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.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_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 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)