def do_inference(args):
prepare_output_path(args.output_dir, args.overwrite_output_dir)
device, n_gpus = setup_backend(args.no_cuda)
args.task_name = args.task_name.lower()
task = get_glue_task(args.task_name, data_dir=args.data_dir)
args.batch_size = args.per_gpu_eval_batch_size * max(1, n_gpus)
classifier = TransformerSequenceClassifier.load_model(
model_path=args.model_path,
model_type=args.model_type,
task_type=task.task_type,
metric_fn=get_metric_fn(task.name),
do_lower_case=args.do_lower_case,
load_quantized=args.load_quantized_model,
)
classifier.to(device, n_gpus)
examples = task.get_dev_examples(
) if args.evaluate else task.get_test_examples()
preds = classifier.inference(examples,
args.max_seq_length,
args.batch_size,
evaluate=args.evaluate)
with io.open(os.path.join(args.output_dir, "output.txt"),
"w",
encoding="utf-8") as fw:
for p in preds:
fw.write("{}\n".format(p))
def do_inference(args):
prepare_output_path(args.output_dir, args.overwrite_output_dir)
device, n_gpus = setup_backend(args.no_cuda)
args.batch_size = args.per_gpu_eval_batch_size * max(1, n_gpus)
inference_examples = process_inference_input(args.data_file)
classifier = NeuralTagger.load_model(model_path=args.model_dir)
classifier.to(device, n_gpus)
output = classifier.inference(inference_examples, args.b)
write_column_tagged_file(args.output_dir + os.sep + "output.txt", output)
def do_training(args):
prepare_output_path(args.output_dir, args.overwrite_output_dir)
device, n_gpus = setup_backend(args.no_cuda)
# Set seed
set_seed(args.seed, n_gpus)
# Prepare GLUE task
args.task_name = args.task_name.lower()
task = get_glue_task(args.task_name, data_dir=args.data_dir)
classifier = TransformerSequenceClassifier(
model_type=args.model_type,
model_name_or_path=args.model_name_or_path,
labels=task.get_labels(),
task_type=task.task_type,
metric_fn=get_metric_fn(task.name),
config_name=args.config_name,
tokenizer_name=args.tokenizer_name,
do_lower_case=args.do_lower_case,
output_path=args.output_dir,
device=device,
n_gpus=n_gpus)
train_batch_size = args.per_gpu_train_batch_size * max(1, n_gpus)
train_ex = task.get_train_examples()
dev_ex = task.get_dev_examples()
train_dataset = classifier.convert_to_tensors(train_ex,
args.max_seq_length)
dev_dataset = classifier.convert_to_tensors(dev_ex, args.max_seq_length)
train_sampler = RandomSampler(train_dataset)
dev_sampler = SequentialSampler(dev_dataset)
train_dl = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=train_batch_size)
dev_dl = DataLoader(dev_dataset,
sampler=dev_sampler,
batch_size=args.per_gpu_eval_batch_size)
total_steps, _ = classifier.get_train_steps_epochs(
args.max_steps, args.num_train_epochs, args.per_gpu_train_batch_size,
len(train_dataset))
classifier.setup_default_optimizer(weight_decay=args.weight_decay,
learning_rate=args.learning_rate,
adam_epsilon=args.adam_epsilon,
warmup_steps=args.warmup_steps,
total_steps=total_steps)
classifier.train(
train_dl,
dev_dl,
None,
gradient_accumulation_steps=args.gradient_accumulation_steps,
per_gpu_train_batch_size=args.per_gpu_train_batch_size,
max_steps=args.max_steps,
num_train_epochs=args.num_train_epochs,
max_grad_norm=args.max_grad_norm,
logging_steps=args.logging_steps,
save_steps=args.save_steps)
classifier.save_model(args.output_dir, args=args)
def do_inference(args):
prepare_output_path(args.output_dir, args.overwrite_output_dir)
device, n_gpus = setup_backend(args.no_cuda)
args.batch_size = args.per_gpu_eval_batch_size * max(1, n_gpus)
inference_examples = process_inference_input(args.data_file)
classifier = TransformerTokenClassifier.load_model(model_path=args.model_path,
model_type=args.model_type,
do_lower_case=args.do_lower_case,
load_quantized=args.load_quantized_model)
classifier.to(device, n_gpus)
output = classifier.inference(inference_examples, args.max_seq_length, args.batch_size)
write_column_tagged_file(args.output_dir + os.sep + "output.txt", output)
def do_inference(args):
prepare_output_path(args.output_dir, args.overwrite_output_dir)
device, n_gpus = setup_backend(args.no_cuda)
args.task_name = args.task_name.lower()
task = get_glue_task(args.task_name, data_dir=args.data_dir)
args.batch_size = args.per_gpu_eval_batch_size * max(1, n_gpus)
classifier = TransformerSequenceClassifier.load_model(
model_path=args.model_path, model_type=args.model_type)
classifier.to(device, n_gpus)
preds = classifier.inference(task.get_test_examples(), args.batch_size)
with io.open(os.path.join(args.output_dir, "output.txt"),
"w",
encoding="utf-8") as fw:
for p in preds:
fw.write("{}\n".format(p))
def do_kd_training(args):
prepare_output_path(args.output_dir, args.overwrite_output_dir)
device, n_gpus = setup_backend(args.no_cuda)
# Set seed
set_seed(args.seed, n_gpus)
# prepare data
processor = TokenClsProcessor(args.data_dir, tag_col=args.tag_col)
train_ex = processor.get_train_examples()
dev_ex = processor.get_dev_examples()
test_ex = processor.get_test_examples()
vocab = processor.get_vocabulary()
vocab_size = len(vocab) + 1
num_labels = len(processor.get_labels()) + 1
# create an embedder
embedder_cls = MODEL_TYPE[args.model_type]
if args.config_file is not None:
embedder_model = embedder_cls.from_config(vocab_size, num_labels,
args.config_file)
else:
embedder_model = embedder_cls(vocab_size, num_labels)
# load external word embeddings if present
if args.embedding_file is not None:
emb_dict = load_embedding_file(args.embedding_file)
emb_mat = get_embedding_matrix(emb_dict, vocab)
emb_mat = torch.tensor(emb_mat, dtype=torch.float)
embedder_model.load_embeddings(emb_mat)
classifier = NeuralTagger(embedder_model,
word_vocab=vocab,
labels=processor.get_labels(),
use_crf=args.use_crf,
device=device,
n_gpus=n_gpus)
train_batch_size = args.b * max(1, n_gpus)
train_dataset = classifier.convert_to_tensors(
train_ex,
max_seq_length=args.max_sentence_length,
max_word_length=args.max_word_length)
teacher = TransformerTokenClassifier.load_model(
model_path=args.teacher_model_path, model_type=args.teacher_model_type)
teacher.to(device, n_gpus)
teacher_dataset = teacher.convert_to_tensors(train_ex,
args.max_sentence_length,
False)
train_dataset = ParallelDataset(train_dataset, teacher_dataset)
train_sampler = RandomSampler(train_dataset)
train_dl = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=train_batch_size)
if dev_ex is not None:
dev_dataset = classifier.convert_to_tensors(
dev_ex,
max_seq_length=args.max_sentence_length,
max_word_length=args.max_word_length)
dev_sampler = SequentialSampler(dev_dataset)
dev_dl = DataLoader(dev_dataset,
sampler=dev_sampler,
batch_size=args.b)
if test_ex is not None:
test_dataset = classifier.convert_to_tensors(
test_ex,
max_seq_length=args.max_sentence_length,
max_word_length=args.max_word_length)
test_sampler = SequentialSampler(test_dataset)
test_dl = DataLoader(test_dataset,
sampler=test_sampler,
batch_size=args.b)
if args.lr is not None:
opt = classifier.get_optimizer(lr=args.lr)
distiller = TeacherStudentDistill(teacher, args.kd_temp, args.kd_dist_w,
args.kd_student_w, args.kd_loss_fn)
classifier.train(train_dl,
dev_dl,
test_dl,
epochs=args.e,
batch_size=args.b,
logging_steps=args.logging_steps,
save_steps=args.save_steps,
save_path=args.output_dir,
optimizer=opt if opt is not None else None,
distiller=distiller)
classifier.save_model(args.output_dir)
def do_training(args):
prepare_output_path(args.output_dir, args.overwrite_output_dir)
device, n_gpus = setup_backend(args.no_cuda)
# Set seed
args.seed = set_seed(args.seed, n_gpus)
# prepare data
processor = TokenClsProcessor(args.data_dir)
classifier = TransformerTokenClassifier(
model_type=args.model_type,
model_name_or_path=args.model_name_or_path,
labels=processor.get_labels(),
config_name=args.config_name,
tokenizer_name=args.tokenizer_name,
do_lower_case=args.do_lower_case,
output_path=args.output_dir,
device=device,
n_gpus=n_gpus,
)
train_ex = processor.get_train_examples(filename=args.train_file_name)
if train_ex is None:
raise Exception("No train examples found, quitting.")
dev_ex = processor.get_dev_examples()
test_ex = processor.get_test_examples()
train_batch_size = args.per_gpu_train_batch_size * max(1, n_gpus)
train_dataset = classifier.convert_to_tensors(
train_ex, max_seq_length=args.max_seq_length)
train_sampler = RandomSampler(train_dataset)
train_dl = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=train_batch_size)
dev_dl = None
test_dl = None
if dev_ex is not None:
dev_dataset = classifier.convert_to_tensors(
dev_ex, max_seq_length=args.max_seq_length)
dev_sampler = SequentialSampler(dev_dataset)
dev_dl = DataLoader(dev_dataset,
sampler=dev_sampler,
batch_size=args.per_gpu_eval_batch_size)
if test_ex is not None:
test_dataset = classifier.convert_to_tensors(
test_ex, max_seq_length=args.max_seq_length)
test_sampler = SequentialSampler(test_dataset)
test_dl = DataLoader(test_dataset,
sampler=test_sampler,
batch_size=args.per_gpu_eval_batch_size)
total_steps, _ = classifier.get_train_steps_epochs(
args.max_steps, args.num_train_epochs, args.per_gpu_train_batch_size,
len(train_dataset))
classifier.setup_default_optimizer(
weight_decay=args.weight_decay,
learning_rate=args.learning_rate,
adam_epsilon=args.adam_epsilon,
warmup_steps=args.warmup_steps,
total_steps=total_steps,
)
classifier.train(
train_dl,
dev_dl,
test_dl,
gradient_accumulation_steps=args.gradient_accumulation_steps,
per_gpu_train_batch_size=args.per_gpu_train_batch_size,
max_steps=args.max_steps,
num_train_epochs=args.num_train_epochs,
max_grad_norm=args.max_grad_norm,
logging_steps=args.logging_steps,
save_steps=args.save_steps,
)
classifier.save_model(args.output_dir, args=args)
def do_kd_pseudo_training(args):
prepare_output_path(args.output_dir, args.overwrite_output_dir)
device, n_gpus = setup_backend(args.no_cuda)
# Set seed
args.seed = set_seed(args.seed, n_gpus)
# prepare data
processor = TokenClsProcessor(
args.data_dir, tag_col=args.tag_col, ignore_token=args.ignore_token
)
train_labeled_ex = processor.get_train_examples(filename=args.train_filename)
train_unlabeled_ex = processor.get_train_examples(filename=args.unlabeled_filename)
dev_ex = processor.get_dev_examples(filename=args.dev_filename)
test_ex = processor.get_test_examples(filename=args.test_filename)
vocab = processor.get_vocabulary(train_labeled_ex + train_unlabeled_ex + dev_ex + test_ex)
vocab_size = len(vocab) + 1
num_labels = len(processor.get_labels()) + 1
# create an embedder
embedder_cls = MODEL_TYPE[args.model_type]
if args.config_file is not None:
embedder_model = embedder_cls.from_config(vocab_size, num_labels, args.config_file)
else:
embedder_model = embedder_cls(vocab_size, num_labels)
# load external word embeddings if present
if args.embedding_file is not None:
emb_dict = load_embedding_file(args.embedding_file, dim=embedder_model.word_embedding_dim)
emb_mat = get_embedding_matrix(emb_dict, vocab)
emb_mat = torch.tensor(emb_mat, dtype=torch.float)
embedder_model.load_embeddings(emb_mat)
classifier = NeuralTagger(
embedder_model,
word_vocab=vocab,
labels=processor.get_labels(),
use_crf=args.use_crf,
device=device,
n_gpus=n_gpus,
)
train_batch_size = args.b * max(1, n_gpus)
train_labeled_dataset = classifier.convert_to_tensors(
train_labeled_ex,
max_seq_length=args.max_sentence_length,
max_word_length=args.max_word_length,
)
train_unlabeled_dataset = classifier.convert_to_tensors(
train_unlabeled_ex,
max_seq_length=args.max_sentence_length,
max_word_length=args.max_word_length,
include_labels=False,
)
if args.parallel_batching:
# # concat labeled+unlabeled dataset
# train_dataset = ConcatTensorDataset(train_labeled_dataset, [train_unlabeled_dataset])
# match sizes of labeled/unlabeled train data for parallel batching
larger_ds, smaller_ds = (
(train_labeled_dataset, train_unlabeled_dataset)
if len(train_labeled_dataset) > len(train_unlabeled_dataset)
else (train_unlabeled_dataset, train_labeled_dataset)
)
concat_smaller_ds = smaller_ds
while len(concat_smaller_ds) < len(larger_ds):
concat_smaller_ds = ConcatTensorDataset(concat_smaller_ds, [smaller_ds])
if len(concat_smaller_ds[0]) == 4:
train_unlabeled_dataset = concat_smaller_ds
else:
train_labeled_dataset = concat_smaller_ds
else:
train_dataset = CombinedTensorDataset([train_labeled_dataset, train_unlabeled_dataset])
# load saved teacher args if exist
if os.path.exists(args.teacher_model_path + os.sep + "training_args.bin"):
t_args = torch.load(args.teacher_model_path + os.sep + "training_args.bin")
t_device, t_n_gpus = setup_backend(t_args.no_cuda)
teacher = TransformerTokenClassifier.load_model(
model_path=args.teacher_model_path,
model_type=args.teacher_model_type,
config_name=t_args.config_name,
tokenizer_name=t_args.tokenizer_name,
do_lower_case=t_args.do_lower_case,
output_path=t_args.output_dir,
device=t_device,
n_gpus=t_n_gpus,
)
else:
teacher = TransformerTokenClassifier.load_model(
model_path=args.teacher_model_path, model_type=args.teacher_model_type
)
teacher.to(device, n_gpus)
teacher_labeled_dataset = teacher.convert_to_tensors(train_labeled_ex, args.teacher_max_seq_len)
teacher_unlabeled_dataset = teacher.convert_to_tensors(
train_unlabeled_ex, args.teacher_max_seq_len, False
)
if args.parallel_batching:
# # concat teacher labeled+unlabeled dataset
# teacher_dataset = ConcatTensorDataset(teacher_labeled_dataset, [teacher_unlabeled_dataset])
# match sizes of labeled/unlabeled teacher train data for parallel batching
larger_ds, smaller_ds = (
(teacher_labeled_dataset, teacher_unlabeled_dataset)
if len(teacher_labeled_dataset) > len(teacher_unlabeled_dataset)
else (teacher_unlabeled_dataset, teacher_labeled_dataset)
)
concat_smaller_ds = smaller_ds
while len(concat_smaller_ds) < len(larger_ds):
concat_smaller_ds = ConcatTensorDataset(concat_smaller_ds, [smaller_ds])
if len(concat_smaller_ds[0]) == 4:
teacher_unlabeled_dataset = concat_smaller_ds
else:
teacher_labeled_dataset = concat_smaller_ds
train_all_dataset = ParallelDataset(
train_labeled_dataset,
teacher_labeled_dataset,
train_unlabeled_dataset,
teacher_unlabeled_dataset,
)
train_all_sampler = RandomSampler(train_all_dataset)
# this way must use same batch size for both labeled/unlabeled sets
train_dl = DataLoader(
train_all_dataset, sampler=train_all_sampler, batch_size=train_batch_size
)
else:
teacher_dataset = CombinedTensorDataset(
[teacher_labeled_dataset, teacher_unlabeled_dataset]
)
train_dataset = ParallelDataset(train_dataset, teacher_dataset)
train_sampler = RandomSampler(train_dataset)
train_dl = DataLoader(train_dataset, sampler=train_sampler, batch_size=train_batch_size)
if dev_ex is not None:
dev_dataset = classifier.convert_to_tensors(
dev_ex, max_seq_length=args.max_sentence_length, max_word_length=args.max_word_length
)
dev_sampler = SequentialSampler(dev_dataset)
dev_dl = DataLoader(dev_dataset, sampler=dev_sampler, batch_size=args.b)
if test_ex is not None:
test_dataset = classifier.convert_to_tensors(
test_ex, max_seq_length=args.max_sentence_length, max_word_length=args.max_word_length
)
test_sampler = SequentialSampler(test_dataset)
test_dl = DataLoader(test_dataset, sampler=test_sampler, batch_size=args.b)
if args.lr is not None:
opt = classifier.get_optimizer(lr=args.lr)
distiller = TeacherStudentDistill(
teacher, args.kd_temp, args.kd_dist_w, args.kd_student_w, args.kd_loss_fn
)
classifier.train(
train_dl,
dev_dl,
test_dl,
epochs=args.e,
batch_size=args.b,
logging_steps=args.logging_steps,
save_steps=args.save_steps,
save_path=args.output_dir,
optimizer=opt if opt is not None else None,
best_result_file=args.best_result_file,
distiller=distiller,
word_dropout=args.word_dropout,
)
classifier.save_model(args.output_dir)
