def do_train(args):
paddle.enable_static() if not args.eager_run else None
paddle.set_device("gpu" if args.n_gpu else "cpu")
if paddle.distributed.get_world_size() > 1:
paddle.distributed.init_parallel_env()
set_seed(args)
args.task_name = args.task_name.lower()
dataset_class, metric_class = TASK_CLASSES[args.task_name]
args.model_type = args.model_type.lower()
model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
train_dataset, dev_dataset = dataset_class.get_datasets(["train", "dev"])
tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path)
trans_func = partial(convert_example,
tokenizer=tokenizer,
label_list=train_dataset.get_labels(),
max_seq_length=args.max_seq_length)
train_dataset = train_dataset.apply(trans_func, lazy=True)
# train_batch_sampler = SamplerHelper(train_dataset).shuffle().batch(
# batch_size=args.batch_size).shard()
train_batch_sampler = paddle.io.DistributedBatchSampler(
# train_dataset, batch_size=args.batch_size, shuffle=True)
train_dataset,
batch_size=args.batch_size,
shuffle=False)
batchify_fn = lambda samples, fn=Tuple(
Pad(axis=0, pad_val=tokenizer.vocab[tokenizer.pad_token]), # input
Pad(axis=0, pad_val=tokenizer.vocab[tokenizer.pad_token]), # segment
Stack(), # length
Stack(dtype="int64"
if train_dataset.get_labels() else "float32") # label
): [data for i, data in enumerate(fn(samples)) if i != 2]
train_data_loader = DataLoader(dataset=train_dataset,
batch_sampler=train_batch_sampler,
collate_fn=batchify_fn,
num_workers=0,
return_list=True)
dev_dataset = dev_dataset.apply(trans_func, lazy=True)
# dev_batch_sampler = SamplerHelper(dev_dataset).batch(
# batch_size=args.batch_size)
dev_batch_sampler = paddle.io.BatchSampler(dev_dataset,
batch_size=args.batch_size,
shuffle=False)
dev_data_loader = DataLoader(dataset=dev_dataset,
batch_sampler=dev_batch_sampler,
collate_fn=batchify_fn,
num_workers=0,
return_list=True)
# model = model_class.from_pretrained(
# args.model_name_or_path,) num_classes=len(train_dataset.get_labels()))
model = BertForPretraining(
BertModel(**model_class.pretrained_init_configuration[
args.model_name_or_path]))
if paddle.distributed.get_world_size() > 1:
model = paddle.DataParallel(model)
num_training_steps = args.max_steps if args.max_steps > 0 else len(
train_data_loader) * args.num_train_epochs
lr_scheduler = LinearDecayWithWarmup(args.learning_rate,
num_training_steps, args.warmup_steps)
# Generate parameter names needed to perform weight decay.
# All bias and LayerNorm parameters are excluded.
decay_params = [
p.name for n, p in model.named_parameters()
if not any(nd in n for nd in ["bias", "norm"])
]
optimizer = paddle.optimizer.AdamW(
learning_rate=lr_scheduler,
epsilon=args.adam_epsilon,
parameters=model.parameters(),
weight_decay=args.weight_decay,
apply_decay_param_fun=lambda x: x in decay_params)
loss_fct = paddle.nn.loss.CrossEntropyLoss() if train_dataset.get_labels(
) else paddle.nn.loss.MSELoss()
metric = metric_class()
### TODO: use hapi
# trainer = paddle.hapi.Model(model)
# trainer.prepare(optimizer, loss_fct, paddle.metric.Accuracy())
# trainer.fit(train_data_loader,
# dev_data_loader,
# log_freq=args.logging_steps,
# epochs=args.num_train_epochs,
# save_dir=args.output_dir)
model.eval()
param_names = list(model.state_dict().keys())
import pickle
with open(args.params_pd_path, "rb") as f:
np_params = pickle.load(f)
model.set_state_dict(dict(zip(param_names, np_params)))
paddle.save(model.state_dict(), "%s.pdparams" % args.model_name_or_path)
for data in train_data_loader():
print(model(*data[:-1]))
exit(0)
global_step = 0
tic_train = time.time()
for epoch in range(args.num_train_epochs):
for step, batch in enumerate(train_data_loader):
input_ids, segment_ids, labels = batch
logits = model(input_ids, segment_ids)
loss = loss_fct(logits, labels)
if global_step % args.logging_steps == 0:
print(
"global step %d, epoch: %d, batch: %d, loss: %f, speed: %.2f step/s"
% (global_step, epoch, step, loss, args.logging_steps /
(time.time() - tic_train)))
tic_train = time.time()
loss.backward()
optimizer.step()
lr_scheduler.step()
optimizer.clear_grad()
if global_step % args.save_steps == 0:
evaluate(model, loss_fct, metric, dev_data_loader)
if (not args.n_gpu > 1) or paddle.distributed.get_rank() == 0:
paddle.save(
model.state_dict(),
os.path.join(args.output_dir,
"model_%d.pdparams" % global_step))
global_step += 1
def do_train(args):
paddle.set_device(args.device)
if paddle.distributed.get_world_size() > 1:
paddle.distributed.init_parallel_env()
set_seed(args)
args.task_name = args.task_name.lower()
metric_class = METRIC_CLASSES[args.task_name]
args.model_type = args.model_type.lower()
model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
train_ds = load_dataset('glue', args.task_name, splits="train")
tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path)
trans_func = partial(convert_example,
tokenizer=tokenizer,
label_list=train_ds.label_list,
max_seq_length=args.max_seq_length)
train_ds = train_ds.map(trans_func, lazy=True)
train_batch_sampler = paddle.io.DistributedBatchSampler(
train_ds, batch_size=args.batch_size,
shuffle=False) # for same data when converting
batchify_fn = lambda samples, fn=Tuple(
Pad(axis=0, pad_val=tokenizer.pad_token_id), # input
Pad(axis=0, pad_val=tokenizer.pad_token_type_id), # segment
Stack(dtype="int64" if train_ds.label_list else "float32") # label
): fn(samples)
train_data_loader = DataLoader(dataset=train_ds,
batch_sampler=train_batch_sampler,
collate_fn=batchify_fn,
num_workers=0,
return_list=True)
if args.task_name == "mnli":
dev_ds_matched, dev_ds_mismatched = load_dataset(
'glue', args.task_name, splits=["dev_matched", "dev_mismatched"])
dev_ds_matched = dev_ds_matched.map(trans_func, lazy=True)
dev_ds_mismatched = dev_ds_mismatched.map(trans_func, lazy=True)
dev_batch_sampler_matched = paddle.io.BatchSampler(
dev_ds_matched, batch_size=args.batch_size, shuffle=False)
dev_data_loader_matched = DataLoader(
dataset=dev_ds_matched,
batch_sampler=dev_batch_sampler_matched,
collate_fn=batchify_fn,
num_workers=0,
return_list=True)
dev_batch_sampler_mismatched = paddle.io.BatchSampler(
dev_ds_mismatched, batch_size=args.batch_size, shuffle=False)
dev_data_loader_mismatched = DataLoader(
dataset=dev_ds_mismatched,
batch_sampler=dev_batch_sampler_mismatched,
collate_fn=batchify_fn,
num_workers=0,
return_list=True)
else:
dev_ds = load_dataset('glue', args.task_name, splits='dev')
dev_ds = dev_ds.map(trans_func, lazy=True)
dev_batch_sampler = paddle.io.BatchSampler(dev_ds,
batch_size=args.batch_size,
shuffle=False)
dev_data_loader = DataLoader(dataset=dev_ds,
batch_sampler=dev_batch_sampler,
collate_fn=batchify_fn,
num_workers=0,
return_list=True)
num_classes = 1 if train_ds.label_list == None else len(
train_ds.label_list)
# model = model_class.from_pretrained(
# args.model_name_or_path, num_classes=num_classes)
model = BertForPretraining(
BertModel(**model_class.pretrained_init_configuration[
args.model_name_or_path]))
if paddle.distributed.get_world_size() > 1:
model = paddle.DataParallel(model)
num_training_steps = args.max_steps if args.max_steps > 0 else (
len(train_data_loader) * args.num_train_epochs)
warmup = args.warmup_steps if args.warmup_steps > 0 else args.warmup_proportion
lr_scheduler = LinearDecayWithWarmup(args.learning_rate,
num_training_steps, warmup)
# Generate parameter names needed to perform weight decay.
# All bias and LayerNorm parameters are excluded.
decay_params = [
p.name for n, p in model.named_parameters()
if not any(nd in n for nd in ["bias", "norm"])
]
optimizer = paddle.optimizer.AdamW(
learning_rate=lr_scheduler,
beta1=0.9,
beta2=0.999,
epsilon=args.adam_epsilon,
parameters=model.parameters(),
weight_decay=args.weight_decay,
apply_decay_param_fun=lambda x: x in decay_params)
loss_fct = paddle.nn.loss.CrossEntropyLoss(
) if train_ds.label_list else paddle.nn.loss.MSELoss()
metric = metric_class()
if args.use_amp:
scaler = paddle.amp.GradScaler(init_loss_scaling=args.scale_loss)
# load converted model and run once to compare
model.eval()
param_names = list(model.state_dict().keys())
import pickle
with open(args.params_pd_path, "rb") as f:
np_params = pickle.load(f)
model.set_state_dict(dict(zip(param_names, np_params)))
paddle.save(model.state_dict(), "%s.pdparams" % args.model_name_or_path)
for data in train_data_loader():
print(model(*data[:-1]))
exit(0)
global_step = 0
tic_train = time.time()
for epoch in range(args.num_train_epochs):
for step, batch in enumerate(train_data_loader):
global_step += 1
input_ids, segment_ids, labels = batch
with paddle.amp.auto_cast(
args.use_amp,
custom_white_list=["layer_norm", "softmax", "gelu"]):
logits = model(input_ids, segment_ids)
loss = loss_fct(logits, labels)
if args.use_amp:
scaler.scale(loss).backward()
scaler.minimize(optimizer, loss)
else:
loss.backward()
optimizer.step()
lr_scheduler.step()
optimizer.clear_grad()
if global_step % args.logging_steps == 0:
print(
"global step %d/%d, epoch: %d, batch: %d, rank_id: %s, loss: %f, lr: %.10f, speed: %.4f step/s"
% (global_step, num_training_steps, epoch, step,
paddle.distributed.get_rank(), loss, optimizer.get_lr(),
args.logging_steps / (time.time() - tic_train)))
tic_train = time.time()
if global_step % args.save_steps == 0 or global_step == num_training_steps:
tic_eval = time.time()
if args.task_name == "mnli":
evaluate(model, loss_fct, metric, dev_data_loader_matched)
evaluate(model, loss_fct, metric,
dev_data_loader_mismatched)
print("eval done total : %s s" % (time.time() - tic_eval))
else:
evaluate(model, loss_fct, metric, dev_data_loader)
print("eval done total : %s s" % (time.time() - tic_eval))
if paddle.distributed.get_rank() == 0:
output_dir = os.path.join(
args.output_dir, "%s_ft_model_%d.pdparams" %
(args.task_name, global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
# Need better way to get inner model of DataParallel
model_to_save = model._layers if isinstance(
model, paddle.DataParallel) else model
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
