def _init_network(self):
def init_weights(m):
if type(m) == nn.Conv2d:
nn.init.normal_(m.weight.data, 0.0, 0.02)
if type(m) == nn.BatchNorm2d:
nn.init.normal_(m.weight.data, 0.0, 0.02)
nn.init.constant_(m.bias.data, 0.0)
generator = Generator(DCGAN_Config).to(self.device)
discriminator = Discriminator(DCGAN_Config).to(self.device)
if self.device.type =='cuda' and DCGAN_Config['ngpu']>1:
generator = nn.DataParallel(generator, list(range(DCGAN_Config['ngpu'])))
discriminator = nn.DataParallel(discriminator, list(range(DCGAN_Config['ngpu'])))
generator.apply(init_weights)
discriminator.apply(init_weights)
print(generator)
print(discriminator)
return generator, discriminator
# img=dataset.__getitem__(10)
# print(img['A'].size())
netG_A2B = Generator(opt.input_nc, opt.output_nc, opt.ngf)
netG_B2A = Generator(opt.input_nc, opt.output_nc, opt.ngf)
netD_A = Discriminator(opt.input_nc, opt.ndf)
netD_B = Discriminator(opt.input_nc, opt.ndf)
summary(netG_A2B, input_size=(3, 256, 256), device='cpu')
summary(netD_A, input_size=(3, 256, 256), device='cpu')
#instialize weights
netG_A2B.apply(weights_init)
netG_B2A.apply(weights_init)
netD_A.apply(weights_init)
netD_B.apply(weights_init)
#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(),
classname = m.__class__.__name__
if classname.find('Conv') != -1:
m.weight.data.normal_(0.0, 0.02)
elif classname.find('BatchNorm') != -1:
m.weight.data.normal_(1.0, 0.02)
m.bias.data.fill_(0)
NetG = Decoder(nc, ngf, nz).to(device)
NetD = Discriminator(imageSize, nc, ndf, nz).to(device)
NetE = Encoder(imageSize, nc, ngf, nz).to(device)
Sampler = Sampler().to(device)
NetE.apply(weights_init)
NetG.apply(weights_init)
NetD.apply(weights_init)
# load weights
if opt.netE != '':
NetE.load_state_dict(torch.load(opt.netE))
if opt.netG != '':
NetG.load_state_dict(torch.load(opt.netG))
if opt.netD != '':
NetD.load_state_dict(torch.load(opt.netD))
optimizer_encorder = optim.RMSprop(params=NetE.parameters(),
lr=lr,
alpha=0.9,
eps=1e-8,
weight_decay=0,
momentum=0,
D_A = Discriminator().to(device)
D_B = Discriminator().to(device)
def weights_init(m):
if isinstance(m, nn.Conv2d):
torch.nn.init.normal_(m.weight.data, 0.0, 0.02)
if opt.load_pretrained:
G_A2B.load_state_dict(torch.load('pretrained/G_A2B.pth'))
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)
#BUILDING NETWORK STRUCTURE
generator = Generator(128)
discriminator = Discriminator(128)
generator.cuda()
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)
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)
# Decide which device we want to run on
device = torch.device("cuda:0" if (torch.cuda.is_available() and args.ngpu > 0) else "cpu")
# Create the generator
netG = Generator(args.ngpu, args.nc, args.nz, args.ndf, args.ngf).to(device)
if (device.type == 'cuda') and (args.ngpu > 1):
netG = nn.DataParallel(netG, list(range(args.ngpu)))
netG.apply(weights_init)
print(netG)
# Create the Discriminator
netD = Discriminator(args.ngpu, args.nc, args.nz, args.ndf, args.ngf).to(device)
if (device.type == 'cuda') and (args.ngpu > 1):
netD = nn.DataParallel(netD, list(range(args.ngpu)))
netD.apply(weights_init)
print(netD)
# 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
from torch.utils.data import DataLoader
import torchvision.transforms as transforms
from utils import imshow_grid, mse_loss, reparameterize, l1_loss
from network import Generator, Discriminator
from torchvision.utils import save_image
batch_size = 8
num_epochs = 500
image_size = 128
generator = Generator(nc_dim=80)
generator.apply(weights_init)
generator = generator.cuda()
discriminator = Discriminator()
discriminator.apply(weights_init)
discriminator = discriminator.cuda()
ones_label = torch.ones(batch_size)
ones_label = Variable(ones_label.cuda())
zeros_label = torch.zeros(batch_size)
zeros_label = Variable(zeros_label.cuda())
loss = nn.BCEWithLogitsLoss()
def to_img(x):
x = x.clamp(0, 1)
x = x.view(x.size(0), 3, 128, 128)
return x
parser = argparse.ArgumentParser()
parser.add_argument('--model_name', default='test')
parser.add_argument('--batch_size', default=64)
parser.add_argument('--lr', type=float, default=5e-3)
parser.add_argument('--l1', type=float, default=1e1)
parser.add_argument('--num_iters', type=int, default=100)
parser.add_argument("--train", dest="train", default=False, action="store_true") # noqa
parser.add_argument('--test_save_path', default='test')
args = parser.parse_args()
train_loader,val_loader,test_loader = read_data(args.batch_size)
network = Network()
network.apply(weights_init)
d = Discriminator()
d.apply(weights_init)
bce_loss = torch.nn.BCEWithLogitsLoss()
true_crit, fake_crit = torch.ones(args.batch_size, 1, device='cuda'), torch.zeros(args.batch_size, 1, device='cuda')
print(network)
train_loss = []
val_loss = []
if args.train:
optimizer = torch.optim.Adam(network.parameters(), lr=args.lr)
d_optimizer = torch.optim.Adam(d.parameters(), lr=args.lr)
network.cuda()
d.cuda()
for epoch in range(args.num_iters):
train_loader,val_loader,test_loader = read_data(args.batch_size)
network.train()
for idx, x in enumerate(train_loader):