def label(self, report_pairs, reports_path, extract_impression):
"""Label the provided report(s)."""
# Load reports in place.
loader = Loader(report_pairs, reports_path, extract_impression)
loader.load()
# Extract observation mentions in place.
self.extractor.extract(loader.collection)
# Classify mentions in place.
self.classifier.classify(loader.collection)
# output mentions/categories/negation/attributes
attributes = self.aggregator.getAttributeOutput(loader.collection)
# # Aggregate mentions to obtain one set of labels for each report.
# labels = aggregator.aggregate(loader.collection)
return loader.reports, attributes
def main(output_dir, model_name, batch_size, num_workers, augment_epoch,
unaugment_epoch, device, label_type, confidence, num_classes,
epochs_per_save, teacher_noise, data_config, model_config):
assert label_type in ["soft", "hard"]
# setup for directory and log
output_dir = os.path.join(output_dir, model_name)
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
os.chmod(output_dir, 0o775)
if hasattr(print, "set_log_dir"):
print.set_log_dir(output_dir)
print(f"==========Settings==========")
print(f"batch size: {batch_size}")
print(f"augment epoch: {augment_epoch}")
print(f"unaugment epoch: {unaugment_epoch}")
print(f"label type: {label_type}")
print(f"confidence: {confidence}")
print(f"models: {model_config['models']}")
print(f"epochs: {model_config['epochs']}")
print(f"ratio: {model_config['ratio']}")
print(f"learning rate: {model_config['learning_rate']}")
print(f"lr_decay_rate: {model_config['lr_decay_rate']}")
print(f"lr_decay_epoch: {model_config['lr_decay_epoch']}")
print(f"N, M: {data_config['N']}, {data_config['M']}")
print(f"teacher noise: {teacher_noise}")
print(f"dropout_prob: {model_config['dropout_prob']}")
print(f"stoc_depth_prob: {model_config['stochastic_depth_prob']}")
print("============================")
logger = Logger(os.path.join(output_dir, 'logs'))
# dataset, dataloader
dataset = load_dataset(**data_config)
dataloaders = {}
dataloaders['train_unaugmented'] = DataLoader(dataset['train_unaugmented'],
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
dataloaders['train_augmented'] = DataLoader(dataset['train_augmented'],
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
dataloaders['test'] = DataLoader(dataset['test'],
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
# load model
loader = Loader(**model_config, device=device, num_classes=num_classes)
(teacher, student, optimizer, lr_scheduler, start_epoch, end_epoch, phase,
best_acc, best_epoch, best_model) = loader.load(0)
model_name = model_config["models"][phase]
ratio = model_config["ratio"][phase]
save_path = os.path.join(output_dir, f"{model_name}.pt")
logger.set_model_name(model_name)
############################
###### Training phase ######
############################
if teacher is None:
dataloaders['train'] = dataloaders[
'train_augmented'] if teacher_noise else dataloaders[
'train_unaugmented']
else:
dataloaders['train'] = NS_DataLoader(dataloaders, dataset, teacher,
device, label_type, confidence,
ratio, num_workers, batch_size,
num_classes, print)
while True:
teacher = train_model(student, optimizer, lr_scheduler, dataloaders,
start_epoch, end_epoch, best_acc, best_epoch,
best_model, device, logger, phase,
epochs_per_save, save_path, augment_epoch,
unaugment_epoch)
phase += 1
if phase < len(model_config["models"]):
(_, student, optimizer, lr_scheduler, start_epoch, end_epoch, _,
best_acc, best_epoch, best_model) = loader.load(phase)
model_name = model_config["models"][phase]
ratio = model_config["ratio"][phase]
logger.set_model_name(model_name)
save_path = os.path.join(output_dir, f"{model_name}.pt")
dataloaders['train'] = NS_DataLoader(dataloaders, dataset, teacher,
device, label_type,
confidence, ratio,
num_workers, batch_size,
num_classes, print)
del teacher
else:
break
