def main(args):
distributed, device_ids = main_util.init_distributed_mode(
args.world_size, args.dist_url)
config = yaml_util.load_yaml_file(args.config)
if args.json is not None:
main_util.overwrite_config(config, args.json)
device = torch.device(args.device)
print(args)
print('Loading data')
train_config = config['train']
train_sampler, train_data_loader, val_data_loader, test_data_loader =\
data_util.get_coco_data_loaders(config['dataset'], train_config['batch_size'], distributed)
print('Creating model')
model_config = config['model']
model = get_model(model_config, device)
print('Model Created')
if distributed:
model = nn.parallel.DistributedDataParallel(model,
device_ids=device_ids)
if args.train:
print('Start training')
start_time = time.time()
train(model, train_sampler, train_data_loader, val_data_loader, device,
distributed, config, args, model_config['ckpt'])
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
main_util.evaluate(model, test_data_loader, device=device)
def run(args):
distributed, device_ids = main_util.init_distributed_mode(
args.world_size, args.dist_url)
device = torch.device(args.device if torch.cuda.is_available() else 'cpu')
if device.type == 'cuda':
cudnn.benchmark = True
print(args)
config = yaml_util.load_yaml_file(args.config)
input_shape = config['input_shape']
ckpt_file_path = config['autoencoder']['ckpt']
train_loader, valid_loader, test_loader = main_util.get_data_loaders(
config, distributed)
if not args.test_only:
train(train_loader, valid_loader, input_shape, config, device,
distributed, device_ids)
autoencoder, _ = ae_util.get_autoencoder(config, device)
resume_from_ckpt(ckpt_file_path, autoencoder)
extended_model, model = ae_util.get_extended_model(autoencoder, config,
input_shape, device)
if not args.extended_only:
if device.type == 'cuda':
model = DistributedDataParallel(model, device_ids=device_ids) if distributed \
else DataParallel(model)
evaluate(model, test_loader, device, title='[Original model]')
if device.type == 'cuda':
extended_model = DistributedDataParallel(extended_model, device_ids=device_ids) if distributed \
else DataParallel(extended_model)
evaluate(extended_model, test_loader, device, title='[Mimic model]')
def main(args):
config = yaml_util.load_yaml_file(args.config)
if args.json is not None:
main_util.overwrite_config(config, args.json)
distributed, device_ids = main_util.init_distributed_mode(
args.world_size, args.dist_url)
device = torch.device(args.device if torch.cuda.is_available() else 'cpu')
teacher_model = get_model(config['teacher_model'], device)
module_util.freeze_module_params(teacher_model)
student_model_config = config['student_model']
student_model = get_model(student_model_config, device)
freeze_modules(student_model, student_model_config)
print('Updatable parameters: {}'.format(
module_util.get_updatable_param_names(student_model)))
distill_backbone_only = student_model_config['distill_backbone_only']
train_config = config['train']
train_sampler, train_data_loader, val_data_loader, test_data_loader = \
data_util.get_coco_data_loaders(config['dataset'], train_config['batch_size'], distributed)
if distributed:
teacher_model = DataParallel(teacher_model, device_ids=device_ids)
student_model = DistributedDataParallel(student_model,
device_ids=device_ids)
if args.distill:
distill(teacher_model, student_model, train_sampler, train_data_loader,
val_data_loader, device, distributed, distill_backbone_only,
config, args)
load_ckpt(
config['student_model']['ckpt'],
model=student_model.module if isinstance(
student_model, DistributedDataParallel) else student_model)
evaluate(teacher_model, student_model, test_data_loader, device,
args.skip_teacher_eval, args.transform_bottleneck)
def run(args):
distributed, device_ids = main_util.init_distributed_mode(
args.world_size, args.dist_url)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
if device.type == 'cuda':
cudnn.benchmark = True
print(args)
config = yaml_util.load_yaml_file(args.config)
train_loader, valid_loader, test_loader = main_util.get_data_loaders(
config, distributed)
if 'mimic_model' in config:
model = mimic_util.get_mimic_model_easily(config, device)
model_config = config['mimic_model']
else:
model = module_util.get_model(config, device)
model_config = config['model']
model_type, best_valid_acc, start_epoch, ckpt_file_path =\
module_util.resume_from_ckpt(model, model_config, args.init)
train_config = config['train']
criterion_config = train_config['criterion']
criterion = func_util.get_loss(criterion_config['type'],
criterion_config['params'])
if not args.evaluate:
train(model, train_loader, valid_loader, best_valid_acc, criterion,
device, distributed, device_ids, train_config, args.epoch,
start_epoch, args.lr, ckpt_file_path, model_type)
test(model, test_loader, device)
def main(args):
distributed, device_ids = main_util.init_distributed_mode(
args.world_size, args.dist_url)
config = yaml_util.load_yaml_file(args.config)
if args.json is not None:
main_util.overwrite_config(config, args.json)
device = torch.device(args.device)
print(args)
print('Loading data')
train_config = config['train']
train_sampler, train_data_loader, val_data_loader, test_data_loader =\
data_util.get_coco_data_loaders(config['dataset'], train_config['batch_size'], distributed)
print('Creating model')
model_config = config['model']
model = get_model(model_config, device, strict=False)
module_util.freeze_module_params(model)
ext_classifier = model.get_ext_classifier()
module_util.unfreeze_module_params(ext_classifier)
print('Updatable parameters: {}'.format(
module_util.get_updatable_param_names(model)))
model.train_ext()
if distributed:
model = nn.parallel.DistributedDataParallel(model,
device_ids=device_ids)
if args.train:
print('Start training')
start_time = time.time()
ckpt_file_path = model_config['backbone']['ext_config']['ckpt']
train(model, ext_classifier, train_sampler, train_data_loader,
val_data_loader, device, distributed, config, args,
ckpt_file_path)
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
load_ckpt(ckpt_file_path, model=ext_classifier)
evaluate(model,
test_data_loader,
device=device,
min_recall=args.min_recall,
split_name='Test')
def run(args):
distributed, device_ids = main_util.init_distributed_mode(
args.world_size, args.dist_url)
device = torch.device(args.device if torch.cuda.is_available() else 'cpu')
if torch.cuda.is_available():
cudnn.benchmark = True
print(args)
config = yaml_util.load_yaml_file(args.config)
dataset_config = config['dataset']
input_shape = config['input_shape']
train_config = config['train']
test_config = config['test']
train_loader, valid_loader, test_loader =\
dataset_util.get_data_loaders(dataset_config, batch_size=train_config['batch_size'],
rough_size=train_config['rough_size'], reshape_size=input_shape[1:3],
test_batch_size=test_config['batch_size'], jpeg_quality=-1,
distributed=distributed)
teacher_model_config = config['teacher_model']
if not args.test_only:
distill(train_loader, valid_loader, input_shape, args.aux, config,
device, distributed, device_ids)
org_model, teacher_model_type = mimic_util.get_org_model(
teacher_model_config, device)
if not args.student_only:
if distributed:
org_model = DataParallel(org_model, device_ids=device_ids)
evaluate(org_model, test_loader, device, title='[Original model]')
mimic_model = mimic_util.get_mimic_model(config, org_model,
teacher_model_type,
teacher_model_config, device)
mimic_model_without_dp = mimic_model.module if isinstance(
mimic_model, DataParallel) else mimic_model
file_util.save_pickle(mimic_model_without_dp,
config['mimic_model']['ckpt'])
if distributed:
mimic_model = DistributedDataParallel(mimic_model_without_dp,
device_ids=device_ids)
evaluate(mimic_model, test_loader, device, title='[Mimic model]')
def main(args):
config = yaml_util.load_yaml_file(args.config)
if args.json is not None:
main_util.overwrite_config(config, args.json)
distributed, device_ids = main_util.init_distributed_mode(args.world_size, args.dist_url)
device = torch.device(args.device if torch.cuda.is_available() else 'cpu')
teacher_model = get_model(config['teacher_model'], device)
module_util.freeze_module_params(teacher_model)
student_model_config = config['student_model']
student_model = get_model(student_model_config, device)
freeze_modules(student_model, student_model_config)
ckpt_file_path = config['student_model']['ckpt']
train_config = config['train']
optim_config = train_config['optimizer']
optimizer = func_util.get_optimizer(student_model, optim_config['type'], optim_config['params'])
scheduler_config = train_config['scheduler']
lr_scheduler = func_util.get_scheduler(optimizer, scheduler_config['type'], scheduler_config['params'])
if file_util.check_if_exists(ckpt_file_path):
best_val_map, _, _ = load_ckpt(ckpt_file_path, optimizer=optimizer, lr_scheduler=lr_scheduler)
save_ckpt(student_model, optimizer, lr_scheduler, best_val_map, config, args, ckpt_file_path)
def main(args):
if args.apex:
if sys.version_info < (3, 0):
raise RuntimeError(
'Apex currently only supports Python 3. Aborting.')
if amp is None:
raise RuntimeError(
'Failed to import apex. Please install apex from https://www.github.com/nvidia/apex '
'to enable mixed-precision training.')
distributed, device_ids = main_util.init_distributed_mode(
args.world_size, args.dist_url)
print(args)
if torch.cuda.is_available():
torch.backends.cudnn.benchmark = True
config = yaml_util.load_yaml_file(args.config)
device = torch.device(args.device if torch.cuda.is_available() else 'cpu')
dataset_config = config['dataset']
input_shape = config['input_shape']
train_config = config['train']
test_config = config['test']
train_data_loader, val_data_loader, test_data_loader =\
dataset_util.get_data_loaders(dataset_config, batch_size=train_config['batch_size'],
rough_size=train_config['rough_size'], reshape_size=input_shape[1:3],
jpeg_quality=-1, test_batch_size=test_config['batch_size'],
distributed=distributed)
teacher_model_config = config['teacher_model']
teacher_model, teacher_model_type = mimic_util.get_org_model(
teacher_model_config, device)
module_util.freeze_module_params(teacher_model)
student_model = mimic_util.get_mimic_model_easily(config, device)
student_model_config = config['mimic_model']
optim_config = train_config['optimizer']
optimizer = func_util.get_optimizer(student_model, optim_config['type'],
optim_config['params'])
use_apex = args.apex
if use_apex:
student_model, optimizer = amp.initialize(
student_model, optimizer, opt_level=args.apex_opt_level)
if distributed:
teacher_model = DataParallel(teacher_model, device_ids=device_ids)
student_model = DistributedDataParallel(student_model,
device_ids=device_ids)
start_epoch = args.start_epoch
if not args.test_only:
distill(teacher_model, student_model, train_data_loader,
val_data_loader, device, distributed, start_epoch, config,
args)
student_model_without_ddp =\
student_model.module if isinstance(student_model, DistributedDataParallel) else student_model
load_ckpt(student_model_config['ckpt'],
model=student_model_without_ddp,
strict=True)
if not args.student_only:
evaluate(teacher_model,
test_data_loader,
device,
title='[Teacher: {}]'.format(teacher_model_type))
evaluate(student_model,
test_data_loader,
device,
title='[Student: {}]'.format(student_model_config['type']))
