def main():
"""
Main body of script.
"""
args = parse_args()
args.seed = init_rand(seed=args.seed)
_, log_file_exist = initialize_logging(
logging_dir_path=args.save_dir,
logging_file_name=args.logging_file_name,
script_args=args,
log_packages=args.log_packages,
log_pip_packages=args.log_pip_packages)
use_cuda, batch_size = prepare_pt_context(num_gpus=args.num_gpus,
batch_size=args.batch_size)
net = prepare_model(model_name=args.model,
use_pretrained=args.use_pretrained,
pretrained_model_file_path=args.resume.strip(),
use_cuda=use_cuda)
real_net = net.module if hasattr(net, "module") else net
assert (hasattr(real_net, "num_classes"))
num_classes = real_net.num_classes
ds_metainfo = get_dataset_metainfo(dataset_name=args.dataset)
ds_metainfo.update(args=args)
train_data = get_train_data_source(ds_metainfo=ds_metainfo,
batch_size=batch_size,
num_workers=args.num_workers)
val_data = get_val_data_source(ds_metainfo=ds_metainfo,
batch_size=batch_size,
num_workers=args.num_workers)
optimizer, lr_scheduler, start_epoch = prepare_trainer(
net=net,
optimizer_name=args.optimizer_name,
wd=args.wd,
momentum=args.momentum,
lr_mode=args.lr_mode,
lr=args.lr,
lr_decay_period=args.lr_decay_period,
lr_decay_epoch=args.lr_decay_epoch,
lr_decay=args.lr_decay,
num_epochs=args.num_epochs,
state_file_path=args.resume_state)
if args.save_dir and args.save_interval:
param_names = ds_metainfo.val_metric_capts + ds_metainfo.train_metric_capts + [
"Train.Loss", "LR"
]
lp_saver = TrainLogParamSaver(
checkpoint_file_name_prefix="{}_{}".format(ds_metainfo.short_label,
args.model),
last_checkpoint_file_name_suffix="last",
best_checkpoint_file_name_suffix=None,
last_checkpoint_dir_path=args.save_dir,
best_checkpoint_dir_path=None,
last_checkpoint_file_count=2,
best_checkpoint_file_count=2,
checkpoint_file_save_callback=save_params,
checkpoint_file_exts=(".pth", ".states"),
save_interval=args.save_interval,
num_epochs=args.num_epochs,
param_names=param_names,
acc_ind=ds_metainfo.saver_acc_ind,
# bigger=[True],
# mask=None,
score_log_file_path=os.path.join(args.save_dir, "score.log"),
score_log_attempt_value=args.attempt,
best_map_log_file_path=os.path.join(args.save_dir, "best_map.log"))
else:
lp_saver = None
train_net(batch_size=batch_size,
num_epochs=args.num_epochs,
start_epoch1=args.start_epoch,
train_data=train_data,
val_data=val_data,
net=net,
optimizer=optimizer,
lr_scheduler=lr_scheduler,
lp_saver=lp_saver,
log_interval=args.log_interval,
num_classes=num_classes,
val_metric=get_composite_metric(
ds_metainfo.val_metric_names,
ds_metainfo.val_metric_extra_kwargs),
train_metric=get_composite_metric(
ds_metainfo.train_metric_names,
ds_metainfo.train_metric_extra_kwargs),
use_cuda=use_cuda)
def main():
"""
Main body of script.
"""
args = parse_args()
args.seed = init_rand(seed=args.seed)
# args.num_gpus = 1
# args.num_epochs = 500
# args.dataset = "CIFAR10"
# args.model = "resnet20_cifar10"
args.use_pretrained = False
num_non_res = 3
pretrained_model_file_path = '/exdrive/resnet20-cifar10/unfrozen-training/by-stack/non-res-stack-3-models/cifar10-non-res-stack-2-teacher.pth'
_, log_file_exist = initialize_logging(
logging_dir_path=args.save_dir,
logging_file_name=args.logging_file_name,
script_args=args,
log_packages=args.log_packages,
log_pip_packages=args.log_pip_packages)
use_cuda, batch_size = prepare_pt_context(num_gpus=args.num_gpus,
batch_size=args.batch_size)
# Get model
net = prepare_model(
model_name=args.model,
use_pretrained=args.use_pretrained,
# pretrained_model_file_path=pretrained_model_file_path,
pretrained_model_file_path=args.resume.strip(),
use_cuda=use_cuda,
remove_module=True
) # True if using our own trained model waits; o.w. False
real_net = net.module if hasattr(net, "module") else net
assert (hasattr(real_net, "num_classes"))
num_classes = real_net.num_classes
ds_metainfo = get_dataset_metainfo(dataset_name=args.dataset)
ds_metainfo.update(args=args)
train_data = get_train_data_source(ds_metainfo=ds_metainfo,
batch_size=batch_size,
num_workers=args.num_workers)
val_data = get_val_data_source(ds_metainfo=ds_metainfo,
batch_size=batch_size,
num_workers=args.num_workers)
optimizer, lr_scheduler, start_epoch = prepare_trainer(
net=net,
optimizer_name=args.optimizer_name,
wd=args.wd,
momentum=args.momentum,
lr_mode=args.lr_mode,
lr=args.lr,
lr_decay_period=args.lr_decay_period,
lr_decay_epoch=args.lr_decay_epoch,
lr_decay=args.lr_decay,
num_epochs=args.num_epochs,
state_file_path=args.resume_state)
if args.save_dir and args.save_interval:
param_names = ds_metainfo.val_metric_capts + ds_metainfo.train_metric_capts + [
"Train.Loss", "LR"
]
lp_saver = TrainLogParamSaver(
checkpoint_file_name_prefix="{}_{}".format(ds_metainfo.short_label,
args.model),
last_checkpoint_file_name_suffix="last",
best_checkpoint_file_name_suffix=None,
last_checkpoint_dir_path=args.save_dir,
best_checkpoint_dir_path=None,
last_checkpoint_file_count=2,
best_checkpoint_file_count=2,
checkpoint_file_save_callback=save_params,
checkpoint_file_exts=(".pth", ".states"),
save_interval=args.save_interval,
num_epochs=args.num_epochs,
param_names=param_names,
acc_ind=ds_metainfo.saver_acc_ind,
# bigger=[True],
# mask=None,
score_log_file_path=os.path.join(args.save_dir, "score.log"),
score_log_attempt_value=args.attempt,
best_map_log_file_path=os.path.join(args.save_dir, "best_map.log"))
else:
lp_saver = None
print(net)
# summary(net, (3, 32, 32))
if torch.cuda.is_available():
print('cuda available - sending net to gpu')
print('training using cuda = ', use_cuda)
net.cuda()
print('\n\nTraining nonresnet20 on cifar10 with num_non_res=', num_non_res)
train_net(batch_size=batch_size,
num_epochs=args.num_epochs,
start_epoch1=args.start_epoch,
train_data=train_data,
val_data=val_data,
net=net,
optimizer=optimizer,
lr_scheduler=lr_scheduler,
lp_saver=lp_saver,
log_interval=args.log_interval,
num_classes=num_classes,
val_metric=get_composite_metric(
ds_metainfo.val_metric_names,
ds_metainfo.val_metric_extra_kwargs),
train_metric=get_composite_metric(
ds_metainfo.train_metric_names,
ds_metainfo.train_metric_extra_kwargs),
use_cuda=use_cuda)