def run():
df = pd.read_csv(config.TRAIN_PATH)
kfold = KFold(n_splits=5, random_state=config.SEED, shuffle=True)
fold_losses = []
for i, (train_idx, val_idx) in enumerate(kfold.split(df)):
print("-------------------------------------------------------")
print(f"Training fold {i}")
print("-------------------------------------------------------")
train = df.iloc[train_idx]
validation = df.iloc[val_idx]
train_dataset = PicDataset(train)
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=config.BATCH_SIZE
)
val_dataset = PicDataset(validation)
val_data_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=config.BATCH_SIZE
)
device = 'cuda:0' if torch.cuda.is_available() else "cpu"
model = DNN()
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=config.LR)
loss = 0
for _ in range(config.EPOCHS):
engine.train_fn(train_data_loader, model, optimizer, device)
loss = engine.eval_fn(val_data_loader, model, device)
print(f"Loss on fold {i} is {loss}")
fold_losses.append(loss)
torch.save(model.state_dict(), f'./models/model_{i}.bin')
print(f"Average loss on cross validation is {sum(fold_losses) / 5}")
def train(args, config, io):
train_loader, validation_loader = get_loader(args, config)
device = torch.device("cuda" if args.cuda else "cpu")
# print(len(train_loader), len(validation_loader))
#Try to load models
model = DNN(args).to(device)
"""if device == torch.device("cuda"):
model = nn.DataParallel(model)"""
if args.model_path != "":
model.load_state_dict(torch.load(args.model_path))
# for para in list(model.parameters())[:-5]:
# para.requires_grad=False
# print(model)
if args.use_sgd:
# print("Use SGD")
opt = optim.SGD(model.parameters(),
lr=args.lr * 100,
momentum=args.momentum,
weight_decay=1e-4)
else:
# print("Use Adam")
opt = optim.Adam(model.parameters(), lr=args.lr, weight_decay=1e-4)
"""opt = optim.Adam([
{'params': list(model.parameters())[:-1], 'lr':args.lr/50, 'weight_decay': 1e-4},
{'params': list(model.parameters())[-1], 'lr':args.lr, 'weight_decay': 1e-4}
])
"""
scheduler = CosineAnnealingLR(opt, args.epochs, eta_min=args.lr)
criterion = nn.MSELoss()
best_test_loss = 9999999.
for epoch in range(args.epochs):
startTime = time.time()
####################
# Train
####################
train_loss = 0.0
train_dis = 0.0
count = 0.0
model.train()
for data, label in train_loader:
data, label = data.to(device), label.to(device)
data = drop(jitter(data, device), device)
# data = jitter(data, device, delta=0.05)
batch_size = data.shape[0]
logits = model(data)
loss = criterion(logits, label)
opt.zero_grad()
loss.backward()
opt.step()
dis = distance(logits, label)
count += batch_size
train_loss += loss.item() * batch_size
train_dis += dis.item() * batch_size
scheduler.step()
outstr = 'Train %d, loss: %.6f, distance: %.6f' % (
epoch, train_loss * 1.0 / count, train_dis * 1.0 / count)
io.cprint(outstr)
####################
# Evaluation
####################
test_loss = 0.0
test_dis = 0.0
count = 0.0
model.eval()
with torch.no_grad():
for data, label in validation_loader:
data, label = data.to(device), label.to(device)
batch_size = data.shape[0]
logits = model(data)
loss = criterion(logits, label)
dis = distance(logits, label)
count += batch_size
test_loss += loss.item() * batch_size
test_dis += dis.item() * batch_size
outstr = 'Test %d, loss: %.6f, distance: %.6f' % (
epoch, test_loss * 1.0 / count, test_dis * 1.0 / count)
io.cprint(outstr)
if test_loss <= best_test_loss:
best_test_loss = test_loss
torch.save(model.state_dict(),
'checkpoints/%s/models/model.t7' % args.exp_name)
torch.save(model, (config.root + config.model_path))
io.cprint('Time: %.3f sec' % (time.time() - startTime))
# print(correct, total)
loss = criterion(y_pred, y.long())
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (idx + 1) % 100 == 0:
print(
'Epoch: [{}/{}], Step: [{}/{}], Loss: {:.4f}, Acc: {:.4f}'.
format(epoch + 1, args.epochs, idx + 1, len(train_loader),
loss.item(), 100 * correct / total))
# 保存模型参数
torch.save(
model.state_dict(),
os.path.join(
'./log', '{}_{}_{}.ckpt'.format(args.model, args.dataset,
args.epochs)))
model.eval()
with torch.no_grad():
correct = 0
total = 0
for idx, (x, y) in enumerate(test_loader):
x, y = x.to(device), y.to(device)
y_pred = model(x)
_, y_pred = torch.max(y_pred.data, 1)
total += y.size(0)
correct += (y_pred == y).sum().item()
# print(result)
if idx % 100 == 0:
def train(args, config, io):
train_loader, validation_loader, unlabelled_loader = get_loader(
args, config)
device = torch.device("cuda" if args.cuda else "cpu")
#Try to load models
model = DNN(args).to(device)
ema_model = DNN(args).to(device)
for param in ema_model.parameters():
param.detach_()
if device == torch.device("cuda"):
model = nn.DataParallel(model)
ema_model = nn.DataParallel(ema_model)
if args.model_path != "":
model.load_state_dict(torch.load(args.model_path))
ema_model.load_state_dict(torch.load(args.model_path))
if args.use_sgd:
print("Use SGD")
opt = optim.SGD(model.parameters(),
lr=args.lr * 100,
momentum=args.momentum,
weight_decay=1e-4)
else:
print("Use Adam")
opt = optim.Adam(model.parameters(), lr=args.lr, weight_decay=1e-4)
scheduler = CosineAnnealingLR(opt, args.epochs, eta_min=args.lr)
criterion = nn.MSELoss()
consistency_criterion = nn.MSELoss()
best_test_loss = 9999999.
global_step = 0
for epoch in range(args.epochs):
startTime = time.time()
####################
# Train
####################
train_loss = 0.0
count = 0.0
model.train()
ema_model.train()
i = -1
for (data, label), (u, _) in zip(cycle(train_loader),
unlabelled_loader):
i = i + 1
if data.shape[0] != u.shape[0]:
bt_size = np.minimum(data.shape[0], u.shape[0])
data = data[0:bt_size]
label = label[0:bt_size]
u = u[0:bt_size]
data, label, u = data.to(device), label.to(device), u.to(device)
batch_size = data.shape[0]
logits = model(data)
class_loss = criterion(logits, label)
u_student = jitter(u, device)
u_teacher = jitter(u, device)
logits_unlabeled = model(u_student)
ema_logits_unlabeled = ema_model(u_teacher)
ema_logits_unlabeled = Variable(ema_logits_unlabeled.detach().data,
requires_grad=False)
consistency_loss = consistency_criterion(logits_unlabeled,
ema_logits_unlabeled)
if epoch < args.consistency_rampup_starts:
consistency_weight = 0.0
else:
consistency_weight = get_current_consistency_weight(
args, args.final_consistency, epoch, i,
len(unlabelled_loader))
consistency_loss = consistency_weight * consistency_loss
loss = class_loss + consistency_loss
opt.zero_grad()
loss.backward()
opt.step()
global_step += 1
# print(global_step)
update_ema_variables(model, ema_model, args.ema_decay, global_step)
count += batch_size
train_loss += loss.item() * batch_size
scheduler.step()
outstr = 'Train %d, loss: %.6f' % (epoch, train_loss * 1.0 / count)
io.cprint(outstr)
####################
# Evaluation
####################
test_loss = 0.0
count = 0.0
model.eval()
ema_model.eval()
for data, label in validation_loader:
data, label = data.to(device), label.to(device)
batch_size = data.shape[0]
logits = ema_model(data)
loss = criterion(logits, label)
count += batch_size
test_loss += loss.item() * batch_size
outstr = 'Test %d, loss: %.6f' % (epoch, test_loss * 1.0 / count)
io.cprint(outstr)
if test_loss <= best_test_loss:
best_test_loss = test_loss
torch.save(ema_model.state_dict(),
'checkpoints/%s/models/model.t7' % args.exp_name)
torch.save(ema_model, (config.root + config.model_path))
io.cprint('Time: %.3f sec' % (time.time() - startTime))
def main():
print('> Starting execution...')
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
group = parser.add_mutually_exclusive_group()
group.add_argument('--fit',
action='store_true',
help='fit the tuned model on digits 0-4')
group.add_argument('--transfer',
action='store_true',
help='train a pretrained model on digits 5-9')
parser.add_argument('--batch-size',
type=int,
default=256,
metavar='N',
help='input batch size for training (default: 256)')
parser.add_argument('--epochs',
type=int,
default=50,
metavar='E',
help='number of epochs to train (default: 50)')
parser.add_argument('--lr',
type=float,
default=1e-3,
metavar='L',
help='learning rate (default: 1e-3)')
parser.add_argument('--early-stopping',
type=int,
default=7,
metavar='E',
help='early stopping (default: 7 epochs)')
parser.add_argument(
'--size',
type=int,
default=100,
metavar='S',
help='size of the training data for transfer learning (default: 100)')
parser.add_argument('--seed',
type=int,
default=23,
metavar='S',
help='random seed (default: 23)')
args = parser.parse_args()
use_cuda = torch.cuda.is_available() # use cuda if available
device = torch.device("cuda" if use_cuda else "cpu")
torch.manual_seed(args.seed) # random seed
print('> Loading MNIST data')
train_set = datasets.MNIST(MNIST_DATA_DIR,
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
]))
test_set = datasets.MNIST(MNIST_DATA_DIR,
train=False,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
]))
train_digits_04 = np.where(train_set.train_labels < 5)[0]
train_digits_59 = np.where(train_set.train_labels > 4)[0]
test_digits_04 = np.where(test_set.test_labels < 5)[0]
test_digits_59 = np.where(test_set.test_labels > 4)[0]
if args.fit:
# Training the tuned model on digits 0-4
print('> Training a new model on MNIST digits 0-4')
X_train_04, y_train_04, X_valid_04, y_valid_04 = data_to_numpy(
train_set, test_set, INPUT_DIM, train_digits_04, test_digits_04)
torch.manual_seed(args.seed)
print('> Initializing the model')
model = DNN(INPUT_DIM, OUTPUT_DIM, HIDDEN_DIM, batch_norm=True)
model.apply(init_he_normal) # He initialization
model = model.to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=args.lr)
print('> Training the model')
model, _, _ = train_model(model,
device,
X_train_04,
y_train_04,
criterion,
optimizer,
X_valid=X_valid_04,
y_valid=y_valid_04,
batch_size=args.batch_size,
n_epochs=args.epochs,
early_stopping=args.early_stopping)
print(f'> Saving the model state at {MODEL_04_PATH}')
torch.save(model.state_dict(), MODEL_04_PATH)
elif args.transfer:
# Transfer learning
print(
'> Training a model on MNIST digits 5-9 from a pretrained model for digits 0-4'
)
if os.path.isfile(MODEL_04_PATH):
print('> Loading the pretrained model')
model = DNN(INPUT_DIM, OUTPUT_DIM, HIDDEN_DIM,
batch_norm=True).to(device)
model.load_state_dict(torch.load(MODEL_04_PATH))
for param in model.parameters():
param.requires_grad = False
# Parameters of newly constructed modules have requires_grad=True by default
model.fc4 = nn.Linear(HIDDEN_DIM, HIDDEN_DIM)
model.fc5 = nn.Linear(HIDDEN_DIM, HIDDEN_DIM)
model.out = nn.Linear(HIDDEN_DIM, OUTPUT_DIM)
print('> Using saved model state')
else:
print(
'> Model state file is not found, fit a model before the transfer learning'
)
print('> Stopping execution')
return
X_train_59, y_train_59, X_valid_59, y_valid_59 = data_to_numpy(
train_set, test_set, INPUT_DIM, train_digits_59[:args.size],
test_digits_59)
# fixing the issues with labels
y_train_59 = y_train_59 - 5
y_valid_59 = y_valid_59 - 5
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=args.lr)
print('> Training the model')
model, _, _ = train_model(model,
device,
X_train_59,
y_train_59,
criterion,
optimizer,
X_valid=X_valid_59,
y_valid=y_valid_59,
batch_size=args.batch_size,
n_epochs=args.epochs,
early_stopping=args.early_stopping)
print(f'> Saving the model state at {MODEL_59_PATH}')
torch.save(model.state_dict(), MODEL_59_PATH)
else:
print('> Incorrect mode, try either `--fit` or `--transfer`')
print('> Stopping execution')
noisy_batch_var = noisy_batch_var.to(device)
outputs = model(noisy_batch_var)
loss = MSE(outputs, clean_batch_var)
# back-propagate and update
optimizer.zero_grad()
loss.backward()
optimizer.step()
tbwriter.add_scalar('loss', loss.item(), total_steps)
if (i + 1) % 100 == 0:
print(
'Epoch {}\t'
'Step {}\t'
'loss {:.5f}'
.format(epoch + 1, i + 1, loss.item()))
# print(outputs)
# print(clean_batch_var)
# record scalar data for tensorboard
total_steps += 1
# save various states
state_path = os.path.join(checkpoint_path, 'state-{}.pkl'.format(epoch + 1))
state = {
'DNN': model.state_dict(),
'optimizer': optimizer.state_dict(),
}
torch.save(state, state_path)
tbwriter.close()
print('Finished Training!')