shuffle=True)
############### MODEL ####################
ndf = opt.ndf
ngf = opt.ngf
nc = 1
netD = Discriminator(nc, ndf)
netG = Generator(nc, ngf, opt.nz)
#if(opt.cuda):
netD.cuda()
netG.cuda()
########### LOSS & OPTIMIZER ##########
criterion = nn.BCELoss()
optimizerD = torch.optim.Adam(netD.parameters(),
lr=opt.lr,
betas=(opt.beta1, 0.999))
optimizerG = torch.optim.Adam(netG.parameters(),
lr=opt.lr,
betas=(opt.beta1, 0.999))
########## GLOBAL VARIABLES ###########
#noise_all = torch.FloatTensor(20,opt.nz,1,1)
noise = torch.FloatTensor(opt.batchSize, opt.nz, 1, 1)
real = torch.FloatTensor(opt.batchSize, nc, opt.imageSize, opt.imageSize)
label = torch.FloatTensor(opt.batchSize)
real_label = 1
fake_label = 0
#noise_all = Variable(noise_all)
noise = Variable(noise)
def main(args):
#===========================================================================
# Set the file name format
FILE_NAME_FORMAT = "{0}_{1}_{2:d}_{3:d}_{4:d}_{5:f}{6}".format(
args.model, args.dataset, args.epochs, args.obj_step, args.batch_size,
args.lr, args.flag)
# Set the results file path
RESULT_FILE_NAME = FILE_NAME_FORMAT + '_results.pkl'
RESULT_FILE_PATH = os.path.join(RESULTS_PATH, RESULT_FILE_NAME)
# Set the checkpoint file path
CHECKPOINT_FILE_NAME = FILE_NAME_FORMAT + '.ckpt'
CHECKPOINT_FILE_PATH = os.path.join(CHECKPOINT_PATH, CHECKPOINT_FILE_NAME)
BEST_CHECKPOINT_FILE_NAME = FILE_NAME_FORMAT + '_best.ckpt'
BEST_CHECKPOINT_FILE_PATH = os.path.join(CHECKPOINT_PATH,
BEST_CHECKPOINT_FILE_NAME)
# Set the random seed same for reproducibility
random.seed(190811)
torch.manual_seed(190811)
torch.cuda.manual_seed_all(190811)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Step1 ====================================================================
# Load dataset
if args.dataset == 'CelebA':
dataloader = CelebA_Dataloader()
else:
assert False, "Please select the proper dataset."
train_loader = dataloader.get_train_loader(batch_size=args.batch_size,
num_workers=args.num_workers)
print('==> DataLoader ready.')
# Step2 ====================================================================
# Make the model
if args.model in ['WGAN', 'DCGAN']:
generator = Generator(BN=True)
discriminator = Discriminator(BN=True)
elif args.model in ['WGAN_noBN', 'DCGAN_noBN']:
generator = Generator(BN=False)
discriminator = Discriminator(BN=False)
else:
assert False, "Please select the proper model."
# Check DataParallel available
if torch.cuda.device_count() > 1:
generator = nn.DataParallel(generator)
discriminator = nn.DataParallel(discriminator)
# Check CUDA available
if torch.cuda.is_available():
generator.cuda()
discriminator.cuda()
print('==> Model ready.')
# Step3 ====================================================================
# Set loss function and optimizer
if args.model in ['DCGAN', 'DCGAN_noBN']:
criterion = nn.BCELoss()
else:
criterion = None
optimizer_G = torch.optim.RMSprop(generator.parameters(), lr=args.lr)
optimizer_D = torch.optim.RMSprop(discriminator.parameters(), lr=args.lr)
step_counter = StepCounter(args.obj_step)
print('==> Criterion and optimizer ready.')
# Step4 ====================================================================
# Train and validate the model
start_epoch = 0
best_metric = float("inf")
validate_noise = torch.randn(args.batch_size, 100, 1, 1)
# Initialize the result lists
train_loss_G = []
train_loss_D = []
train_distance = []
if args.resume:
assert os.path.exists(CHECKPOINT_FILE_PATH), 'No checkpoint file!'
checkpoint = torch.load(CHECKPOINT_FILE_PATH)
generator.load_state_dict(checkpoint['generator_state_dict'])
discriminator.load_state_dict(checkpoint['discriminator_state_dict'])
optimizer_G.load_state_dict(checkpoint['optimizer_G_state_dict'])
optimizer_D.load_state_dict(checkpoint['optimizer_D_state_dict'])
start_epoch = checkpoint['epoch']
step_counter.current_step = checkpoint['current_step']
train_loss_G = checkpoint['train_loss_G']
train_loss_D = checkpoint['train_loss_D']
train_distance = checkpoint['train_distance']
best_metric = checkpoint['best_metric']
# Save the training information
result_data = {}
result_data['model'] = args.model
result_data['dataset'] = args.dataset
result_data['target_epoch'] = args.epochs
result_data['batch_size'] = args.batch_size
# Check the directory of the file path
if not os.path.exists(os.path.dirname(RESULT_FILE_PATH)):
os.makedirs(os.path.dirname(RESULT_FILE_PATH))
if not os.path.exists(os.path.dirname(CHECKPOINT_FILE_PATH)):
os.makedirs(os.path.dirname(CHECKPOINT_FILE_PATH))
print('==> Train ready.')
# Validate before training (step 0)
val(generator, validate_noise, step_counter, FILE_NAME_FORMAT)
for epoch in range(args.epochs):
# strat after the checkpoint epoch
if epoch < start_epoch:
continue
print("\n[Epoch: {:3d}/{:3d}]".format(epoch + 1, args.epochs))
epoch_time = time.time()
#=======================================================================
# train the model (+ validate the model)
tloss_G, tloss_D, tdist = train(generator, discriminator, train_loader,
criterion, optimizer_G, optimizer_D,
args.clipping, args.num_critic,
step_counter, validate_noise,
FILE_NAME_FORMAT)
train_loss_G.extend(tloss_G)
train_loss_D.extend(tloss_D)
train_distance.extend(tdist)
#=======================================================================
current = time.time()
# Calculate average loss
avg_loss_G = sum(tloss_G) / len(tloss_G)
avg_loss_D = sum(tloss_D) / len(tloss_D)
avg_distance = sum(tdist) / len(tdist)
# Save the current result
result_data['current_epoch'] = epoch
result_data['train_loss_G'] = train_loss_G
result_data['train_loss_D'] = train_loss_D
result_data['train_distance'] = train_distance
# Save result_data as pkl file
with open(RESULT_FILE_PATH, 'wb') as pkl_file:
pickle.dump(result_data,
pkl_file,
protocol=pickle.HIGHEST_PROTOCOL)
# Save the best checkpoint
# if avg_distance < best_metric:
# best_metric = avg_distance
# torch.save({
# 'epoch': epoch+1,
# 'generator_state_dict': generator.state_dict(),
# 'discriminator_state_dict': discriminator.state_dict(),
# 'optimizer_G_state_dict': optimizer_G.state_dict(),
# 'optimizer_D_state_dict': optimizer_D.state_dict(),
# 'current_step': step_counter.current_step,
# 'best_metric': best_metric,
# }, BEST_CHECKPOINT_FILE_PATH)
# Save the current checkpoint
torch.save(
{
'epoch': epoch + 1,
'generator_state_dict': generator.state_dict(),
'discriminator_state_dict': discriminator.state_dict(),
'optimizer_G_state_dict': optimizer_G.state_dict(),
'optimizer_D_state_dict': optimizer_D.state_dict(),
'current_step': step_counter.current_step,
'train_loss_G': train_loss_G,
'train_loss_D': train_loss_D,
'train_distance': train_distance,
'best_metric': best_metric,
}, CHECKPOINT_FILE_PATH)
# Print the information on the console
print("model : {}".format(args.model))
print("dataset : {}".format(args.dataset))
print("batch_size : {}".format(args.batch_size))
print("current step : {:d}".format(step_counter.current_step))
print("current lrate : {:f}".format(args.lr))
print("gen/disc loss : {:f}/{:f}".format(
avg_loss_G, avg_loss_D))
print("distance metric : {:f}".format(avg_distance))
print("epoch time : {0:.3f} sec".format(current -
epoch_time))
print("Current elapsed time : {0:.3f} sec".format(current - start))
# If iteration step has been satisfied
if step_counter.exit_signal:
break
print('==> Train done.')
print(' '.join(['Results have been saved at', RESULT_FILE_PATH]))
print(' '.join(['Checkpoints have been saved at', CHECKPOINT_FILE_PATH]))
model.load_state_dict(torch.load(load_path))
print('model loaded')
torch.cuda.empty_cache()
model = model.cuda()
discriminator = discriminator.cuda()
vgg_model = models.vgg16(pretrained=True)
vgg_model.cuda()
loss_network = utils.LossNetwork(vgg_model)
loss_network.eval()
# Optimizers & LR schedulers
optimizer_G = optim.Adam(model.parameters(), lr=opt.lr, betas=(0.5, 0.999))
optimizer_D = optim.Adam(discriminator.parameters(),
lr=opt.lr,
betas=(0.5, 0.999))
# Dataset loader
trainloader = loaddata.getTrainingData(opt.batchSize, size=128)
testloader = loaddata.getTestingData(1, size=128)
criteria_recon_l2 = nn.MSELoss()
criteria_recon_l1 = nn.L1Loss()
###################################
###### Training ######
with open('./output_' + opt.output_str + '/loss.csv', 'w',
newline='') as csvfile:
from model.Discriminator import Discriminator
from model.Generator import Generator
from loaders.MNISTLoader import MNIST
from util.ImageUtil import ImageUtil
# create model objects
discriminator = Discriminator()
generator = Generator()
# set data loader
dataLoader = MNIST()
train_loader, test_loader = dataLoader.train_loader, dataLoader.test_loader
# optimizer
D_optimizer = Adam(params=discriminator.parameters(), lr=0.001)
G_optimizer = Adam(params=generator.parameters(), lr=0.001)
# loss function
D_loss_function = nn.BCELoss() # Binary Cross Entropy loss
G_loss_function = nn.BCELoss() # Binary Cross Entropy loss
imageUtil = ImageUtil()
epoch_size = 10000
for epoch in range(epoch_size):
for i, data in enumerate(train_loader):
real_data, real_label = data
real_data, real_label = Variable(real_data), Variable(real_label)
if torch.cuda.is_available():
G_BA.apply(weights_init)
D_A.apply(weights_init)
D_B.apply(weights_init)
if(cuda):
D_A.cuda()
D_B.cuda()
G_AB.cuda()
G_BA.cuda()
########### LOSS & OPTIMIZER ##########
criterionMSE = nn.L1Loss()
criterion = nn.MSELoss()
# chain is used to update two generators simultaneously
optimizerD_A = torch.optim.Adam(D_A.parameters(),lr=0.0002, betas=(0.5, 0.999), weight_decay=1e-4)
optimizerD_B = torch.optim.Adam(D_B.parameters(),lr=0.0002, betas=(0.5, 0.999), weight_decay=1e-4)
optimizerG = torch.optim.Adam(chain(G_AB.parameters(),G_BA.parameters()),lr=0.0002, betas=(0.5, 0.999))
real_A = torch.FloatTensor(batchSize, input_nc, fineSize, fineSize)
AB = torch.FloatTensor(batchSize, input_nc, fineSize, fineSize)
real_B = torch.FloatTensor(batchSize, output_nc, fineSize, fineSize)
BA = torch.FloatTensor(batchSize, output_nc, fineSize, fineSize)
label = torch.FloatTensor(batchSize)
if(cuda):
real_A = real_A.cuda()
real_B = real_B.cuda()
label = label.cuda()
AB = AB.cuda()
G_BA.apply(weights_init)
if (opt.cuda):
D_A.cuda()
D_B.cuda()
G_AB.cuda()
G_BA.cuda()
D_A.apply(weights_init)
D_B.apply(weights_init)
########### LOSS & OPTIMIZER ##########
criterionMSE = nn.MSELoss()
criterion = nn.BCELoss()
# chain is used to update two generators simultaneously
optimizerD = torch.optim.Adam(chain(D_A.parameters(), D_B.parameters()),
lr=opt.lr,
betas=(opt.beta1, 0.999),
weight_decay=opt.weight_decay)
optimizerG = torch.optim.Adam(chain(G_AB.parameters(), G_BA.parameters()),
lr=opt.lr,
betas=(opt.beta1, 0.999))
########### GLOBAL VARIABLES ###########
input_nc = opt.input_nc
output_nc = opt.output_nc
fineSize = opt.fineSize
real_A = torch.FloatTensor(opt.batchSize, input_nc, fineSize, fineSize)
real_B = torch.FloatTensor(opt.batchSize, output_nc, fineSize, fineSize)
label = torch.FloatTensor(opt.batchSize)
D_A.cuda()
D_B.cuda()
G_AB.cuda()
G_BA.cuda()
D_A.apply(weights_init)
D_B.apply(weights_init)
########### LOSS & OPTIMIZER ##########
criterionMSE = nn.L1Loss()
if (opt.loss_type == 'bce'):
criterion = nn.BCELoss()
else:
criterion = nn.MSELoss()
# chain is used to update two generators simultaneously
optimizerD_A = torch.optim.Adam(D_A.parameters(),
lr=opt.lr,
betas=(opt.beta1, 0.999),
weight_decay=opt.weight_decay)
optimizerD_B = torch.optim.Adam(D_B.parameters(),
lr=opt.lr,
betas=(opt.beta1, 0.999),
weight_decay=opt.weight_decay)
optimizerG = torch.optim.Adam(chain(G_AB.parameters(), G_BA.parameters()),
lr=opt.lr,
betas=(opt.beta1, 0.999))
########### GLOBAL VARIABLES ###########
input_nc = opt.input_nc
output_nc = opt.output_nc
fineSize = opt.fineSize
D_A.cuda()
D_B.cuda()
G_AB.cuda()
G_BA.cuda()
D_A.apply(weights_init)
D_B.apply(weights_init)
########### LOSS & OPTIMIZER ##########
criterionMSE = nn.L1Loss()
if (opt.loss_type == 'bce'):
criterion = nn.BCELoss()
else:
criterion = nn.MSELoss()
# chain is used to update two generators simultaneously
optimizerD_A = torch.optim.Adam(D_A.parameters(),
lr=opt.lr,
betas=(opt.beta1, 0.999),
weight_decay=opt.weight_decay)
optimizerD_B = torch.optim.Adam(D_B.parameters(),
lr=opt.lr,
betas=(opt.beta1, 0.999),
weight_decay=opt.weight_decay)
optimizerG = torch.optim.Adam(chain(G_AB.parameters(), G_BA.parameters()),
lr=opt.lr,
betas=(opt.beta1, 0.999))
########### GLOBAL VARIABLES ###########
input_nc = opt.input_nc
output_nc = opt.output_nc
fineSize = opt.fineSize