def main(args):
log_file_path = args.log
if is_main_process() and log_file_path is not None:
setup_log_file(os.path.expanduser(log_file_path))
distributed, device_ids = init_distributed_mode(args.world_size,
args.dist_url)
logger.info(args)
cudnn.benchmark = True
set_seed(args.seed)
config = yaml_util.load_yaml_file(os.path.expanduser(args.config))
device = torch.device(args.device)
dataset_dict = util.get_all_datasets(config['datasets'])
models_config = config['models']
teacher_model_config = models_config.get('teacher_model', None)
teacher_model = load_model(
teacher_model_config,
device) if teacher_model_config is not None else None
student_model_config =\
models_config['student_model'] if 'student_model' in models_config else models_config['model']
ckpt_file_path = student_model_config['ckpt']
student_model = load_model(student_model_config, device)
if args.log_config:
logger.info(config)
if not args.test_only:
train(teacher_model, student_model, dataset_dict, ckpt_file_path,
device, device_ids, distributed, config, args)
student_model_without_ddp =\
student_model.module if module_util.check_if_wrapped(student_model) else student_model
load_ckpt(student_model_config['ckpt'],
model=student_model_without_ddp,
strict=True)
test_config = config['test']
test_data_loader_config = test_config['test_data_loader']
test_data_loader = util.build_data_loader(
dataset_dict[test_data_loader_config['dataset_id']],
test_data_loader_config, distributed)
log_freq = test_config.get('log_freq', 1000)
iou_types = args.iou_types
if not args.student_only and teacher_model is not None:
evaluate(teacher_model,
test_data_loader,
iou_types,
device,
device_ids,
distributed,
log_freq=log_freq,
title='[Teacher: {}]'.format(teacher_model_config['name']))
evaluate(student_model,
test_data_loader,
iou_types,
device,
device_ids,
distributed,
log_freq=log_freq,
title='[Student: {}]'.format(student_model_config['name']))
def main(args):
log_file_path = args.log
if is_main_process() and log_file_path is not None:
setup_log_file(os.path.expanduser(log_file_path))
distributed, device_ids = init_distributed_mode(args.world_size,
args.dist_url)
logger.info(args)
cudnn.benchmark = True
config = yaml_util.load_yaml_file(os.path.expanduser(args.config))
device = torch.device(args.device)
dataset_dict = util.get_all_dataset(config['datasets'])
models_config = config['models']
teacher_model_config = models_config['teacher_model']
teacher_model = get_model(teacher_model_config, device, distributed, False)
student_model_config = models_config['student_model']
student_model = get_model(student_model_config, device, distributed,
args.sync_bn)
if not args.test_only:
distill(teacher_model, student_model, dataset_dict, device, device_ids,
distributed, config, args)
student_model_without_ddp =\
student_model.module if module_util.check_if_wrapped(student_model) else student_model
load_ckpt(student_model_config['ckpt'],
model=student_model_without_ddp,
strict=True)
test_config = config['test']
test_data_loader_config = test_config['test_data_loader']
test_data_loader = util.build_data_loader(
dataset_dict[test_data_loader_config['dataset_id']],
test_data_loader_config, distributed)
if not args.student_only:
evaluate(teacher_model,
test_data_loader,
device,
device_ids,
distributed,
title='[Teacher: {}]'.format(teacher_model_config['name']))
evaluate(student_model,
test_data_loader,
device,
device_ids,
distributed,
title='[Student: {}]'.format(student_model_config['name']))
def main(args):
log_file_path = args.log
if is_main_process() and log_file_path is not None:
setup_log_file(os.path.expanduser(log_file_path))
world_size = args.world_size
logger.info(args)
cudnn.benchmark = True
set_seed(args.seed)
config = yaml_util.load_yaml_file(os.path.expanduser(args.config))
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
distributed = accelerator.state.distributed_type == DistributedType.MULTI_GPU
device_ids = [accelerator.device.index]
if distributed:
setup_for_distributed(is_main_process())
logger.info(accelerator.state)
device = accelerator.device
# Setup logging, we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(
logging.INFO if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# Load pretrained model and tokenizer
task_name = args.task_name
models_config = config['models']
teacher_model_config = models_config.get('teacher_model', None)
teacher_tokenizer, teacher_model = (None, None) if teacher_model_config is None \
else load_tokenizer_and_model(teacher_model_config, task_name, True)
student_model_config =\
models_config['student_model'] if 'student_model' in models_config else models_config['model']
student_tokenizer, student_model = load_tokenizer_and_model(
student_model_config, task_name)
ckpt_dir_path = student_model_config['ckpt']
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or specify a GLUE benchmark task (the dataset will be downloaded automatically from the datasets Hub).
dataset_dict, label_names_dict, is_regression = \
get_all_datasets(config['datasets'], task_name, student_tokenizer, student_model)
# Update config with dataset size len(data_loader)
customize_lr_config(config, dataset_dict, world_size)
# register collate function
register_collate_func(
DataCollatorWithPadding(
student_tokenizer,
pad_to_multiple_of=(8 if accelerator.use_fp16 else None)))
# Get the metric function
metric = get_metrics(task_name)
if not args.test_only:
train(teacher_model, student_model, dataset_dict, is_regression,
ckpt_dir_path, metric, device, device_ids, distributed, config,
args, accelerator)
student_tokenizer.save_pretrained(ckpt_dir_path)
test_config = config['test']
test_data_loader_config = test_config['test_data_loader']
test_data_loader = util.build_data_loader(
dataset_dict[test_data_loader_config['dataset_id']],
test_data_loader_config, distributed)
test_data_loader = accelerator.prepare(test_data_loader)
if not args.student_only and teacher_model is not None:
teacher_model = teacher_model.to(accelerator.device)
evaluate(teacher_model,
test_data_loader,
metric,
is_regression,
accelerator,
title='[Teacher: {}]'.format(teacher_model_config['name']))
# Reload best checkpoint based on validation result
student_tokenizer, student_model = load_tokenizer_and_model(
student_model_config, task_name, True)
student_model = accelerator.prepare(student_model)
evaluate(student_model,
test_data_loader,
metric,
is_regression,
accelerator,
title='[Student: {}]'.format(student_model_config['name']))
# Output prediction for private dataset(s) if both the config and output dir path are given
private_configs = config.get('private', None)
private_output_dir_path = args.private_output
if private_configs is not None and private_output_dir_path is not None and is_main_process(
):
predict_private(student_model, dataset_dict, label_names_dict,
is_regression, accelerator, private_configs,
private_output_dir_path)