if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--ckpt_path',
type=str,
default='work_dir/bert_ner.pt')
parser.add_argument('--input_path',
type=str,
default='work_dir/selected_paras.json')
parser.add_argument('--output_path',
type=str,
default='work_dir/entities.json')
parser.add_argument('--batch_size', type=int, default=32)
args = parser.parse_args()
eval_dataset = EvalDataset(args.input_path, debug=False)
eval_iter = data.DataLoader(dataset=eval_dataset,
batch_size=args.batch_size,
shuffle=False,
collate_fn=pad)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = Net(top_rnns=False,
vocab_size=len(VOCAB),
device=device,
finetuning=True).cuda()
model = nn.DataParallel(model)
model.load_state_dict(torch.load(args.ckpt_path))
model.eval()
eval_para(model, eval_iter, eval_dataset.sent_id, args.output_path)
optimizer.zero_grad()
confusion_matrix = net(captions,
{'face': face, 'audio': audio, 'visual': video, 'motion': flow}, ind, True)
loss = max_margin(confusion_matrix)
loss.backward()
optimizer.step()
running_loss += loss.data[0]
if (i_batch+1) % n_display == 0:
print 'Epoch %d, Epoch status: %.2f, Training loss: %.4f'%(epoch + 1,
args.batch_size*float(i_batch)/dataset_size,running_loss/n_display)
running_loss = 0.0
print 'evaluating epoch %d ...'%(epoch+1)
net.eval()
if args.MSRVTT:
retrieval_samples = dataset.getRetrievalSamples()
video = Variable(retrieval_samples['video'].cuda(), volatile=True)
captions = Variable(retrieval_samples['text'].cuda(), volatile=True)
audio = Variable(retrieval_samples['audio'].cuda(), volatile=True)
flow = Variable(retrieval_samples['flow'].cuda(), volatile=True)
face = Variable(retrieval_samples['face'].cuda(), volatile=True)
face_ind = retrieval_samples['face_ind']
ind = {}
ind['face'] = face_ind
ind['visual'] = np.ones((len(face_ind)))
ind['motion'] = np.ones((len(face_ind)))
eta_min=0.005)
# Training
n_epoch = int(config["Train"]["epoch"])
save_every_epoch = int(config["Train"]["SaveEveryEpoch"])
if save_every_epoch == -1:
save_every_epoch = n_epoch + 1
dst_dir = os.path.join(ABS_DIR, config["Train"]["RecordDestination"])
for t in range(n_epoch):
result = []
for phase in ['train', 'val']:
if phase == 'train':
net.train()
else:
net.eval()
# keep track of training and validation loss
running_loss = 0.0
running_batch = 0
running_confusion = np.zeros(shape=(n_class, n_class), dtype=np.float)
for data, target in data_loader[phase]:
data, target = data.to(device), target.to(device)
# print(target.shape)
with torch.set_grad_enabled(phase == 'train'):
# feed the input
preds = net(data)
# calculate the loss
loss = loss_func(preds, target)
if phase == 'train':
optimizer.zero_grad()
class FSRCNNTrainer(object):
def __init__(self, config, training_loader, testing_loader):
super(FSRCNNTrainer, self).__init__()
self.CUDA = torch.cuda.is_available()
self.device = torch.device('cuda' if self.CUDA else 'cpu')
self.model = None
self.lr = config.lr
self.nEpochs = config.nEpochs
self.criterion = None
self.optimizer = None
self.scheduler = None
self.seed = config.seed
self.upscale_factor = config.upscale_factor
self.training_loader = training_loader
self.testing_loader = testing_loader
configure('logs', flush_secs=5)
def build_model(self):
self.model = Net(num_channels=1, upscale_factor=self.upscale_factor).to(self.device)
self.model.weight_init(mean=0.0, std=0.2)
self.criterion = torch.nn.MSELoss()
torch.manual_seed(self.seed)
if self.CUDA:
torch.cuda.manual_seed(self.seed) # 随机种子
cudnn.benchmark = True
self.criterion.cuda()
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=self.lr)
self.scheduler = torch.optim.lr_scheduler.MultiStepLR(self.optimizer, milestones=[50, 75, 100],
gamma=0.5) # lr decay
def save_model(self):
model_out_path = "./checkpoints/model_path.pth"
torch.save(self.model, model_out_path)
print("Checkpoint saved to {}".format(model_out_path))
def train(self, epoch):
self.model.train()
train_loss = 0
for batch_num, (data, target) in enumerate(self.training_loader):
data, target = data.to(self.device), target.to(self.device)
self.optimizer.zero_grad()
loss = self.criterion(self.model(data), target)
train_loss += loss.item()
loss.backward()
self.optimizer.step()
progress_bar(batch_num, len(self.training_loader), 'Loss: %.4f' % (train_loss / (batch_num + 1)))
print(" Average Loss: {:.4f}".format(train_loss / len(self.training_loader)))
log_value('train_loss', train_loss / (batch_num + 1), epoch)
def test(self, epoch):
self.model.eval()
avg_psnr = 0
with torch.no_grad():
for batch_num, (data, target) in enumerate(self.testing_loader):
data, target = data.to(self.device), target.to(self.device)
prediction = self.model(data)
mse = self.criterion(prediction, target)
psnr = 10 * log10(1 / mse.item())
avg_psnr += psnr
progress_bar(batch_num, len(self.testing_loader), 'PSNR: %.4f' % (avg_psnr / (batch_num + 1)))
print(" Average PSNR: {:.4f} dB".format(avg_psnr / len(self.testing_loader)))
log_value('PSNR', avg_psnr / (batch_num + 1), epoch)
def run(self):
self.build_model()
for epoch in range(1, self.nEpochs + 1):
print("\n===> Epoch {} starts:".format(epoch))
self.train(epoch)
self.test(epoch)
self.scheduler.step(epoch)
if epoch == self.nEpochs:
self.save_model()
class Solver(object):
def __init__(self, config, source_loader, source_val_loader, target_loader, target_val_loader):
self.source_loader = source_loader
self.source_val_loader = source_val_loader
self.target_loader = target_loader
self.target_val_loader = target_val_loader
self.net = None
self.net_optimizer = None
self.net_d = None
self.net_optimizer_d = None
self.beta1 = config.beta1
self.beta2 = config.beta2
self.train_iters = config.train_iters
self.pretrain_iters = config.pretrain_iters
self.batch_size = config.batch_size
self.lr = config.lr
self.lr_d = config.lr_d
self.alpha_s = config.alpha_s
self.alpha_t = config.alpha_t
self.beta_c = config.beta_c
self.beta_sep = config.beta_sep
self.beta_p = config.beta_p
self.log_step = config.log_step
self.model_path = config.model_path
self.num_classes = config.num_classes
self.build_model()
def build_model(self):
"""Builds a generator and a discriminator."""
self.net = Net()
self.net_d = Net_D()
net_params = list(self.net.parameters())
net_d_params = list(self.net_d.parameters())
self.net_optimizer = optim.Adam(net_params, self.lr, [self.beta1, self.beta2])
self.net_optimizer_d = optim.Adam(net_d_params, self.lr_d, [self.beta1, self.beta2])
if torch.cuda.is_available():
self.net.cuda()
self.net_d.cuda()
def to_var(self, x):
"""Converts numpy to variable."""
if torch.cuda.is_available():
x = x.cuda()
return Variable(x, requires_grad=False)
def to_data(self, x):
"""Converts variable to numpy."""
if torch.cuda.is_available():
x = x.cpu()
return x.data.numpy()
def reset_grad(self):
"""Zeros the gradient buffers."""
self.net_optimizer.zero_grad()
self.net_optimizer_d.zero_grad()
def separability_loss(self, labels, latents, imbalance_parameter=1):
criteria = torch.nn.modules.loss.CosineEmbeddingLoss()
loss_up = 0
one_cuda = torch.ones(1).cuda()
mean = torch.mean(latents, dim=0).cuda().view(1, -1)
loss_down = 0
for i in range(self.num_classes):
indexes = labels.eq(i)
mean_i = torch.mean(latents[indexes], dim=0).view(1, -1)
if str(mean_i.norm().item()) != 'nan':
for latent in latents[indexes]:
loss_up += criteria(latent.view(1, -1), mean_i, one_cuda)
loss_down += criteria(mean, mean_i, one_cuda)
loss = (loss_up / loss_down) * imbalance_parameter
return loss
def initialisation(self):
self.net.apply(xavier_weights_init)
self.net_d.apply(xavier_weights_init)
source_iter = iter(self.source_loader)
target_iter = iter(self.target_loader)
source_val_loader = iter(self.source_val_loader)
target_val_iter = iter(self.target_val_loader)
source_per_epoch = len(source_iter)
target_per_epoch = len(target_iter)
targetval_per_epoch = len(target_val_iter)
print(source_per_epoch, target_per_epoch, targetval_per_epoch)
criterion = nn.CrossEntropyLoss()
f_labels = torch.LongTensor(128)
f_labels[...] = 10
t_labels = torch.LongTensor(128)
t_labels[...] = 1
# pretrain
log_pre = 50
source_iter = iter(self.source_loader)
source_val_iter = iter(self.source_val_loader)
target_iter = iter(self.target_loader)
return criterion, source_per_epoch, target_per_epoch, target_iter, source_iter, log_pre, source_val_iter
def train(self):
criterion, source_per_epoch, target_per_epoch, target_iter, source_iter, log_pre, source_val_iter = self.initialisation()
pre_train = not os.path.exists(os.path.join(self.model_path, 'pre_train.pth'))
print("Pretrain:\n*********")
if pre_train:
for step in range(self.pretrain_iters + 1):
# ============ Initialization ============#
# refresh
if (step + 1) % (source_per_epoch) == 0:
source_iter = iter(self.source_loader)
if (step + 1) % (target_per_epoch) == 0:
target_iter = iter(self.target_loader)
# load the data
source, s_labels = source_iter.next()
target, t_labels = target_iter.next()
target_rgb = target
target, t_labels = self.to_var(target_rgb), self.to_var(t_labels).long().squeeze()
source, s_labels = self.to_var(source), self.to_var(s_labels).long().squeeze()
# ============ Training ============ #
self.reset_grad()
# forward
latent, c = self.net(source)
# loss
loss_source_class = criterion(c, s_labels)
# one step
loss_source_class.backward()
self.net_optimizer.step()
self.reset_grad()
# ============ Validation ============ #
if (step + 1) % log_pre == 0:
_, c_source = self.net(source)
_, c_target = self.net(target)
print("[%d/20000] classification loss: %.4f" % (
step + 1, loss_source_class.item()))
print("source accuracy %.4f; target accuracy %.4f" % (
accuracy(s_labels, c_source),
accuracy(t_labels, c_target)))
self.save_model()
else:
self.load_model()
# ============ Initialization ============ #
source_iter = iter(self.source_loader)
target_iter = iter(self.target_loader)
source_val_iter = iter(self.source_val_loader)
maxacc = 0.0
maximum_acc = 0.0
max_iter = 0
net_params = list(self.net.parameters())
net_d_params = list(self.net_d.parameters())
self.net_optimizer = optim.Adam(net_params, self.lr, [self.beta1, self.beta2])
self.net_optimizer_d = optim.Adam(net_d_params, self.lr_d, [self.beta1, self.beta2])
print("Second:\n******")
# self.validate_source()
self.validate_target()
for step in range(self.train_iters):
# ============ Initialization ============#
# refresh
if (step + 1) % (target_per_epoch) == 0:
target_iter = iter(self.target_loader)
if (step + 1) % (source_per_epoch) == 0:
source_iter = iter(self.source_loader)
source_val_iter = iter(self.source_val_loader)
# load the data
source, s_labels = source_iter.next()
source, s_labels = self.to_var(source), self.to_var(s_labels).long().squeeze() # must squeeze
# source_val, s_val_labels = source_val_iter.next()
# source_val, s_val_labels = self.to_var(source_val), self.to_var(s_val_labels).long().squeeze()
target, t_labels = target_iter.next()
target_rgb = target
target, t_labels = self.to_var(target_rgb), self.to_var(t_labels).long().squeeze()
# ============ train D ============#
self.reset_grad()
latent_source, c = self.net(source)
d = self.net_d(latent_source)
loss_d_s1 = F.binary_cross_entropy(d, torch.ones_like(d, dtype=torch.float32))
loss_d_s0 = F.binary_cross_entropy(d, torch.zeros_like(d, dtype=torch.float32))
loss_c_source = criterion(c, s_labels)
latent_target, c = self.net(target)
d = self.net_d(latent_target)
loss_d_t0 = F.binary_cross_entropy(d, torch.zeros_like(d, dtype=torch.float32))
loss_p = loss_d_s0
loss_d = loss_d_s1 + loss_d_t0
# ============ train pseudo labeling ============#
chosen_target, pseudo_labels, indexes, imbalance_parameter = pseudo_labeling(target, c)
if chosen_target is not None:
loss_c_target = criterion(c[indexes], pseudo_labels)
latent_target = latent_target[indexes]
# ============ class loss ============#
loss_sep = self.separability_loss(torch.cat((s_labels, pseudo_labels)),
torch.cat((latent_source, latent_target)),
imbalance_parameter=imbalance_parameter)
else:
loss_c_target = 0
loss_sep = 0
loss = self.beta_c * (self.alpha_s * loss_c_source + self.alpha_t * loss_c_target) + \
self.beta_p * loss_p + \
self.beta_sep * loss_sep
loss.backward(retain_graph=True)
self.net_optimizer.step()
loss_d.backward()
self.net_optimizer_d.step()
self.reset_grad()
# ============ Validation ============ #
if (step + 1) % self.log_step == 0:
print("max accuracy:", colored(maximum_acc, "green"), "iteration", max_iter)
_, c_source = self.net(source)
_, c_target = self.net(target)
print("source accuracy (svhn_train) %.4f; target accuracy %.4f" % (
accuracy(s_labels, c_source), accuracy(t_labels, c_target)))
self.validate_source()
acc = self.validate_target()
if acc > maximum_acc:
maximum_acc = acc
max_iter = step
if acc > maxacc:
maxacc = acc
torch.save(self.net, "./model_c_" + str(step) + '_' + str(acc) + ".pth")
torch.save(self.net_d, "./model_d_" + str(step) + '_' + str(acc) + ".pth")
self.reset_grad()
# ============ Save the model ============ #
torch.save(self.net, "./model_c_final.pth")
torch.save(self.net_d, "./model_d_final.pth")
def validate_target(self, ):
class_correct = [0] * self.num_classes
class_total = [0.] * self.num_classes
classes = [str(i) for i in range(self.num_classes)]
self.net.eval() # prep model for evaluation
for data, target in self.target_val_loader:
# forward pass: compute predicted outputs by passing inputs to the model
data, target = self.to_var(data), self.to_var(target).long().squeeze()
data = data.cuda()
target = target.cuda()
latent, output = self.net(data)
_, pred = torch.max(output, 1)
correct = np.squeeze(pred.eq(target.data.view_as(pred)))
# calculate test accuracy for each object class
for i in range(len(target.data)):
label = target.data[i]
class_correct[label] += correct[i].item()
class_total[label] += 1
for i in range(self.num_classes):
if class_total[i] > 0:
print('Test Accuracy (mnist-test) of %5s: %2d%% (%2d/%2d)' % (
str(i), 100 * class_correct[i] / class_total[i],
np.sum(class_correct[i]), np.sum(class_total[i])))
else:
print('Test Accuracy (mnist-test) of %5s: N/A (no training examples)' % (classes[i]))
print("\nTest Accuracy (mnist-test) (Overall): ", end="")
print(colored('%2d%% ' % (100. * np.sum(class_correct) / np.sum(class_total)), "red"), end="")
print("(", end="")
print(colored(str(int(np.sum(class_correct))), "red"), end=" ")
print('/%2d)' % (np.sum(class_total)))
self.net.train()
return 100. * np.sum(class_correct) / np.sum(class_total)
def validate_source(self, ):
class_correct = [0] * self.num_classes
class_total = [0.] * self.num_classes
classes = [str(i) for i in range(self.num_classes)]
self.net.eval() # prep model for evaluation
for data, target in self.source_val_loader:
# forward pass: compute predicted outputs by passing inputs to the model
data, target = self.to_var(data), self.to_var(target).long().squeeze()
data = data.cuda()
target = target.cuda()
latent, output = self.net(data)
_, pred = torch.max(output, 1)
correct = np.squeeze(pred.eq(target.data.view_as(pred)))
# calculate test accuracy for each object class
for i in range(len(target.data)):
label = target.data[i]
class_correct[label] += correct[i].item()
class_total[label] += 1
for i in range(self.num_classes):
if class_total[i] > 0:
print('Test Accuracy (svhn-test) of %5s: %2d%% (%2d/%2d)' % (
str(i), 100 * class_correct[i] / class_total[i],
np.sum(class_correct[i]), np.sum(class_total[i])))
else:
print('Test Accuracy (svhn_test) of %5s: N/A (no training examples)' % (classes[i]))
print("\nTest Accuracy (svhn-test) (Overall): ", end="")
print(colored('%2d%% ' % (100. * np.sum(class_correct) / np.sum(class_total)), "red"), end="")
print("(", end="")
print(colored(str(int(np.sum(class_correct))), "red"), end=" ")
print('/%2d)' % (np.sum(class_total)))
self.net.train()
return 100. * np.sum(class_correct) / np.sum(class_total)
def save_model(self):
torch.save(self.net, os.path.join(self.model_path, 'pre_train.pth'))
def load_model(self):
self.net = torch.load(os.path.join(self.model_path, 'pre_train.pth'))
class Trainer:
def __init__(self, cfg):
self.cfg = cfg
self.init_env()
self.init_device()
self.init_data()
self.init_model()
self.init_optimizer()
def init_env(self):
self.exp_dir = Path(
self.cfg.train_log_root).expanduser().joinpath(self.cfg.exp_id)
self.exp_dir.mkdir(parents=True, exist_ok=True)
self.log_dir = self.exp_dir.joinpath(self.cfg.log_subdir)
self.tb_dir = self.exp_dir.joinpath(self.cfg.tb_subdir)
self.ckpt_dir = self.exp_dir.joinpath(self.cfg.ckpt_subdir)
self.logger = get_logger(__name__, self.log_dir)
self.tb = SummaryWriter(self.tb_dir)
torch.manual_seed(self.cfg.seed)
self.epoch = 0
self.acc = 0.
self.logger.info('Train log location: {}'.format(self.exp_dir))
def init_device(self):
self.use_cuda = not self.cfg.no_cuda and torch.cuda.is_available()
if self.use_cuda:
self.device = torch.device('cuda')
self.logger.info('Using gpu')
else:
self.device = torch.device('cpu')
self.logger.info('Using cpu')
def init_data(self):
self.logger.info('Initializing data loader...')
kwargs = {
'num_workers': 1, 'pin_memory': True} if self.use_cuda else {}
self.train_loader = torch.utils.data.DataLoader(
datasets.MNIST(
self.cfg.data_root, train=True, download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))])),
batch_size=self.cfg.batch_size, shuffle=True, **kwargs)
self.logger.info('Train loader has been initialized.')
self.val_loader = torch.utils.data.DataLoader(
datasets.MNIST(
self.cfg.data_root, train=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))])),
batch_size=self.cfg.val_batch_size, shuffle=True, **kwargs)
self.logger.info('Test loader has been initialized.')
def init_model(self):
self.model = Net()
data, target = next(iter(self.train_loader))
self.tb.add_graph(self.model, data)
self.model = self.model.to(self.device)
self.logger.info('Model has been initialized.')
def init_optimizer(self):
cfg_optim = self.cfg.optim
optim_func = getattr(optim, cfg_optim.type)
self.optimizer = optim_func(
self.model.parameters(), **dict(self.cfg.optim.args))
self.logger.info('Optimizer has been initialized.')
def train(self):
self.model.train()
for batch_idx, (data, target) in enumerate(self.train_loader):
data, target = data.to(self.device), target.to(self.device)
self.optimizer.zero_grad()
output = self.model(data)
loss = F.nll_loss(output, target)
loss.backward()
self.optimizer.step()
self.train_loss = loss.item()
if batch_idx % self.cfg.log_interval == 0:
self.logger.info(
'{:2d}, {}/{} loss: {:.6f}, test acc: {:.2f}%'.format(
self.epoch, batch_idx * len(data),
len(self.train_loader.dataset), loss.item(), self.acc))
total_iter = self.epoch * len(self.train_loader) + batch_idx
self.tb.add_scalar('train/loss', loss.item(), total_iter)
def test(self):
self.model.eval()
test_loss = 0
correct = 0
with torch.no_grad():
for data, target in self.val_loader:
data, target = data.to(self.device), target.to(self.device)
output = self.model(data)
test_loss += F.nll_loss(output, target, reduction='sum').item()
pred = output.argmax(dim=1, keepdim=True)
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(self.val_loader.dataset)
self.acc = 100. * correct / len(self.val_loader.dataset)
self.logger.info(
'{:2d}, test loss: {:.4f}, test acc: {}/{} ({:.2f}%)'.format(
self.epoch, test_loss, correct, len(self.val_loader.dataset),
self.acc))
self.tb.add_scalar('test/acc', self.acc, self.epoch)
self.tb.add_scalar('test/loss', test_loss, self.epoch)
def load(self, for_resuming_training=True, label='latest'):
ckpt_path = self.ckpt_dir.joinpath('{}.pt'.format(label))
if ckpt_path.is_file():
self.logger.info('Loading model from {}'.format(ckpt_path))
ckpt = torch.load(ckpt_path, map_location=self.device)
self.model.load_state_dict(ckpt['model_state_dict'])
if for_resuming_training:
self.optimizer.load_state_dict(ckpt['optimizer_state_dict'])
self.epoch = ckpt['epoch'] + 1
self.acc = ckpt['acc']
self.logger.info(
'Model of epoch {} loaded.'.format(ckpt['epoch']))
else:
self.logger.info('No checkpoint found.')
def save(self, label='latest'):
self.logger.info('Saving model...')
self.ckpt_dir.mkdir(exist_ok=True, parents=True)
ckpt_path = self.ckpt_dir.joinpath('{}.pt'.format(label))
torch.save({
'epoch': self.epoch,
'model_state_dict': self.model.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict(),
'train_loss': self.train_loss,
'acc': self.acc
}, ckpt_path)
self.logger.info('Model saved to {}.'.format(ckpt_path))
def start(self):
self.load(for_resuming_training=True)
if self.epoch > 0:
self.logger.info('Training start from epoch {}'.format(self.epoch))
try:
for self.epoch in range(self.epoch, self.cfg.epochs):
self.train()
self.test()
self.logger.info('Training is finished.')
except KeyboardInterrupt:
self.logger.warning('Keyboard Interrupted.')
except Exception as e:
self.logger.exception(repr(e))
finally:
if self.epoch > 0:
self.save()
def train(args):
check_paths(args)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
try:
with open(param_path, 'r') as tc:
trainingParams = json.load(tc)
ngf = int(trainingParams.get('ngf', args.ngf))
epochs = int(trainingParams.get('epochs', args.epochs))
batch_size = int(trainingParams.get('batch_size', args.batch_size))
log_interval = int(
trainingParams.get('log_interval', args.log_interval))
learning_rate = float(
trainingParams.get('learning_rate', args.learning_rate))
cuda = int(trainingParams.get('cuda', args.cuda))
if cuda:
logger.info("Using CUDA")
torch.cuda.manual_seed(args.seed)
kwargs = {'num_workers': 8, 'pin_memory': True}
logger.info("Using kwarguments: \n" + str(kwargs))
else:
kwargs = {}
transform = transforms.Compose([
transforms.Scale(args.image_size),
transforms.CenterCrop(args.image_size),
transforms.ToTensor(),
transforms.Lambda(lambda x: x.mul(255))
])
train_dataset = datasets.ImageFolder(args.dataset, transform)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
**kwargs)
style_model = Net(ngf=ngf)
print(style_model)
optimizer = Adam(style_model.parameters(), learning_rate)
mse_loss = torch.nn.MSELoss()
vgg = Vgg16()
utils.mod_utils.init_vgg16(args.vgg_model_dir)
vgg.load_state_dict(
torch.load(os.path.join(args.vgg_model_dir, "vgg16.weight")))
if cuda:
style_model.cuda()
vgg.cuda()
style_loader = StyleLoader(args.style_folder, args.style_size)
for e in range(epochs):
style_model.train()
agg_content_loss = 0.
agg_style_loss = 0.
count = 0
for batch_id, (x, _) in enumerate(train_loader):
n_batch = len(x)
count += n_batch
optimizer.zero_grad()
x = Variable(preprocess_batch(x))
if cuda:
x.cuda()
style_v = style_loader.get(batch_id)
style_model.setTarget(style_v)
style_v = utils.img_utils.subtract_imagenet_mean_batch(
style_v)
features_style = vgg(style_v)
gram_style = [
utils.img_utils.gram_matrix(y) for y in features_style
]
y = style_model(x.cuda())
xc = Variable(x.data.clone(), volatile=True)
y = utils.img_utils.subtract_imagenet_mean_batch(y)
xc = utils.img_utils.subtract_imagenet_mean_batch(xc)
features_y = vgg(y)
features_xc = vgg(xc.cuda())
f_xc_c = Variable(features_xc[1].data, requires_grad=False)
content_loss = args.content_weight * \
mse_loss(features_y[1], f_xc_c)
style_loss = 0.
for m in range(len(features_y)):
gram_y = utils.img_utils.gram_matrix(features_y[m])
gram_s = Variable(gram_style[m].data,
requires_grad=False).repeat(
args.batch_size, 1, 1, 1)
style_loss += args.style_weight * mse_loss(
gram_y, gram_s[:n_batch, :, :])
total_loss = content_loss + style_loss
total_loss.backward()
optimizer.step()
agg_content_loss += content_loss.data[0]
agg_style_loss += style_loss.data[0]
if (batch_id + 1) % log_interval == 0:
msg = "{}\tEpoch {}:\t[{}/{}]\tcontent: {:.6f}\tstyle: {:.6f}\ttotal: {:.6f}".format(
time.ctime(), e + 1, count, len(train_dataset),
agg_content_loss / (batch_id + 1),
agg_style_loss / (batch_id + 1),
(agg_content_loss + agg_style_loss) /
(batch_id + 1))
print(msg)
if (batch_id + 1) % (20 * log_interval) == 0:
# save model
style_model.eval()
style_model.cpu()
save_model_filename = "Epoch_" + str(e) + "_" +\
"iters_" + str(count) + \
"_" + str(time.ctime()).replace(' ','_') + \
"_" + str(args.content_weight) + "_" + \
str(args.style_weight) + ".model"
save_model_path = os.path.join(temp_save_model_dir,
save_model_filename)
torch.save(style_model.state_dict(), save_model_path)
style_model.train()
style_model.cuda()
logger.info("Checkpoint, trained model saved at " +
str(save_model_path))
# save the final model
style_model.eval()
style_model.cpu()
save_final_model_path = os.path.join(model_path,
final_model_filename)
torch.save(style_model.state_dict(), save_final_model_path)
logger.info("Done, trained model saved at " +
save_final_model_path)
# Write out the success file
with open(os.path.join(output_path, 'success'), 'w') as s:
s.write('Done')
except Exception as e:
with open(os.path.join(output_path, 'failure'), 'w') as s:
trc = traceback.format_exc()
logger.info('Exception during training: ' + str(e) + '\n' + trc)
s.write('Exception during training: ' + str(e) + '\n' + trc)