class _3DGAN(object):
def __init__(self, args, config=config):
self.args = args
self.attribute = args.attribute
self.gpu = args.gpu
self.mode = args.mode
self.restore = args.restore
# init dataset and networks
self.config = config
self.dataset = ShapeNet(self.attribute)
self.G = Generator()
self.D = Discriminator()
self.adv_criterion = torch.nn.BCELoss()
self.set_mode_and_gpu()
self.restore_from_file()
def set_mode_and_gpu(self):
if self.mode == 'train':
self.G.train()
self.D.train()
if self.gpu:
with torch.cuda.device(self.gpu[0]):
self.G.cuda()
self.D.cuda()
self.adv_criterion.cuda()
if len(self.gpu) > 1:
self.G = torch.nn.DataParallel(self.G, device_ids=self.gpu)
self.D = torch.nn.DataParallel(self.D, device_ids=self.gpu)
elif self.mode == 'test':
self.G.eval()
self.D.eval()
if self.gpu:
with torch.cuda.device(self.gpu[0]):
self.G.cuda()
self.D.cuda()
if len(self.gpu) > 1:
self.G = torch.nn.DataParallel(self.G, device_ids=self.gpu)
self.D = torch.nn.DataParallel(self.D, device_ids=self.gpu)
else:
raise NotImplementationError()
def restore_from_file(self):
if self.restore is not None:
ckpt_file_G = os.path.join(
self.config.model_dir,
'G_iter_{:06d}.pth'.format(self.restore))
assert os.path.exists(ckpt_file_G)
self.G.load_state_dict(torch.load(ckpt_file_G))
if self.mode == 'train':
ckpt_file_D = os.path.join(
self.config.model_dir,
'D_iter_{:06d}.pth'.format(self.restore))
assert os.path.exists(ckpt_file_D)
self.D.load_state_dict(torch.load(ckpt_file_D))
self.start_step = self.restore + 1
else:
self.start_step = 1
def save_log(self):
scalar_info = {
'loss_D': self.loss_D,
'loss_G': self.loss_G,
'G_lr': self.G_lr_scheduler.get_lr()[0],
'D_lr': self.D_lr_scheduler.get_lr()[0],
}
for key, value in self.G_loss.items():
scalar_info['G_loss/' + key] = value
for key, value in self.D_loss.items():
scalar_info['D_loss/' + key] = value
for tag, value in scalar_info.items():
self.writer.add_scalar(tag, value, self.step)
def save_img(self, save_num=5):
for i in range(save_num):
mdict = {'instance': self.fake_X[i, 0].data.cpu().numpy()}
sio.savemat(
os.path.join(self.config.img_dir,
'{:06d}_{:02d}.mat'.format(self.step, i)), mdict)
def save_model(self):
torch.save(
{key: val.cpu()
for key, val in self.G.state_dict().items()},
os.path.join(self.config.model_dir,
'G_iter_{:06d}.pth'.format(self.step)))
torch.save(
{key: val.cpu()
for key, val in self.D.state_dict().items()},
os.path.join(self.config.model_dir,
'D_iter_{:06d}.pth'.format(self.step)))
def train(self):
self.writer = SummaryWriter(self.config.log_dir)
self.opt_G = torch.optim.Adam(self.G.parameters(),
lr=self.config.G_lr,
betas=(0.5, 0.999))
self.opt_D = torch.optim.Adam(self.D.parameters(),
lr=self.config.D_lr,
betas=(0.5, 0.999))
self.G_lr_scheduler = torch.optim.lr_scheduler.StepLR(
self.opt_G,
step_size=self.config.step_size,
gamma=self.config.gamma)
self.D_lr_scheduler = torch.optim.lr_scheduler.StepLR(
self.opt_D,
step_size=self.config.step_size,
gamma=self.config.gamma)
# start training
for step in range(self.start_step, 1 + self.config.max_iter):
self.step = step
self.G_lr_scheduler.step()
self.D_lr_scheduler.step()
self.real_X = next(self.dataset.gen(True))
self.noise = torch.randn(self.config.nchw[0], 200)
if len(self.gpu):
with torch.cuda.device(self.gpu[0]):
self.real_X = self.real_X.cuda()
self.noise = self.noise.cuda()
self.fake_X = self.G(self.noise)
# update D
self.D_real = self.D(self.real_X)
self.D_fake = self.D(self.fake_X.detach())
self.D_loss = {
'adv_real':
self.adv_criterion(self.D_real, torch.ones_like(self.D_real)),
'adv_fake':
self.adv_criterion(self.D_fake, torch.zeros_like(self.D_fake)),
}
self.loss_D = sum(self.D_loss.values())
self.opt_D.zero_grad()
self.loss_D.backward()
self.opt_D.step()
# update G
self.D_fake = self.D(self.fake_X)
self.G_loss = {
'adv_fake':
self.adv_criterion(self.D_fake, torch.ones_like(self.D_fake))
}
self.loss_G = sum(self.G_loss.values())
self.opt_G.zero_grad()
self.loss_G.backward()
self.opt_G.step()
print('step: {:06d}, loss_D: {:.6f}, loss_G: {:.6f}'.format(
self.step,
self.loss_D.data.cpu().numpy(),
self.loss_G.data.cpu().numpy()))
if self.step % 100 == 0:
self.save_log()
if self.step % 1000 == 0:
self.save_img()
self.save_model()
print('Finished training!')
self.writer.close()
class GAN_CLS(object):
def __init__(self, args, data_loader, SUPERVISED=True):
"""
Arguments :
----------
args : Arguments defined in Argument Parser
data_loader = An instance of class DataLoader for loading our dataset in batches
SUPERVISED :
"""
self.data_loader = data_loader
self.num_epochs = args.num_epochs
self.batch_size = args.batch_size
self.log_step = config.log_step
self.sample_step = config.sample_step
self.log_dir = args.log_dir
self.checkpoint_dir = args.checkpoint_dir
self.sample_dir = config.sample_dir
self.final_model = args.final_model
self.dataset = args.dataset
self.model_name = args.model_name
self.img_size = args.img_size
self.z_dim = args.z_dim
self.text_embed_dim = args.text_embed_dim
self.text_reduced_dim = args.text_reduced_dim
self.learning_rate = args.learning_rate
self.beta1 = args.beta1
self.beta2 = args.beta2
self.l1_coeff = args.l1_coeff
self.resume_epoch = args.resume_epoch
self.SUPERVISED = SUPERVISED
# Logger setting
self.logger = logging.getLogger('__name__')
self.logger.setLevel(logging.INFO)
self.formatter = logging.Formatter('%(asctime)s:%(levelname)s:%(message)s')
self.file_handler = logging.FileHandler(self.log_dir)
self.file_handler.setFormatter(self.formatter)
self.logger.addHandler(self.file_handler)
self.build_model()
def build_model(self):
""" A function of defining following instances :
----- Generator
----- Discriminator
----- Optimizer for Generator
----- Optimizer for Discriminator
----- Defining Loss functions
"""
# ---------------------------------------------------------------------
# 1. Network Initialization
# ---------------------------------------------------------------------
self.gen = Generator(batch_size=self.batch_size,
img_size=self, img_size,
z_dim=self.z_dim,
text_embed_dim=self.text_embed_dim,
text_reduced_dim=self.text_reduced_dim)
self.disc = Discriminator(batch_size=self.batch_size,
img_size=self, img_size,
text_embed_dim=self.text_embed_dim,
text_reduced_dim=self.text_reduced_dim)
self.gen_optim = optim.Adam(self.gen.parameters(),
lr=self.learning_rate,
betas=(self.beta1, self.beta2))
self.disc_optim = optim.Adam(self.disc.parameters(),
lr=self.learning_rate,
betas=(self.beta1, self.beta2))
self.cls_gan_optim = optim.Adam(itertools.chain(self.gen.parameters(),
self.disc.parameters()),
lr=self.learning_rate,
betas=(self.beta1, self.beta2))
print ('------------- Generator Model Info ---------------')
self.print_network(self.gen, 'G')
print ('------------------------------------------------')
print ('------------- Discriminator Model Info ---------------')
self.print_network(self.disc, 'D')
print ('------------------------------------------------')
self.gen.cuda()
self.disc.cuda()
self.criterion = nn.BCELoss().cuda()
# self.CE_loss = nn.CrossEntropyLoss().cuda()
# self.MSE_loss = nn.MSELoss().cuda()
self.gen.train()
self.disc.train()
def print_network(self, model, name):
""" A function for printing total number of model parameters """
num_params = 0
for p in model.parameters():
num_params += p.numel()
print(model)
print(name)
print("Total number of parameters: {}".format(num_params))
def load_checkpoints(self, resume_epoch):
"""Restore the trained generator and discriminator."""
print('Loading the trained models from step {}...'.format(resume_epoch))
G_path = os.path.join(self.checkpoint_dir, '{}-G.ckpt'.format(resume_epoch))
D_path = os.path.join(self.checkpoint_dir, '{}-D.ckpt'.format(resume_epoch))
self.gen.load_state_dict(torch.load(G_path, map_location=lambda storage, loc: storage))
self.disc.load_state_dict(torch.load(D_path, map_location=lambda storage, loc: storage))
def train_model(self):
data_loader = self.data_loader
start_epoch = 0
if self.resume_epoch:
start_epoch = self.resume_epoch
self.load_checkpoints(self.resume_epoch)
print ('--------------- Model Training Started ---------------')
start_time = time.time()
for epoch in range(start_epoch, self.num_epochs):
for idx, batch in enumerate(data_loader):
true_imgs = batch['true_imgs']
true_embed = batch['true_embed']
false_imgs = batch['false_imgs']
real_labels = torch.ones(true_imgs.size(0))
fake_labels = torch.zeros(true_imgs.size(0))
smooth_real_labels = torch.FloatTensor(Utils.smooth_label(real_labels.numpy(), -0.1))
true_imgs = Variable(true_imgs.float()).cuda()
true_embed = Variable(true_embed.float()).cuda()
false_imgs = Variable(false_imgs.float()).cuda()
real_labels = Variable(real_labels).cuda()
smooth_real_labels = Variable(smooth_real_labels).cuda()
fake_labels = Variable(fake_labels).cuda()
# ---------------------------------------------------------------
# 2. Training the generator
# ---------------------------------------------------------------
self.gen.zero_grad()
z = Variable(torch.randn(true_imgs.size(0), self.z_dim)).cuda()
fake_imgs = self.gen(true_embed, z)
fake_out, fake_logit = self.disc(fake_imgs, true_embed)
true_out, true_logit = self.disc(true_imgs, true_embed)
gen_loss = self.criterion(fake_out, real_labels) +
self.l1_coeff * nn.L1Loss(fake_imgs, true_imgs)
gen_loss.backward()
self.gen_optim.step()
# ---------------------------------------------------------------
# 3. Training the discriminator
# ---------------------------------------------------------------
self.disc.zero_grad()
false_out, false_logit = self.disc(false_imgs, true_embed)
disc_loss = self.criterion(true_out, smooth_real_labels) +
self.criterion(fake_out, fake_labels) + self.criterion(false_out, fake_labels)
disc_loss.backward()
self.disc_optim.step()
# self.cls_gan_optim.step()
# Logging
loss = {}
loss['G_loss'] = gen_loss.item()
loss['D_loss'] = disc_loss.item()
# ---------------------------------------------------------------
# 4. Logging INFO into log_dir
# ---------------------------------------------------------------
if (idx + 1) % self.log_step == 0:
end_time = time.time() - start_time
end_time = datetime.timedelta(seconds=end_time)
log = "Elapsed [{}], Epoch [{}/{}], Idx [{}]".format(end_time, epoch + 1,
self.num_epochs, idx)
for net, loss_value in loss.items():
log += ", {}: {:.4f}".format(net, loss_value)
self.logger.info(log)
print (log)
# ---------------------------------------------------------------
# 5. Saving generated images
# ---------------------------------------------------------------
if (idx + 1) % self.sample_step == 0:
concat_imgs = torch.cat((true_imgs, fake_imgs), 2) # ??????????
save_path = os.path.join(self.sample_dir, '{}-images.jpg'.format(idx + 1))
cocat_imgs = (cocat_imgs + 1) / 2
# out.clamp_(0, 1)
save_image(concat_imgs.data.cpu(), self.sample_dir, nrow=1, padding=0)
print ('Saved real and fake images into {}...'.format(self.sample_dir))
# ---------------------------------------------------------------
# 6. Saving the checkpoints & final model
# ---------------------------------------------------------------
if (idx + 1) % self.model_save_step == 0:
G_path = os.path.join(self.checkpoint_dir, '{}-G.ckpt'.format(idx + 1))
D_path = os.path.join(self.checkpoint_dir, '{}-D.ckpt'.format(idx + 1))
torch.save(self.gen.state_dict(), G_path)
torch.save(self.disc.state_dict(), D_path)
print('Saved model checkpoints into {}...'.format(self.checkpoint_dir))
else:
print('not implemented')
exit()
trn_dloader = torch.utils.data.DataLoader(dataset=trn_dataset, batch_size=14, shuffle=True)
val_dloader = torch.utils.data.DataLoader(dataset=val_dataset, batch_size=1, shuffle=False)
hmaps_ch, pmaps_ch = trn_dataset.num_channels()
# load networks
G = FSRNet(hmaps_ch, pmaps_ch)
G = nn.DataParallel(G)
G = G.cuda()
D = Discriminator(input_shape=(3, 128, 128))
D = nn.DataParallel(D)
D = D.cuda()
F = FeatureExtractor().cuda()
F.eval()
# settings
a = 1
b = 1
r_c = 1e-3
r_p = 1e-1
learning_rate = 2.5e-4
criterion_MSE = nn.MSELoss()
criterion_BCE = nn.BCELoss()
optimizer_G = optim.RMSprop(G.parameters(), lr=learning_rate)
optimizer_D = optim.RMSprop(D.parameters(), lr=learning_rate)
# nn.LeakyReLU(0.2, inplace=True),
# nn.Linear(1024, 784),
# nn.Tanh()
# )
# def forward(self, x):
# x = x.view(x.size(0), 100)
# out = self.model(x)
# return out
from nets import Generator, Discriminator
G = Generator((100, 500, 28 * 28), 'relu')
D = Discriminator((28 * 28, 500, 1), 'relu')
discriminator = D.cuda()
generator = G.cuda()
criterion = nn.BCELoss()
lr = 0.0002
d_optimizer = torch.optim.Adam(discriminator.parameters(), lr=lr)
g_optimizer = torch.optim.Adam(generator.parameters(), lr=lr)
def train_discriminator(discriminator, images, real_labels, fake_images,
fake_labels):
discriminator.zero_grad()
# real_outputs = discriminator(images.reshape(-1, 28*28))
# real_loss = criterion(outputs, real_labels)
# real_score = real_outputs