def main(argv):
"""Run the training script with command line arguments @argv."""
args = parse_args(argv)
utils.general_setup(args.save, args.gpus)
logging.info("Arguments parsed.\n{}".format(pprint.pformat(vars(args))))
# Create the train and the validation data loaders.
train_loader = imagenet.get_train_loader(args.imagenet, args.batch_size,
args.num_workers, args.image_size)
val_loader = imagenet.get_val_loader(args.imagenet, args.batch_size,
args.num_workers, args.image_size)
# Create model with optional teachers.
model, loss = model_factory.create_model(
args.model, args.student_state_file, args.gpus, args.teacher_model,
args.teacher_state_file, False)
logging.info("Model:\n{}".format(model))
discriminator_loss, update_parameters = create_discriminator_criterion(args)
if args.lr_regime is None:
lr_regime = model.LR_REGIME
else:
lr_regime = args.lr_regime
regime = LearningRateRegime(lr_regime)
# Train and test for needed number of epochs.
optimizer = create_optimizer(model, update_parameters, args.momentum, args.weight_decay)
for epoch in range(args.start_epoch, args.epochs):
lr = regime.get_lr(epoch)
_set_learning_rate(optimizer, lr)
train_for_one_epoch(model, loss, discriminator_loss, train_loader, optimizer, epoch, args)
test.test_for_one_epoch(model, loss, val_loader, epoch)
save_checkpoint(args.save, model, optimizer, epoch)
def main(argv):
"""Run the training script with command line arguments @argv."""
args = parse_args(argv)
utils.general_setup(args.save, args.gpus)
logging.info("Arguments parsed.\n{}".format(pprint.pformat(vars(args))))
# convert to TRUE number of loading-images since we use multiple crops from the same image within a minbatch
args.batch_size = math.ceil(args.batch_size / args.num_crops)
# Create the train and the validation data loaders.
train_loader = imagenet.get_train_loader_FKD(
args.imagenet, args.batch_size, args.num_workers, args.image_size,
args.num_crops, args.softlabel_path)
val_loader = imagenet.get_val_loader(args.imagenet, args.batch_size,
args.num_workers, args.image_size)
# Create model with optional teachers.
model, loss = model_factory.create_model(args.model,
args.student_state_file,
args.gpus, args.teacher_model,
args.teacher_state_file, True)
logging.info("Model:\n{}".format(model))
discriminator_loss, update_parameters = create_discriminator_criterion(
args)
optimizer = create_optimizer(model, update_parameters, args.momentum,
args.weight_decay)
for epoch in range(args.start_epoch, args.epochs, args.num_crops):
adjust_learning_rate(optimizer, epoch, args)
train_for_one_epoch(model, loss, discriminator_loss, train_loader,
optimizer, epoch, args)
test.test_for_one_epoch(model, loss, val_loader, epoch)
save_checkpoint(args.save, model, optimizer, epoch)
def main(argv):
"""Run the training script with command line arguments @argv."""
args = parse_args(argv)
utils.general_setup(args.save, args.gpus)
logging.info("Arguments parsed.\n{}".format(pprint.pformat(vars(args))))
# used the same imagenet loader for the first run with the image folders
# Create the train and the validation data loaders.
# train_loader = imagenet.get_train_loader(args.imagenet, args.batch_size,
# args.num_workers, args.image_size)
# val_loader = imagenet.get_val_loader(args.imagenet, args.batch_size,
# args.num_workers, args.image_size)
train_loader = AliProductDataLoader(
data_dir= '/media/School/Datasets/AliProducts/',
data_list='/media/School/Datasets/AliProducts/lists/train_list_250.txt',
batch_size=args.batch_size,
image_size=args.image_size,
sample='balance',
num_workers=args.num_workers
)
val_loader = AliProductDataLoader(
data_dir= '/media/School/Datasets/AliProducts/',
data_list='/media/School/Datasets/AliProducts/lists/new_valid_list.txt',
batch_size=args.batch_size,
image_size=args.image_size,
num_workers=args.num_workers
)
# Create model with optional teachers.
model, loss = model_factory.create_model(
args.model, args.student_state_file, args.gpus, args.teacher_model,
args.teacher_state_file, load_from_file=True)
logging.info("Model:\n{}".format(model))
discriminator_loss, update_parameters = create_discriminator_criterion(args)
if args.lr_regime is None:
lr_regime = model.LR_REGIME
else:
lr_regime = args.lr_regime
regime = LearningRateRegime(lr_regime)
# Train and test for needed number of epochs.
optimizer = create_optimizer(model, update_parameters, args.momentum, args.weight_decay)
for epoch in range(args.start_epoch, args.epochs):
lr = regime.get_lr(epoch)
_set_learning_rate(optimizer, lr)
train_for_one_epoch(model, loss, discriminator_loss, train_loader, optimizer, epoch)
test.test_for_one_epoch(model, loss, val_loader, epoch)
save_checkpoint(args.save, model, optimizer, epoch)
def main2(argv):
"""Run the training script with command line arguments @argv."""
args = parse_args(argv)
utils.general_setup(args.save, args.gpus)
logging.info("Arguments parsed.\n{}".format(pprint.pformat(vars(args))))
# Create the train and the validation data loaders.
train_loader = torch.utils.data.DataLoader(DataLmdb(
"/kaggle/working/Low_Test/Train-Low_lmdb",
db_size=1464004,
crop_size=128,
flip=True,
scale=0.00390625),
batch_size=256,
shuffle=True)
val_loader = torch.utils.data.DataLoader(DataLmdb(
"/kaggle/working/Low_Test/Valid-Low_lmdb",
db_size=6831,
crop_size=128,
flip=False,
scale=0.00390625,
random=False),
batch_size=256,
shuffle=False)
# Create model with optional teachers.
model, loss = model_factory.create_model(args.model,
args.student_state_file,
args.gpus, args.teacher_model,
args.teacher_state_file)
# logging.info("Model:\n{}".format(model))
discriminator_loss, update_parameters = create_discriminator_criterion(
args)
if args.lr_regime is None:
lr_regime = model.LR_REGIME
else:
lr_regime = args.lr_regime
regime = LearningRateRegime(lr_regime)
# Train and test for needed number of epochs.
optimizer = create_optimizer(model, update_parameters, args.momentum,
args.weight_decay)
optimizer.load_state_dict(torch.load('models/optim_state_03.pytar'))
for epoch in range(args.start_epoch, args.epochs):
lr = regime.get_lr(epoch)
_set_learning_rate(optimizer, lr)
train_for_one_epoch(model, loss, discriminator_loss, train_loader,
optimizer, epoch)
test.test_for_one_epoch(model, loss, val_loader, epoch)
save_checkpoint(args.save, model, optimizer, epoch)
def main(argv):
"""Run the training script with command line arguments @argv."""
args = parse_args(argv)
if args.label_refinery_model is None:
if args.coslinear:
save_dir = args.save+'/'+args.model+'_cos_'+name_time
else:
save_dir = args.save+'/'+args.model+'_'+name_time
else:
if args.coslinear:
save_dir = args.save+'/'+args.model+'_cos_rfn_'+name_time
else:
save_dir = args.save+'/'+args.model+'_rfn_'+name_time
if not os.path.isdir(save_dir):
os.mkdir(save_dir)
utils.general_setup(save_dir, args.gpus)
logging.info("Arguments parsed.\n{}".format(pprint.pformat(vars(args))))
train_loader = mul_cifar100.mul_CIFAR100DataLoader(root=args.data_dir,
image_size=32, train=True, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers)
val_loader = mul_cifar100.mul_CIFAR100DataLoader(root=args.data_dir,
image_size=32, train=False, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers)
# Create model with optional label refinery.
model, loss = model_factory.create_model(
args.model, args.model_state_file, args.gpus, args.label_refinery_model,
args.label_refinery_state_file, args.coslinear, args.s)
if args.lr_regime is None:
lr_regime = model.LR_REGIME
else:
lr_regime = args.lr_regime
regime = LearningRateRegime(lr_regime)
# Train and test for needed number of epochs.
optimizer = create_optimizer(model, args.momentum, args.weight_decay)
for epoch in range(1, regime.num_epochs + 1):
lr = regime.get_lr(epoch)
_set_learning_rate(optimizer, lr)
train_for_one_epoch(model, loss, train_loader, optimizer, epoch)
test.test_for_one_epoch(model, loss, val_loader, epoch)
save_checkpoint(save_dir, model, optimizer, epoch)
def main(argv):
"""Run the training script with command line arguments @argv."""
args = parse_args(argv)
if args.label_refinery_model is None:
if args.coslinear:
save_dir = args.save + '/' + args.model + '_cos_' + name_time
else:
save_dir = args.save + '/' + args.model + '_' + name_time
else:
if args.coslinear:
save_dir = args.save + '/' + args.model + '_cos_rfn_' + name_time
else:
save_dir = args.save + '/' + args.model + '_rfn_' + name_time
if not os.path.isdir(save_dir):
os.mkdir(save_dir)
utils.general_setup(save_dir, args.gpus)
logging.info("Arguments parsed.\n{}".format(pprint.pformat(vars(args))))
# Create the train and the validation data loaders.
train_loader = mul_cifar100.mul_CIFAR100DataLoader(
root=args.data_dir,
image_size=32,
train=True,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers)
val_loader = mul_cifar100.mul_CIFAR100DataLoader(
root=args.data_dir,
image_size=32,
train=False,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers)
# Create model with optional label refinery.
model, loss = model_factory.create_model(args.model, args.model_state_file,
args.gpus,
args.label_refinery_model,
args.label_refinery_state_file,
args.coslinear, args.s)
# logging.info("Model:\n{}".format(model))
if args.lr_regime is None:
lr_regime = model.LR_REGIME
else:
lr_regime = args.lr_regime
regime = LearningRateRegime(lr_regime)
# Train and test for needed number of epochs.
optimizer = create_optimizer(model, args.momentum, args.weight_decay)
for epoch in range(1, int(regime.num_epochs) + 1):
lr = regime.get_lr(epoch)
_set_learning_rate(optimizer, lr)
train_for_one_epoch(model, loss, train_loader, optimizer, epoch)
############# Print results ########
# weights = [ p for n,p in model.named_parameters() if 'weight' in n and 'se' not in n and 'conv' in n and len(p.size())==4]
# name = [ n for n,p in model.named_parameters() if 'weight' in n and 'se' not in n and 'conv' in n and len(p.size())==4]
# j = 0
# for wt in weights:
# zr_ch = 0
# rs = wt.pow(2).sum(dim=[0,2,3]).pow(1/2.)
# for i in range(len(rs)):
# if rs[i]<=1e-15:
# zr_ch += 1
# csize = list(wt.size())
# num_ch = csize[1]
# print('Number of zero channels: '+str(zr_ch)+'/'+str(num_ch)+' in :'+name[j])
# j += 1
####################################
test.test_for_one_epoch(model, loss, val_loader, epoch)
save_checkpoint(save_dir, model, optimizer, epoch)