def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
finetune_opts(parser)
parser.add_argument(
"--max_choices_num",
default=4,
type=int,
help=
"The maximum number of cadicate answer, shorter than this will be padded."
)
parser.add_argument(
"--tokenizer",
choices=["bert", "char", "space"],
default="bert",
help="Specify the tokenizer."
"Original Google BERT uses bert tokenizer on Chinese corpus."
"Char tokenizer segments sentences into characters."
"Space tokenizer segments sentences into words according to space.")
args = parser.parse_args()
args.labels_num = args.max_choices_num
# Load the hyperparameters from the config file.
args = load_hyperparam(args)
set_seed(args.seed)
# Build tokenizer.
args.tokenizer = str2tokenizer[args.tokenizer](args)
# Build multiple choice model.
model = MultipleChoice(args)
# Load or initialize parameters.
load_or_initialize_parameters(args, model)
args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(args.device)
# Training phase.
trainset = read_dataset(args, args.train_path)
random.shuffle(trainset)
instances_num = len(trainset)
batch_size = args.batch_size
src = torch.LongTensor([example[0] for example in trainset])
tgt = torch.LongTensor([example[1] for example in trainset])
seg = torch.LongTensor([example[2] for example in trainset])
args.train_steps = int(instances_num * args.epochs_num / batch_size) + 1
print("Batch size: ", batch_size)
print("The number of training instances:", instances_num)
optimizer, scheduler = build_optimizer(args, model)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
args.amp = amp
if torch.cuda.device_count() > 1:
print("{} GPUs are available. Let's use them.".format(
torch.cuda.device_count()))
model = torch.nn.DataParallel(model)
args.model = model
total_loss, result, best_result = 0.0, 0.0, 0.0
print("Start training.")
for epoch in range(1, args.epochs_num + 1):
model.train()
for i, (src_batch, tgt_batch, seg_batch,
_) in enumerate(batch_loader(batch_size, src, tgt, seg)):
loss = train_model(args, model, optimizer, scheduler, src_batch,
tgt_batch, seg_batch)
total_loss += loss.item()
if (i + 1) % args.report_steps == 0:
print("Epoch id: {}, Training steps: {}, Avg loss: {:.3f}".
format(epoch, i + 1, total_loss / args.report_steps))
total_loss = 0.0
result = evaluate(args, read_dataset(args, args.dev_path))
if result[0] > best_result:
best_result = result[0]
save_model(model, args.output_model_path)
# Evaluation phase.
if args.test_path is not None:
print("Test set evaluation.")
if torch.cuda.device_count() > 1:
model.module.load_state_dict(torch.load(args.output_model_path))
else:
model.load_state_dict(torch.load(args.output_model_path))
evaluate(args, read_dataset(args, args.test_path))
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
finetune_opts(parser)
parser.add_argument("--train_answer_path",
type=str,
required=True,
help="Path of the answers for trainset.")
parser.add_argument("--dev_answer_path",
type=str,
required=True,
help="Path of the answers for devset.")
parser.add_argument(
"--max_choices_num",
default=10,
type=int,
help=
"The maximum number of cadicate answer, shorter than this will be padded."
)
args = parser.parse_args()
args.labels_num = args.max_choices_num
if args.output_model_path == None:
args.output_model_path = "./models/chid_model.bin"
# Load the hyperparameters from the config file.
args = load_hyperparam(args)
set_seed(args.seed)
# Build tokenizer.
args.tokenizer = CharTokenizer(args)
# Build multiple choice model.
model = MultipleChoice(args)
# Load or initialize parameters.
load_or_initialize_parameters(args, model)
args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(args.device)
# Training phase.
trainset = read_dataset(args, args.train_path, args.train_answer_path)
random.shuffle(trainset)
instances_num = len(trainset)
batch_size = args.batch_size
src = torch.LongTensor([example[0] for example in trainset])
tgt = torch.LongTensor([example[1] for example in trainset])
seg = torch.LongTensor([example[2] for example in trainset])
args.train_steps = int(instances_num * args.epochs_num / batch_size) + 1
print("Batch size: ", batch_size)
print("The number of training instances:", instances_num)
optimizer, scheduler = build_optimizer(args, model)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
args.amp = amp
if torch.cuda.device_count() > 1:
print("{} GPUs are available. Let's use them.".format(
torch.cuda.device_count()))
model = torch.nn.DataParallel(model)
args.model = model
total_loss, result, best_result = 0., 0., 0.
print("Start training.")
for epoch in range(1, args.epochs_num + 1):
model.train()
for i, (src_batch, tgt_batch, seg_batch,
_) in enumerate(batch_loader(batch_size, src, tgt, seg)):
loss = train_model(args, model, optimizer, scheduler, src_batch,
tgt_batch, seg_batch)
total_loss += loss.item()
if (i + 1) % args.report_steps == 0:
print("Epoch id: {}, Training steps: {}, Avg loss: {:.3f}".
format(epoch, i + 1, total_loss / args.report_steps))
total_loss = 0.
result = evaluate(
args, read_dataset(args, args.dev_path, args.dev_answer_path))
if result[0] > best_result:
best_result = result[0]
save_model(model, args.output_model_path)
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
finetune_opts(parser)
parser.add_argument(
"--doc_stride",
default=128,
type=int,
help=
"When splitting up a long document into chunks, how much stride to take between chunks."
)
args = parser.parse_args()
# Load the hyperparameters from the config file.
args = load_hyperparam(args)
set_seed(args.seed)
# Build tokenizer.
args.tokenizer = CharTokenizer(args)
# Build machine reading comprehension model.
model = MachineReadingComprehension(args)
# Load or initialize parameters.
load_or_initialize_parameters(args, model)
args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(args.device)
# Build tokenizer.
args.tokenizer = CharTokenizer(args)
# Training phase.
batch_size = args.batch_size
print("Batch size: ", batch_size)
trainset, _ = read_dataset(args, args.train_path)
random.shuffle(trainset)
instances_num = len(trainset)
src = torch.LongTensor([sample[0] for sample in trainset])
seg = torch.LongTensor([sample[1] for sample in trainset])
start_position = torch.LongTensor([sample[2] for sample in trainset])
end_position = torch.LongTensor([sample[3] for sample in trainset])
args.train_steps = int(instances_num * args.epochs_num / batch_size) + 1
print("The number of training instances:", instances_num)
optimizer, scheduler = build_optimizer(args, model)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
if torch.cuda.device_count() > 1:
print("{} GPUs are available. Let's use them.".format(
torch.cuda.device_count()))
model = torch.nn.DataParallel(model)
args.model = model
total_loss = 0.0
result = 0.0
best_result = 0.0
print("Start training.")
for epoch in range(1, args.epochs_num + 1):
model.train()
for i, (src_batch, seg_batch, start_position_batch,
end_position_batch) in enumerate(
batch_loader(batch_size, src, seg, start_position,
end_position)):
loss = train(args, model, optimizer, scheduler, src_batch,
seg_batch, start_position_batch, end_position_batch)
total_loss += loss.item()
if (i + 1) % args.report_steps == 0:
print("Epoch id: {}, Training steps: {}, Avg loss: {:.3f}".
format(epoch, i + 1, total_loss / args.report_steps))
total_loss = 0.0
result = evaluate(args, *read_dataset(args, args.dev_path))
if result > best_result:
best_result = result
save_model(model, args.output_model_path)
# Evaluation phase.
if args.test_path is not None:
print("Test set evaluation.")
if torch.cuda.device_count() > 1:
model.module.load_state_dict(torch.load(args.output_model_path))
else:
model.load_state_dict(torch.load(args.output_model_path))
evaluate(args, *read_dataset(args, args.test_path))
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
finetune_opts(parser)
parser.add_argument("--label2id_path",
type=str,
required=True,
help="Path of the label2id file.")
args = parser.parse_args()
# Load the hyperparameters of the config file.
args = load_hyperparam(args)
set_seed(args.seed)
args.begin_ids = []
with open(args.label2id_path, mode="r", encoding="utf-8") as f:
l2i = json.load(f)
print("Labels: ", l2i)
l2i["[PAD]"] = len(l2i)
for label in l2i:
if label.startswith("B"):
args.begin_ids.append(l2i[label])
args.l2i = l2i
args.labels_num = len(l2i)
args.tokenizer = SpaceTokenizer(args)
# Build sequence labeling model.
model = NerTagger(args)
# Load or initialize parameters.
load_or_initialize_parameters(args, model)
args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(args.device)
# Training phase.
instances = read_dataset(args, args.train_path)
src = torch.LongTensor([ins[0] for ins in instances])
tgt = torch.LongTensor([ins[1] for ins in instances])
seg = torch.LongTensor([ins[2] for ins in instances])
instances_num = src.size(0)
batch_size = args.batch_size
args.train_steps = int(instances_num * args.epochs_num / batch_size) + 1
print("Batch size: ", batch_size)
print("The number of training instances:", instances_num)
optimizer, scheduler = build_optimizer(args, model)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
if torch.cuda.device_count() > 1:
print("{} GPUs are available. Let's use them.".format(
torch.cuda.device_count()))
model = torch.nn.DataParallel(model)
args.model = model
total_loss, f1, best_f1 = 0.0, 0.0, 0.0
print("Start training.")
for epoch in range(1, args.epochs_num + 1):
model.train()
for i, (src_batch, tgt_batch,
seg_batch) in enumerate(batch_loader(batch_size, src, tgt,
seg)):
loss = train(args, model, optimizer, scheduler, src_batch,
tgt_batch, seg_batch)
total_loss += loss.item()
if (i + 1) % args.report_steps == 0:
print("Epoch id: {}, Training steps: {}, Avg loss: {:.3f}".
format(epoch, i + 1, total_loss / args.report_steps))
total_loss = 0.0
f1 = evaluate(args, read_dataset(args, args.dev_path))
if f1 > best_f1:
best_f1 = f1
save_model(model, args.output_model_path)
else:
continue
# Evaluation phase.
if args.test_path is not None:
print("Test set evaluation.")
if torch.cuda.device_count() > 1:
args.model.module.load_state_dict(
torch.load(args.output_model_path))
else:
args.model.load_state_dict(torch.load(args.output_model_path))
evaluate(args, read_dataset(args, args.test_path))
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
finetune_opts(parser)
parser.add_argument("--pooling",
choices=["mean", "max", "first", "last"],
default="first",
help="Pooling type.")
tokenizer_opts(parser)
parser.add_argument("--soft_targets",
action='store_true',
help="Train model with logits.")
parser.add_argument("--soft_alpha",
type=float,
default=0.5,
help="Weight of the soft targets loss.")
args = parser.parse_args()
# Load the hyperparameters from the config file.
args = load_hyperparam(args)
set_seed(args.seed)
# Count the number of labels.
args.labels_num = count_labels_num(args.train_path)
# Build tokenizer.
args.tokenizer = str2tokenizer[args.tokenizer](args)
# Build classification model.
model = Classifier(args)
# Load or initialize parameters.
load_or_initialize_parameters(args, model)
args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(args.device)
# Training phase.
trainset = read_dataset(args, args.train_path)
random.shuffle(trainset)
instances_num = len(trainset)
batch_size = args.batch_size
src = torch.LongTensor([example[0] for example in trainset])
tgt = torch.LongTensor([example[1] for example in trainset])
seg = torch.LongTensor([example[2] for example in trainset])
if args.soft_targets:
soft_tgt = torch.FloatTensor([example[3] for example in trainset])
else:
soft_tgt = None
args.train_steps = int(instances_num * args.epochs_num / batch_size) + 1
print("Batch size: ", batch_size)
print("The number of training instances:", instances_num)
optimizer, scheduler = build_optimizer(args, model)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
args.amp = amp
if torch.cuda.device_count() > 1:
print("{} GPUs are available. Let's use them.".format(
torch.cuda.device_count()))
model = torch.nn.DataParallel(model)
args.model = model
total_loss, result, best_result = 0.0, 0.0, 0.0
print("Start training.")
for epoch in range(1, args.epochs_num + 1):
model.train()
for i, (src_batch, tgt_batch, seg_batch, soft_tgt_batch) in enumerate(
batch_loader(batch_size, src, tgt, seg, soft_tgt)):
loss = train_model(args, model, optimizer, scheduler, src_batch,
tgt_batch, seg_batch, soft_tgt_batch)
total_loss += loss.item()
if (i + 1) % args.report_steps == 0:
print("Epoch id: {}, Training steps: {}, Avg loss: {:.3f}".
format(epoch, i + 1, total_loss / args.report_steps))
total_loss = 0.0
result = evaluate(args, read_dataset(args, args.dev_path))
if result[0] > best_result:
best_result = result[0]
save_model(model, args.output_model_path)
# Evaluation phase.
if args.test_path is not None:
print("Test set evaluation.")
if torch.cuda.device_count() > 1:
args.model.module.load_state_dict(
torch.load(args.output_model_path))
else:
args.model.load_state_dict(torch.load(args.output_model_path))
evaluate(args, read_dataset(args, args.test_path), True)
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
finetune_opts(parser)
tokenizer_opts(parser)
args = parser.parse_args()
# Load the hyperparameters from the config file.
args = load_hyperparam(args)
set_seed(args.seed)
# Count the number of labels.
args.labels_num = count_labels_num(args.train_path)
# Build tokenizer.
args.tokenizer = str2tokenizer[args.tokenizer](args)
# Build classification model.
model = SiameseClassifier(args)
# Load or initialize parameters.
load_or_initialize_parameters(args, model)
# Get logger.
args.logger = init_logger(args)
args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(args.device)
# Training phase.
trainset = read_dataset(args, args.train_path)
instances_num = len(trainset)
batch_size = args.batch_size
args.train_steps = int(instances_num * args.epochs_num / batch_size) + 1
args.logger.info("Batch size: {}".format(batch_size))
args.logger.info(
"The number of training instances: {}".format(instances_num))
optimizer, scheduler = build_optimizer(args, model)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
args.amp = amp
if torch.cuda.device_count() > 1:
args.logger.info("{} GPUs are available. Let's use them.".format(
torch.cuda.device_count()))
model = torch.nn.DataParallel(model)
args.model = model
total_loss, result, best_result = 0.0, 0.0, 0.0
args.logger.info("Start training.")
for epoch in range(1, args.epochs_num + 1):
random.shuffle(trainset)
src_a = torch.LongTensor([example[0][0] for example in trainset])
src_b = torch.LongTensor([example[0][1] for example in trainset])
tgt = torch.LongTensor([example[1] for example in trainset])
seg_a = torch.LongTensor([example[2][0] for example in trainset])
seg_b = torch.LongTensor([example[2][1] for example in trainset])
model.train()
for i, (src_batch, tgt_batch, seg_batch) in enumerate(
batch_loader(batch_size, (src_a, src_b), tgt, (seg_a, seg_b))):
loss = train_model(args, model, optimizer, scheduler, src_batch,
tgt_batch, seg_batch)
total_loss += loss.item()
if (i + 1) % args.report_steps == 0:
args.logger.info(
"Epoch id: {}, Training steps: {}, Avg loss: {:.3f}".
format(epoch, i + 1, total_loss / args.report_steps))
total_loss = 0.0
result = evaluate(args, read_dataset(args, args.dev_path))
if result[0] > best_result:
best_result = result[0]
save_model(model, args.output_model_path)
# Evaluation phase.
if args.test_path is not None:
args.logger.info("Test set evaluation.")
if torch.cuda.device_count() > 1:
args.model.module.load_state_dict(
torch.load(args.output_model_path))
else:
args.model.load_state_dict(torch.load(args.output_model_path))
evaluate(args, read_dataset(args, args.test_path))
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
finetune_opts(parser)
tokenizer_opts(parser)
parser.add_argument("--pooling",
choices=["mean", "max", "first", "last"],
default="first",
help="Pooling type.")
parser.add_argument("--temperature", type=float, default=0.05)
parser.add_argument("--eval_steps",
type=int,
default=200,
help="Evaluate frequency.")
args = parser.parse_args()
# Load the hyperparameters from the config file.
args = load_hyperparam(args)
set_seed(args.seed)
# Build tokenizer.
args.tokenizer = str2tokenizer[args.tokenizer](args)
# Build classification model.
model = SimCSE(args)
# Load or initialize parameters.
load_or_initialize_parameters(args, model)
args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(args.device)
# Training phase.
trainset = read_dataset(args, args.train_path)
random.shuffle(trainset)
instances_num = len(trainset)
batch_size = args.batch_size
src_a = torch.LongTensor([example[0][0] for example in trainset])
src_b = torch.LongTensor([example[0][1] for example in trainset])
tgt = torch.FloatTensor([example[1] for example in trainset])
seg_a = torch.LongTensor([example[2][0] for example in trainset])
seg_b = torch.LongTensor([example[2][1] for example in trainset])
args.train_steps = int(instances_num * args.epochs_num / batch_size) + 1
print("Batch size: ", batch_size)
print("The number of training instances:", instances_num)
optimizer, scheduler = build_optimizer(args, model)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
args.amp = amp
if torch.cuda.device_count() > 1:
print("{} GPUs are available. Let's use them.".format(
torch.cuda.device_count()))
model = torch.nn.DataParallel(model)
args.model = model
total_loss, result, best_result = 0.0, 0.0, 0.0
print("Start training.")
for epoch in range(1, args.epochs_num + 1):
model.train()
for i, (src_batch, tgt_batch, seg_batch) in enumerate(
batch_loader(batch_size, (src_a, src_b), tgt, (seg_a, seg_b))):
model.zero_grad()
src_a_batch, src_b_batch = src_batch
seg_a_batch, seg_b_batch = seg_batch
src_a_batch = src_a_batch.to(args.device)
src_b_batch = src_b_batch.to(args.device)
seg_a_batch = seg_a_batch.to(args.device)
seg_b_batch = seg_b_batch.to(args.device)
features_0, features_1 = model((src_a_batch, src_b_batch),
(seg_a_batch, seg_b_batch))
similarity_matrix = similarity(features_0, features_1,
args.temperature)
tgt_batch = torch.arange(similarity_matrix.size(0),
device=similarity_matrix.device,
dtype=torch.long)
loss = nn.CrossEntropyLoss()(similarity_matrix, tgt_batch)
if args.fp16:
with args.amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
scheduler.step()
total_loss += loss.item()
if (i + 1) % args.report_steps == 0:
print("Epoch id: {}, Training steps: {}, Avg loss: {:.3f}".
format(epoch, i + 1, total_loss / args.report_steps))
total_loss = 0.0
if (i + 1) % args.eval_steps == 0 or (i + 1) == math.ceil(
instances_num / batch_size):
result = evaluate(args, read_dataset(args, args.dev_path))
print(
"Epoch id: {}, Training steps: {}, Evaluate result: {}, Best result: {}"
.format(epoch, i + 1, result, best_result))
if result > best_result:
best_result = result
save_model(model, args.output_model_path)
print(
"It is the best model until now. Save it to {}".format(
args.output_model_path))