim=latent_dim, output_size=img_size)
optimizer_G = getattr(optim, optimizer_name)(generator.parameters(),
lr=lr,
betas=(0.5, 0.999))
optimizer_D = getattr(optim, optimizer_name)(discriminator.parameters(),
lr=lr,
betas=(0.5, 0.999))
# lr_scheduler = lr_scheduler.StepLR(optimizer, step_size=step_size, gamma=gamma)
# if load_checkpoint:
# load_model("model_checkpoint/check_point.pkl", model, optimizer, lr_scheduler)
if mutil_gpu:
discriminator = nn.DataParallel(discriminator, device_ids, device)
generator = nn.DataParallel(generator, device_ids, device)
discriminator = discriminator.to(device=device)
generator = generator.to(device=device)
mnist = torchvision.datasets.CIFAR10(root="cifar10",
train=True,
download=True,
class GAN:
def __init__(self):
self.generator = Generator()
self.generator.to(Params.Device)
self.discriminator = Discriminator()
self.discriminator.to(Params.Device)
self.loss_fn = nn.BCELoss()
self.optimizer_g = torch.optim.Adam(self.generator.parameters(), Params.LearningRateG, betas=(Params.Beta, 0.999))
self.optimizer_d = torch.optim.Adam(self.discriminator.parameters(), Params.LearningRateD, betas=(Params.Beta, 0.999))
self.exemplar_latent_vectors = torch.randn( (64, Params.LatentVectorSize), device=Params.Device )
def load_image_set(self):
transforms = torchvision.transforms.Compose( [
torchvision.transforms.Resize(Params.ImageSize),
torchvision.transforms.CenterCrop(Params.ImageSize),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize( (0.5,0.5,0.5), (0.5,0.5,0.5) )
] )
self.dataset = torchvision.datasets.ImageFolder(
Params.ImagePath,
transforms
)
self.loader = torch.utils.data.DataLoader(
self.dataset,
batch_size = Params.BatchSize,
shuffle=True,
num_workers=Params.NumWorkers
)
def train(self, start_epoch=0):
self.generator.train()
self.discriminator.train()
criterion = nn.BCELoss()
self.loss_record = {
'D': [],
'G': []
}
for cur_epoch in range(start_epoch, Params.NumEpochs):
tic = time.perf_counter()
self.train_epoch(cur_epoch)
toc = time.perf_counter()
print( f"Last epoch took: {toc - tic:.0f} seconds" )
if cur_epoch % Params.CheckpointEvery == 0:
self.save_checkpoint(f"epoch{cur_epoch+1:03d}")
self.save_checkpoint(f"Final ({Params.NumEpochs} epochs)")
def train_epoch(self, cur_epoch):
for (i, real_images) in enumerate(self.loader):
# Discriminator training step
self.discriminator.zero_grad()
real_images = real_images[0].to(Params.Device)
batch_size = real_images.size(0)
labels = torch.full(
(batch_size,),
1,
dtype=torch.float, device=Params.Device
)
D_real = self.discriminator(real_images).view(-1)
D_loss_real = self.loss_fn(D_real, labels)
D_loss_real.backward()
latent_vectors = torch.randn( (batch_size, Params.LatentVectorSize), device=Params.Device )
fake_images = self.generator(latent_vectors)
D_fake = self.discriminator(fake_images.detach()).view(-1)
labels.fill_(0)
D_loss_fake = self.loss_fn(D_fake, labels)
D_loss_fake.backward()
self.optimizer_d.step()
# Generator training step
self.generator.zero_grad()
labels.fill_(1)
D_fake = self.discriminator(fake_images).view(-1)
G_loss = self.loss_fn(D_fake, labels)
G_loss.backward()
self.optimizer_g.step()
D_loss = D_loss_real.item() + D_loss_fake.item()
G_loss = G_loss.item()
if (i % Params.ReportEvery == 0) and i>0:
print( f"Epoch[{cur_epoch}/{Params.NumEpochs}] Batch[{i}/{len(self.loader)}]" )
print( f"\tD Loss: {D_loss}\tG Loss: {G_loss}\n" )
self.loss_record['D'].append(D_loss)
self.loss_record['G'].append(G_loss)
def save_checkpoint(self, checkpoint_name):
dir_path = os.path.join(Params.CheckpointPath, checkpoint_name)
os.makedirs(dir_path, exist_ok=True)
print( f"Saving snapshot to: {dir_path}" )
save_state = dict()
save_state['g_state'] = self.generator.state_dict()
save_state['g_optimizer_state'] = self.optimizer_g.state_dict()
save_state['d_state'] = self.discriminator.state_dict()
save_state['d_optimizer_state'] = self.optimizer_d.state_dict()
torch.save(
save_state,
os.path.join(dir_path, 'model_params.pt')
)
with torch.no_grad():
self.generator.eval()
gen_images = self.generator(self.exemplar_latent_vectors)
self.generator.train()
torchvision.utils.save_image(
gen_images,
os.path.join(dir_path, 'gen_images.png'),
nrow = 8,
normalize=True
)
class tag2pix(object):
def __init__(self, args):
if args.model == 'tag2pix':
from network import Generator
elif args.model == 'senet':
from model.GD_senet import Generator
elif args.model == 'resnext':
from model.GD_resnext import Generator
elif args.model == 'catconv':
from model.GD_cat_conv import Generator
elif args.model == 'catall':
from model.GD_cat_all import Generator
elif args.model == 'adain':
from model.GD_adain import Generator
elif args.model == 'seadain':
from model.GD_seadain import Generator
else:
raise Exception('invalid model name: {}'.format(args.model))
self.args = args
self.epoch = args.epoch
self.batch_size = args.batch_size
self.gpu_mode = not args.cpu
self.input_size = args.input_size
self.color_revert = ColorSpace2RGB(args.color_space)
self.layers = args.layers
[self.cit_weight, self.cvt_weight] = args.cit_cvt_weight
self.load_dump = (args.load is not "")
self.load_path = Path(args.load)
self.l1_lambda = args.l1_lambda
self.guide_beta = args.guide_beta
self.adv_lambda = args.adv_lambda
self.save_freq = args.save_freq
self.two_step_epoch = args.two_step_epoch
self.brightness_epoch = args.brightness_epoch
self.save_all_epoch = args.save_all_epoch
self.iv_dict, self.cv_dict, self.id_to_name = get_tag_dict(
args.tag_dump)
cvt_class_num = len(self.cv_dict.keys())
cit_class_num = len(self.iv_dict.keys())
self.class_num = cvt_class_num + cit_class_num
self.start_epoch = 1
#### load dataset
if not args.test:
self.train_data_loader, self.test_data_loader = get_dataset(args)
self.result_path = Path(args.result_dir) / time.strftime(
'%y%m%d-%H%M%S', time.localtime())
if not self.result_path.exists():
self.result_path.mkdir()
self.test_images = self.get_test_data(self.test_data_loader,
args.test_image_count)
else:
self.test_data_loader = get_dataset(args)
self.result_path = Path(args.result_dir)
##### initialize network
self.net_opt = {
'guide': not args.no_guide,
'relu': args.use_relu,
'bn': not args.no_bn,
'cit': not args.no_cit
}
if self.net_opt['cit']:
self.Pretrain_ResNeXT = se_resnext_half(
dump_path=args.pretrain_dump,
num_classes=cit_class_num,
input_channels=1)
else:
self.Pretrain_ResNeXT = nn.Sequential()
self.G = Generator(input_size=args.input_size,
layers=args.layers,
cv_class_num=cvt_class_num,
iv_class_num=cit_class_num,
net_opt=self.net_opt)
self.D = Discriminator(input_dim=3,
output_dim=1,
input_size=self.input_size,
cv_class_num=cvt_class_num,
iv_class_num=cit_class_num)
for param in self.Pretrain_ResNeXT.parameters():
param.requires_grad = False
if args.test:
for param in self.G.parameters():
param.requires_grad = False
for param in self.D.parameters():
param.requires_grad = False
self.Pretrain_ResNeXT = nn.DataParallel(self.Pretrain_ResNeXT)
self.G = nn.DataParallel(self.G)
self.D = nn.DataParallel(self.D)
self.G_optimizer = optim.Adam(self.G.parameters(),
lr=args.lrG,
betas=(args.beta1, args.beta2))
self.D_optimizer = optim.Adam(self.D.parameters(),
lr=args.lrD,
betas=(args.beta1, args.beta2))
self.BCE_loss = nn.BCELoss()
self.CE_loss = nn.CrossEntropyLoss()
self.L1Loss = nn.L1Loss()
self.device = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
print("gpu mode: ", self.gpu_mode)
print("device: ", self.device)
print(torch.cuda.device_count(), "GPUS!")
if self.gpu_mode:
self.Pretrain_ResNeXT.to(self.device)
self.G.to(self.device)
self.D.to(self.device)
self.BCE_loss.to(self.device)
self.CE_loss.to(self.device)
self.L1Loss.to(self.device)
def train(self):
self.train_hist = {}
self.train_hist['D_loss'] = []
self.train_hist['G_loss'] = []
self.train_hist['per_epoch_time'] = []
self.train_hist['total_time'] = []
self.y_real_, self.y_fake_ = torch.ones(self.batch_size,
1), torch.zeros(
self.batch_size, 1)
if self.gpu_mode:
self.y_real_, self.y_fake_ = self.y_real_.to(
self.device), self.y_fake_.to(self.device)
if self.load_dump:
self.load(self.load_path)
print("continue training!!!!")
else:
self.end_epoch = self.epoch
self.print_params()
self.D.train()
print('training start!!')
start_time = time.time()
for epoch in range(self.start_epoch, self.end_epoch + 1):
print("EPOCH: {}".format(epoch))
self.G.train()
epoch_start_time = time.time()
if epoch == self.brightness_epoch:
print('changing brightness ...')
self.train_data_loader.dataset.enhance_brightness(
self.input_size)
max_iter = self.train_data_loader.dataset.__len__(
) // self.batch_size
for iter, (original_, sketch_, iv_tag_, cv_tag_) in enumerate(
tqdm(self.train_data_loader, ncols=80)):
if iter >= max_iter:
break
if self.gpu_mode:
sketch_, original_, iv_tag_, cv_tag_ = sketch_.to(
self.device), original_.to(self.device), iv_tag_.to(
self.device), cv_tag_.to(self.device)
# update D network
self.D_optimizer.zero_grad()
with torch.no_grad():
feature_tensor = self.Pretrain_ResNeXT(sketch_)
if self.gpu_mode:
feature_tensor = feature_tensor.to(self.device)
D_real, CIT_real, CVT_real = self.D(original_)
D_real_loss = self.BCE_loss(D_real, self.y_real_)
G_f, _ = self.G(sketch_, feature_tensor, cv_tag_)
if self.gpu_mode:
G_f = G_f.to(self.device)
D_f_fake, CIT_f_fake, CVT_f_fake = self.D(G_f)
D_f_fake_loss = self.BCE_loss(D_f_fake, self.y_fake_)
if self.two_step_epoch == 0 or epoch >= self.two_step_epoch:
CIT_real_loss = self.BCE_loss(
CIT_real, iv_tag_) if self.net_opt['cit'] else 0
CVT_real_loss = self.BCE_loss(CVT_real, cv_tag_)
C_real_loss = self.cvt_weight * CVT_real_loss + self.cit_weight * CIT_real_loss
CIT_f_fake_loss = self.BCE_loss(
CIT_f_fake, iv_tag_) if self.net_opt['cit'] else 0
CVT_f_fake_loss = self.BCE_loss(CVT_f_fake, cv_tag_)
C_f_fake_loss = self.cvt_weight * CVT_f_fake_loss + self.cit_weight * CIT_f_fake_loss
else:
C_real_loss = 0
C_f_fake_loss = 0
D_loss = self.adv_lambda * (D_real_loss + D_f_fake_loss) + (
C_real_loss + C_f_fake_loss)
self.train_hist['D_loss'].append(D_loss.item())
D_loss.backward()
self.D_optimizer.step()
# update G network
self.G_optimizer.zero_grad()
G_f, G_g = self.G(sketch_, feature_tensor, cv_tag_)
if self.gpu_mode:
G_f, G_g = G_f.to(self.device), G_g.to(self.device)
D_f_fake, CIT_f_fake, CVT_f_fake = self.D(G_f)
D_f_fake_loss = self.BCE_loss(D_f_fake, self.y_real_)
if self.two_step_epoch == 0 or epoch >= self.two_step_epoch:
CIT_f_fake_loss = self.BCE_loss(
CIT_f_fake, iv_tag_) if self.net_opt['cit'] else 0
CVT_f_fake_loss = self.BCE_loss(CVT_f_fake, cv_tag_)
C_f_fake_loss = self.cvt_weight * CVT_f_fake_loss + self.cit_weight * CIT_f_fake_loss
else:
C_f_fake_loss = 0
L1_D_f_fake_loss = self.L1Loss(G_f, original_)
L1_D_g_fake_loss = self.L1Loss(
G_g, original_) if self.net_opt['guide'] else 0
G_loss = (D_f_fake_loss + C_f_fake_loss) + \
(L1_D_f_fake_loss + L1_D_g_fake_loss * self.guide_beta) * self.l1_lambda
self.train_hist['G_loss'].append(G_loss.item())
G_loss.backward()
self.G_optimizer.step()
if ((iter + 1) % 100) == 0:
print(
"Epoch: [{:2d}] [{:4d}/{:4d}] D_loss: {:.8f}, G_loss: {:.8f}"
.format(epoch, (iter + 1), max_iter, D_loss.item(),
G_loss.item()))
self.train_hist['per_epoch_time'].append(time.time() -
epoch_start_time)
with torch.no_grad():
self.visualize_results(epoch)
utils.loss_plot(self.train_hist, self.result_path, epoch)
if epoch >= self.save_all_epoch > 0:
self.save(epoch)
elif self.save_freq > 0 and epoch % self.save_freq == 0:
self.save(epoch)
print("Training finish!... save training results")
if self.save_freq == 0 or epoch % self.save_freq != 0:
if self.save_all_epoch <= 0 or epoch < self.save_all_epoch:
self.save(epoch)
self.train_hist['total_time'].append(time.time() - start_time)
print(
"Avg one epoch time: {:.2f}, total {} epochs time: {:.2f}".format(
np.mean(self.train_hist['per_epoch_time']), self.epoch,
self.train_hist['total_time'][0]))
def test(self):
self.load_test(self.args.load)
self.D.eval()
self.G.eval()
load_path = self.load_path
result_path = self.result_path / load_path.stem
if not result_path.exists():
result_path.mkdir()
with torch.no_grad():
for sketch_, index_, _, cv_tag_ in tqdm(self.test_data_loader,
ncols=80):
if self.gpu_mode:
sketch_, cv_tag_ = sketch_.to(self.device), cv_tag_.to(
self.device)
with torch.no_grad():
feature_tensor = self.Pretrain_ResNeXT(sketch_)
if self.gpu_mode:
feature_tensor = feature_tensor.to(self.device)
# D_real, CIT_real, CVT_real = self.D(original_)
G_f, _ = self.G(sketch_, feature_tensor, cv_tag_)
G_f = self.color_revert(G_f.cpu())
for ind, result in zip(index_.cpu().numpy(), G_f):
save_path = result_path / f'{ind}.png'
if save_path.exists():
for i in range(100):
save_path = result_path / f'{ind}_{i}.png'
if not save_path.exists():
break
img = Image.fromarray(result)
img.save(save_path)
def visualize_results(self, epoch, fix=True):
if not self.result_path.exists():
self.result_path.mkdir()
self.G.eval()
# test_data_loader
original_, sketch_, iv_tag_, cv_tag_ = self.test_images
image_frame_dim = int(np.ceil(np.sqrt(len(original_))))
# iv_tag_ to feature tensor 16 * 16 * 256 by pre-reained Sketch.
with torch.no_grad():
feature_tensor = self.Pretrain_ResNeXT(sketch_)
if self.gpu_mode:
original_, sketch_, iv_tag_, cv_tag_, feature_tensor = original_.to(
self.device), sketch_.to(self.device), iv_tag_.to(
self.device), cv_tag_.to(
self.device), feature_tensor.to(self.device)
G_f, G_g = self.G(sketch_, feature_tensor, cv_tag_)
if self.gpu_mode:
G_f = G_f.cpu()
G_g = G_g.cpu()
G_f = self.color_revert(G_f)
G_g = self.color_revert(G_g)
utils.save_images(
G_f[:image_frame_dim * image_frame_dim, :, :, :],
[image_frame_dim, image_frame_dim],
self.result_path / 'tag2pix_epoch{:03d}_G_f.png'.format(epoch))
utils.save_images(
G_g[:image_frame_dim * image_frame_dim, :, :, :],
[image_frame_dim, image_frame_dim],
self.result_path / 'tag2pix_epoch{:03d}_G_g.png'.format(epoch))
def save(self, save_epoch):
if not self.result_path.exists():
self.result_path.mkdir()
with (self.result_path / 'arguments.txt').open('w') as f:
f.write(pprint.pformat(self.args.__dict__))
save_dir = self.result_path
torch.save(
{
'G': self.G.state_dict(),
'D': self.D.state_dict(),
'G_optimizer': self.G_optimizer.state_dict(),
'D_optimizer': self.D_optimizer.state_dict(),
'finish_epoch': save_epoch,
'result_path': str(save_dir)
}, str(save_dir / 'tag2pix_{}_epoch.pkl'.format(save_epoch)))
with (save_dir /
'tag2pix_{}_history.pkl'.format(save_epoch)).open('wb') as f:
pickle.dump(self.train_hist, f)
print("============= save success =============")
print("epoch from {} to {}".format(self.start_epoch, save_epoch))
print("save result path is {}".format(str(self.result_path)))
def load_test(self, checkpoint_path):
checkpoint = torch.load(str(checkpoint_path))
self.G.load_state_dict(checkpoint['G'])
def load(self, checkpoint_path):
checkpoint = torch.load(str(checkpoint_path))
self.G.load_state_dict(checkpoint['G'])
self.D.load_state_dict(checkpoint['D'])
self.G_optimizer.load_state_dict(checkpoint['G_optimizer'])
self.D_optimizer.load_state_dict(checkpoint['D_optimizer'])
self.start_epoch = checkpoint['finish_epoch'] + 1
self.finish_epoch = self.args.epoch + self.start_epoch - 1
print("============= load success =============")
print("epoch start from {} to {}".format(self.start_epoch,
self.finish_epoch))
print("previous result path is {}".format(checkpoint['result_path']))
def get_test_data(self, test_data_loader, count):
test_count = 0
original_, sketch_, iv_tag_, cv_tag_ = [], [], [], []
for orig, sket, ivt, cvt in test_data_loader:
original_.append(orig)
sketch_.append(sket)
iv_tag_.append(ivt)
cv_tag_.append(cvt)
test_count += len(orig)
if test_count >= count:
break
original_ = torch.cat(original_, 0)
sketch_ = torch.cat(sketch_, 0)
iv_tag_ = torch.cat(iv_tag_, 0)
cv_tag_ = torch.cat(cv_tag_, 0)
self.save_tag_tensor_name(iv_tag_, cv_tag_,
self.result_path / "test_image_tags.txt")
image_frame_dim = int(np.ceil(np.sqrt(len(original_))))
if self.gpu_mode:
original_ = original_.cpu()
sketch_np = sketch_.data.numpy().transpose(0, 2, 3, 1)
original_np = self.color_revert(original_)
utils.save_images(
original_np[:image_frame_dim * image_frame_dim, :, :, :],
[image_frame_dim, image_frame_dim],
self.result_path / 'tag2pix_original.png')
utils.save_images(
sketch_np[:image_frame_dim * image_frame_dim, :, :, :],
[image_frame_dim, image_frame_dim],
self.result_path / 'tag2pix_sketch.png')
return original_, sketch_, iv_tag_, cv_tag_
def save_tag_tensor_name(self, iv_tensor, cv_tensor, save_file_path):
'''iv_tensor, cv_tensor: batched one-hot tag tensors'''
iv_dict_inverse = {
tag_index: tag_id
for (tag_id, tag_index) in self.iv_dict.items()
}
cv_dict_inverse = {
tag_index: tag_id
for (tag_id, tag_index) in self.cv_dict.items()
}
with open(save_file_path, 'w') as f:
f.write("CIT tags\n")
for tensor_i, batch_unit in enumerate(iv_tensor):
tag_list = []
f.write(f'{tensor_i} : ')
for i, is_tag in enumerate(batch_unit):
if is_tag:
tag_name = self.id_to_name[iv_dict_inverse[i]]
tag_list.append(tag_name)
f.write(f"{tag_name}, ")
f.write("\n")
f.write("\nCVT tags\n")
for tensor_i, batch_unit in enumerate(cv_tensor):
tag_list = []
f.write(f'{tensor_i} : ')
for i, is_tag in enumerate(batch_unit):
if is_tag:
tag_name = self.id_to_name[cv_dict_inverse[i]]
tag_list.append(self.id_to_name[cv_dict_inverse[i]])
f.write(f"{tag_name}, ")
f.write("\n")
def print_params(self):
params_cnt = [0, 0, 0]
for param in self.G.parameters():
params_cnt[0] += param.numel()
for param in self.D.parameters():
params_cnt[1] += param.numel()
for param in self.Pretrain_ResNeXT.parameters():
params_cnt[2] += param.numel()
print(
f'Parameter #: G - {params_cnt[0]} / D - {params_cnt[1]} / Pretrain - {params_cnt[2]}'
)
def train_network(args):
device = torch.device('cuda' if args.gpu_no >= 0 else 'cpu')
G_A2B = Generator().to(device).train()
G_B2A = Generator().to(device).train()
print("Load Generator", G_A2B)
D_A = Discriminator().to(device).train()
D_B = Discriminator().to(device).train()
print("Load Discriminator", D_A)
optimizer_G_A2B = torch.optim.Adam(G_A2B.parameters(),
lr=args.lr,
betas=(args.beta1, args.beta2))
optimizer_G_B2A = torch.optim.Adam(G_B2A.parameters(),
lr=args.lr,
betas=(args.beta1, args.beta2))
optimizer_D_A = torch.optim.Adam(D_A.parameters(),
lr=args.lr,
betas=(args.beta1, args.beta2))
optimizer_D_B = torch.optim.Adam(D_B.parameters(),
lr=args.lr,
betas=(args.beta1, args.beta2))
mse_criterion = nn.MSELoss()
l1_criterion = nn.L1Loss()
dataset = FacadeFolder(args.data_A, args.data_B, args.imsize,
args.cropsize, args.cencrop)
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers)
loss_seq = {'G_A2B': [], 'G_B2A': [], 'D_A': [], 'D_B': [], 'Cycle': []}
buffer_fake_A = BufferN(50)
buffer_fake_B = BufferN(50)
for epoch in range(args.epoch):
"""
learning rate schedule
"""
for iteration, (item_A, item_B) in enumerate(dataloader, 1):
item_A, item_B = item_A.to(device), item_B.to(device)
"""
train discriminators
"""
with torch.no_grad():
fake_B = G_A2B(item_A)
fake_A = G_B2A(item_B)
buffer_fake_B.push(fake_B.detach())
N_fake_B_discrimination = D_B(
torch.cat(buffer_fake_B.get_buffer(), dim=0))
real_B_discrimination = D_B(item_B)
buffer_fake_A.push(fake_A.detach())
N_fake_A_discrimination = D_A(
torch.cat(buffer_fake_A.get_buffer(), dim=0))
real_A_discrimination = D_A(item_A)
loss_D_B_fake = mse_criterion(
N_fake_B_discrimination,
torch.zeros_like(N_fake_B_discrimination).to(device))
loss_D_B_real = mse_criterion(
real_B_discrimination,
torch.ones_like(real_B_discrimination).to(device))
loss_D_B = (loss_D_B_fake +
loss_D_B_real) * args.discriminator_weight
loss_seq['D_B'].append(loss_D_B.item())
loss_D_A_fake = mse_criterion(
N_fake_A_discrimination,
torch.zeros_like(N_fake_A_discrimination).to(device))
loss_D_A_real = mse_criterion(
real_A_discrimination,
torch.ones_like(real_A_discrimination).to(device))
loss_D_A = (loss_D_A_fake +
loss_D_A_real) * args.discriminator_weight
loss_seq['D_A'].append(loss_D_A.item())
optimizer_D_B.zero_grad()
loss_D_B.backward()
optimizer_D_B.step()
optimizer_D_A.zero_grad()
loss_D_A.backward()
optimizer_D_A.step()
"""
train generators
"""
fake_B = G_A2B(item_A)
fake_A = G_B2A(item_B)
fake_B_discrimination = D_B(fake_B)
fake_A_discrimination = D_A(fake_A)
B_from_fake_A = G_A2B(fake_A)
A_from_fake_B = G_B2A(fake_B)
loss_G_A2B_gan = mse_criterion(
fake_B_discrimination,
torch.ones_like(fake_B_discrimination).to(device))
loss_G_B2A_gan = mse_criterion(
fake_A_discrimination,
torch.ones_like(fake_A_discrimination).to(device))
loss_seq['G_A2B'].append(loss_G_A2B_gan.item())
loss_seq['G_B2A'].append(loss_G_B2A_gan.item())
loss_cyc = l1_criterion(B_from_fake_A, item_B) + l1_criterion(
A_from_fake_B, item_A)
loss_seq['Cycle'].append(loss_cyc.item())
loss_G = loss_G_A2B_gan + loss_G_B2A_gan + args.cyc_weight * loss_cyc
optimizer_G_A2B.zero_grad()
optimizer_G_B2A.zero_grad()
loss_G.backward()
optimizer_G_A2B.step()
optimizer_G_B2A.step()
"""
check training loss
"""
if iteration % args.check_iter == 0:
check_str = "%s: epoch:[%d/%d]\titeration:[%d/%d]" % (
time.ctime(), epoch, args.epoch, iteration,
len(dataloader))
for key, value in loss_seq.items():
check_str += "\t%s: %2.2f" % (
key, lastest_arverage_value(value))
print(check_str)
imsave(torch.cat([item_A, item_B, fake_B], dim=0),
args.save_path + 'training_image_A2B.jpg')
imsave(torch.cat([item_B, item_A, fake_A], dim=0),
args.save_path + 'training_image_B2A.jpg')
# save networks
torch.save(
{
'iteration': iteration,
'G_A2B_state_dict': G_A2B.state_dict(),
'G_B2A_state_dict': G_B2A.state_dict(),
'D_A_state_dict': D_A.state_dict(),
'D_B_state_dict': D_B.state_dict(),
'loss_seq': loss_seq
}, args.save_path + 'check_point.pth')
return None
# shift models to cuda if possible
if torch.cuda.is_available():
G12.cuda()
G21.cuda()
D1.cuda()
D2.cuda()
# %%
# optimizer and loss
LGAN = MSELoss()
LCYC = L1Loss()
LIdentity = L1Loss()
optimizer_G = Adam(itertools.chain(G12.parameters(), G21.parameters()),
lr=0.001)
optimizer_D1 = Adam(D1.parameters(), lr=0.001)
optimizer_D2 = Adam(D2.parameters(), lr=0.001)
# %%
# train models
real_label = torch.full((32, ), 1, device="cuda:0")
false_label = torch.full((32, ), 0, device="cuda:0")
bufD1 = Buffer(50)
bufD2 = Buffer(50)
num_epochs = 100
learning_rate = 0.01
for epoch in range(num_epochs):
for i, (realA, realB) in enumerate(dataloader):
if torch.cuda.is_available():
realA.cuda()
pass
net = net.to(device)
print(net)
NZ = opt.imageSize // 2**nDep
noise = torch.FloatTensor(opt.batchSize, nz, NZ, NZ)
real_label = 1
fake_label = 0
noise = noise.to(device)
fixnoise = fixnoise.to(device)
targetMosaic = targetMosaic.to(device)
# setup optimizer
optimizerD = optim.Adam(netD.parameters(), lr=opt.lr,
betas=(opt.beta1, 0.999)) #netD.parameters()
optimizerU = optim.Adam(
[param for net in Gnets for param in list(net.parameters())],
lr=opt.lr,
betas=(opt.beta1, 0.999))
def ganGeneration(content, noise, templates=None, bVis=False):
x = netMix(content, noise)
if bVis:
return x, 0, 0, 0
return x
buf = []
def train(data_path,
crop_size=128,
final_size=64,
batch_size=16,
alternating_step=10000,
ncritic=1,
lambda_gp=0.1,
debug_step=100):
# define networks
generator = Generator(final_size=final_size)
generator.generate_network()
g_optimizer = Adam(generator.parameters())
discriminator = Discriminator(final_size=final_size)
discriminator.generate_network()
d_optimizer = Adam(discriminator.parameters())
num_channels = min(generator.num_channels, generator.max_channels)
# get debugging vectors
N = (5, 10)
debug_vectors = torch.randn(N[0] * N[1], num_channels, 1, 1).to(device)
global upsample
upsample = [
Upsample(scale_factor=2**i)
for i in reversed(range(generator.num_blocks))
]
# get loader
loader = get_loader(data_path, crop_size, batch_size)
# training loop
start_time = time.time()
for index in range(generator.num_blocks):
loader.dataset.set_transform_by_index(index)
data_iterator = iter(loader)
for phase in ('fade', 'stabilize'):
if index == 0 and phase == 'fade': continue
print("index: {}, size: {}x{}, phase: {}".format(
index, 2**(index + 2), 2**(index + 2), phase))
for i in range(alternating_step):
print(i)
try:
batch = next(data_iterator)
except:
data_iterator = iter(loader)
batch = next(data_iterator)
alpha = i / alternating_step if phase == "fade" else 1.0
batch = batch.to(device)
d_loss_real = -torch.mean(discriminator(batch, index, alpha))
latent = torch.randn(batch_size, num_channels, 1, 1).to(device)
fake_batch = generator(latent, index, alpha).detach()
d_loss_fake = torch.mean(
discriminator(fake_batch, index, alpha))
d_loss = d_loss_real + d_loss_fake
d_optimizer.zero_grad()
d_loss.backward() # if retain_graph=True
# then gp works but I'm not sure it's right
d_optimizer.step()
# Compute gradient penalty
alpha_gp = torch.rand(batch.size(0), 1, 1, 1).to(device)
# mind that x_hat must be both detached from the previous
# gradient graph (from fake_barch) and with
# requires_graph=True so that the gradient can be computed
x_hat = (alpha_gp * batch +
(1 - alpha_gp) * fake_batch).requires_grad_(True)
# x_hat = torch.cuda.FloatTensor(x_hat).requires_grad_(True)
out = discriminator(x_hat, index, alpha)
grad = torch.autograd.grad(
outputs=out,
inputs=x_hat,
grad_outputs=torch.ones_like(out).to(device),
retain_graph=True,
create_graph=True,
only_inputs=True)[0]
grad = grad.view(grad.size(0), -1) #is this the same as
# detach?
l2norm = torch.sqrt(torch.sum(grad**2, dim=1))
d_loss_gp = torch.mean((l2norm - 1)**2)
d_loss_gp *= lambda_gp
d_optimizer.zero_grad()
d_loss_gp.backward()
d_optimizer.step()
if (i + 1) % ncritic == 0:
latent = torch.randn(batch_size, num_channels, 1,
1).to(device)
fake_batch = generator(latent, index, alpha)
g_loss = -torch.mean(
discriminator(fake_batch, index, alpha))
g_optimizer.zero_grad()
g_loss.backward()
g_optimizer.step()
# print debugging images
if (i + 1) % debug_step == 0:
print_debugging_images(generator, debug_vectors, N, index,
alpha, i)
print (e,"weightinit")
pass
net=net.to(device)
print(net)
NZ = opt.imageSize//2**nDep
noise = torch.FloatTensor(opt.batchSize, nz, NZ,NZ)
real_label = 1
fake_label = 0
noise=noise.to(device)
fixnoise=fixnoise.to(device)
# setup optimizer
optimizerD = optim.Adam(netD.parameters(), lr=opt.lr, betas=(opt.beta1, 0.999))#netD.parameters()
optimizerU = optim.Adam([param for net in Gnets for param in list(net.parameters())], lr=opt.lr, betas=(opt.beta1, 0.999))
def famosGeneration(content, noise, templatePatch, bVis = False):
if opt.multiScale >0:
x = netMix(content, noise,templatePatch)
else:
x = netMix(content,noise)
a5=x[:,-5:]
A =4*nn.functional.tanh(x[:,:-5])##smooths probs somehow
A = nn.functional.softmax(1*(A - A.detach().max()), dim=1)
mixed = getTemplateMixImage(A, templatePatch)
alpha = nn.functional.sigmoid(a5[:,3:4])
beta = nn.functional.sigmoid(a5[:, 4:5])
fake = blend(nn.functional.tanh(a5[:,:3]),mixed,alpha,beta)
class Solver(object):
def __init__(self, configuration):
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# retrieve configuration variables
self.data_path = configuration.data_path
self.crop_size = configuration.crop_size
self.final_size = configuration.final_size
self.batch_size = configuration.batch_size
self.alternating_step = configuration.alternating_step
self.ncritic = configuration.ncritic
self.lambda_gp = configuration.lambda_gp
self.debug_step = configuration.debug_step
self.save_step = configuration.save_step
self.max_checkpoints = configuration.max_checkpoints
self.log_step = configuration.log_step
# self.tflogger = Logger(configuration.log_dir)
## directoriess
self.train_dir = configuration.train_dir
self.img_dir = configuration.img_dir
self.models_dir = configuration.models_dir
## variables
self.eps_drift = 0.001
self.resume_training = configuration.resume_training
self._initialise_networks()
def _initialise_networks(self):
self.generator = Generator(final_size=self.final_size)
self.generator.generate_network()
self.g_optimizer = Adam(self.generator.parameters(), lr=0.001, betas=(0, 0.99))
self.discriminator = Discriminator(final_size=self.final_size)
self.discriminator.generate_network()
self.d_optimizer = Adam(self.discriminator.parameters(), lr=0.001, betas=(0, 0.99))
self.num_channels = min(self.generator.num_channels,
self.generator.max_channels)
self.upsample = [Upsample(scale_factor=2**i)
for i in reversed(range(self.generator.num_blocks))]
def print_debugging_images(self, generator, latent_vectors, shape, index,
alpha, iteration):
with torch.no_grad():
columns = []
for i in range(shape[0]):
row = []
for j in range(shape[1]):
img_ij = generator(latent_vectors[i * shape[1] +
j].unsqueeze_(0),
index, alpha)
img_ij = self.upsample[index](img_ij)
row.append(img_ij)
columns.append(torch.cat(row, dim=3))
debugging_image = torch.cat(columns, dim=2)
# denorm
debugging_image = (debugging_image + 1) / 2
debugging_image.clamp_(0, 1)
save_image(debugging_image.data,
os.path.join(self.img_dir, "debug_{}_{}.png".format(index,
iteration)))
def save_trained_networks(self, block_index, phase, step):
models_file = os.path.join(self.models_dir, "models.json")
if os.path.isfile(models_file):
with open(models_file, 'r') as file:
models_config = json.load(file)
else:
models_config = json.loads('{ "checkpoints": [] }')
generator_save_name = "generator_{}_{}_{}.pth".format(
block_index, phase, step
)
torch.save(self.generator.state_dict(),
os.path.join(self.models_dir, generator_save_name))
discriminator_save_name = "discriminator_{}_{}_{}.pth".format(
block_index, phase, step
)
torch.save(self.discriminator.state_dict(),
os.path.join(self.models_dir, discriminator_save_name))
models_config["checkpoints"].append(OrderedDict({
"block_index": block_index,
"phase": phase,
"step": step,
"generator": generator_save_name,
"discriminator": discriminator_save_name
}))
if len(models_config["checkpoints"]) > self.max_checkpoints:
old_save = models_config["checkpoints"][0]
os.remove(os.path.join(self.models_dir, old_save["generator"]))
os.remove(os.path.join(self.models_dir, old_save["discriminator"]))
models_config["checkpoints"] = models_config["checkpoints"][1:]
with open(os.path.join(self.models_dir, "models.json"), 'w') as file:
json.dump(models_config, file, indent=4)
def load_trained_networks(self):
models_file = os.path.join(self.models_dir, "models.json")
if os.path.isfile(models_file):
with open(models_file, 'r') as file:
models_config = json.load(file)
else:
raise FileNotFoundError("File 'models.json' not found in {"
"}".format(self.models_dir))
last_checkpoint = models_config["checkpoints"][-1]
block_index = last_checkpoint["block_index"]
phase = last_checkpoint["phase"]
step = last_checkpoint["step"]
generator_save_name = os.path.join(
self.models_dir, last_checkpoint["generator"])
discriminator_save_name = os.path.join(
self.models_dir, last_checkpoint["discriminator"])
self.generator.load_state_dict(torch.load(generator_save_name))
self.discriminator.load_state_dict(torch.load(discriminator_save_name))
return block_index, phase, step
def train(self):
# get debugging vectors
N = (5, 10)
debug_vectors = torch.randn(N[0] * N[1], self.num_channels, 1,
1).to(self.device)
# get loader
loader = get_loader(self.data_path, self.crop_size, self.batch_size)
losses = {
"d_loss_real": None,
"d_loss_fake": None,
"g_loss": None
}
# resume training if needed
if self.resume_training:
start_index, start_phase, start_step = self.load_trained_networks()
else:
start_index, start_phase, start_step = (0, "fade", 0)
# training loop
start_time = time.time()
absolute_step = -1
for index in range(start_index, self.generator.num_blocks):
loader.dataset.set_transform_by_index(index)
data_iterator = iter(loader)
for phase in ('fade', 'stabilize'):
if index == 0 and phase == 'fade': continue
if self.resume_training and \
index == start_index and \
phase is not start_phase:
continue #
if phase == 'phade': self.alternating_step = 10000 #FIXME del
print("index: {}, size: {}x{}, phase: {}".format(
index, 2 ** (index + 2), 2 ** (index + 2), phase))
if self.resume_training and \
phase == start_phase and \
index == start_index:
step_range = range(start_step, self.alternating_step)
else:
step_range = range(self.alternating_step)
for i in step_range:
absolute_step += 1
try:
batch = next(data_iterator)
except:
data_iterator = iter(loader)
batch = next(data_iterator)
alpha = i / self.alternating_step if phase == "fade" else 1.0
batch = batch.to(self.device)
d_loss_real = - torch.mean(
self.discriminator(batch, index, alpha))
losses["d_loss_real"] = torch.mean(d_loss_real).data[0]
latent = torch.randn(
batch.size(0), self.num_channels, 1, 1).to(self.device)
fake_batch = self.generator(latent, index, alpha).detach()
d_loss_fake = torch.mean(
self.discriminator(fake_batch, index, alpha))
losses["d_loss_fake"] = torch.mean(d_loss_fake).data[0]
# drift factor
drift = d_loss_real.pow(2) + d_loss_fake.pow(2)
d_loss = d_loss_real + d_loss_fake + self.eps_drift * drift
self.d_optimizer.zero_grad()
d_loss.backward() # if retain_graph=True
# then gp works but I'm not sure it's right
self.d_optimizer.step()
# Compute gradient penalty
alpha_gp = torch.rand(batch.size(0), 1, 1, 1).to(self.device)
# mind that x_hat must be both detached from the previous
# gradient graph (from fake_barch) and with
# requires_graph=True so that the gradient can be computed
x_hat = (alpha_gp * batch + (1 - alpha_gp) *
fake_batch).requires_grad_(True)
# x_hat = torch.cuda.FloatTensor(x_hat).requires_grad_(True)
out = self.discriminator(x_hat, index, alpha)
grad = torch.autograd.grad(
outputs=out,
inputs=x_hat,
grad_outputs=torch.ones_like(out).to(self.device),
retain_graph=True,
create_graph=True,
only_inputs=True
)[0]
grad = grad.view(grad.size(0), -1) # is this the same as
# detach?
l2norm = torch.sqrt(torch.sum(grad ** 2, dim=1))
d_loss_gp = torch.mean((l2norm - 1) ** 2)
d_loss_gp *= self.lambda_gp
self.d_optimizer.zero_grad()
d_loss_gp.backward()
self.d_optimizer.step()
# train generator
if (i + 1) % self.ncritic == 0:
latent = torch.randn(
self.batch_size, self.num_channels, 1, 1).to(self.device)
fake_batch = self.generator(latent, index, alpha)
g_loss = - torch.mean(self.discriminator(
fake_batch, index, alpha))
losses["g_loss"] = torch.mean(g_loss).data[0]
self.g_optimizer.zero_grad()
g_loss.backward()
self.g_optimizer.step()
# tensorboard logging
if (i + 1) % self.log_step == 0:
elapsed = time.time() - start_time
elapsed = str(datetime.timedelta(seconds=elapsed))
print("{}:{}:{}/{} time {}, d_loss_real {}, "
"d_loss_fake {}, "
"g_loss {}, alpha {}".format(index, phase, i,
self.alternating_step,
elapsed,
d_loss_real,
d_loss_fake,
g_loss, alpha))
for name, value in losses.items():
self.tflogger.scalar_summary(name, value, absolute_step)
# print debugging images
if (i + 1) % self.debug_step == 0:
self.print_debugging_images(
self.generator, debug_vectors, N, index, alpha, i)
# save trained networks
if (i + 1) % self.save_step == 0:
self.save_trained_networks(index, phase, i)
def main():
data_loader = load_data()
G = Generator(C_dim_disc + C_dim_cont + Z_dim)
D = Discriminator(C_dim_disc=C_dim_disc, C_dim_cont=C_dim_cont)
G_optimizer = optim.Adam(G.parameters(), G_lr, [G_beta1, G_beta2])
D_optimizer = optim.Adam(D.parameters(), D_lr, [G_beta1, D_beta2])
for epoch in range(epochs):
for it, (image, _label) in enumerate(data_loader):
# Sample data
X_numpy = np.array(image)
X_numpy = -X_numpy * 2
X = Variable(torch.from_numpy(X_numpy))
C_disc, C_cont, Z = (
sample_multinomial(size=[batch_size], n=C_dim_disc),
sample_normal([batch_size, C_dim_cont], scale=0.5),
sample_normal([batch_size, Z_dim], scale=0.5),
)
C = torch.cat([C_disc, C_cont], dim=1)
G.train(True)
# Dicriminator forward-loss-backward-update
G_sample = G(torch.cat([C, Z], dim=1))
D_real, D_result = D(X), D(G_sample, cal_Q=True)
D_fake, C_given_X = D_result
C_disc_given_x, C_cont_given_x = C_given_X
# D_result = D(torch.cat([X, G_sample], dim=0), cal_Q=True)
# D_real, D_fake = D_result[0][:mb_size], D_result[0][mb_size:]
# C_disc_given_x, C_cont_given_x = D_result[1][0][mb_size:], D_result[1][1][mb_size:]
Q_loss = -torch.mean(torch.sum(
C_disc * C_disc_given_x, 1)) + torch.mean(
(((C_cont - C_cont_given_x) / 0.5)**2))
D_loss = -torch.mean(
torch.log(D_real + 1e-8) +
torch.log(1 - D_fake + 1e-8)) + Q_loss
D_optimizer.zero_grad()
D_loss.backward(retain_graph=True)
D_optimizer.step()
# Generator forward-loss-backward-update
G_loss = -torch.mean(torch.log(D_fake + 1e-8)) + Q_loss
G_optimizer.zero_grad()
G_loss.backward()
G_optimizer.step()
if it % 200 == 0:
G.train(False)
print(f"epoch: {epoch}, iter: {it}")
print(
f"D_loss: {D_loss.data.numpy():.6f}, G_loss: {G_loss.data.numpy():.6f}, Prob diff: {D_real.mean() - D_fake.mean():.6f}"
)
C_sample_cont = torch.zeros([1, C_dim_cont])
Z_sample = torch.zeros([1, Z_dim])
C_sample_disc, C_sample_cont = (
sample_multinomial(size=[100], n=C_dim_disc),
sample_normal([100, C_dim_cont], scale=0.5),
)
G_sample = G(torch.cat([C_sample_disc, C_sample_cont], dim=1))
for k in range(2):
gs = gridspec.GridSpec(10, 10)
fig = plt.figure(figsize=(10, 10))
plt.suptitle(f"epoch: {epoch}, iter: {it}\n{get_time()}")
gs.update(wspace=0.05, hspace=0.05)
k = 0
for i in range(10):
j_space = np.linspace(-1, 1,
10) if k == 0 else range(10)
for j in j_space:
C_sample_disc = torch.zeros([1, 10])
if k == 0:
C_sample_disc[0, i] = 1
C_sample_cont[0, 0] = j
else:
C_sample_disc[0, i] = 1
C_sample_disc[0, j] = 1
G_sample = G(
torch.cat(
[C_sample_disc, C_sample_cont, Z_sample],
dim=1))[0]
ax = plt.subplot(gs[k])
plt.axis("off")
ax.set_xticklabels([])
ax.set_yticklabels([])
ax.set_aspect("equal")
plt.imshow(G_sample.reshape(28, 28).data.numpy(),
cmap="Greys",
vmin=-1,
vmax=1)
fig.subplots_adjust(top=0.92)
plt.tight_layout()
plt.savefig(f"output/{epoch:02d}_{it:04d}_{k}.png",
bbox_inches="tight")
plt.close(fig)
torch.save(G.state_dict(), "model/G.pkl")
torch.save(D.state_dict(), "model/D.pkl")
criterion = torch.nn.CrossEntropyLoss()
models=[]
zzz=np.arange(len(AllLabel)*2).reshape((len(AllLabel), 2))
X=FinalData
y=np.array(AllLabel)
skf = StratifiedKFold(n_splits=5,shuffle=True, random_state=random.randint(0,99))
for train_index,test_index in skf.split(zzz,y):
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = y[train_index], y[test_index]
torch_dataset = torch.utils.data.TensorDataset(torch.FloatTensor(X_train), torch.LongTensor(y_train))
data_loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=True, drop_last=True)
loss_classifier_list = []
for epoch in range(1, num_epochs + 1):
learning_rate = base_lr * math.pow(0.9, epoch / lr_step)
gamma_rate = 2 / (1 + math.exp(-10 * (epoch) / num_epochs)) - 1
optimizer = torch.optim.Adam([{'params': discriminator.parameters()},], lr=learning_rate, weight_decay=l2_decay)
discriminator.train()
iter_data = iter(data_loader)
num_iter = len(data_loader)
#print(num_iter)
total_clas_loss = 0
num_batches = 0
for it in range(0, num_iter):
data, label = iter_data.next()
if it % len(data_loader) == 0:
iter_data = iter(data_loader)
data = Variable(torch.FloatTensor(data))
label = Variable(torch.LongTensor(label))
Disc_a = discriminator(data)
def main(params):
if params['load_dataset']:
dataset = load_pkl(params['load_dataset'])
elif params['dataset_class']:
dataset = globals()[params['dataset_class']](**params[params['dataset_class']])
if params['save_dataset']:
save_pkl(params['save_dataset'], dataset)
else:
raise Exception('One of either load_dataset (path to pkl) or dataset_class needs to be specified.')
result_dir = create_result_subdir(params['result_dir'], params['exp_name'])
losses = ['G_loss', 'D_loss', 'D_real', 'D_fake']
stats_to_log = [
'tick_stat',
'kimg_stat',
]
if params['progressive_growing']:
stats_to_log.extend([
'depth',
'alpha',
'lod',
'minibatch_size'
])
stats_to_log.extend([
'time',
'sec.tick',
'sec.kimg'
] + losses)
logger = TeeLogger(os.path.join(result_dir, 'log.txt'), stats_to_log, [(1, 'epoch')])
logger.log(params_to_str(params))
if params['resume_network']:
G, D = load_models(params['resume_network'], params['result_dir'], logger)
else:
G = Generator(dataset.shape, **params['Generator'])
D = Discriminator(dataset.shape, **params['Discriminator'])
if params['progressive_growing']:
assert G.max_depth == D.max_depth
G.cuda()
D.cuda()
latent_size = params['Generator']['latent_size']
logger.log(str(G))
logger.log('Total nuber of parameters in Generator: {}'.format(
sum(map(lambda x: reduce(lambda a, b: a*b, x.size()), G.parameters()))
))
logger.log(str(D))
logger.log('Total nuber of parameters in Discriminator: {}'.format(
sum(map(lambda x: reduce(lambda a, b: a*b, x.size()), D.parameters()))
))
def get_dataloader(minibatch_size):
return DataLoader(dataset, minibatch_size, sampler=InfiniteRandomSampler(dataset),
num_workers=params['num_data_workers'], pin_memory=False, drop_last=True)
def rl(bs):
return lambda: random_latents(bs, latent_size)
# Setting up learning rate and optimizers
opt_g = Adam(G.parameters(), params['G_lr_max'], **params['Adam'])
opt_d = Adam(D.parameters(), params['D_lr_max'], **params['Adam'])
def rampup(cur_nimg):
if cur_nimg < params['lr_rampup_kimg'] * 1000:
p = max(0.0, 1 - cur_nimg / (params['lr_rampup_kimg'] * 1000))
return np.exp(-p * p * 5.0)
else:
return 1.0
lr_scheduler_d = LambdaLR(opt_d, rampup)
lr_scheduler_g = LambdaLR(opt_g, rampup)
mb_def = params['minibatch_size']
D_loss_fun = partial(wgan_gp_D_loss, return_all=True, iwass_lambda=params['iwass_lambda'],
iwass_epsilon=params['iwass_epsilon'], iwass_target=params['iwass_target'])
G_loss_fun = wgan_gp_G_loss
trainer = Trainer(D, G, D_loss_fun, G_loss_fun,
opt_d, opt_g, dataset, iter(get_dataloader(mb_def)), rl(mb_def), **params['Trainer'])
# plugins
if params['progressive_growing']:
max_depth = min(G.max_depth, D.max_depth)
trainer.register_plugin(DepthManager(get_dataloader, rl, max_depth, **params['DepthManager']))
for i, loss_name in enumerate(losses):
trainer.register_plugin(EfficientLossMonitor(i, loss_name))
checkpoints_dir = params['checkpoints_dir'] if params['checkpoints_dir'] else result_dir
trainer.register_plugin(SaverPlugin(checkpoints_dir, **params['SaverPlugin']))
def subsitute_samples_path(d):
return {k:(os.path.join(result_dir, v) if k == 'samples_path' else v) for k,v in d.items()}
postprocessors = [ globals()[x](**subsitute_samples_path(params[x])) for x in params['postprocessors'] ]
trainer.register_plugin(OutputGenerator(lambda x: random_latents(x, latent_size),
postprocessors, **params['OutputGenerator']))
trainer.register_plugin(AbsoluteTimeMonitor(params['resume_time']))
trainer.register_plugin(LRScheduler(lr_scheduler_d, lr_scheduler_g))
trainer.register_plugin(logger)
init_comet(params, trainer)
trainer.run(params['total_kimg'])
dataset.close()
DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Set up the GAN.
discriminator_model = Discriminator().to(DEVICE)
generator_model = Generator().to(DEVICE)
# Load pre-trained models if they are provided.
if args.load_discriminator_model_path:
discriminator_model.load_state_dict(
torch.load(args.load_discriminator_model_path))
if args.load_generator_model_path:
generator_model.load_state_dict(torch.load(args.load_generator_model_path))
# Set up Adam optimizers for both models.
discriminator_optimizer = optim.Adam(discriminator_model.parameters(),
lr=args.learning_rate,
betas=(0, 0.9))
generator_optimizer = optim.Adam(generator_model.parameters(),
lr=args.learning_rate,
betas=(0, 0.9))
# Create a random batch of latent space vectors that will be used to visualize the progression of the generator.
fixed_latent_space_vectors = torch.randn(64, 512, 1, 1, device=DEVICE)
# Load and preprocess images.
images = load_images(args.data_dir, args.training_set_size)
# Add network architectures for Discriminator and Generator to TensorBoard.
if not args.dry_run:
WRITER.add_graph(discriminator_model,
n_classes = 10
init_epoch = 0
max_epoch = 10
batch_size = 128
save_epoch = 1
in_channels = z_dim + n_classes
out_channels = n_classes + 1 # for fake class
lr_init = 0.001
train_loader = get_dataset(root, batch_size)
gen = Generator(in_channels).to(device)
dis = Discriminator(out_channels).to(device)
g_optimizer = torch.optim.RMSprop(gen.parameters(), lr=lr_init, alpha=0.9)
d_optimizer = torch.optim.RMSprop(dis.parameters(), lr=lr_init, alpha=0.9)
gen_loss = []
dis_loss = []
for index in range(init_epoch, max_epoch):
for i, data in enumerate(train_loader):
# lr reduced by factor of 10 after 8 epoch
if index >= 8:
for g in g_optimizer.param_groups:
g['lr'] = lr_init / 10
for g in d_optimizer.param_groups:
g['lr'] = lr_init / 10
# Define input variables
x_n = torch.tensor(np.array(data[0], dtype=np.float32), device=device)
y = torch.tensor(np.array(data[1], dtype=np.float32), device=device)
#print(netG_A2B)
#print(netD_A)
# Lossess
gan_loss = torch.nn.MSELoss()
cycle_consistency_loss = torch.nn.L1Loss()
#optimizers
optimizerG_A2B = optim.Adam(netG_A2B.parameters(),
lr=opt.lr,
betas=(opt.beta1, 0.999))
optimizerG_B2A = optim.Adam(netG_B2A.parameters(),
lr=opt.lr,
betas=(opt.beta1, 0.999))
optimizerD_A = optim.Adam(netD_A.parameters(),
lr=opt.lr,
betas=(opt.beta1, 0.999))
optimizerD_B = optim.Adam(netD_B.parameters(),
lr=opt.lr,
betas=(opt.beta1, 0.999))
lr_scheduler_G_A2B = optim.lr_scheduler.LambdaLR(optimizerG_A2B,
lr_lambda=updateLr(
opt.num_epochs,
opt.decay_epoch).update)
lr_scheduler_G_B2A = optim.lr_scheduler.LambdaLR(optimizerG_B2A,
lr_lambda=updateLr(
opt.num_epochs,
opt.decay_epoch).update)
lr_scheduler_D_A = optim.lr_scheduler.LambdaLR(optimizerD_A,
def main(params):
if params['load_dataset']:
dataset = load_pkl(params['load_dataset'])
elif params['dataset_class']:
dataset = globals()[params['dataset_class']](
**params[params['dataset_class']])
if params['save_dataset']:
save_pkl(params['save_dataset'], dataset)
else:
raise Exception(
'One of either load_dataset (path to pkl) or dataset_class needs to be specified.'
)
result_dir = create_result_subdir(params['result_dir'], params['exp_name'])
losses = ['G_loss', 'D_loss', 'D_real', 'D_fake']
stats_to_log = [
'tick_stat',
'kimg_stat',
]
if params['progressive_growing']:
stats_to_log.extend(['depth', 'alpha', 'lod', 'minibatch_size'])
stats_to_log.extend(['time', 'sec.tick', 'sec.kimg'] + losses)
logger = TeeLogger(os.path.join(result_dir, 'log.txt'), stats_to_log,
[(1, 'epoch')])
logger.log(params_to_str(params))
if params['resume_network']:
G, D = load_models(params['resume_network'], params['result_dir'],
logger)
else:
G = Generator(dataset.shape, **params['Generator'])
D = Discriminator(dataset.shape, **params['Discriminator'])
if params['progressive_growing']:
assert G.max_depth == D.max_depth
G.cuda()
D.cuda()
latent_size = params['Generator']['latent_size']
logger.log(str(G))
logger.log('Total nuber of parameters in Generator: {}'.format(
sum(map(lambda x: reduce(lambda a, b: a * b, x.size()),
G.parameters()))))
logger.log(str(D))
logger.log('Total nuber of parameters in Discriminator: {}'.format(
sum(map(lambda x: reduce(lambda a, b: a * b, x.size()),
D.parameters()))))
def get_dataloader(minibatch_size):
return DataLoader(dataset,
minibatch_size,
sampler=InfiniteRandomSampler(dataset),
num_workers=params['num_data_workers'],
pin_memory=False,
drop_last=True)
def rl(bs):
return lambda: random_latents(bs, latent_size)
# Setting up learning rate and optimizers
opt_g = Adam(G.parameters(), params['G_lr_max'], **params['Adam'])
opt_d = Adam(D.parameters(), params['D_lr_max'], **params['Adam'])
def rampup(cur_nimg):
if cur_nimg < params['lr_rampup_kimg'] * 1000:
p = max(0.0, 1 - cur_nimg / (params['lr_rampup_kimg'] * 1000))
return np.exp(-p * p * 5.0)
else:
return 1.0
lr_scheduler_d = LambdaLR(opt_d, rampup)
lr_scheduler_g = LambdaLR(opt_g, rampup)
mb_def = params['minibatch_size']
D_loss_fun = partial(wgan_gp_D_loss,
return_all=True,
iwass_lambda=params['iwass_lambda'],
iwass_epsilon=params['iwass_epsilon'],
iwass_target=params['iwass_target'])
G_loss_fun = wgan_gp_G_loss
trainer = Trainer(D, G, D_loss_fun, G_loss_fun, opt_d, opt_g, dataset,
iter(get_dataloader(mb_def)), rl(mb_def),
**params['Trainer'])
# plugins
if params['progressive_growing']:
max_depth = min(G.max_depth, D.max_depth)
trainer.register_plugin(
DepthManager(get_dataloader, rl, max_depth,
**params['DepthManager']))
for i, loss_name in enumerate(losses):
trainer.register_plugin(EfficientLossMonitor(i, loss_name))
checkpoints_dir = params['checkpoints_dir'] if params[
'checkpoints_dir'] else result_dir
trainer.register_plugin(
SaverPlugin(checkpoints_dir, **params['SaverPlugin']))
def subsitute_samples_path(d):
return {
k: (os.path.join(result_dir, v) if k == 'samples_path' else v)
for k, v in d.items()
}
postprocessors = [
globals()[x](**subsitute_samples_path(params[x]))
for x in params['postprocessors']
]
trainer.register_plugin(
OutputGenerator(lambda x: random_latents(x, latent_size),
postprocessors, **params['OutputGenerator']))
trainer.register_plugin(AbsoluteTimeMonitor(params['resume_time']))
trainer.register_plugin(LRScheduler(lr_scheduler_d, lr_scheduler_g))
trainer.register_plugin(logger)
init_comet(params, trainer)
trainer.run(params['total_kimg'])
dataset.close()
