def create_optimizer(args, model: Net):
from torch.optim import Adam
from torch.optim.lr_scheduler import StepLR
opt1 = Adam(model.partial_parameters(True), lr=args.lr_ddr)
opt2 = Adam(model.partial_parameters(False), lr=args.lr)
lrs1 = StepLR(opt1, args.lr_steps, args.lr_gamma)
lrs1.last_epoch = args.start_epoch - 1
lrs2 = StepLR(opt2, args.lr_steps, args.lr_gamma)
lrs2.last_epoch = args.start_epoch - 1
return opt1, opt2, lrs1, lrs2
def train(args):
""" The function to run the training loop.
Args:
dataset: The dataset is provided by ElasticDL for the elastic training.
Now, the dataset if tf.data.Dataset and we need to convert
the data in dataset to torch.tensor. Later, ElasticDL will
pass a torch.utils.data.DataLoader.
elastic_controller: The controller for elastic training.
"""
use_cuda = not args.no_cuda and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
train_data = torchvision.datasets.ImageFolder(args.training_data)
test_data = torchvision.datasets.ImageFolder(args.validation_data)
allreduce_controller = create_elastic_controller(
batch_size=args.batch_size,
dataset_size=len(train_data.imgs),
num_epochs=args.num_epochs,
shuffle=True,
)
train_dataset = ElasticDataset(
train_data.imgs, allreduce_controller.data_shard_service
)
train_loader = DataLoader(
dataset=train_dataset, batch_size=args.batch_size, num_workers=2
)
test_dataset = ElasticDataset(test_data.imgs)
test_loader = DataLoader(
dataset=test_dataset, batch_size=args.batch_size, num_workers=2
)
model = Net()
optimizer = optim.SGD(model.parameters(), lr=args.learning_rate)
optimizer = DistributedOptimizer(optimizer, fixed_global_batch_size=True)
scheduler = StepLR(optimizer, step_size=1, gamma=0.5)
# Set the model and optimizer to broadcast.
allreduce_controller.set_broadcast_model(model)
allreduce_controller.set_broadcast_optimizer(optimizer)
epoch = 0
# Use the elastic function to wrap the training function with a batch.
elastic_train_one_batch = allreduce_controller.elastic_run(train_one_batch)
if torch.cuda.is_available():
model.cuda()
with allreduce_controller.scope():
for batch_idx, (data, target) in enumerate(train_loader):
model.train()
target = target.type(torch.LongTensor)
data, target = data.to(device), target.to(device)
loss = elastic_train_one_batch(model, optimizer, data, target)
print("loss = {}, step = {}".format(loss, batch_idx))
new_epoch = allreduce_controller.get_current_epoch()
if new_epoch > epoch:
epoch = new_epoch
# Set epoch of the scheduler
scheduler.last_epoch = epoch - 1
scheduler.step()
test(model, device, test_loader)
def create_lr_scheduler(optimizer, configs):
"""Create learning rate scheduler for training process"""
if configs.lr_type == 'step_lr':
lr_scheduler = StepLR(optimizer,
step_size=configs.lr_step_size,
gamma=configs.lr_factor)
elif configs.lr_type == 'plateau':
lr_scheduler = ReduceLROnPlateau(optimizer,
factor=configs.lr_factor,
patience=configs.lr_patience)
elif configs.optimizer_type == 'cosin':
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
lf = lambda x: (((1 + math.cos(x * math.pi / configs.num_epochs)) / 2)
**1.0) * 0.9 + 0.1 # cosine
lr_scheduler = LambdaLR(optimizer, lr_lambda=lf)
lr_scheduler.last_epoch = configs.start_epoch - 1 # do not move
# https://discuss.pytorch.org/t/a-problem-occured-when-resuming-an-optimizer/28822
# plot_lr_scheduler(optimizer, scheduler, epochs)
else:
raise TypeError
return lr_scheduler
def _train(train_img_path, train_txt_path, val_img_path, val_txt_path,
path_to_log_dir, path_to_restore_checkpoint_file, training_options):
batch_size = training_options['batch_size']
initial_learning_rate = training_options['learning_rate']
initial_patience = training_options['patience']
num_steps_to_show_loss = 100
num_steps_to_check = 1000
step = 0
patience = initial_patience
best_accuracy = 0.0
duration = 0.0
model = Model(21)
model.cuda()
transform = transforms.Compose([
transforms.Resize([285, 285]),
transforms.ToTensor(),
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
])
train_loader = torch.utils.data.DataLoader(BarcodeDataset(
train_img_path, train_txt_path, transform),
batch_size=batch_size,
shuffle=True,
num_workers=4,
pin_memory=True)
evaluator = Evaluator(val_img_path, val_txt_path)
optimizer = optim.SGD(model.parameters(),
lr=initial_learning_rate,
momentum=0.9,
weight_decay=0.0005)
scheduler = StepLR(optimizer,
step_size=training_options['decay_steps'],
gamma=training_options['decay_rate'])
if path_to_restore_checkpoint_file is not None:
assert os.path.isfile(
path_to_restore_checkpoint_file
), '%s not found' % path_to_restore_checkpoint_file
step = model.restore(path_to_restore_checkpoint_file)
scheduler.last_epoch = step
print('Model restored from file: %s' % path_to_restore_checkpoint_file)
path_to_losses_npy_file = os.path.join(path_to_log_dir, 'losses.npy')
if os.path.isfile(path_to_losses_npy_file):
losses = np.load(path_to_losses_npy_file)
else:
losses = np.empty([0], dtype=np.float32)
while True:
for batch_idx, (images, digits_labels) in enumerate(train_loader):
start_time = time.time()
images, digits_labels = images.cuda(), [
digit_label.cuda() for digit_label in digits_labels
]
digit2_logits, digit3_logits, digit4_logits, digit5_logits, digit6_logits, digit7_logits, digit8_logits, digit9_logits, digit10_logits, digit11_logits, digit12_logits, digit13_logits = model.train(
)(images)
loss = _loss(digit2_logits, digit3_logits, digit4_logits,
digit5_logits, digit6_logits, digit7_logits,
digit8_logits, digit9_logits, digit10_logits,
digit11_logits, digit12_logits, digit13_logits,
digits_labels)
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
step += 1
duration += time.time() - start_time
if step % num_steps_to_show_loss == 0:
examples_per_sec = batch_size * num_steps_to_show_loss / duration
duration = 0.0
print(
'=> %s: step %d, loss = %f, learning_rate = %f (%.1f examples/sec)'
% (datetime.now(), step, loss.item(),
scheduler.get_lr()[0], examples_per_sec))
if step % num_steps_to_check != 0:
continue
losses = np.append(losses, loss.item())
np.save(path_to_losses_npy_file, losses)
print('=> Evaluating on validation dataset...')
accuracy = evaluator.evaluate(model)
print('==> accuracy = %f, best accuracy %f' %
(accuracy, best_accuracy))
if accuracy > best_accuracy:
path_to_checkpoint_file = model.store(path_to_log_dir,
step=step)
print('=> Model saved to file: %s' % path_to_checkpoint_file)
patience = initial_patience
best_accuracy = accuracy
else:
patience -= 1
print('=> patience = %d' % patience)
if patience == 0:
return
def _train(path_to_train_lmdb_dir, path_to_val_lmdb_dir, path_to_log_dir,
path_to_restore_checkpoint_file, training_options, max_steps):
batch_size = training_options['batch_size']
initial_learning_rate = training_options['learning_rate']
initial_patience = training_options['patience']
num_steps_to_show_loss = 100
num_steps_to_check = training_options["validation_interval"]
step = 0
patience = initial_patience
best_accuracy = 0.0
duration = 0.0
model = Model()
model.cuda()
transform = transforms.Compose([
transforms.RandomCrop([54, 54]),
transforms.ToTensor(),
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
])
train_loader = torch.utils.data.DataLoader(Dataset(path_to_train_lmdb_dir, transform),
batch_size=batch_size, shuffle=True,
num_workers=0, pin_memory=True)
evaluator = Evaluator(path_to_val_lmdb_dir)
optimizer = optim.SGD(model.parameters(), lr=initial_learning_rate, momentum=0.9, weight_decay=0.0005)
scheduler = StepLR(optimizer, step_size=training_options['decay_steps'], gamma=training_options['decay_rate'])
if path_to_restore_checkpoint_file is not None:
assert os.path.isfile(path_to_restore_checkpoint_file), '%s not found' % path_to_restore_checkpoint_file
step = model.restore(path_to_restore_checkpoint_file)
scheduler.last_epoch = step
print('Model restored from file: %s' % path_to_restore_checkpoint_file)
path_to_losses_npy_file = os.path.join(path_to_log_dir, 'losses.npy')
if os.path.isfile(path_to_losses_npy_file):
losses = np.load(path_to_losses_npy_file)
else:
losses = np.empty([0], dtype=np.float32)
path_to_test_losses_npy_file = os.path.join(path_to_log_dir, 'test_losses.npy')
if os.path.isfile(path_to_test_losses_npy_file):
test_losses = np.load(path_to_test_losses_npy_file)
else:
test_losses = np.empty([0], dtype=np.float32)
train_loss_array = []
val_loss_array = []
model_checkpoints = []
model_saved = False
# Used to save model (checkpoint) every 2 epochs
model_save_counter = 0
while True:
for batch_idx, (images, length_labels, digits_labels, _) in enumerate(train_loader):
start_time = time.time()
images, length_labels, digits_labels = images.cuda(), length_labels.cuda(), [digit_labels.cuda() for digit_labels in digits_labels]
length_logits, digit1_logits, digit2_logits, digit3_logits, digit4_logits, digit5_logits = model.train()(images)
loss = _loss(length_logits, digit1_logits, digit2_logits, digit3_logits, digit4_logits, digit5_logits, length_labels, digits_labels)
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
step += 1
duration += time.time() - start_time
if step % num_steps_to_show_loss == 0:
examples_per_sec = batch_size * num_steps_to_show_loss / duration
duration = 0.0
print('=> %s: step %d, loss = %f, learning_rate = %f (%.1f examples/sec)' % (
datetime.now(), step, loss.item(), scheduler.get_lr()[0], examples_per_sec))
if step % num_steps_to_check != 0:
continue
model_save_counter += 1
losses = np.append(losses, loss.item())
np.save(path_to_losses_npy_file, losses)
train_loss_array.append((step, loss.item()))
print('=> Evaluating on validation dataset...')
accuracy, test_loss_args = evaluator.evaluate(model)
test_loss = _loss(*test_loss_args)
val_loss_array.append((step, test_loss.item()))
print('==> accuracy = %f, best accuracy %f' % (accuracy, best_accuracy))
# print(f'==> loss = {test_loss}')
# Save model every 2 epochs
if model_save_counter >= 2 or step in [1000, 2000, 3000, 4000, 5000]:
path_to_checkpoint_file = model.store(path_to_log_dir, step=step)
print('=> Model saved to file: %s' % path_to_checkpoint_file)
model_save_counter = 0
model_saved = True
model_checkpoints.append((step, f"model-{step}.pth"))
if accuracy > best_accuracy:
patience = initial_patience
best_accuracy = accuracy
else:
patience -= 1
print("Train losses: ", train_loss_array)
print("Saved Model Checkpoints: ", model_checkpoints)
print('=> patience = %d' % patience)
if patience == 0 or step >= max_steps:
if not model_saved:
path_to_checkpoint_file = model.store(path_to_log_dir, step=step)
print('=> Model MANUALLY saved to file: %s' % path_to_checkpoint_file)
model_checkpoints.append((step, f"model-{step}.pth"))
training_output = {
"model_checkpoints": model_checkpoints,
"train_loss": train_loss_array,
"val_loss": val_loss_array,
}
print("TRAINING OUTPUT -----------------------------")
print(training_output)
return training_output