def main():
# init random seed
init_random_seed(params.manual_seed)
# Load dataset
mnist_data_loader = get_mnist(train=True, download=True)
mnist_data_loader_eval = get_mnist(train=False, download=True)
usps_data_loader = get_usps(train=True, download=True)
usps_data_loader_eval = get_usps(train=False, download=True)
# Model init Revgard
tgt_encoder = model_init(Encoder(), params.tgt_encoder_revgrad_path)
critic = model_init(Discriminator(), params.disc_revgard_path)
clf = model_init(Classifier(), params.clf_revgrad_path)
# Train models
print("====== Training source encoder and classifier in MNIST and USPS domains ======")
if not (tgt_encoder.pretrained and clf.pretrained and critic.pretrained and params.model_trained):
tgt_encoder, clf, critic = train_revgrad(tgt_encoder, clf, critic,
mnist_data_loader, usps_data_loader, robust=False)
# Eval target encoder on test set of target dataset
print("====== Evaluating classifier for encoded MNIST and USPS domain ======")
print("-------- MNIST domain --------")
eval_tgt(tgt_encoder, clf, mnist_data_loader_eval)
print("-------- USPS adaption --------")
eval_tgt(tgt_encoder, clf, usps_data_loader_eval)
def main():
# init random seed
init_random_seed(params.manual_seed)
# Load dataset
svhn_data_loader = get_svhn(split='train', download=True)
svhn_data_loader_eval = get_svhn(split='test', download=True)
mnist_data_loader = get_mnist(train=True, download=True)
mnist_data_loader_eval = get_mnist(train=False, download=True)
# Model init WDGRL
tgt_encoder = model_init(Encoder(), params.encoder_wdgrl_path)
critic = model_init(Discriminator(in_dims=params.d_in_dims,
h_dims=params.d_h_dims,
out_dims=params.d_out_dims),
params.disc_wdgrl_path)
clf = model_init(Classifier(), params.clf_wdgrl_path)
# Train critic to optimality
print("====== Training critic ======")
if not (critic.pretrained and params.model_trained):
critic = train_critic_wdgrl(tgt_encoder, critic, svhn_data_loader, mnist_data_loader)
# Train target encoder
print("====== Training encoder for both SVHN and MNIST domains ======")
if not (tgt_encoder.pretrained and clf.pretrained and params.model_trained):
tgt_encoder, clf = train_tgt_wdgrl(tgt_encoder, clf, critic,
svhn_data_loader, mnist_data_loader, robust=False)
# Eval target encoder on test set of target dataset
print("====== Evaluating classifier for encoded SVHN and MNIST domains ======")
print("-------- SVHN domain --------")
eval_tgt(tgt_encoder, clf, svhn_data_loader_eval)
print("-------- MNIST adaption --------")
eval_tgt(tgt_encoder, clf, mnist_data_loader_eval)
def main():
# init random seed
init_random_seed(params.manual_seed)
# Load dataset
mnist_data_loader = get_mnist(train=True, download=True)
mnist_data_loader_eval = get_mnist(train=False, download=True)
usps_data_loader = get_usps(train=True, download=True)
usps_data_loader_eval = get_usps(train=False, download=True)
# Model init WDGRL
tgt_encoder = model_init(Encoder(), params.encoder_wdgrl_rb_path)
critic = model_init(Discriminator(), params.disc_wdgrl_rb_path)
clf = model_init(Classifier(), params.clf_wdgrl_rb_path)
# Train target encoder
print("====== Robust Training encoder for both MNIST and USPS domains ======")
if not (tgt_encoder.pretrained and clf.pretrained and params.model_trained):
tgt_encoder, clf = train_tgt_wdgrl(tgt_encoder, clf, critic,
mnist_data_loader, usps_data_loader, usps_data_loader_eval, robust=True)
# Eval target encoder on test set of target dataset
print("====== Evaluating classifier for encoded MNIST and USPS domains ======")
print("-------- MNIST domain --------")
eval_tgt_robust(tgt_encoder, clf, mnist_data_loader_eval)
print("-------- USPS adaption --------")
eval_tgt_robust(tgt_encoder, clf, usps_data_loader_eval)
def test(model_config):
mode = 'test'
batch_size = 1
dataset = ShakespeareModern(train_shakespeare_path, test_shakespeare_path, train_modern_path, test_modern_path, mode=mode)
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=False)
shakespeare_disc = Discriminator(model_config['embedding_size'], model_config['hidden_dim'], len(dataset.vocab)).cuda()
shakespeare_disc.load_state_dict(torch.load('./shakespeare_disc.pth'))
shakespeare_disc.eval()
num_correct = 0
total_samples = 0
for idx, (s, s_addn_feats, m, m_addn_feats) in tqdm(enumerate(dataloader)):
s = s.transpose(0, 1)
m = m.transpose(0, 1)
total_samples += 2
s = Variable(s).cuda()
s_output = shakespeare_disc(s, s_addn_feats)
if round(s_output.item()) == 1.0:
num_correct += 1
m = Variable(m).cuda()
m_output = shakespeare_disc(m, m_addn_feats)
if round(m_output.item()) == 0.0:
num_correct += 1
print ('Accuracy: {}'.format(num_correct/total_samples))
def main():
# init random seed
init_random_seed(params.manual_seed)
# Load dataset
mnist_data_loader = get_usps(train=True, download=True)
mnist_data_loader_eval = get_usps(train=False, download=True)
usps_data_loader = get_usps(train=True, download=True)
usps_data_loader_eval = get_usps(train=False, download=True)
# Model init ADDA
src_encoder = model_init(Encoder(), params.src_encoder_adda_rb_path)
tgt_encoder = model_init(Encoder(), params.tgt_encoder_adda_rb_path)
critic = model_init(Discriminator(), params.disc_adda_rb_path)
clf = model_init(Classifier(), params.clf_adda_rb_path)
# Train source model for adda
print(
"====== Robust training source encoder and classifier in MNIST domain ======"
)
if not (src_encoder.pretrained and clf.pretrained
and params.model_trained):
src_encoder, clf = train_src_robust(src_encoder, clf,
mnist_data_loader)
# Eval source model
print("====== Evaluating classifier for MNIST domain ======")
eval_tgt(src_encoder, clf, mnist_data_loader_eval)
# Train target encoder
print("====== Robust training encoder for USPS domain ======")
# Initialize target encoder's weights with those of the source encoder
if not tgt_encoder.pretrained:
tgt_encoder.load_state_dict(src_encoder.state_dict())
if not (tgt_encoder.pretrained and critic.pretrained
and params.model_trained):
tgt_encoder = train_tgt_adda(src_encoder,
tgt_encoder,
clf,
critic,
mnist_data_loader,
usps_data_loader,
usps_data_loader_eval,
robust=True)
# Eval target encoder on test set of target dataset
print("====== Ealuating classifier for encoded USPS domain ======")
print("-------- Source only --------")
eval_tgt(src_encoder, clf, usps_data_loader_eval)
print("-------- Domain adaption --------")
eval_tgt(tgt_encoder, clf, usps_data_loader_eval)
def main():
opt = get_opt()
print(opt)
print("Start to train stage: %s" % (opt.stage))
# create dataset
if opt.stage == "Shape":
dataset = PolyDatasetShape(128)
train_loader = DataLoader(dataset,
batch_size=opt.b,
shuffle=False,
num_workers=opt.j,
drop_last=True,
pin_memory=True)
elif opt.stage == "Stitch":
dataset = PolyDatasetStitch(128)
train_loader = DataLoader(dataset,
batch_size=opt.b,
shuffle=False,
num_workers=opt.j,
drop_last=True,
pin_memory=True)
elif opt.stage == "Refine":
dataset = PolyDatasetRefine(128)
train_loader = DataLoader(dataset,
batch_size=opt.b,
shuffle=False,
num_workers=opt.j,
drop_last=True,
pin_memory=True)
else:
sys.exit("Please mention the Stage from [Shape, Stitch, Refine]")
if not os.path.exists(opt.results):
os.makedirs(opt.results)
netG = GeneratorCoarse(opt.input_channel, 3)
netD = Discriminator()
# create model & train & save the final checkpoint
netG.cuda()
netD.cuda()
netG.apply(weights_init_normal)
netD.apply(weights_init_normal)
train(opt, train_loader, netG, netD)
print('Finished training %s!' % (opt.stage))
def main():
# init random seed
init_random_seed(params.manual_seed)
# Load dataset
svhn_data_loader = get_svhn(split='train', download=True)
svhn_data_loader_eval = get_svhn(split='test', download=True)
mnist_data_loader = get_mnist(train=True, download=True)
mnist_data_loader_eval = get_mnist(train=False, download=True)
# Model init DANN
tgt_encoder = model_init(Encoder(), params.tgt_encoder_dann_rb_path)
critic = model_init(
Discriminator(in_dims=params.d_in_dims,
h_dims=params.d_h_dims,
out_dims=params.d_out_dims), params.disc_dann_rb_path)
clf = model_init(Classifier(), params.clf_dann_rb_path)
# Train models
print(
"====== Training source encoder and classifier in SVHN and MNIST domains ======"
)
if not (tgt_encoder.pretrained and clf.pretrained and critic.pretrained
and params.model_trained):
tgt_encoder, clf, critic = train_dann(tgt_encoder,
clf,
critic,
svhn_data_loader,
mnist_data_loader,
mnist_data_loader_eval,
robust=True)
# Eval target encoder on test set of target dataset
print(
"====== Evaluating classifier for encoded SVHN and MNIST domains ======"
)
print("-------- SVHN domain --------")
eval_tgt_robust(tgt_encoder, clf, svhn_data_loader_eval)
print("-------- MNIST adaption --------")
eval_tgt_robust(tgt_encoder, clf, mnist_data_loader_eval)
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
quant = True if 'GS8' in args.base_model_str else False
subnet_model_path = os.path.join('subnet_structures', args.dataset, args.task,
args.base_model_str, 'pth')
## Networks
# G:
dim_lst_path = os.path.join(subnet_model_path, 'epoch%d_netG.npy' % 199)
netG = Generator(args.input_nc,
args.output_nc,
dim_lst=np.load(dim_lst_path),
quant=quant).cuda()
# D:
netD = Discriminator(args.input_nc).cuda()
## results_dir:
optimizer_str = 'adam_lr%s_wd%s' % (args.lr, args.wd)
loss_str = 'beta%s_%s' % (args.beta, args.lc)
results_dir = os.path.join('finetune_results', args.dataset, args.task,
args.base_model_str,
'%s_%s' % (optimizer_str, loss_str))
img_dir = os.path.join(results_dir, 'img')
pth_dir = os.path.join(results_dir, 'pth')
create_dir(img_dir), create_dir(pth_dir)
# Optimizers
optimizer_G = torch.optim.Adam(netG.parameters(),
lr=args.lr,
weight_decay=args.wd,
parser.add_argument("--lambda_gp",
type=int,
default=10,
help="lambda for gradient penalty")
opt = parser.parse_args()
exp_folder = "{}_{}".format(opt.exp_folder, opt.target_set)
os.makedirs("./exps/" + exp_folder, exist_ok=True)
# Loss function
adversarial_loss = torch.nn.BCEWithLogitsLoss()
distance_loss = torch.nn.L1Loss()
# Initialize generator and discriminator
generator = Generator()
discriminator = Discriminator()
if torch.cuda.is_available():
device = torch.device('cuda:0')
generator.to(device)
discriminator.to(device)
adversarial_loss.to(device)
# Visualize a single batch
def visualizeSingleBatch(fp_loader_test,
opt,
exp_folder,
batches_done,
batch_size=8):
print('Loading saved model ... \n{}'.format(
'./checkpoints/{}_{}.pth'.format(exp_folder, batches_done)))
try:
model.trans.load_state_dict(torch.load(sys.argv[2]))
model.atmos.load_state_dict(torch.load(sys.argv[3]))
'''
for param in model.trans.parameters():
param.requires_grad = False
for param in model.atmos.parameters():
param.requires_grad = False
'''
except Exception as e:
try:
model.load_state_dict(torch.load(sys.argv[2]))
except Exception as e:
print("No weights. Training from scratch.")
if MODE == 'GAN':
model_d = Discriminator().to(device)
optimizer_d = torch.optim.Adam(model_d.parameters(), lr=learning_rate)
try:
model_d.load_state_dict(torch.load(sys.argv[3]))
if opt['parallel']:
model_d = nn.DataParallel(model_d)
except Exception as e:
print("No weights. Training from scratch discrim.")
else:
print('MODE INCORRECT : TRANS or ATMOS or FAST or DUAL or GAN')
exit()
# Wrap in Data Parallel for multi-GPU use
if opt['parallel']:
model = nn.DataParallel(model)
iter_number = (int)(INPUT_SIZE / BATCH_SIZE) + 1
image_root = utils.download_images()
train_ds = utils.load_data(image_root)
train_ds = utils.prepare_train_ds(train_ds,
BATCH_SIZE,
BUFFER_SIZE,
image_size=128)
# Guild generator and discriminator model.
generator_net = Generator(dtype=flags.FLAGS.dtype)
generator_optimizer = tf.train.AdamOptimizer(
learning_rate=flags.FLAGS.learning_rate_generator,
beta1=flags.FLAGS.beta1,
beta2=flags.FLAGS.beta2)
discriminator_net = Discriminator(alpha=flags.FLAGS.alpha,
dtype=flags.FLAGS.dtype)
discriminator_optimizer = tf.train.AdamOptimizer(
learning_rate=flags.FLAGS.learning_rate_discriminator,
beta1=flags.FLAGS.beta1,
beta2=flags.FLAGS.beta2)
# Print the network structure to show that the model is well built.
generator_net.build(input_shape=(None, 128))
discriminator_net.build(input_shape=(None, 128, 128, 3))
generator_net.summary()
discriminator_net.summary()
# Save model parameters and tensorboard information the adversarial network
basepath = "./mnist/" + str(flags.FLAGS.model_id)
logdir = os.path.join(basepath, "logs")
opt = parser.parse_args()
print(opt)
random.seed(opt.seed)
torch.manual_seed(opt.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(opt.seed)
# Networks
if opt.upsample == 'ori':
netG_A2B = Generator_ori(opt.input_nc, opt.output_nc)
netG_B2A = Generator_ori(opt.output_nc, opt.input_nc)
else:
netG_A2B = Generator(opt.input_nc, opt.output_nc)
netG_B2A = Generator(opt.output_nc, opt.input_nc)
netD_A = Discriminator(opt.input_nc)
netD_B = Discriminator(opt.output_nc)
netG_A2B.cuda()
netG_B2A.cuda()
netD_A.cuda()
netD_B.cuda()
netG_A2B.apply(weights_init_normal)
netG_B2A.apply(weights_init_normal)
netD_A.apply(weights_init_normal)
netD_B.apply(weights_init_normal)
torch.save(netG_A2B.state_dict(),
"initial_weights/netG_A2B_seed_{}.pth.tar".format(opt.seed))
torch.save(netG_B2A.state_dict(),
print(args)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
quant = True if 'GS8' in args.base_model_str else False
subnet_model_path = os.path.join('subnet_structures', args.dataset, args.task, args.base_model_str, 'pth')
## Networks
# G:
dim_lst_path = os.path.join(subnet_model_path, 'epoch%d_netG.npy' % 199)
netG = Generator(args.input_nc, args.output_nc, dim_lst=np.load(dim_lst_path), quant=quant).cuda()
# D:
netD = Discriminator(args.input_nc).cuda()
## results_dir:
optimizer_str = 'adam_lr%s_wd%s' % (args.lr, args.wd)
loss_str = 'beta%s_%s' % (args.beta, args.lc)
results_dir = os.path.join('finetune_results', args.dataset, args.task, args.base_model_str,
'%s_%s' % (optimizer_str, loss_str))
img_dir = os.path.join(results_dir, 'img')
pth_dir = os.path.join(results_dir, 'pth')
create_dir(img_dir), create_dir(pth_dir)
# Optimizers
optimizer_G = torch.optim.Adam(netG.parameters(), lr=args.lr, weight_decay=args.wd, betas=(0.5, 0.999)) # lr=1e-3
optimizer_D = torch.optim.Adam(netD.parameters(), lr=args.lr, weight_decay=args.wd, betas=(0.5, 0.999)) # lr=1e-3
# LR schedulers:
return syn_feature, syn_label, syn_att
""""pre-train a classifier on seen classes"""
trc = utils.train_cla(data.train_feature, data.train_label, CLA, device=opts.device, )
trc.run(50, data.test_seen_feature, data.test_seen_label, save_path='./cla_model')
# load best classifier
pre_cla = torch.load("./cla_model/model.pt")
for p in pre_cla.parameters(): # set requires_grad to False
p.requires_grad = False
if(opts.GD == 1):
netG = Generator(opts).to(opts.device)
netD = Discriminator(opts).to(opts.device)
else:
netG = Generator1(opts).to(opts.device)
netD = Discriminator1(opts).to(opts.device)
# seen reconstructor
netRS = Reconstructor(opts).to(opts.device)
# unseen reconstructor
netRU = Reconstructor(opts).to(opts.device)
if opts.optimizer == "ADAM":
optimzerF = optim.Adam
else:
optimzerF = optim.RMSprop
#train setup
loss_str = 'alpha%s_beta%s_%s' % (args.alpha, args.beta, args.lc)
results_dir = os.path.join(
'distill_results', args.dataset, args.task,
'%s_%s_%s_%s' % (method_str, loss_str, opt_str, W_optimizer_str))
img_dir = os.path.join(results_dir, 'img')
pth_dir = os.path.join(results_dir, 'pth')
create_dir(img_dir), create_dir(pth_dir)
## Networks
# G:
netG = Generator(args.input_nc,
args.output_nc,
quant=args.quant,
alpha=args.alpha).cuda()
# D:
netD = Discriminator(args.input_nc).cuda()
# FLOPs for G:
netG.cpu()
count_ops = measure_model(netG, 256, 256)
print("#parameters: %s" % model_param_num(netG))
f = open(os.path.join(results_dir, 'model_size.txt'), 'a+')
f.write('count_ops: {:.6f}M'.format(count_ops / 1024. / 1024.))
f.write("#parameters: %s" % model_param_num(netG))
f.close()
netG.cuda()
# Optimizers:
optimizer_G = torch.optim.Adam(netG.parameters(),
lr=args.lrw,
weight_decay=args.wd,
opt = parser.parse_args()
cuda = True if torch.cuda.is_available() else False
lambda_gp = 10
multi_gpu = False
exp_folder = "{}_{}".format(opt.exp_folder, opt.target_set)
os.makedirs("./exps/" + exp_folder, exist_ok=True)
# Loss function
adversarial_loss = torch.nn.BCEWithLogitsLoss()
distance_loss = torch.nn.L1Loss()
# Initialize generator and discriminator
generator = Generator()
discriminator = Discriminator()
if cuda:
generator.cuda()
discriminator.cuda()
adversarial_loss.cuda()
# Visualize a single batch
def visualizeSingleBatch(fp_loader_test,
opt,
exp_folder,
batches_done,
batch_size=8):
print('Loading saved model ... \n{}'.format(
'./checkpoints/{}_{}.pth'.format(exp_folder, batches_done)))
generatorTest = Generator()
gamma_optimizer_str = 'sgd_mom%s_lrgamma%s' % (args.momentum, args.lrgamma)
W_optimizer_str = 'adam_lrw%s_wd%s' % (args.lrw, args.wd)
opt_str = 'e%d-b%d' % (args.epochs, args.batch_size)
loss_str = 'rho%s_beta%s_%s' % (args.rho, args.beta, args.lc)
results_dir = os.path.join(
'results', args.dataset, args.task, '%s_%s_%s_%s_%s' %
(method_str, loss_str, opt_str, gamma_optimizer_str, W_optimizer_str))
img_dir = os.path.join(results_dir, 'img')
pth_dir = os.path.join(results_dir, 'pth')
create_dir(img_dir), create_dir(pth_dir)
## Networks
# G:
netG = Generator(args.input_nc, args.output_nc, quant=args.quant).cuda()
# D:
netD = Discriminator(args.input_nc).cuda()
# param list:
parameters_G, parameters_D, parameters_gamma = [], [], []
for name, para in netG.named_parameters():
if 'weight' in name and para.ndimension() == 1:
parameters_gamma.append(para)
else:
parameters_G.append(para)
for name, para in netD.named_parameters():
# print(name, para.size(), para.ndimension())
parameters_D.append(para)
print('parameters_gamma:', len(parameters_gamma))
# Optimizers:
optimizer_gamma = torch.optim.SGD(parameters_gamma,
def train(model_config, train_config):
mode = 'train'
dataset = ShakespeareModern(train_shakespeare_path, test_shakespeare_path, train_modern_path, test_modern_path, mode=mode)
dataloader = DataLoader(dataset, batch_size=train_config['batch_size'], shuffle=False)
print(dataset.domain_A_max_len)
shakespeare_disc = Discriminator(model_config['embedding_size'], model_config['hidden_dim'], len(dataset.vocab), batch_size=train_config['batch_size']).cuda()
shakespeare_disc.train()
if train_config['continue_train']:
shakespeare_disc.load_state_dict(torch.load(train_config['model_path']))
criterion = nn.BCELoss().cuda()
optimizer = torch.optim.Adam(shakespeare_disc.parameters(), lr=train_config['base_lr'],
weight_decay=1e-5)
real_label = torch.ones((train_config['batch_size'], 1)).cuda()
fake_label = torch.zeros((train_config['batch_size'], 1)).cuda()
for epoch in range(train_config['num_epochs']):
for idx, (s, s_addn_feats, m, m_addn_feats) in tqdm(enumerate(dataloader)):
s = s.transpose(0, 1)
m = m.transpose(0, 1)
s = Variable(s).cuda()
s_output = shakespeare_disc(s, s_addn_feats)
s_loss = criterion(s_output, real_label)
s_loss = 100 * s_loss
optimizer.zero_grad()
s_loss.backward()
optimizer.step()
shakespeare_disc.hidden = shakespeare_disc.init_hidden()
m = Variable(m).cuda()
m_output = shakespeare_disc(m, m_addn_feats)
m_loss = criterion(m_output, fake_label)
m_loss = 100 * m_loss
optimizer.zero_grad()
m_loss.backward()
optimizer.step()
shakespeare_disc.hidden = shakespeare_disc.init_hidden()
if idx % 100 == 0:
print('\tepoch [{}/{}], iter: {}, s_loss: {:.4f}, m_loss: {:.4f}, preds: s: {}, {}, m: {}, {}'
.format(epoch+1, train_config['num_epochs'], idx, s_loss.item(), m_loss.item(), s_output.item(), round(s_output.item()), m_output.item(), round(m_output.item())))
print('\tepoch [{}/{}]'.format(epoch+1, train_config['num_epochs']))
torch.save(shakespeare_disc.state_dict(), './shakespeare_disc.pth')
def main():
parser = argparse.ArgumentParser(
description='Train Cartoon avatar Gan models')
parser.add_argument('--crop_size',
default=64,
type=int,
help='Training images crop size')
parser.add_argument('--num_epochs',
default=50,
type=int,
help='Train epoch number')
parser.add_argument('--data_root',
default='data/cartoon',
help='Root directory for dataset')
parser.add_argument('--worker',
default=2,
type=int,
help='Number of workers for dataloader')
parser.add_argument('--batch_size',
default=16,
type=int,
help='Batch size during training')
parser.add_argument('--channels',
default=3,
type=int,
help='Number of channels in the training images')
parser.add_argument('--nz',
default=100,
type=int,
help='Size of generator input')
parser.add_argument('--ngf',
default=64,
type=int,
help='Size of feature maps in generator')
parser.add_argument('--ndf',
default=64,
type=int,
help='Size of feature maps in descriminator')
parser.add_argument('--lr',
default=0.0002,
type=float,
help='Learning rate for optimizer')
parser.add_argument('--beta1',
default=0.5,
type=float,
help='Beta1 hyperparam for Adam optimizers')
parser.add_argument('--beta2',
default=0.999,
type=float,
help='Beta2 hyperparam for Adam optimizers')
parser.add_argument('--ngpu',
default=1,
type=int,
help='Number of GPUs , use 0 for CPU mode')
parser.add_argument(
'--latent_vector_num',
default=8,
type=int,
help=
'latent vectors that we will use to visualize , 8 means that it will visualize 8 images during training'
)
opt = parser.parse_args()
dataroot = opt.data_root
workers = opt.worker
batch_size = opt.batch_size
image_size = opt.crop_size
nc = opt.channels
nz = opt.nz
ngf = opt.ngf
ndf = opt.ndf
num_epochs = opt.num_epochs
lr = opt.lr
beta1 = opt.beta1
beta2 = opt.beta2
ngpu = opt.ngpu
latent_vector_num = opt.latent_vector_num
# Create the dataset
dataset = dset.ImageFolder(root=dataroot,
transform=transforms.Compose([
transforms.Resize(image_size),
transforms.CenterCrop(image_size),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5),
(0.5, 0.5, 0.5)),
]))
# Create the dataloader
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
shuffle=True,
num_workers=workers)
# Decide which device we want to run on
device = torch.device("cuda:0" if (
torch.cuda.is_available() and ngpu > 0) else "cpu")
# Create the generator
netG = Generator(ngpu, nz, ngf, nc).to(device)
# Create the Discriminator
netD = Discriminator(ngpu, nc, ndf).to(device)
# Handle multi-gpu if desired
if (device.type == 'cuda') and (ngpu > 1):
netG = nn.DataParallel(netG, list(range(ngpu)))
netD = nn.DataParallel(netD, list(range(ngpu)))
# Apply the weights_init function to randomly initialize all weights
# to mean=0, stdev=0.2.
netG.apply(weights_init)
netD.apply(weights_init)
# Setup Adam optimizers for both G and D
optimizerD = optim.Adam(netD.parameters(), lr=lr, betas=(beta1, 0.999))
optimizerG = optim.Adam(netG.parameters(), lr=lr, betas=(beta1, 0.999))
# Print models
print(netG)
print(netD)
# Initialize BCELoss function
criterion = nn.BCELoss()
# Create batch of latent vectors that we will use to visualize
fixed_noise = torch.randn(latent_vector_num, nz, 1, 1, device=device)
#real and fake labels during training
real_label = 1
fake_label = 0
# Lists to keep track of progress
img_list = []
G_losses = []
D_losses = []
iters = 0
print("Starting Training ...")
for epoch in range(num_epochs):
for i, data in enumerate(dataloader, 0):
############################
# (1) Update D network: maximize log(D(x)) + log(1 - D(G(z)))
###########################
## Train with all-real batch
netD.zero_grad()
# Format batch
real_cpu = data[0].to(device)
b_size = real_cpu.size(0)
label = torch.full((b_size, ), real_label, device=device)
# Forward pass real batch through D
output = netD(real_cpu).view(-1)
# Calculate loss on all-real batch
errD_real = criterion(output, label)
# Calculate gradients for D in backward pass
errD_real.backward()
D_x = output.mean().item()
## Train with all-fake batch
# Generate batch of latent vectors
noise = torch.randn(b_size, nz, 1, 1, device=device)
# Generate fake image batch with G
fake = netG(noise)
label.fill_(fake_label)
# Classify all fake batch with D
output = netD(fake.detach()).view(-1)
# Calculate D's loss on the all-fake batch
errD_fake = criterion(output, label)
# Calculate the gradients for this batch
errD_fake.backward()
D_G_z1 = output.mean().item()
# Add the gradients from the all-real and all-fake batches
errD = errD_real + errD_fake
# Update D
optimizerD.step()
############################
# (2) Update G network: maximize log(D(G(z)))
###########################
netG.zero_grad()
label.fill_(real_label)
# fake labels are real for generator cost
# Since we just updated D, perform another forward pass of all-fake batch through D
output = netD(fake).view(-1)
# Calculate G's loss based on this output
errG = criterion(output, label)
# Calculate gradients for G
errG.backward()
D_G_z2 = output.mean().item()
# Update G
optimizerG.step()
# Output training stats
if i % 50 == 0:
# Save model data
torch.save(netG.state_dict(),
'pretrained_model/netG_epoch_%d.pth' % (iters))
torch.save(netD.state_dict(),
'pretrained_model/netD_epoch_%d.pth' % (iters))
# Print training stats
print(
'[%d/%d][%d/%d]\tLoss_D: %.4f\tLoss_G: %.4f\tD(x): %.4f\tD(G(z)): %.4f / %.4f'
% (epoch, num_epochs, i, len(dataloader), errD.item(),
errG.item(), D_x, D_G_z1, D_G_z2))
# Save Losses for plotting later
G_losses.append(errG.item())
D_losses.append(errD.item())
# Check how the generator is doing by saving G's output on fixed_noise
if (iters % 650 == 0) or ((epoch == num_epochs - 1) and
(i == len(dataloader) - 1)):
with torch.no_grad():
fake = netG(fixed_noise).detach().cpu()
img_list.append(
vutils.make_grid(fake, padding=2, normalize=True))
iters += 1
# Display and Save samples GIF
fig = plt.figure(figsize=(8, 8))
plt.axis("off")
ims = [[plt.imshow(np.transpose(i, (1, 2, 0)), animated=True)]
for i in img_list]
ani = animation.ArtistAnimation(fig,
ims,
interval=1000,
repeat_delay=1000,
blit=True)
ani.save('output/samples.gif', writer='imagemagick', fps=100)
gamma_optimizer_str = 'sgd_mom%s_lrgamma%s' % (args.momentum, args.lrgamma)
W_optimizer_str = 'adam_lrw%s_wd%s' % (args.lrw, args.wd)
opt_str = 'e%d-b%d' % (args.epochs, args.batch_size)
loss_str = 'rho%s_beta%s_%s' % (args.rho, args.beta, args.lc)
results_dir = os.path.join(
'results', args.dataset, args.task, '%s_%s_%s_%s_%s' %
(method_str, loss_str, opt_str, gamma_optimizer_str, W_optimizer_str))
img_dir = os.path.join(results_dir, 'img')
pth_dir = os.path.join(results_dir, 'pth')
create_dir(img_dir), create_dir(pth_dir)
## Networks
# G:
netG = Generator(args.input_nc, args.output_nc, quant=args.quant).cuda()
# D:
netD = Discriminator(args.input_nc).cuda()
# param list:
parameters_G, parameters_D, parameters_gamma = [], [], []
for name, para in netG.named_parameters():
if 'weight' in name and para.ndimension() == 1:
parameters_gamma.append(para)
else:
parameters_G.append(para)
for name, para in netD.named_parameters():
# print(name, para.size(), para.ndimension())
parameters_D.append(para)
print('parameters_gamma:', len(parameters_gamma))
# Optimizers:
optimizer_gamma = torch.optim.SGD(parameters_gamma,
def main():
# init random seed
init_random_seed(params.manual_seed)
# Load dataset
mnist_data_loader = get_mnist(train=True, download=True)
mnist_data_loader_eval = get_mnist(train=False, download=True)
usps_data_loader = get_usps(train=True, download=True)
usps_data_loader_eval = get_usps(train=False, download=True)
# Model init DANN
tgt_encoder = model_init(Encoder(), params.tgt_encoder_dann_rb_path)
critic = model_init(Discriminator(), params.disc_dann_rb_path)
clf = model_init(Classifier(), params.clf_dann_rb_path)
# Train models
print(
"====== Robust Training source encoder and classifier in MNIST and USPS domains ======"
)
if not (tgt_encoder.pretrained and clf.pretrained and critic.pretrained
and params.model_trained):
tgt_encoder, clf, critic = train_dann(tgt_encoder,
clf,
critic,
mnist_data_loader,
usps_data_loader,
usps_data_loader_eval,
robust=False)
# Eval target encoder on test set of target dataset
print(
"====== Evaluating classifier for encoded MNIST and USPS domains ======"
)
print("-------- MNIST domain --------")
eval_tgt_robust(tgt_encoder, clf, critic, mnist_data_loader_eval)
print("-------- USPS adaption --------")
eval_tgt_robust(tgt_encoder, clf, critic, usps_data_loader_eval)
print("====== Pseudo labeling on USPS domain ======")
pseudo_label(tgt_encoder, clf, "usps_train_pseudo", usps_data_loader)
# Init a new model
tgt_encoder = model_init(Encoder(), params.tgt_encoder_path)
clf = model_init(Classifier(), params.clf_path)
# Load pseudo labeled dataset
usps_pseudo_loader = get_usps(train=True, download=True, get_pseudo=True)
print("====== Standard training on USPS domain with pseudo labels ======")
if not (tgt_encoder.pretrained and clf.pretrained):
train_src_adda(tgt_encoder, clf, usps_pseudo_loader, mode='ADV')
print("====== Evaluating on USPS domain with real labels ======")
eval_tgt(tgt_encoder, clf, usps_data_loader_eval)
tgt_encoder = model_init(Encoder(), params.tgt_encoder_rb_path)
clf = model_init(Classifier(), params.clf_rb_path)
print("====== Robust training on USPS domain with pseudo labels ======")
if not (tgt_encoder.pretrained and clf.pretrained):
train_src_robust(tgt_encoder, clf, usps_pseudo_loader, mode='ADV')
print("====== Evaluating on USPS domain with real labels ======")
eval_tgt(tgt_encoder, clf, usps_data_loader_eval)
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
quant = True if 'GS8' in args.base_model_str else False
subnet_model_path = os.path.join('subnet_structures_ticket', args.dataset,
args.task, args.base_model_str, 'pth')
## Networks
# G:
dim_lst_path = os.path.join(subnet_model_path, 'epoch%d_netG.npy' % 199)
netG = Generator(args.input_nc,
args.output_nc,
dim_lst=np.load(dim_lst_path),
quant=quant).cuda()
# D:
netD = Discriminator(args.input_nc).cuda()
## results_dir:
optimizer_str = 'adam_lr%s_wd%s' % (args.lr, args.wd)
loss_str = ''
results_dir = os.path.join('cp_finetune_results', args.dataset, args.task,
args.base_model_str,
'%s_%s' % (optimizer_str, loss_str))
img_dir = os.path.join(results_dir, 'img')
pth_dir = os.path.join(results_dir, 'pth')
create_dir(img_dir), create_dir(pth_dir)
# Optimizers
optimizer_G = torch.optim.Adam(netG.parameters(),
lr=args.lr,
weight_decay=args.wd,
else:
source_str, target_str = 'B', 'A'
foreign_dir = '/home/haotao/PyTorch-CycleGAN/'
print(args)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
## Networks
# G:
dim_lst_path = os.path.join(subnet_model_path, 'epoch%d_netG.npy' % 199)
netG = Generator(args.input_nc, args.output_nc, dim_lst=np.load(dim_lst_path), quant=quant).cuda()
# D:
netD = Discriminator(args.input_nc).cuda()
# load sub G extracted from latest.pth
g_path = os.path.join(args.base_model_str, 'epoch%d_netG.pth' % 199)
netG.load_state_dict(torch.load(g_path))
print('load G from %s' % g_path)
# load full D directly from latest.pth
d_path = os.path.join('cp_results', args.dataset, args.task, args.base_model_str, 'pth', 'latest.pth')
netD.load_state_dict(torch.load(d_path)['netD'])
print('load D from %s' % d_path)
start_epoch = 0
best_FID = 1e9
loss_G_lst, loss_G_perceptual_lst, loss_G_GAN_lst, loss_D_lst = [], [], [], []
# Dataset loader: img shape=(256,256)
