def get_tpu_sampler(dataset: torch.utils.data.dataset.Dataset):
if xm.xrt_world_size() <= 1:
return RandomSampler(dataset)
return DistributedSampler(dataset,
num_replicas=xm.xrt_world_size(),
rank=xm.get_ordinal())
def evaluate(args, model, tokenizer, prefix=""):
# Loop to handle MNLI double evaluation (matched, mis-matched)
eval_task_names = (args.task_name, )
eval_outputs_dirs = (args.output_dir, )
results = {}
for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
eval_dataset = load_and_cache_examples(args,
eval_task,
tokenizer,
evaluate=True)
if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(eval_output_dir)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(
1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(
eval_dataset) if args.local_rank == -1 else DistributedSampler(
eval_dataset)
eval_dataloader = DataLoader(eval_dataset,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
# Eval!
logger.info("***** Running evaluation {} *****".format(prefix))
logger.info(" Num examples = %d", len(eval_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
eval_loss = 0.0
nb_eval_steps = 0
preds = None
out_label_ids = None
for batch in tqdm(eval_dataloader, desc="Evaluating"):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {
'input_ids':
batch[0],
'attention_mask':
batch[1],
'token_type_ids':
batch[2] if args.model_type in ['bert', 'xlnet'] else
None, # XLM don't use segment_ids
'labels':
batch[3]
}
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if preds is None:
preds = logits.detach().cpu().numpy()
out_label_ids = inputs['labels'].detach().cpu().numpy()
else:
preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
out_label_ids = np.append(
out_label_ids,
inputs['labels'].detach().cpu().numpy(),
axis=0)
eval_loss = eval_loss / nb_eval_steps
if args.output_mode == "classification":
preds = np.argmax(preds, axis=1)
elif args.output_mode == "regression":
preds = np.squeeze(preds)
result = compute_metrics(eval_task, preds, out_label_ids)
results.update(result)
output_eval_file = os.path.join(eval_output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results {} *****".format(prefix))
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
return results
def train(args, train_dataset, model, tokenizer):
""" Train the model """
if args.local_rank in [-1, 0]:
tb_writer = SummaryWriter()
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = RandomSampler(
train_dataset) if args.local_rank == -1 else DistributedSampler(
train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (
len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0
},
]
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=t_total)
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(
args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler.pt")):
# Load in optimizer and scheduler states
optimizer.load_state_dict(
torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
scheduler.load_state_dict(
torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
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)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 1
epochs_trained = 0
steps_trained_in_current_epoch = 0
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path):
try:
# set global_step to gobal_step of last saved checkpoint from model path
checkpoint_suffix = args.model_name_or_path.split("-")[-1].split(
"/")[0]
global_step = int(checkpoint_suffix)
epochs_trained = global_step // (len(train_dataloader) //
args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (
len(train_dataloader) // args.gradient_accumulation_steps)
logger.info(
" Continuing training from checkpoint, will skip to saved global_step"
)
logger.info(" Continuing training from epoch %d", epochs_trained)
logger.info(" Continuing training from global step %d",
global_step)
logger.info(" Will skip the first %d steps in the first epoch",
steps_trained_in_current_epoch)
except ValueError:
logger.info(" Starting fine-tuning.")
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
train_iterator = trange(epochs_trained,
int(args.num_train_epochs),
desc="Epoch",
disable=args.local_rank not in [-1, 0])
# Added here for reproductibility
set_seed(args)
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"token_type_ids": batch[2],
"start_positions": batch[3],
"end_positions": batch[4],
}
if args.model_type in [
"xlm", "roberta", "distilbert", "camembert"
]:
del inputs["token_type_ids"]
if args.model_type in ["xlnet", "xlm"]:
inputs.update({"cls_index": batch[5], "p_mask": batch[6]})
if args.version_2_with_negative:
inputs.update({"is_impossible": batch[7]})
if hasattr(model, "config") and hasattr(
model.config, "lang2id"):
inputs.update({
"langs":
(torch.ones(batch[0].shape, dtype=torch.int64) *
args.lang_id).to(args.device)
})
outputs = model(**inputs)
# model outputs are always tuple in transformers (see doc)
loss = outputs[0]
if args.n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu parallel (not distributed) training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
# Log metrics
if args.local_rank in [
-1, 0
] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
# Only evaluate when single GPU otherwise metrics may not average well
if args.local_rank == -1 and args.evaluate_during_training:
results = evaluate(args, model, tokenizer)
for key, value in results.items():
tb_writer.add_scalar("eval_{}".format(key), value,
global_step)
tb_writer.add_scalar("lr",
scheduler.get_lr()[0], global_step)
tb_writer.add_scalar("loss", (tr_loss - logging_loss) /
args.logging_steps, global_step)
logging_loss = tr_loss
# Save model checkpoint
if args.local_rank in [
-1, 0
] and args.save_steps > 0 and global_step % args.save_steps == 0:
output_dir = os.path.join(
args.output_dir, "checkpoint-{}".format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
# Take care of distributed/parallel training
model_to_save = model.module if hasattr(
model, "module") else model
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
torch.save(args,
os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
torch.save(optimizer.state_dict(),
os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(),
os.path.join(output_dir, "scheduler.pt"))
logger.info("Saving optimizer and scheduler states to %s",
output_dir)
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
if args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
if args.local_rank in [-1, 0]:
tb_writer.close()
return global_step, tr_loss / global_step
def f_load_data_yelp_restaurant(self, args):
self.m_data_name = args.data_name
# self.m_vocab_file = self.m_data_name+".vocab.json"
self.m_vocab_file = args.vocab_file
self.m_item_boa_file = args.item_boa_file
self.m_user_boa_file = args.user_boa_file
train_data_file = args.data_dir + "/new_train.pickle"
valid_data_file = args.data_dir + "/new_valid.pickle"
test_data_file = args.data_dir + "/new_valid.pickle"
# train_data_file = args.data_dir+"/train.pickle"
# valid_data_file = args.data_dir+"/valid.pickle"
# test_data_file = args.data_dir+"/test.pickle"
train_df = pd.read_pickle(train_data_file)
valid_df = pd.read_pickle(valid_data_file)
test_df = pd.read_pickle(test_data_file)
user_num = train_df.userid.nunique()
print("user num", user_num)
with open(os.path.join(args.data_dir, self.m_vocab_file),
'r',
encoding='utf8') as f:
vocab = json.loads(f.read())
with open(os.path.join(args.data_dir, self.m_item_boa_file),
'r',
encoding='utf8') as f:
item_boa_dict = json.loads(f.read())
# with open(os.path.join(args.data_dir, self.m_user_boa_file), 'r', encoding='utf8') as f:
# user_boa_dict = json.loads(f.read())
user_boa_dict = {}
vocab_obj = _Vocab()
vocab_obj.f_set_vocab(vocab['a2i'], vocab['i2a'])
vocab_obj.f_set_user_num(user_num)
global_user2iid = vocab['user_index']
global_item2iid = vocab['item_index']
vocab_obj.f_set_user(global_user2iid)
vocab_obj.f_set_item(global_item2iid)
print("vocab size", vocab_obj.m_vocab_size)
train_data = _YELP_RESTAURANT(args, vocab_obj, train_df, item_boa_dict,
user_boa_dict)
valid_data = _YELP_RESTAURANT(args, vocab_obj, valid_df, item_boa_dict,
user_boa_dict)
batch_size = args.batch_size
if args.parallel:
train_sampler = DistributedSampler(dataset=train_data)
train_loader = DataLoader(dataset=train_data,
batch_size=batch_size,
sampler=train_sampler,
num_workers=8,
collate_fn=train_data.collate)
else:
train_loader = DataLoader(dataset=train_data,
batch_size=batch_size,
shuffle=True,
num_workers=8,
collate_fn=train_data.collate)
test_loader = DataLoader(dataset=valid_data,
batch_size=batch_size,
shuffle=False,
num_workers=4,
collate_fn=valid_data.collate)
return train_loader, test_loader, vocab_obj
def __init__(self,
data_dir,
label_dir,
tokenizer,
train_file='train.csv',
val_file='val.csv',
test_data=None,
label_file='labels.csv',
text_col='text',
label_col='label',
bs=32,
maxlen=512,
multi_gpu=True,
multi_label=False,
backend="nccl"):
self.data_dir = data_dir
self.tokenizer = tokenizer
self.maxlen = maxlen
self.bs = bs
self.train_dl = None
self.val_dl = None
self.test_dl = None
self.multi_label = multi_label
if multi_label:
processor = MultiLabelTextProcessor(data_dir, label_dir)
else:
processor = TextProcessor(data_dir, label_dir)
self.labels = processor.get_labels(label_file)
if train_file:
# Train DataLoader
train_examples = processor.get_train_examples(train_file,
text_col=text_col,
label_col=label_col)
train_features = convert_examples_to_features(
train_examples,
label_list=self.labels,
tokenizer=tokenizer,
max_seq_length=maxlen)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor(
[f.input_mask for f in train_features], dtype=torch.long)
all_segment_ids = torch.tensor(
[f.segment_ids for f in train_features], dtype=torch.long)
if multi_label:
all_label_ids = torch.tensor(
[f.label_id for f in train_features], dtype=torch.float)
else:
all_label_ids = torch.tensor(
[f.label_id for f in train_features], dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if multi_gpu:
train_sampler = RandomSampler(train_data)
else:
torch.distributed.init_process_group(
backend=backend,
init_method="tcp://localhost:23459",
rank=0,
world_size=1)
#torch.distributed.init_process_group(backend='nccl')
train_sampler = DistributedSampler(train_data)
self.train_dl = DataLoader(train_data,
sampler=train_sampler,
batch_size=bs)
if val_file:
# Validation DataLoader
val_examples = processor.get_dev_examples(val_file,
text_col=text_col,
label_col=label_col)
val_features = convert_examples_to_features(val_examples,
label_list=self.labels,
tokenizer=tokenizer,
max_seq_length=maxlen)
all_input_ids = torch.tensor([f.input_ids for f in val_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in val_features],
dtype=torch.long)
all_segment_ids = torch.tensor(
[f.segment_ids for f in val_features], dtype=torch.long)
if multi_label:
all_label_ids = torch.tensor(
[f.label_id for f in val_features], dtype=torch.float)
else:
all_label_ids = torch.tensor(
[f.label_id for f in val_features], dtype=torch.long)
val_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
val_sampler = SequentialSampler(val_data)
self.val_dl = DataLoader(val_data,
sampler=val_sampler,
batch_size=bs)
if test_data:
test_examples = []
input_data = []
for index, text in enumerate(test_data):
test_examples.append(InputExample(index, text))
input_data.append({'id': index, 'text': text})
test_features = convert_examples_to_features(
test_examples,
label_list=self.labels,
tokenizer=tokenizer,
max_seq_length=maxlen)
all_input_ids = torch.tensor([f.input_ids for f in test_features],
dtype=torch.long)
all_input_mask = torch.tensor(
[f.input_mask for f in test_features], dtype=torch.long)
all_segment_ids = torch.tensor(
[f.segment_ids for f in test_features], dtype=torch.long)
test_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids)
test_sampler = SequentialSampler(test_data)
self.test_dl = DataLoader(test_data,
sampler=test_sampler,
batch_size=bs)
def evaluate(args, model, tokenizer, prefix=""):
metric = SeqEntityScore(args.id2label, markup=args.markup)
eval_output_dir = args.output_dir
if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(eval_output_dir)
eval_dataset = load_and_cache_examples(args,
args.task_name,
tokenizer,
data_type='dev')
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(
eval_dataset) if args.local_rank == -1 else DistributedSampler(
eval_dataset)
eval_dataloader = DataLoader(eval_dataset,
sampler=eval_sampler,
batch_size=args.eval_batch_size,
collate_fn=collate_fn)
# Eval!
logger.info("***** Running evaluation %s *****", prefix)
logger.info(" Num examples = %d", len(eval_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
eval_loss = 0.0
nb_eval_steps = 0
pbar = ProgressBar(n_total=len(eval_dataloader), desc="Evaluating")
if isinstance(model, nn.DataParallel):
model = model.module
for step, batch in enumerate(eval_dataloader):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"labels": batch[3],
'input_lens': batch[4]
}
if args.model_type != "distilbert":
# XLM and RoBERTa don"t use segment_ids
inputs["token_type_ids"] = (batch[2] if args.model_type
in ["bert", "xlnet"] else None)
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
tags = model.crf.decode(logits, inputs['attention_mask'])
if args.n_gpu > 1:
tmp_eval_loss = tmp_eval_loss.mean(
) # mean() to average on multi-gpu parallel evaluating
eval_loss += tmp_eval_loss.item()
nb_eval_steps += 1
out_label_ids = inputs['labels'].cpu().numpy().tolist()
input_lens = inputs['input_lens'].cpu().numpy().tolist()
tags = tags.squeeze(0).cpu().numpy().tolist()
for i, label in enumerate(out_label_ids):
temp_1 = []
temp_2 = []
for j, m in enumerate(label):
if j == 0:
continue
elif j == input_lens[i] - 1:
metric.update(pred_paths=[temp_2], label_paths=[temp_1])
break
else:
temp_1.append(args.id2label[out_label_ids[i][j]])
temp_2.append(args.id2label[tags[i][j]])
pbar(step)
logger.info("\n")
eval_loss = eval_loss / nb_eval_steps
eval_info, entity_info = metric.result()
results = {f'{key}': value for key, value in eval_info.items()}
results['loss'] = eval_loss
logger.info("***** Eval results %s *****", prefix)
info = "-".join(
[f' {key}: {value:.4f} ' for key, value in results.items()])
logger.info(info)
logger.info("***** Entity results %s *****", prefix)
for key in sorted(entity_info.keys()):
logger.info("******* %s results ********" % key)
info = "-".join([
f' {key}: {value:.4f} ' for key, value in entity_info[key].items()
])
logger.info(info)
return results
def predict(args, model, tokenizer, prefix=""):
pred_output_dir = args.output_dir
if not os.path.exists(pred_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(pred_output_dir)
test_dataset = load_and_cache_examples(args,
args.task_name,
tokenizer,
data_type='test')
# Note that DistributedSampler samples randomly
test_sampler = SequentialSampler(
test_dataset) if args.local_rank == -1 else DistributedSampler(
test_dataset)
test_dataloader = DataLoader(test_dataset,
sampler=test_sampler,
batch_size=1,
collate_fn=collate_fn)
# Eval!
logger.info("***** Running prediction %s *****", prefix)
logger.info(" Num examples = %d", len(test_dataset))
logger.info(" Batch size = %d", 1)
results = []
output_predict_file = os.path.join(pred_output_dir, prefix,
"test_prediction.json")
pbar = ProgressBar(n_total=len(test_dataloader), desc="Predicting")
if isinstance(model, nn.DataParallel):
model = model.module
for step, batch in enumerate(test_dataloader):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"labels": None,
'input_lens': batch[4]
}
if args.model_type != "distilbert":
# XLM and RoBERTa don"t use segment_ids
inputs["token_type_ids"] = (batch[2] if args.model_type
in ["bert", "xlnet"] else None)
outputs = model(**inputs)
logits = outputs[0]
tags = model.crf.decode(logits, inputs['attention_mask'])
tags = tags.squeeze(0).cpu().numpy().tolist()
preds = tags[0][1:-1] # [CLS]XXXX[SEP]
label_entities = get_entities(preds, args.id2label, args.markup)
json_d = {}
json_d['id'] = step
json_d['tag_seq'] = " ".join([args.id2label[x] for x in preds])
json_d['entities'] = label_entities
results.append(json_d)
pbar(step)
logger.info("\n")
with open(output_predict_file, "w") as writer:
for record in results:
writer.write(json.dumps(record) + '\n')
if args.task_name == 'cluener':
output_submit_file = os.path.join(pred_output_dir, prefix,
"test_submit.json")
test_text = []
with open(os.path.join(args.data_dir, "test.json"), 'r') as fr:
for line in fr:
test_text.append(json.loads(line))
test_submit = []
for x, y in zip(test_text, results):
json_d = {}
json_d['id'] = x['id']
json_d['label'] = {}
entities = y['entities']
words = list(x['text'])
if len(entities) != 0:
for subject in entities:
tag = subject[0]
start = subject[1]
end = subject[2]
word = "".join(words[start:end + 1])
if tag in json_d['label']:
if word in json_d['label'][tag]:
json_d['label'][tag][word].append([start, end])
else:
json_d['label'][tag][word] = [[start, end]]
else:
json_d['label'][tag] = {}
json_d['label'][tag][word] = [[start, end]]
test_submit.append(json_d)
json_to_text(output_submit_file, test_submit)
def main():
parser = ArgumentParser()
parser.add_argument("--file_num", type=int, default=10,
help="Number of pregenerate file")
parser.add_argument("--reduce_memory", action="store_true",
help="Store training data as on-disc memmaps to massively reduce memory usage")
parser.add_argument("--epochs", type=int, default=4,
help="Number of epochs to train for")
parser.add_argument('--num_eval_steps', default=2000)
parser.add_argument('--num_save_steps', default=5000)
parser.add_argument("--local_rank", type=int, default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument("--no_cuda", action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.")
parser.add_argument("--train_batch_size", default=18, type=int,
help="Total batch size for training.")
parser.add_argument('--loss_scale', type=float, default=0,
help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument("--warmup_proportion", default=0.1, type=float,
help="Linear warmup over warmup_steps.")
parser.add_argument("--adam_epsilon", default=1e-8, type=float,
help="Epsilon for Adam optimizer.")
parser.add_argument('--max_grad_norm', default=1.0, type=float)
parser.add_argument("--learning_rate", default=2e-4, type=float,
help="The initial learning rate for Adam.")
parser.add_argument('--seed', type=int, default=42,
help="random seed for initialization")
parser.add_argument('--fp16_opt_level', type=str, default='O2',
help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
parser.add_argument('--fp16', action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
args = parser.parse_args()
pregenerated_data = config['data_dir'] / "corpus/train"
assert pregenerated_data.is_dir(), \
"--pregenerated_data should point to the folder of files made by prepare_lm_data_mask.py!"
samples_per_epoch = 0
for i in range(args.file_num):
data_file = pregenerated_data / f"file_{i}.json"
metrics_file = pregenerated_data / f"file_{i}_metrics.json"
if data_file.is_file() and metrics_file.is_file():
metrics = json.loads(metrics_file.read_text())
samples_per_epoch += metrics['num_training_examples']
else:
if i == 0:
exit("No training data was found!")
print(f"Warning! There are fewer epochs of pregenerated data ({i}) than training epochs ({args.epochs}).")
print("This script will loop over the available data, but training diversity may be negatively impacted.")
break
logger.info(f"samples_per_epoch: {samples_per_epoch}")
if args.local_rank == -1 or args.no_cuda:
device = torch.device(f"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
args.n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
f"device: {device} , distributed training: {bool(args.local_rank != -1)}, 16-bits training: {args.fp16}")
if args.gradient_accumulation_steps < 1:
raise ValueError(
f"Invalid gradient_accumulation_steps parameter: {args.gradient_accumulation_steps}, should be >= 1")
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
seed_everything(args.seed)
tokenizer = BertTokenizer(vocab_file=config['checkpoint_dir'] / 'vocab.txt')
total_train_examples = samples_per_epoch * args.epochs
num_train_optimization_steps = int(
total_train_examples / args.train_batch_size / args.gradient_accumulation_steps)
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
args.warmup_steps = int(num_train_optimization_steps * args.warmup_proportion)
# Prepare model
with open(str(config['checkpoint_dir'] / 'config.json'), "r", encoding='utf-8') as reader:
json_config = json.loads(reader.read())
print(json_config)
bert_config = BertConfig.from_json_file(str(config['checkpoint_dir'] / 'config.json'))
model = BertForMaskedLM(config=bert_config)
# model = BertForMaskedLM.from_pretrained(config['checkpoint_dir'] / 'checkpoint-580000')
model.to(device)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
lr_scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=num_train_optimization_steps)
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 args.n_gpu > 1:
model = torch.nn.DataParallel(model)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
output_device=args.local_rank)
global_step = 0
metric = LMAccuracy()
tr_acc = AverageMeter()
tr_loss = AverageMeter()
train_logs = {}
logger.info("***** Running training *****")
logger.info(f" Num examples = {total_train_examples}")
logger.info(f" Batch size = {args.train_batch_size}")
logger.info(f" Num steps = {num_train_optimization_steps}")
logger.info(f" warmup_steps = {args.warmup_steps}")
seed_everything(args.seed) # Added here for reproducibility
for epoch in range(args.epochs):
for idx in range(args.file_num):
epoch_dataset = PregeneratedDataset(file_id=idx, training_path=pregenerated_data, tokenizer=tokenizer,
reduce_memory=args.reduce_memory)
if args.local_rank == -1:
train_sampler = RandomSampler(epoch_dataset)
else:
train_sampler = DistributedSampler(epoch_dataset)
train_dataloader = DataLoader(epoch_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
model.train()
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(train_dataloader):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids = batch
outputs = model(input_ids=input_ids, token_type_ids=segment_ids,
attention_mask=input_mask, masked_lm_labels=lm_label_ids)
pred_output = outputs[1]
loss = outputs[0]
metric(logits=pred_output.view(-1, bert_config.vocab_size), target=lm_label_ids.view(-1))
if args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
nb_tr_steps += 1
tr_acc.update(metric.value(), n=input_ids.size(0))
tr_loss.update(loss.item(), n=1)
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
lr_scheduler.step()
optimizer.step()
optimizer.zero_grad()
global_step += 1
if global_step % args.num_eval_steps == 0:
train_logs['loss'] = tr_loss.avg
train_logs['acc'] = tr_acc.avg
show_info = f'\n[Training]:[{epoch}/{args.epochs}]{global_step}/{num_train_optimization_steps} ' + "-".join(
[f' {key}: {value:.4f} ' for key, value in train_logs.items()])
logger.info(show_info)
tr_acc.reset()
tr_loss.reset()
if global_step % args.num_save_steps == 0:
if args.local_rank in [-1, 0] and args.num_save_steps > 0:
# Save model checkpoint
output_dir = config['checkpoint_dir'] / f'lm-checkpoint-{global_step}'
if not output_dir.exists():
output_dir.mkdir()
# save model
model_to_save = model.module if hasattr(model,
'module') else model # Take care of distributed/parallel training
model_to_save.save_pretrained(str(output_dir))
torch.save(args, str(output_dir / 'training_args.bin'))
logger.info("Saving model checkpoint to %s", output_dir)
# save config
output_config_file = output_dir / CONFIG_NAME
with open(str(output_config_file), 'w') as f:
f.write(model_to_save.config.to_json_string())
# save vocab
tokenizer.save_vocabulary(output_dir)
def train(argv):
parser = argparse.ArgumentParser(
description="ListenAttendSpell AM with fully supervised training")
# for training
parser.add_argument('--data-path',
default='./data',
type=str,
help="dataset path to use in training")
parser.add_argument('--min-len',
default=1.,
type=float,
help="min length of utterance to use in secs")
parser.add_argument('--max-len',
default=10.,
type=float,
help="max length of utterance to use in secs")
parser.add_argument(
'--batch-size',
default=32,
type=int,
help="number of images (and labels) to be considered in a batch")
parser.add_argument('--num-workers',
default=32,
type=int,
help="number of dataloader workers")
parser.add_argument('--num-epochs',
default=100,
type=int,
help="number of epochs to run")
parser.add_argument('--init-lr',
default=1e-4,
type=float,
help="initial learning rate for Adam optimizer")
parser.add_argument('--max-norm',
default=1e-2,
type=int,
help="norm cutoff to prevent explosion of gradients")
# optional
parser.add_argument('--use-cuda',
default=False,
action='store_true',
help="use cuda")
parser.add_argument('--fp16',
default=False,
action='store_true',
help="use FP16 model")
parser.add_argument('--visdom',
default=False,
action='store_true',
help="use visdom logging")
parser.add_argument('--visdom-host',
default="127.0.0.1",
type=str,
help="visdom server ip address")
parser.add_argument('--visdom-port',
default=8097,
type=int,
help="visdom server port")
parser.add_argument('--tensorboard',
default=False,
action='store_true',
help="use tensorboard logging")
parser.add_argument(
'--slack',
default=False,
action='store_true',
help=
"use slackclient logging (need to set SLACK_API_TOKEN and SLACK_API_USER env_var"
)
parser.add_argument('--seed',
default=None,
type=int,
help="seed for controlling randomness in this example")
parser.add_argument('--log-dir',
default='./logs_las',
type=str,
help="filename for logging the outputs")
parser.add_argument('--model-prefix',
default='las',
type=str,
help="model file prefix to store")
parser.add_argument('--checkpoint',
default=False,
action='store_true',
help="save checkpoint")
parser.add_argument('--continue-from',
default=None,
type=str,
help="model file path to make continued from")
parser.add_argument('--opt-type',
default="adamw",
type=str,
help=f"optimizer type in {OPTIMIZER_TYPES}")
parser.add_argument(
'--local_rank',
default=-1,
type=int,
help=f"set rank 0..n-1 when pytorch's launcher is used")
args = parser.parse_args(argv)
init_distributed(args.use_cuda, local_rank=args.local_rank)
init_logger(log_file="train.log", rank=get_rank(), **vars(args))
set_seed(args.seed)
# prepare trainer object
input_folding = 3
model = ListenAttendSpell(label_vec_size=p.NUM_CTC_LABELS,
input_folding=input_folding)
amp_handle = get_amp_handle(args)
trainer = LASTrainer(model, amp_handle, **vars(args))
labeler = trainer.decoder.labeler
train_datasets = [
NonSplitTrainDataset(labeler=labeler,
manifest_file=f"{args.data_path}/swbd/train.csv",
stride=input_folding),
NonSplitTrainDataset(
labeler=labeler,
manifest_file=f"{args.data_path}/aspire/train.csv",
stride=input_folding),
NonSplitTrainDataset(labeler=labeler,
manifest_file=f"{args.data_path}/aspire/dev.csv",
stride=input_folding),
NonSplitTrainDataset(labeler=labeler,
manifest_file=f"{args.data_path}/aspire/test.csv",
stride=input_folding),
]
datasets = {
"warmup":
AudioSubset(train_datasets[0],
data_size=0,
min_len=args.min_len,
max_len=args.max_len),
"train":
ConcatDataset([
AudioSubset(d,
data_size=0,
min_len=args.min_len,
max_len=args.max_len) for d in train_datasets
]),
"dev":
AudioSubset(NonSplitTrainDataset(
labeler=labeler,
manifest_file=f"{args.data_path}/swbd/eval2000.csv",
stride=input_folding),
data_size=0),
"test":
NonSplitTrainDataset(labeler=labeler,
manifest_file=f"{args.data_path}/swbd/rt03.csv",
stride=input_folding),
}
dataloaders = {
"warmup":
NonSplitTrainDataLoader(datasets["warmup"],
sort=True,
sampler=(DistributedSampler(datasets["warmup"])
if is_distributed() else None),
batch_size=args.batch_size,
num_workers=args.num_workers,
shuffle=(not is_distributed()),
pin_memory=args.use_cuda),
"train":
NonSplitTrainDataLoader(datasets["train"],
sort=True,
sampler=(DistributedSampler(datasets["train"])
if is_distributed() else None),
batch_size=args.batch_size,
num_workers=args.num_workers,
shuffle=(not is_distributed()),
pin_memory=args.use_cuda),
"dev":
NonSplitTrainDataLoader(datasets["dev"],
sort=True,
batch_size=16,
num_workers=8,
shuffle=False,
pin_memory=args.use_cuda),
"test":
NonSplitTrainDataLoader(datasets["test"],
sort=True,
batch_size=16,
num_workers=8,
shuffle=False,
pin_memory=args.use_cuda),
}
# run inference for a certain number of epochs
for i in range(trainer.epoch, args.num_epochs):
trainer.train_epoch(dataloaders["warmup"])
#trainer.train_epoch(dataloaders["train"])
trainer.validate(dataloaders["dev"])
# final test to know WER
trainer.test(dataloaders["test"])
def evaluate(args,
model,
tokenizer,
labels,
pad_token_label_id,
mode,
prefix=""):
eval_dataset = load_and_cache_examples(args,
tokenizer,
labels,
pad_token_label_id,
mode=mode)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(
eval_dataset) if args.local_rank == -1 else DistributedSampler(
eval_dataset)
eval_dataloader = DataLoader(eval_dataset,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
# multi-gpu evaluate
if args.n_gpu > 1:
model = torch.nn.DataParallel(model, device_ids=[0, 1, 2, 3])
# Eval!
logger.info("***** Running evaluation %s *****", prefix)
logger.info(" Num examples = %d", len(eval_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
eval_loss = 0.0
nb_eval_steps = 0
preds = []
out_label_ids = None
model.eval()
for batch in tqdm(eval_dataloader, desc="Evaluating"):
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
input_ids = batch[0]
attention_mask = batch[1]
label_ids = batch[3]
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"labels": batch[3],
"device": args.device
}
if args.model_type != "distilbert":
inputs["token_type_ids"] = (
batch[2] if args.model_type in ["bert", "xlnet"] else None
) # XLM and RoBERTa don"t use segment_ids
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
if args.n_gpu > 1:
tmp_eval_loss = tmp_eval_loss.mean(
) # mean() to average on multi-gpu parallel evaluating
eval_loss += tmp_eval_loss.item()
for i in range(logits.size(0)):
non_masked_idx = attention_mask[i] == 1
non_masked_labels = label_ids[i][non_masked_idx]
non_masked_logits = logits[i][non_masked_idx]
# ignore label index -100
keep_index = non_masked_labels != -100
active_labels = non_masked_labels[keep_index]
active_logits = non_masked_logits[keep_index]
# Apply Viterbi decoding
pred_score, pred = model.module._viterbi_decode(
active_logits, args.device)
preds.append(pred)
nb_eval_steps += 1
if nb_eval_steps == 1:
out_label_ids = inputs["labels"].detach().cpu().numpy()
else:
out_label_ids = np.append(out_label_ids,
inputs["labels"].detach().cpu().numpy(),
axis=0)
# if preds is None:
# preds = logits.detach().cpu().numpy()
# out_label_ids = inputs["labels"].detach().cpu().numpy()
# else:
# preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
# out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
eval_loss = eval_loss / nb_eval_steps
# preds = np.argmax(preds, axis=2)
label_map = {i: label for i, label in enumerate(labels)}
out_label_list = [[] for _ in range(out_label_ids.shape[0])]
preds_list = [[] for _ in range(out_label_ids.shape[0])]
for i in range(out_label_ids.shape[0]):
k = 0
for j in range(out_label_ids.shape[1]):
if out_label_ids[i, j] != pad_token_label_id:
out_label_list[i].append(label_map[out_label_ids[i, j]])
preds_list[i].append(label_map[preds[i][k]])
k += 1
# print(out_label_list[i])
# print(preds_list[i])
# print("---------------------")
results = {
"loss": eval_loss,
"precision": precision_score(out_label_list, preds_list),
"recall": recall_score(out_label_list, preds_list),
"f1": f1_score(out_label_list, preds_list),
}
logger.info("***** Eval results %s *****", prefix)
for key in sorted(results.keys()):
logger.info(" %s = %s", key, str(results[key]))
return results, preds_list
def train(num_gpus, rank, group_name, output_directory, epochs, learning_rate,
sigma, iters_per_checkpoint, batch_size, seed, fp16_run,
checkpoint_path, with_tensorboard, waveglow_config, dist_config,
data_config, train_config, **kwargs):
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
# =====START: ADDED FOR DISTRIBUTED======
if num_gpus > 1:
init_distributed(rank, num_gpus, group_name, **dist_config)
# =====END: ADDED FOR DISTRIBUTED======
criterion = WaveGlowLoss(sigma)
model = WaveGlow(**waveglow_config).to(_device)
# =====START: ADDED FOR DISTRIBUTED======
if num_gpus > 1:
model = apply_gradient_allreduce(model)
# =====END: ADDED FOR DISTRIBUTED======
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
if fp16_run:
from apex import amp
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
# Load checkpoint if one exists
iteration = 0
if checkpoint_path != "":
model, optimizer, iteration = load_checkpoint(checkpoint_path, model,
optimizer)
iteration += 1 # next iteration is iteration + 1
iteration_start = iteration
trainset = Mel2Samp(**data_config)
# =====START: ADDED FOR DISTRIBUTED======
train_sampler = DistributedSampler(trainset) if num_gpus > 1 else None
# =====END: ADDED FOR DISTRIBUTED======
train_loader = DataLoader(
trainset,
num_workers=train_config.get('dataloader_num_workers', 8),
shuffle=train_config.get('dataloader_shuffle', True),
sampler=train_sampler,
batch_size=batch_size,
pin_memory=train_config.get('dataloader_pin_memory', False),
drop_last=True)
# Get shared output_directory ready
if rank == 0:
if not os.path.isdir(output_directory):
os.makedirs(output_directory)
os.chmod(output_directory, 0o775)
print("output directory", output_directory)
if with_tensorboard and rank == 0:
from tensorboardX import SummaryWriter
logger = SummaryWriter(os.path.join(os.path.dirname(output_directory),
'tensorboard'),
filename_suffix='.tensorboard')
with open(Path(output_directory).parent.joinpath('metadata', 'train.txt'),
'wt',
encoding='utf8') as fout:
for line in trainset.audio_files:
fpath = os.path.abspath(line)
fout.write(f'{fpath}\n')
model.train()
epoch_offset = max(0, int(iteration / len(train_loader)))
# ================ MAIN TRAINNIG LOOP! ===================
for epoch in range(epoch_offset, epochs):
print("Epoch: {}".format(epoch))
for i, batch in enumerate(
tqdm(train_loader, desc=f"Epoch-{epoch}", ncols=100)):
model.zero_grad()
mel, audio = batch
mel = torch.autograd.Variable(mel.to(_device))
audio = torch.autograd.Variable(audio.to(_device))
outputs = model((mel, audio))
loss = criterion(outputs)
if num_gpus > 1:
reduced_loss = reduce_tensor(loss.data, num_gpus).item()
else:
reduced_loss = loss.item()
if fp16_run:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
# print("{}:\t{:.9f}".format(iteration, reduced_loss))
if with_tensorboard and rank == 0:
logger.add_scalar('training_loss', reduced_loss,
i + len(train_loader) * epoch)
if (iteration % iters_per_checkpoint == 0) or (iteration
== iteration_start):
if rank == 0:
checkpoint_path = "{}/waveglow-{:06d}.pt".format(
output_directory, iteration)
save_checkpoint(model,
optimizer,
learning_rate,
iteration,
checkpoint_path,
waveglow_config=waveglow_config)
info_path = os.path.join(output_directory, 'info.yml')
checkpoint_info = {
'name': os.path.basename(checkpoint_path),
'iteration': iteration,
'loss': reduced_loss
}
keep_n_checkpoints(info_path, checkpoint_info, 5)
if with_tensorboard:
# outputs[0].shape: torch.Size([1, 8, 1000])
with torch.no_grad():
d = model.infer(mel.data[0].unsqueeze(0),
sigma=sigma)
d = d.cpu().squeeze()
pred_audio = (d - d.min()) * 1.98 / (
d.max() - d.min()) - 0.99
logger.add_audio(
"generated/iteration-{}.wav".format(iteration),
pred_audio,
iteration,
sample_rate=trainset.sampling_rate,
)
true_audio = audio.data[0].squeeze()
logger.add_audio(
"original/iteration-{}.wav".format(iteration),
true_audio,
iteration,
sample_rate=trainset.sampling_rate,
)
# 查看频谱，直观了解生成语音的情况
mel_output = trainset.get_mel(pred_audio.cpu())
logger.add_image(
"generated/iteration-{}.png".format(iteration),
plot_spectrogram_to_numpy(
mel_output.data.cpu().numpy()),
iteration,
dataformats='HWC')
mel_input = mel.data[0]
logger.add_image(
"original/iteration-{}.png".format(iteration),
plot_spectrogram_to_numpy(
mel_input.data.cpu().numpy()),
iteration,
dataformats='HWC')
iteration += 1
def main(args=None):
if args is None:
args = model_utils.run_joint_three_models_get_local_args()
print('#start:\t', args.learning_rate, args.train_batch_size, args.num_train_epochs)
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size / args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_predict:
raise ValueError("At least one of `do_train` or `do_predict` must be True.")
if args.do_train:
if not args.train_file:
raise ValueError("If `do_train` is True, then `train_file` must be specified.")
if args.do_predict:
if not args.predict_file:
raise ValueError("If `do_predict` is True, then `predict_file` must be specified.")
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
raise ValueError("Output directory () already exists and is not empty.")
os.makedirs(args.output_dir, exist_ok=True)
#----------------------------------------------
labels_list = ["nmod", "conj", "acl:cl", "acl", "nmod:poss", "advcl", "xcomp"]
num_labels = len(labels_list)
#----------------------------------------------
tokenizer = BertTokenizer.from_pretrained(args.bert_model)
train_examples = None
num_train_steps = None
if args.do_train:
train_examples = read_many_examples(input_file=args.train_file, is_training=True)
num_train_steps = int(len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
model = BertForSpanWithHeadwordWithLabel.from_pretrained(args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.format(args.local_rank), num_labels=num_labels)
print(PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.format(args.local_rank))
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
# hack to remove pooler, which is not used
# thus it produce None grad that break apex
param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}]
t_total = num_train_steps
if args.local_rank != -1:
t_total = t_total // torch.distributed.get_world_size()
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
global_step = 0
if args.do_train:
cached_train_features_file = args.train_file + '_{0}_{1}_{2}_{3}'.format(
args.bert_model, str(args.max_seq_length), str(args.doc_stride), str(args.max_query_length))
train_features = None
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
except:
train_features = convert_examples_to_features(
examples=train_examples,
label_list=labels_list,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=True)
if args.local_rank == -1 or torch.distributed.get_rank() == 0:
# logger.info(" Saving train features into cached file %s", cached_train_features_file)
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
# logger.info("***** Running training *****")
# logger.info(" Num orig examples = %d", len(train_examples))
# logger.info(" Num split examples = %d", len(train_features))
# logger.info(" Batch size = %d", args.train_batch_size)
# logger.info(" Num steps = %d", num_train_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)
all_start_positions = torch.tensor([f.start_position for f in train_features], dtype=torch.long)
all_end_positions = torch.tensor([f.end_position for f in train_features], dtype=torch.long)
all_headword_positions = torch.tensor([f.headword_position for f in train_features], dtype=torch.long)
all_labels = torch.tensor([f.label_id for f in train_features], dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_start_positions,
all_end_positions, all_headword_positions, all_labels)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
if n_gpu == 1:
batch = tuple(t.to(device) for t in batch) # multi-gpu does scattering it-self
input_ids, input_mask, segment_ids, start_positions, end_positions, headword_positions, label_ids = batch
print('headword#####', len(headword_positions), headword_positions)
print('label_ids#####', len(label_ids), label_ids)
loss = model(input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask,
start_positions=start_positions, end_positions=end_positions,
headword_positions=headword_positions, label_ids=label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
# modify learning rate with special warm up BERT uses
lr_this_step = args.learning_rate * span_utils.warmup_linear(global_step / t_total, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
# Save a trained model
model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
torch.save(model_to_save.state_dict(), output_model_file)
# Load a trained model that you have fine-tuned
model_state_dict = torch.load(output_model_file)
model = BertForSpanWithHeadwordWithLabel.from_pretrained(
args.bert_model, state_dict=model_state_dict, num_labels=num_labels)
model.to(device)
if args.do_predict and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
eval_examples = read_many_examples(input_file=args.predict_file, is_training=False)
eval_features = convert_examples_to_features(
examples=eval_examples,
label_list=labels_list,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=False)
# logger.info("***** Running predictions *****")
# logger.info(" Num orig examples = %d", len(eval_examples))
# logger.info(" Num split examples = %d", len(eval_features))
# logger.info(" Batch size = %d", args.predict_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
#all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.predict_batch_size)
model.eval()
all_results = []
# logger.info("Start evaluating")
for input_ids, input_mask, segment_ids, example_indices in tqdm(eval_dataloader, desc="Evaluating"):
# if len(all_results) % 1000 == 0: logger.info("Processing example: %d" % (len(all_results)))
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
#label_ids = label_ids.to(device)
with torch.no_grad():
batch_start_logits, batch_end_logits, batch_headword_logits, batch_label_logits \
= model(input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask)
for i, example_index in enumerate(example_indices):
start_logits = batch_start_logits[i].detach().cpu().tolist()
end_logits = batch_end_logits[i].detach().cpu().tolist()
headword_logits = batch_headword_logits[i].detach().cpu().tolist()
label_logits = batch_label_logits[i].detach().cpu().tolist()
#label_logits_outputs = np.argmax(label_logits, axis=1)
#label_logits_outputs[0]
eval_feature = eval_features[example_index.item()]
unique_id = int(eval_feature.unique_id)
all_results.append(RawResult(unique_id=unique_id,
start_logits=start_logits,
end_logits=end_logits,
headword_logits=headword_logits,
label_logits=label_logits))
output_prediction_file = os.path.join(args.output_dir, "predictions.json")
output_nbest_file = os.path.join(args.output_dir, "nbest_predictions.json")
write_predictions(eval_examples, eval_features, all_results,
args.n_best_size, args.max_answer_length,
args.do_lower_case, output_prediction_file,
output_nbest_file, args.verbose_logging)
def predict(args, model, tokenizer, prefix=""):
metric = SpanEntityScore(args.id2label)
pred_output_dir = args.output_dir
if not os.path.exists(pred_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(pred_output_dir)
test_dataset = load_and_cache_examples(args,
args.task_name,
tokenizer,
data_type='test')
print(len(test_dataset))
# Note that DistributedSampler samples randomly
test_sampler = SequentialSampler(
test_dataset) if args.local_rank == -1 else DistributedSampler(
test_dataset)
test_dataloader = DataLoader(test_dataset,
sampler=test_sampler,
batch_size=1,
collate_fn=collate_fn)
# Eval!
logger.info("***** Running prediction %s *****", prefix)
logger.info(" Num examples = %d", len(test_dataset))
logger.info(" Batch size = %d", 1)
f_results = []
output_predict_file = os.path.join(pred_output_dir, prefix,
"span_test_predict.json")
# pbar = ProgressBar(n_total=len(test_dataloader), desc="Predicting")
for step, batch in enumerate(test_dataloader):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"start_positions": None,
"end_positions": None
}
if args.model_type != "distilbert":
# XLM and RoBERTa don"t use segment_ids
inputs["token_type_ids"] = (batch[2] if args.model_type
in ["bert", "xlnet"] else None)
outputs = model(**inputs)
start_logits, end_logits = outputs[:2]
R = bert_extract_item(start_logits, end_logits)
# T =
if R:
label_entities = [[args.id2label[x[0]], x[1], x[2]] for x in R]
else:
label_entities = []
json_d = {}
json_d['id'] = step
json_d['entities'] = label_entities
f_results.append(json_d)
logger.info("\n")
with open(output_predict_file, "w") as writer:
for record in f_results:
writer.write(json.dumps(record) + '\n')
# eval_loss = eval_loss / nb_eval_steps
# test_info, entity_info = metric.result()
# results = {f'{key}': value for key, value in test_info.items()}
# # results['loss'] = eval_loss
# logger.info("***** Test results %s *****", prefix)
# info = "-".join([f' {key}: {value:.4f} ' for key, value in results.items()])
# logger.info(info)
# logger.info("***** Entity results %s *****", prefix)
# for key in sorted(entity_info.keys()):
# logger.info("******* %s results ********" % key)
# info = "-".join([f' {key}: {value:.4f} ' for key, value in entity_info[key].items()])
# logger.info(info)
# # pbar(step)
if args.task_name == "cluener":
output_submit_file = os.path.join(pred_output_dir, prefix,
"test_submit.json")
test_text = []
with open(os.path.join(args.data_dir, "test.json"), 'r') as fr:
for line in fr:
test_text.append(json.loads(line))
test_submit = []
for x, y in zip(test_text, results):
json_d = {}
json_d['id'] = x['id']
json_d['label'] = {}
entities = y['entities']
words = list(x['text'])
if len(entities) != 0:
for subject in entities:
tag = subject[0]
start = subject[1]
end = subject[2]
word = "".join(words[start:end + 1])
if tag in json_d['label']:
if word in json_d['label'][tag]:
json_d['label'][tag][word].append([start, end])
else:
json_d['label'][tag][word] = [[start, end]]
else:
json_d['label'][tag] = {}
json_d['label'][tag][word] = [[start, end]]
test_submit.append(json_d)
json_to_text(output_submit_file, test_submit)
def train(args, train_dataset, model: PreTrainedModel,
tokenizer: PreTrainedTokenizer) -> Tuple[int, float]:
""" Train the model """
if args.local_rank in [-1, 0]:
tb_writer = SummaryWriter()
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
def collate(examples: List[torch.Tensor]):
if tokenizer._pad_token is None:
return pad_sequence(examples, batch_first=True)
fact_token_ids, fact_embedding_ids = zip(
*[get_inputs(seq, mask) for seq, mask, genre in examples])
seqs = [seq for seq, mask, genre in examples]
pad_seqs = pad_sequence(seqs,
batch_first=True,
padding_value=tokenizer.pad_token_id)
pad_facts = pad_sequence(fact_token_ids,
batch_first=True,
padding_value=tokenizer.pad_token_id)
pad_factsembeds = pad_sequence(fact_embedding_ids,
batch_first=True,
padding_value=tokenizer.pad_token_id)
return list(zip(pad_facts, pad_factsembeds, pad_seqs))
train_sampler = RandomSampler(
train_dataset) if args.local_rank == -1 else DistributedSampler(
train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size,
collate_fn=collate)
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (
len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0
},
]
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=t_total)
# Check if saved optimizer or scheduler states exist
if (args.model_name_or_path and os.path.isfile(
os.path.join(args.model_name_or_path, "optimizer.pt"))
and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler.pt"))):
# Load in optimizer and scheduler states
optimizer.load_state_dict(
torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
scheduler.load_state_dict(
torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
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)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
epochs_trained = 0
steps_trained_in_current_epoch = 0
# Check if continuing training from a checkpoint
if args.model_name_or_path and os.path.exists(args.model_name_or_path):
try:
# set global_step to gobal_step of last saved checkpoint from model path
checkpoint_suffix = args.model_name_or_path.split("-")[-1].split(
"/")[0]
global_step = int(checkpoint_suffix)
epochs_trained = global_step // (len(train_dataloader) //
args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (
len(train_dataloader) // args.gradient_accumulation_steps)
logger.info(
" Continuing training from checkpoint, will skip to saved global_step"
)
logger.info(" Continuing training from epoch %d", epochs_trained)
logger.info(" Continuing training from global step %d",
global_step)
logger.info(" Will skip the first %d steps in the first epoch",
steps_trained_in_current_epoch)
except ValueError:
logger.info(" Starting fine-tuning.")
tr_loss, logging_loss = 0.0, 0.0
model_to_resize = model.module if hasattr(
model,
"module") else model # Take care of distributed/parallel training
model_to_resize.resize_token_embeddings(len(tokenizer))
model.zero_grad()
train_iterator = trange(epochs_trained,
int(args.num_train_epochs),
desc="Epoch",
disable=args.local_rank not in [-1, 0])
set_seed(args) # Added here for reproducibility
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
model.train()
if args.mlm:
inputs, labels = mask_tokens(batch, tokenizer, args)
outputs = model(inputs, masked_lm_labels=labels)
elif args.xlnet:
with torch.no_grad():
pad_facts, pad_factsembeds, pad_seqs = zip(*batch)
tfacts = torch.stack(pad_facts).to(args.device)
tfact_embeds = torch.stack(pad_factsembeds).to(args.device)
facts_padding_masks = torch.where(
tfacts == tokenizer.pad_token_id,
torch.ones_like(tfacts),
torch.zeros_like(tfacts)).to(args.device)
tseqs = torch.stack(pad_seqs).to(args.device)
tseqs_padding_masks = torch.where(
tseqs == tokenizer.pad_token_id,
torch.ones_like(tseqs),
torch.zeros_like(tseqs)).to(args.device)
perm_masks = get_perm_masks(torch.zeros_like(tseqs),
order="L2R")
target_mapping = get_target_mapping(
torch.zeros_like(tseqs), device=args.device)
outputs = model(input_ids=tseqs,
facts_tokens=tfacts,
facts_embeds=tfact_embeds,
input_mask=tseqs_padding_masks,
facts_input_mask=facts_padding_masks,
perm_mask=perm_masks,
target_mapping=target_mapping)
else:
inputs, labels = (batch, batch)
outputs = model(inputs, labels=labels)
loss = outputs[
0] # model outputs are always tuple in transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
if args.local_rank in [
-1, 0
] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
# Log metrics
if (
args.local_rank == -1
and args.evaluate_during_training
): # Only evaluate when single GPU otherwise metrics may not average well
results = evaluate(args, model, tokenizer)
for key, value in results.items():
tb_writer.add_scalar("eval_{}".format(key), value,
global_step)
tb_writer.add_scalar("lr",
scheduler.get_lr()[0], global_step)
tb_writer.add_scalar("loss", (tr_loss - logging_loss) /
args.logging_steps, global_step)
lossval = (tr_loss - logging_loss) / args.logging_steps
print(f"train loss value at step {step} is {lossval}")
logger.info(
f"train loss value at step {step} is {lossval}")
logging_loss = tr_loss
if args.local_rank in [
-1, 0
] and args.save_steps > 0 and global_step % args.save_steps == 0:
checkpoint_prefix = "checkpoint"
# Save model checkpoint
output_dir = os.path.join(
args.output_dir,
"{}-{}".format(checkpoint_prefix, global_step))
os.makedirs(output_dir, exist_ok=True)
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
torch.save(args,
os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
_rotate_checkpoints(args, checkpoint_prefix)
torch.save(optimizer.state_dict(),
os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(),
os.path.join(output_dir, "scheduler.pt"))
logger.info("Saving optimizer and scheduler states to %s",
output_dir)
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
if args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
if args.local_rank in [-1, 0]:
tb_writer.close()
return global_step, tr_loss / global_step
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument("--src_file",
default=None,
type=str,
help="The input data file name.")
parser.add_argument("--tgt_file",
default=None,
type=str,
help="The output data file name.")
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese."
)
parser.add_argument("--config_path",
default=None,
type=str,
help="Bert config file path.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
parser.add_argument(
"--log_dir",
default='',
type=str,
required=True,
help="The output directory where the log will be written.")
parser.add_argument("--model_recover_path",
default=None,
type=str,
required=True,
help="The file of fine-tuned pretraining model.")
parser.add_argument("--optim_recover_path",
default=None,
type=str,
help="The file of pretraining optimizer.")
# Other parameters
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=64,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--label_smoothing",
default=0,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay",
default=0.01,
type=float,
help="The weight decay rate for Adam.")
parser.add_argument("--finetune_decay",
action='store_true',
help="Weight decay to the original weights.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--hidden_dropout_prob",
default=0.1,
type=float,
help="Dropout rate for hidden states.")
parser.add_argument("--attention_probs_dropout_prob",
default=0.1,
type=float,
help="Dropout rate for attention probabilities.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--fp32_embedding',
action='store_true',
help=
"Whether to use 32-bit float precision instead of 16-bit for embeddings"
)
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument('--amp',
action='store_true',
help="Whether to use amp for fp16")
parser.add_argument(
'--from_scratch',
action='store_true',
help=
"Initialize parameters with random values (i.e., training from scratch)."
)
parser.add_argument('--new_segment_ids',
action='store_true',
help="Use new segment ids for bi-uni-directional LM.")
parser.add_argument('--new_pos_ids',
action='store_true',
help="Use new position ids for LMs.")
parser.add_argument('--tokenized_input',
action='store_true',
help="Whether the input is tokenized.")
parser.add_argument('--max_len_a',
type=int,
default=0,
help="Truncate_config: maximum length of segment A.")
parser.add_argument('--max_len_b',
type=int,
default=0,
help="Truncate_config: maximum length of segment B.")
parser.add_argument(
'--trunc_seg',
default='',
help="Truncate_config: first truncate segment A/B (option: a, b).")
parser.add_argument(
'--always_truncate_tail',
action='store_true',
help="Truncate_config: Whether we should always truncate tail.")
parser.add_argument(
"--mask_prob",
default=0.15,
type=float,
help=
"Number of prediction is sometimes less than max_pred when sequence is short."
)
parser.add_argument(
"--mask_prob_eos",
default=0,
type=float,
help=
"Number of prediction is sometimes less than max_pred when sequence is short."
)
parser.add_argument('--max_pred',
type=int,
default=20,
help="Max tokens of prediction.")
parser.add_argument("--num_workers",
default=0,
type=int,
help="Number of workers for the data loader.")
parser.add_argument('--mask_source_words',
action='store_true',
help="Whether to mask source words for training")
parser.add_argument('--skipgram_prb',
type=float,
default=0.0,
help='prob of ngram mask')
parser.add_argument('--skipgram_size',
type=int,
default=1,
help='the max size of ngram mask')
parser.add_argument('--mask_whole_word',
action='store_true',
help="Whether masking a whole word.")
parser.add_argument('--do_l2r_training',
action='store_true',
help="Whether to do left to right training")
parser.add_argument(
'--has_sentence_oracle',
action='store_true',
help="Whether to have sentence level oracle for training. "
"Only useful for summary generation")
parser.add_argument('--max_position_embeddings',
type=int,
default=None,
help="max position embeddings")
parser.add_argument('--relax_projection',
action='store_true',
help="Use different projection layers for tasks.")
parser.add_argument('--ffn_type',
default=0,
type=int,
help="0: default mlp; 1: W((Wx+b) elem_prod x);")
parser.add_argument('--num_qkv',
default=0,
type=int,
help="Number of different <Q,K,V>.")
parser.add_argument('--seg_emb',
action='store_true',
help="Using segment embedding for self-attention.")
parser.add_argument(
'--s2s_special_token',
action='store_true',
help="New special tokens ([S2S_SEP]/[S2S_CLS]) of S2S.")
parser.add_argument('--s2s_add_segment',
action='store_true',
help="Additional segmental for the encoder of S2S.")
parser.add_argument(
'--s2s_share_segment',
action='store_true',
help=
"Sharing segment embeddings for the encoder of S2S (used with --s2s_add_segment)."
)
parser.add_argument('--pos_shift',
action='store_true',
help="Using position shift for fine-tuning.")
args = parser.parse_args()
assert Path(
args.model_recover_path).exists(), "--model_recover_path doesn't exist"
args.output_dir = args.output_dir.replace('[PT_OUTPUT_DIR]',
os.getenv('PT_OUTPUT_DIR', ''))
args.log_dir = args.log_dir.replace('[PT_OUTPUT_DIR]',
os.getenv('PT_OUTPUT_DIR', ''))
os.makedirs(args.output_dir, exist_ok=True)
os.makedirs(args.log_dir, exist_ok=True)
json.dump(args.__dict__,
open(os.path.join(args.output_dir, 'opt.json'), 'w'),
sort_keys=True,
indent=2)
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
dist.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if args.local_rank not in (-1, 0):
# Make sure only the first process in distributed training will download model & vocab
dist.barrier()
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
if args.max_position_embeddings:
tokenizer.max_len = args.max_position_embeddings
data_tokenizer = WhitespaceTokenizer(
) if args.tokenized_input else tokenizer
if args.local_rank == 0:
dist.barrier()
if args.do_train:
print("Loading Train Dataset", args.data_dir)
bi_uni_pipeline = [
seq2seq_loader.Preprocess4Seq2seq(
args.max_pred,
args.mask_prob,
list(tokenizer.vocab.keys()),
tokenizer.convert_tokens_to_ids,
args.max_seq_length,
new_segment_ids=args.new_segment_ids,
truncate_config={
'max_len_a': args.max_len_a,
'max_len_b': args.max_len_b,
'trunc_seg': args.trunc_seg,
'always_truncate_tail': args.always_truncate_tail
},
mask_source_words=args.mask_source_words,
skipgram_prb=args.skipgram_prb,
skipgram_size=args.skipgram_size,
mask_whole_word=args.mask_whole_word,
mode="s2s",
has_oracle=args.has_sentence_oracle,
num_qkv=args.num_qkv,
s2s_special_token=args.s2s_special_token,
s2s_add_segment=args.s2s_add_segment,
s2s_share_segment=args.s2s_share_segment,
pos_shift=args.pos_shift)
]
file_oracle = None
if args.has_sentence_oracle:
file_oracle = os.path.join(args.data_dir, 'train.oracle')
fn_src = os.path.join(args.data_dir,
args.src_file if args.src_file else 'train.src')
fn_tgt = os.path.join(args.data_dir,
args.tgt_file if args.tgt_file else 'train.tgt')
train_dataset = seq2seq_loader.Seq2SeqDataset(
fn_src,
fn_tgt,
args.train_batch_size,
data_tokenizer,
args.max_seq_length,
file_oracle=file_oracle,
bi_uni_pipeline=bi_uni_pipeline)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset, replacement=False)
_batch_size = args.train_batch_size
else:
train_sampler = DistributedSampler(train_dataset)
_batch_size = args.train_batch_size // dist.get_world_size()
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=_batch_size,
sampler=train_sampler,
num_workers=args.num_workers,
collate_fn=seq2seq_loader.batch_list_to_batch_tensors,
pin_memory=False)
# note: args.train_batch_size has been changed to (/= args.gradient_accumulation_steps)
# t_total = int(math.ceil(len(train_dataset.ex_list) / args.train_batch_size)
t_total = int(
len(train_dataloader) * args.num_train_epochs /
args.gradient_accumulation_steps)
amp_handle = None
if args.fp16 and args.amp:
from apex import amp
amp_handle = amp.init(enable_caching=True)
logger.info("enable fp16 with amp")
# Prepare model
recover_step = _get_max_epoch_model(args.output_dir)
cls_num_labels = 2
type_vocab_size = 6 + \
(1 if args.s2s_add_segment else 0) if args.new_segment_ids else 2
num_sentlvl_labels = 2 if args.has_sentence_oracle else 0
relax_projection = 4 if args.relax_projection else 0
if args.local_rank not in (-1, 0):
# Make sure only the first process in distributed training will download model & vocab
dist.barrier()
if (recover_step is None) and (args.model_recover_path is None):
# if _state_dict == {}, the parameters are randomly initialized
# if _state_dict == None, the parameters are initialized with bert-init
_state_dict = {} if args.from_scratch else None
model = BertForPreTrainingLossMask.from_pretrained(
args.bert_model,
state_dict=_state_dict,
num_labels=cls_num_labels,
num_rel=0,
type_vocab_size=type_vocab_size,
config_path=args.config_path,
task_idx=3,
num_sentlvl_labels=num_sentlvl_labels,
max_position_embeddings=args.max_position_embeddings,
label_smoothing=args.label_smoothing,
fp32_embedding=args.fp32_embedding,
relax_projection=relax_projection,
new_pos_ids=args.new_pos_ids,
ffn_type=args.ffn_type,
hidden_dropout_prob=args.hidden_dropout_prob,
attention_probs_dropout_prob=args.attention_probs_dropout_prob,
num_qkv=args.num_qkv,
seg_emb=args.seg_emb)
global_step = 0
else:
if recover_step:
logger.info("***** Recover model: %d *****", recover_step)
model_recover = torch.load(os.path.join(
args.output_dir, "model.{0}.bin".format(recover_step)),
map_location='cpu')
# recover_step == number of epochs
global_step = math.floor(recover_step * t_total /
args.num_train_epochs)
elif args.model_recover_path:
logger.info("***** Recover model: %s *****",
args.model_recover_path)
model_recover = torch.load(args.model_recover_path,
map_location='cpu')
global_step = 0
model = BertForPreTrainingLossMask.from_pretrained(
args.bert_model,
state_dict=model_recover,
num_labels=cls_num_labels,
num_rel=0,
type_vocab_size=type_vocab_size,
config_path=args.config_path,
task_idx=3,
num_sentlvl_labels=num_sentlvl_labels,
max_position_embeddings=args.max_position_embeddings,
label_smoothing=args.label_smoothing,
fp32_embedding=args.fp32_embedding,
relax_projection=relax_projection,
new_pos_ids=args.new_pos_ids,
ffn_type=args.ffn_type,
hidden_dropout_prob=args.hidden_dropout_prob,
attention_probs_dropout_prob=args.attention_probs_dropout_prob,
num_qkv=args.num_qkv,
seg_emb=args.seg_emb)
if args.local_rank == 0:
dist.barrier()
if args.fp16:
model.half()
if args.fp32_embedding:
model.bert.embeddings.word_embeddings.float()
model.bert.embeddings.position_embeddings.float()
model.bert.embeddings.token_type_embeddings.float()
model.to(device)
if args.local_rank != -1:
try:
from torch.nn.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError("DistributedDataParallel")
model = DDP(model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
elif n_gpu > 1:
# model = torch.nn.DataParallel(model)
model = DataParallelImbalance(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
# from apex.optimizers import FP16_Optimizer
from pytorch_pretrained_bert.optimization_fp16 import FP16_Optimizer_State
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer_State(optimizer,
dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer_State(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
if recover_step:
logger.info("***** Recover optimizer: %d *****", recover_step)
optim_recover = torch.load(os.path.join(
args.output_dir, "optim.{0}.bin".format(recover_step)),
map_location='cpu')
if hasattr(optim_recover, 'state_dict'):
optim_recover = optim_recover.state_dict()
optimizer.load_state_dict(optim_recover)
if args.loss_scale == 0:
logger.info("***** Recover optimizer: dynamic_loss_scale *****")
optimizer.dynamic_loss_scale = True
logger.info("***** CUDA.empty_cache() *****")
torch.cuda.empty_cache()
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", t_total)
model.train()
if recover_step:
start_epoch = recover_step + 1
else:
start_epoch = 1
for i_epoch in trange(start_epoch,
int(args.num_train_epochs) + 1,
desc="Epoch",
disable=args.local_rank not in (-1, 0)):
if args.local_rank != -1:
train_sampler.set_epoch(i_epoch)
iter_bar = tqdm(train_dataloader,
desc='Iter (loss=X.XXX)',
disable=args.local_rank not in (-1, 0))
for step, batch in enumerate(iter_bar):
batch = [
t.to(device) if t is not None else None for t in batch
]
if args.has_sentence_oracle:
input_ids, segment_ids, input_mask, mask_qkv, lm_label_ids, masked_pos, masked_weights, is_next, task_idx, oracle_pos, oracle_weights, oracle_labels = batch
else:
input_ids, segment_ids, input_mask, mask_qkv, lm_label_ids, masked_pos, masked_weights, is_next, task_idx = batch
oracle_pos, oracle_weights, oracle_labels = None, None, None
loss_tuple = model(input_ids,
segment_ids,
input_mask,
lm_label_ids,
is_next,
masked_pos=masked_pos,
masked_weights=masked_weights,
task_idx=task_idx,
masked_pos_2=oracle_pos,
masked_weights_2=oracle_weights,
masked_labels_2=oracle_labels,
mask_qkv=mask_qkv)
masked_lm_loss, next_sentence_loss = loss_tuple
if n_gpu > 1: # mean() to average on multi-gpu.
# loss = loss.mean()
masked_lm_loss = masked_lm_loss.mean()
next_sentence_loss = next_sentence_loss.mean()
loss = masked_lm_loss + next_sentence_loss
# logging for each step (i.e., before normalization by args.gradient_accumulation_steps)
iter_bar.set_description('Iter (loss=%5.3f)' % loss.item())
# ensure that accumlated gradients are normalized
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
if amp_handle:
amp_handle._clear_cache()
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
lr_this_step = args.learning_rate * \
warmup_linear(global_step/t_total,
args.warmup_proportion)
if args.fp16:
# modify learning rate with special warm up BERT uses
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
# Save a trained model
if (args.local_rank == -1 or torch.distributed.get_rank() == 0):
logger.info(
"** ** * Saving fine-tuned model and optimizer ** ** * ")
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(
args.output_dir, "model.{0}.bin".format(i_epoch))
torch.save(model_to_save.state_dict(), output_model_file)
output_optim_file = os.path.join(
args.output_dir, "optim.{0}.bin".format(i_epoch))
torch.save(optimizer.state_dict(), output_optim_file)
logger.info("***** CUDA.empty_cache() *****")
torch.cuda.empty_cache()
def evaluate(args, model, eval_dataset=None, tokenizer=None):
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
if not eval_dataset and tokenizer:
eval_dataset = load_and_cache_examples(args, tokenizer, set_type='dev')
if not eval_dataset:
raise ValueError('The eval or test dataset can not be None')
eval_sampler = SequentialSampler(
eval_dataset) if args.local_rank == -1 else DistributedSampler(
eval_dataset)
eval_dataloader = DataLoader(eval_dataset,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
# multi-gpu eval
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Eval!
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
eval_loss = 0.0
nb_eval_steps = 0
preds = None
out_label_ids = None
model.eval()
for batch in tqdm(eval_dataloader, desc="Evaluating"):
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {
'input_ids': batch[0],
'attention_mask': batch[1],
'token_type_ids': batch[2],
'labels': batch[3]
}
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if preds is None:
preds = logits.detach().cpu().numpy()
out_label_ids = inputs["labels"].detach().cpu().numpy()
else:
preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
out_label_ids = np.append(out_label_ids,
inputs["labels"].detach().cpu().numpy(),
axis=0)
eval_loss = eval_loss / nb_eval_steps
preds = np.argmax(preds, axis=1)
result = compute_metrics(out_label_ids, preds, average='binary')
result['eval_loss'] = round(eval_loss, 4)
report = classification_report(out_label_ids, preds, digits=5)
print(report)
# out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
# eval_loss = eval_loss / nb_eval_steps
# preds = np.argmax(preds, axis=1)
# report = classification_report(out_label_ids, preds)
# print(report)
# result = compute_metrics(out_label_ids, preds, average='macro')
# result['eval_loss'] = round(eval_loss, 4)
return result, preds
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument("--pregenerated_data", type=Path, required=True)
parser.add_argument("--teacher_model",
default=None,
type=str,
required=True)
parser.add_argument("--student_model",
default=None,
type=str,
required=True)
parser.add_argument("--output_dir", default=None, type=str, required=True)
# Other parameters
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument(
"--reduce_memory",
action="store_true",
help=
"Store training data as on-disc memmaps to massively reduce memory usage"
)
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument('--weight_decay',
'--wd',
default=1e-4,
type=float,
metavar='W',
help='weight decay')
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--continue_train',
action='store_true',
help='Whether to train from checkpoints')
# Additional arguments
parser.add_argument('--eval_step', type=int, default=1000)
# This is used for running on Huawei Cloud.
parser.add_argument('--data_url', type=str, default="")
args = parser.parse_args()
logger.info('args:{}'.format(args))
samples_per_epoch = []
for i in range(int(args.num_train_epochs)):
epoch_file = args.pregenerated_data / "epoch_{}.json".format(i)
metrics_file = args.pregenerated_data / "epoch_{}_metrics.json".format(
i)
if epoch_file.is_file() and metrics_file.is_file():
metrics = json.loads(metrics_file.read_text())
samples_per_epoch.append(metrics['num_training_examples'])
else:
if i == 0:
exit("No training data was found!")
print(
"Warning! There are fewer epochs of pregenerated data ({}) than training epochs ({})."
.format(i, args.num_train_epochs))
print(
"This script will loop over the available data, but training diversity may be negatively impacted."
)
num_data_epochs = i
break
else:
num_data_epochs = args.num_train_epochs
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.teacher_model,
do_lower_case=args.do_lower_case)
total_train_examples = 0
for i in range(int(args.num_train_epochs)):
# The modulo takes into account the fact that we may loop over limited epochs of data
total_train_examples += samples_per_epoch[i % len(samples_per_epoch)]
num_train_optimization_steps = int(total_train_examples /
args.train_batch_size /
args.gradient_accumulation_steps)
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
if args.continue_train:
student_model = TinyBertForPreTraining.from_pretrained(
args.student_model)
else:
student_model = TinyBertForPreTraining.from_scratch(args.student_model)
teacher_model = BertModel.from_pretrained(args.teacher_model)
# student_model = TinyBertForPreTraining.from_scratch(args.student_model, fit_size=teacher_model.config.hidden_size)
student_model.to(device)
teacher_model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
teacher_model = DDP(teacher_model)
elif n_gpu > 1:
student_model = torch.nn.DataParallel(student_model)
teacher_model = torch.nn.DataParallel(teacher_model)
size = 0
for n, p in student_model.named_parameters():
logger.info('n: {}'.format(n))
logger.info('p: {}'.format(p.nelement()))
size += p.nelement()
logger.info('Total parameters: {}'.format(size))
# Prepare optimizer
param_optimizer = list(student_model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
loss_mse = MSELoss()
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
logging.info("***** Running training *****")
logging.info(" Num examples = {}".format(total_train_examples))
logging.info(" Batch size = %d", args.train_batch_size)
logging.info(" Num steps = %d", num_train_optimization_steps)
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
epoch_dataset = PregeneratedDataset(
epoch=epoch,
training_path=args.pregenerated_data,
tokenizer=tokenizer,
num_data_epochs=num_data_epochs,
reduce_memory=args.reduce_memory)
if args.local_rank == -1:
train_sampler = RandomSampler(epoch_dataset)
else:
train_sampler = DistributedSampler(epoch_dataset)
train_dataloader = DataLoader(epoch_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
tr_loss = 0.
tr_att_loss = 0.
tr_rep_loss = 0.
student_model.train()
nb_tr_examples, nb_tr_steps = 0, 0
with tqdm(total=len(train_dataloader),
desc="Epoch {}".format(epoch)) as pbar:
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration", ascii=True)):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids, is_next = batch
if input_ids.size()[0] != args.train_batch_size:
continue
att_loss = 0.
rep_loss = 0.
student_atts, student_reps = student_model(
input_ids, segment_ids, input_mask)
teacher_reps, teacher_atts, _ = teacher_model(
input_ids, segment_ids, input_mask)
teacher_reps = [
teacher_rep.detach() for teacher_rep in teacher_reps
] # speedup 1.5x
teacher_atts = [
teacher_att.detach() for teacher_att in teacher_atts
]
teacher_layer_num = len(teacher_atts)
student_layer_num = len(student_atts)
assert teacher_layer_num % student_layer_num == 0
layers_per_block = int(teacher_layer_num / student_layer_num)
new_teacher_atts = [
teacher_atts[i * layers_per_block + layers_per_block - 1]
for i in range(student_layer_num)
]
for student_att, teacher_att in zip(student_atts,
new_teacher_atts):
student_att = torch.where(
student_att <= -1e2,
torch.zeros_like(student_att).to(device), student_att)
teacher_att = torch.where(
teacher_att <= -1e2,
torch.zeros_like(teacher_att).to(device), teacher_att)
att_loss += loss_mse(student_att, teacher_att)
new_teacher_reps = [
teacher_reps[i * layers_per_block]
for i in range(student_layer_num + 1)
]
new_student_reps = student_reps
for student_rep, teacher_rep in zip(new_student_reps,
new_teacher_reps):
rep_loss += loss_mse(student_rep, teacher_rep)
loss = att_loss + rep_loss
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_att_loss += att_loss.item()
tr_rep_loss += rep_loss.item()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
pbar.update(1)
mean_loss = tr_loss * args.gradient_accumulation_steps / nb_tr_steps
mean_att_loss = tr_att_loss * args.gradient_accumulation_steps / nb_tr_steps
mean_rep_loss = tr_rep_loss * args.gradient_accumulation_steps / nb_tr_steps
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
global_step += 1
if (global_step + 1) % args.eval_step == 0:
result = {}
result['global_step'] = global_step
result['loss'] = mean_loss
result['att_loss'] = mean_att_loss
result['rep_loss'] = mean_rep_loss
output_eval_file = os.path.join(
args.output_dir, "log.txt")
with open(output_eval_file, "a") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" %
(key, str(result[key])))
# Save a trained model
model_name = "step_{}_{}".format(
global_step, WEIGHTS_NAME)
logging.info(
"** ** * Saving fine-tuned model ** ** * ")
# Only save the model it-self
model_to_save = student_model.module if hasattr(
student_model, 'module') else student_model
output_model_file = os.path.join(
args.output_dir, model_name)
output_config_file = os.path.join(
args.output_dir, CONFIG_NAME)
torch.save(model_to_save.state_dict(),
output_model_file)
model_to_save.config.to_json_file(output_config_file)
tokenizer.save_vocabulary(args.output_dir)
if oncloud:
logging.info(
mox.file.list_directory(args.output_dir,
recursive=True))
logging.info(
mox.file.list_directory('.', recursive=True))
mox.file.copy_parallel(args.output_dir,
args.data_url)
mox.file.copy_parallel('.', args.data_url)
model_name = "step_{}_{}".format(global_step, WEIGHTS_NAME)
logging.info("** ** * Saving fine-tuned model ** ** * ")
model_to_save = student_model.module if hasattr(
student_model, 'module') else student_model
output_model_file = os.path.join(args.output_dir, model_name)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
model_to_save.config.to_json_file(output_config_file)
tokenizer.save_vocabulary(args.output_dir)
if oncloud:
logging.info(
mox.file.list_directory(args.output_dir, recursive=True))
logging.info(mox.file.list_directory('.', recursive=True))
mox.file.copy_parallel(args.output_dir, args.data_url)
mox.file.copy_parallel('.', args.data_url)
def train(args, tokenizer, model):
""" Train the model """
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_dataset = load_and_cache_examples(
args, tokenizer, set_type='train') #load_and_cache_examples 就是加载训练数据集
train_sampler = RandomSampler(
train_dataset) if args.local_rank == -1 else DistributedSampler(
train_dataset) #对数据随机采样
train_dataloader = DataLoader(
train_dataset, sampler=train_sampler, batch_size=args.train_batch_size
) #DataLoader读取数据，计算总共训练步数（梯度累计），warm_up 参数设定，优化器，是否fp16等等，然后一个batch一个batch进行训练
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (
len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
eval_dataset = None
if args.do_eval_during_train:
eval_dataset = load_and_cache_examples(args, tokenizer, set_type='dev')
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ['bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
'weight_decay':
args.weight_decay
}, {
'params': [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0
}]
args.warmup_steps = int(t_total * args.warmup_rate)
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=t_total)
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 args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1))
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
tr_loss, logging_loss = 0.0, 0.0
best_acc = 0
model.zero_grad()
train_iterator = trange(int(args.num_train_epochs),
desc="Epoch",
disable=args.local_rank not in [-1, 0])
set_seed(
args) # Added here for reproductibility (even between python 2 and 3)
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator):
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {
'input_ids': batch[0],
'attention_mask': batch[1],
'token_type_ids': batch[2],
'labels': batch[3]
}
#最核心的代码就是下面的把数据和参数送入到模型中去
outputs = model(**inputs)
loss = outputs[
0] # model outputs are always tuple in transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
epoch_iterator.set_description("loss {}".format(
round(loss.item(), 4)))
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
if args.local_rank in [
-1, 0
] and args.logging_steps > 0 and args.do_eval_during_train and (
global_step % args.logging_steps == 0 or
(global_step + 1) == t_total) and eval_dataset:
eval_result, _ = evaluate(args, model, eval_dataset,
tokenizer)
output_dir = args.output_dir
if not os.path.exists(output_dir):
os.makedirs(output_dir)
current_acc = eval_result['accuracy']
logger.info(" best accuracy : {}".format(best_acc))
logger.info(" current accuracy : {}".format(current_acc))
logger.info(" current step : {}".format(global_step))
logger.info(" ")
for k in eval_result.keys():
logger.info(" eval {} : {}".format(k, eval_result[k]))
if current_acc > best_acc:
best_acc = current_acc
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
logger.info("Saving model checkpoint to %s",
output_dir)
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
if args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
return global_step
def train(args, train_dataset, model, tokenizer):
""" Train the model """
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = RandomSampler(
train_dataset) if args.local_rank == -1 else DistributedSampler(
train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size,
collate_fn=collate_fn)
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (
len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ["bias", "LayerNorm.weight"]
bert_param_optimizer = list(model.bert.named_parameters())
crf_param_optimizer = list(model.crf.named_parameters())
linear_param_optimizer = list(model.classifier.named_parameters())
optimizer_grouped_parameters = [{
'params': [
p for n, p in bert_param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
args.weight_decay,
'lr':
args.learning_rate
}, {
'params': [
p for n, p in bert_param_optimizer
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0,
'lr':
args.learning_rate
}, {
'params': [
p for n, p in crf_param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
args.weight_decay,
'lr':
args.crf_learning_rate
}, {
'params':
[p for n, p in crf_param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0,
'lr':
args.crf_learning_rate
}, {
'params': [
p for n, p in linear_param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
args.weight_decay,
'lr':
args.crf_learning_rate
}, {
'params': [
p for n, p in linear_param_optimizer
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0,
'lr':
args.crf_learning_rate
}]
args.warmup_steps = int(t_total * args.warmup_proportion)
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=t_total)
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(
args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler.pt")):
# Load in optimizer and scheduler states
optimizer.load_state_dict(
torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
scheduler.load_state_dict(
torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
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)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
steps_trained_in_current_epoch = 0
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path
) and "checkpoint" in args.model_name_or_path:
# set global_step to gobal_step of last saved checkpoint from model path
global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
epochs_trained = global_step // (len(train_dataloader) //
args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (
len(train_dataloader) // args.gradient_accumulation_steps)
logger.info(
" Continuing training from checkpoint, will skip to saved global_step"
)
logger.info(" Continuing training from epoch %d", epochs_trained)
logger.info(" Continuing training from global step %d", global_step)
logger.info(" Will skip the first %d steps in the first epoch",
steps_trained_in_current_epoch)
tr_loss, logging_loss = 0.0, 0.0
if args.do_adv:
fgm = FGM(model, emb_name=args.adv_name, epsilon=args.adv_epsilon)
model.zero_grad()
seed_everything(
args.seed
) # Added here for reproductibility (even between python 2 and 3)
for _ in range(int(args.num_train_epochs)):
pbar = ProgressBar(n_total=len(train_dataloader), desc='Training')
for step, batch in enumerate(train_dataloader):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"labels": batch[3],
'input_lens': batch[4]
}
if args.model_type != "distilbert":
# XLM and RoBERTa don"t use segment_ids
inputs["token_type_ids"] = (batch[2] if args.model_type
in ["bert", "xlnet"] else None)
outputs = model(**inputs)
loss = outputs[
0] # model outputs are always tuple in pytorch-transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
pbar(step, {'loss': loss.item()})
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
scheduler.step() # Update learning rate schedule
optimizer.step()
model.zero_grad()
global_step += 1
if args.local_rank in [
-1, 0
] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
# Log metrics
print(" ")
if args.local_rank == -1:
# Only evaluate when single GPU otherwise metrics may not average well
evaluate(args, model, tokenizer)
if args.local_rank in [
-1, 0
] and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(
args.output_dir, "checkpoint-{}".format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
torch.save(args,
os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
tokenizer.save_vocabulary(output_dir)
torch.save(optimizer.state_dict(),
os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(),
os.path.join(output_dir, "scheduler.pt"))
logger.info("Saving optimizer and scheduler states to %s",
output_dir)
logger.info("\n")
if 'cuda' in str(args.device):
torch.cuda.empty_cache()
return global_step, tr_loss / global_step
def train(args, train_dataset, model, tokenizer, criterion):
""" Train the model """
if args.local_rank in [-1, 0]:
tb_writer = SummaryWriter()
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
train_dataloader = DataLoader(
train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size,
collate_fn=collate_fn,
num_workers=args.num_workers,
)
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay,
},
{"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
)
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)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True
)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size
* args.gradient_accumulation_steps
* (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
tr_loss, logging_loss = 0.0, 0.0
best_f1, n_no_improve = 0, 0
model.zero_grad()
train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
set_seed(args) # Added here for reproductibility
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator):
model.train()
batch = tuple(t.to(args.device) for t in batch)
labels = batch[5]
inputs = {
"input_ids": batch[0],
"input_modal": batch[2],
"attention_mask": batch[1],
"modal_start_tokens": batch[3],
"modal_end_tokens": batch[4],
}
outputs = model(**inputs)
logits = outputs[0] # model outputs are always tuple in transformers (see doc)
loss = criterion(logits, labels)
if args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
logs = {}
if (
args.local_rank == -1 and args.evaluate_during_training
): # Only evaluate when single GPU otherwise metrics may not average well
results = evaluate(args, model, tokenizer, criterion)
for key, value in results.items():
eval_key = "eval_{}".format(key)
logs[eval_key] = value
loss_scalar = (tr_loss - logging_loss) / args.logging_steps
learning_rate_scalar = scheduler.get_lr()[0]
logs["learning_rate"] = learning_rate_scalar
logs["loss"] = loss_scalar
logging_loss = tr_loss
for key, value in logs.items():
tb_writer.add_scalar(key, value, global_step)
print(json.dumps({**logs, **{"step": global_step}}))
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
torch.save(model_to_save.state_dict(), os.path.join(output_dir, WEIGHTS_NAME))
torch.save(args, os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
if args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
if args.local_rank == -1:
results = evaluate(args, model, tokenizer, criterion)
if results["micro_f1"] > best_f1:
best_f1 = results["micro_f1"]
n_no_improve = 0
else:
n_no_improve += 1
if n_no_improve > args.patience:
train_iterator.close()
break
if args.local_rank in [-1, 0]:
tb_writer.close()
return global_step, tr_loss / global_step
def train(args, train_dataset, model, tokenizer):
""" Train the model """
if args.local_rank in [-1, 0]:
tb_writer = SummaryWriter()
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = RandomSampler(
train_dataset) if args.local_rank == -1 else DistributedSampler(
train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (
len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ['bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
'weight_decay':
args.weight_decay
}, {
'params': [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0
}]
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=t_total)
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)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Num seq size = %d", len(train_dataset[0])) # 512
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1))
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
train_iterator = trange(int(args.num_train_epochs),
desc="Epoch",
disable=args.local_rank not in [-1, 0])
set_seed(
args) # Added here for reproducibility (even between python 2 and 3)
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator):
inputs, labels = mask_tokens(batch, tokenizer,
args) if args.mlm else (batch, batch)
inputs = inputs.to(args.device)
labels = labels.to(args.device)
model.train()
# import pdb
# pdb.interact
outputs = model(inputs,
masked_lm_labels=labels) if args.mlm else model(
inputs, labels=labels)
loss = outputs[
0] # model outputs are always tuple in transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
if args.local_rank in [
-1, 0
] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
# Log metrics
if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well
results = evaluate(args, model, tokenizer)
for key, value in results.items():
tb_writer.add_scalar('eval_{}'.format(key), value,
global_step)
tb_writer.add_scalar('lr',
scheduler.get_lr()[0], global_step)
tb_writer.add_scalar('loss', (tr_loss - logging_loss) /
args.logging_steps, global_step)
logging_loss = tr_loss
if args.local_rank in [
-1, 0
] and args.save_steps > 0 and global_step % args.save_steps == 0:
checkpoint_prefix = 'checkpoint'
# Save model checkpoint
output_dir = os.path.join(
args.output_dir,
'{}-{}'.format(checkpoint_prefix, global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = model.module if hasattr(
model, 'module'
) else model # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
torch.save(args,
os.path.join(output_dir, 'training_args.bin'))
logger.info("Saving model checkpoint to %s", output_dir)
_rotate_checkpoints(args, checkpoint_prefix)
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
if args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
if args.local_rank in [-1, 0]:
tb_writer.close()
return global_step, tr_loss / global_step
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--input_model_file",
default=None,
type=str,
required=False,
help="The input model file to load if not training, if any.")
parser.add_argument(
"--input_lm_model_file",
default=None,
type=str,
required=False,
help="The input model file to load for masked lm, if any.")
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--do_mask_eval",
action='store_true',
help="Whether to run mask eval on the dev set.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--mask_eval_outfile",
default="mask_eval_out.json",
type=str,
help="Where to store mask eval results (large).")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
args = parser.parse_args()
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mrpc": MrpcProcessor,
"sst": SstProcessor,
"lmrd": LmrdProcessor,
}
num_labels_task = {
"cola": 2,
"mnli": 3,
"mrpc": 2,
"sst": 2,
"lmrd": 2,
}
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval and not args.do_mask_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
os.makedirs(args.output_dir, exist_ok=True)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
num_labels = num_labels_task[task_name]
label_list = processor.get_labels()
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
model = BertForSequenceClassification.from_pretrained(
args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(args.local_rank),
num_labels=num_labels)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
t_total = num_train_steps
if args.local_rank != -1:
t_total = t_total // torch.distributed.get_world_size()
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
global_step = 0
nb_tr_steps = 0
tr_loss = 0
if args.do_train:
get_train_features = convert_examples_to_features(
train_examples, label_list, args.max_seq_length, tokenizer)
train_features = next(get_train_features)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(
" Num inputs = %d",
len(train_features)) # can be different from seq oversize drops
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = model(input_ids, segment_ids, input_mask, label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
# modify learning rate with special warm up BERT uses
lr_this_step = args.learning_rate * warmup_linear(
global_step / t_total, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
# Eval at every epoch if flag is up to track progress against dev set.
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_model(processor, args.data_dir, label_list,
args.max_seq_length, args.eval_batch_size,
tokenizer, device, model, tr_loss, nb_tr_steps,
args.do_train, global_step, args.output_dir, epoch)
# Save a trained model
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
if args.do_train:
torch.save(model_to_save.state_dict(), output_model_file)
# Load a trained model that you have previously fine-tuned
if not args.do_train and args.input_model_file is not None:
logger.info("Loading previously trained model from %s" %
args.input_model_file)
model_state_dict = torch.load(args.input_model_file)
model = BertForSequenceClassification.from_pretrained(
args.bert_model,
state_dict=model_state_dict,
num_labels=num_labels)
model.to(device)
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_model(processor, args.data_dir, label_list, args.max_seq_length,
args.eval_batch_size, tokenizer, device, model, tr_loss,
nb_tr_steps, args.do_train, global_step, args.output_dir,
int(args.num_train_epochs))
if args.do_mask_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
# Load a trained model that you have previously fine-tuned for LM.
if args.input_lm_model_file is not None:
logger.info("Loading previously trained LM model from %s" %
args.input_lm_model_file)
model_state_dict = torch.load(args.input_lm_model_file)
maskedLMModel = BertForMaskedLM.from_pretrained(
args.bert_model, state_dict=model_state_dict)
model.to(device)
else:
maskedLMModel = BertForMaskedLM.from_pretrained(args.bert_model)
mask_eval_model(processor, args.data_dir, label_list,
args.max_seq_length, args.eval_batch_size, tokenizer,
device, model, maskedLMModel, args.mask_eval_outfile)
def evaluate(args, model, tokenizer, prefix=""):
dataset, examples, features = load_and_cache_examples(args,
tokenizer,
evaluate=True,
output_examples=True)
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
os.makedirs(args.output_dir)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(
dataset) if args.local_rank == -1 else DistributedSampler(dataset)
eval_dataloader = DataLoader(dataset,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
# Eval!
logger.info("***** Running evaluation {} *****".format(prefix))
logger.info(" Num examples = %d", len(dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
all_results = []
for batch in tqdm(eval_dataloader, desc="Evaluating"):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {
'input_ids': batch[0],
'attention_mask': batch[1],
'token_type_ids': None if args.model_type == 'xlm' else
batch[2] # XLM don't use segment_ids
}
example_indices = batch[3]
if args.model_type in ['xlnet', 'xlm']:
inputs.update({'cls_index': batch[4], 'p_mask': batch[5]})
outputs = model(**inputs)
for i, example_index in enumerate(example_indices):
eval_feature = features[example_index.item()]
unique_id = int(eval_feature.unique_id)
if args.model_type in ['xlnet', 'xlm']:
# XLNet uses a more complex post-processing procedure
result = RawResultExtended(
unique_id=unique_id,
start_top_log_probs=to_list(outputs[0][i]),
start_top_index=to_list(outputs[1][i]),
end_top_log_probs=to_list(outputs[2][i]),
end_top_index=to_list(outputs[3][i]),
cls_logits=to_list(outputs[4][i]))
else:
result = RawResult(unique_id=unique_id,
start_logits=to_list(outputs[0][i]),
end_logits=to_list(outputs[1][i]))
all_results.append(result)
# Compute predictions
output_prediction_file = os.path.join(args.output_dir,
"predictions_{}.json".format(prefix))
output_nbest_file = os.path.join(
args.output_dir, "nbest_predictions_{}.json".format(prefix))
if args.version_2_with_negative:
output_null_log_odds_file = os.path.join(
args.output_dir, "null_odds_{}.json".format(prefix))
else:
output_null_log_odds_file = None
if args.model_type in ['xlnet', 'xlm']:
# XLNet uses a more complex post-processing procedure
write_predictions_extended(
examples, features, all_results, args.n_best_size,
args.max_answer_length, output_prediction_file, output_nbest_file,
output_null_log_odds_file, args.predict_file,
model.config.start_n_top, model.config.end_n_top,
args.version_2_with_negative, tokenizer, args.verbose_logging)
else:
write_predictions(examples, features, all_results, args.n_best_size,
args.max_answer_length, args.do_lower_case,
output_prediction_file, output_nbest_file,
output_null_log_odds_file, args.verbose_logging,
args.version_2_with_negative,
args.null_score_diff_threshold)
# Evaluate with the official SQuAD script
evaluate_options = EVAL_OPTS(data_file=args.predict_file,
pred_file=output_prediction_file,
na_prob_file=output_null_log_odds_file)
results = evaluate_on_squad(evaluate_options)
return results
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--corpus",
default='yelp',
type=str,
help="The input corpus.")
parser.add_argument(
"--bert_model",
default='bert-base-uncased',
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese."
)
parser.add_argument(
"--output_dir",
default='tmp',
type=str,
help="The output directory where the model checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--max_seq_length",
default=24,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run evaluation.")
parser.add_argument("--train_batch_size",
default=16,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=16,
type=int,
help="Total batch size for training.")
parser.add_argument("--learning_rate",
default=3e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument(
"--on_memory",
action='store_true',
help="Whether to load train samples into memory or use disk")
parser.add_argument(
"--do_lower_case",
action='store_true',
help=
"Whether to lower case the input text. True for uncased models, False for cased models."
)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumualte before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument('-dm',
'--dmax',
type=int,
default=5,
help='Max number of documents')
parser.add_argument('-sm',
'--smax',
type=int,
default=20,
help='Max number of words in each document')
args = parser.parse_args()
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
# load data
print('load data...')
data, length = loadPrepareData(args)
user_length, item_length = length #, user_length2, item_length2 = length
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train:
raise ValueError(
"Training is currently the only implemented execution option. Please set `do_train`."
)
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
#train_examples = None
num_train_optimization_steps = None
if args.do_train:
print("Loading Train Dataset")
# build train dataset
train_dataset = BERTDataset(data.train,
seq_len=args.max_seq_length,
tokenizer=tokenizer,
fa2tokids=data.fa2tokids)
num_train_optimization_steps = int(
len(train_dataset) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
if args.do_eval:
print("Loading Dev Dataset")
# build train dataset
eval_dataset = BERTDataset(data.dev,
seq_len=args.max_seq_length,
tokenizer=tokenizer,
fa2tokids=data.fa2tokids)
# Prepare model
# model = BertForPreTraining.from_pretrained(args.bert_model)
model = BertMLMDecoder.from_pretrained(args.bert_model)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
warmup_linear = WarmupLinearSchedule(
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
# start training
global_step = 0
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
#TODO: check if this works with current data generator from disk that relies on next(file)
# (it doesn't return item back by index)
train_sampler = DistributedSampler(train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
if args.do_eval:
eval_sampler = SequentialSampler(eval_dataset)
eval_dataloader = DataLoader(eval_dataset,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
model.train()
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids, fa_ids, fa_mask = batch
loss = model(input_ids, segment_ids, input_mask, lm_label_ids,
fa_ids, fa_mask)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear.get_lr(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
loss = tr_loss / nb_tr_steps if args.do_train else None
print("Tr loss = {}".format(loss))
if args.do_eval:
model.eval()
eval_loss = 0
nb_eval_examples, nb_eval_steps = 0, 0
for step, batch in enumerate(
tqdm(eval_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids, fa_ids, fa_mask = batch
with torch.no_grad():
tmp_eval_loss = model(input_ids, segment_ids,
input_mask, lm_label_ids, fa_ids,
fa_mask)
eval_loss += tmp_eval_loss.mean().item()
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
print("Eval loss = {}".format(eval_loss))
# Save a trained model
logger.info("** ** * Saving fine - tuned model ** ** * ")
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(
args.output_dir, "pytorch_model_{}.bin".format(epoch))
if args.do_train:
torch.save(model_to_save.state_dict(), output_model_file)
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese."
)
parser.add_argument("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
default=False,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
default=False,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
default=False,
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
default=False,
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
default=False,
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument(
'--log_path',
type=str,
default="./log",
help="The path for saving tensorboard logs. Default is ./log")
args = parser.parse_args()
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
os.makedirs(args.output_dir, exist_ok=True)
processors = {
"qe": MyProcessor,
}
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
model_collections = Collections()
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
model = BertForSequenceClassification.from_pretrained(
args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(args.local_rank))
model.to(device)
# fine-tuning fine-tuning model
# output_config_file = os.path.join(args.bert_model, CONFIG_NAME)
# config = BertConfig(output_config_file)
# model = BertForSequenceClassification(config)
#
# output_model_file = os.path.join(args.bert_model, WEIGHTS_NAME)
# model_state_dict = torch.load(output_model_file)
# model.load_state_dict(model_state_dict)
# model.to(device)
#-----------------------------
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
t_total = num_train_steps
if args.local_rank != -1:
t_total = t_total // torch.distributed.get_world_size()
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
# ignores_names=['classifier.weight','classifier.bias']
#
# base_params = [p for n, p in model.named_parameters() if not any(nd in n for nd in ignores_names)]
# ignores_params=[p for n, p in model.named_parameters() if any(nd in n for nd in ignores_names)]
#
# optimizer = torch.optim.Adam([{'params': base_params},
# {'params': ignores_params, 'lr': args.learning_rate * 10}],
# lr=args.learning_rate)
global_step = 0
if args.do_train:
train_features = convert_examples_to_features(train_examples,
args.max_seq_length,
tokenizer)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.float)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
# Timer for computing speed
timer_for_speed = Timer()
timer_for_speed.tic()
summary_writer = SummaryWriter(log_dir=args.log_path)
is_early_stop = False
disp_freq = 100
loss_valid_freq = 100
early_stop_patience = 10
bad_count = 0
nb_tr_examples, nb_tr_steps = 0, 0
for eidx in trange(int(args.num_train_epochs), desc="Epoch"):
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
model.train()
try:
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
with torch.enable_grad():
loss = model(input_ids, segment_ids, input_mask,
label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
global_step += 1
# model_collections.add_to_collection("train_losses", loss.item())
# summary_writer.add_scalar("train_losses", loss.item(), global_step=nb_tr_steps)
# display some information
if (nb_tr_steps % disp_freq == 0):
lrate = list(optimizer.get_lr())[0]
result = {'train_loss': loss.item(), "lrate": lrate}
logger.info("***** train results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
except RuntimeError as e:
if 'out of memory' in str(e):
print('| WARNING: ran out of memory, skipping batch')
# optimizer.zero_grad()
if hasattr(torch.cuda, 'empty_cache'):
torch.cuda.empty_cache()
else:
raise e
# calculate dev loss
if (nb_tr_steps % loss_valid_freq == 0):
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = processor.get_dev_examples(
args.data_dir)
eval_features = convert_examples_to_features(
eval_examples, args.max_seq_length, tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor(
[f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor(
[f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor(
[f.segment_ids for f in eval_features],
dtype=torch.long)
all_label_ids = torch.tensor(
[f.label_id for f in eval_features],
dtype=torch.float)
eval_data = TensorDataset(all_input_ids,
all_input_mask,
all_segment_ids,
all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(
eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss = 0
nb_eval_steps, nb_eval_examples = 0, 0
for bacth_eval in eval_dataloader:
bacth_eval = tuple(
t.to(device) for t in bacth_eval)
input_ids, input_mask, segment_ids, label_ids = bacth_eval
with torch.no_grad():
tmp_eval_loss = model(input_ids, segment_ids,
input_mask, label_ids)
eval_loss += tmp_eval_loss.mean().item()
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
model_collections.add_to_collection(
"history_losses", eval_loss)
min_history_loss = np.array(
model_collections.get_collection(
"history_losses")).min()
summary_writer.add_scalar("loss",
eval_loss,
global_step=nb_tr_steps)
summary_writer.add_scalar("best_loss",
min_history_loss,
global_step=nb_tr_steps)
lrate = list(optimizer.get_lr())[0]
summary_writer.add_scalar("lrate",
scalar_value=lrate,
global_step=nb_tr_steps)
best_eval_loss = min_history_loss
# If model get new best valid loss
# save model & early stop
if eval_loss <= best_eval_loss:
bad_count = 0
if is_early_stop is False:
# Save a trained model
# Only save the model it-self
model_to_save = model.module if hasattr(
model, 'module') else model
output_model_file = os.path.join(
args.output_dir, "pytorch_model.bin")
torch.save(model_to_save.state_dict(),
output_model_file)
output_config_file = os.path.join(
args.output_dir, CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(
model_to_save.config.to_json_string())
else:
bad_count += 1
# At least one epoch should be traversed
if bad_count >= early_stop_patience and eidx > 0:
is_early_stop = True
logger.info("Early Stop!")
summary_writer.add_scalar("bad_count", bad_count,
nb_tr_steps)
logger.info("{0} Loss: {1:.4f} patience: {2}".format(
nb_tr_steps, eval_loss, bad_count))
if is_early_stop == True:
break
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--cache_dir",
default="",
type=str,
help=
"Where do you want to store the pre-trained models downloaded from s3")
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument('--server_ip',
type=str,
default='',
help="Can be used for distant debugging.")
parser.add_argument('--server_port',
type=str,
default='',
help="Can be used for distant debugging.")
args = parser.parse_args()
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
processors = {
"justification": JustificationProcessor,
}
output_modes = {
"justification": "classification",
}
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
output_mode = output_modes[task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(
args.local_rank))
model = BertForSequenceClassification.from_pretrained(
args.bert_model, cache_dir=cache_dir, num_labels=num_labels)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
nb_tr_steps = 0
tr_loss = 0
if args.do_train:
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer, output_mode)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
if output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.float)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
# define a new function to compute loss values for both output_modes
logits = model(input_ids, segment_ids, input_mask, labels=None)
if output_mode == "classification":
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
elif output_mode == "regression":
loss_fct = MSELoss()
loss = loss_fct(logits.view(-1), label_ids.view(-1))
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
# Save a trained model and the associated configuration
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
# Load a trained model and config that you have fine-tuned
config = BertConfig(output_config_file)
model = BertForSequenceClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
else:
model = BertForSequenceClassification.from_pretrained(
args.bert_model, num_labels=num_labels)
model.to(device)
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
# dev
eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer, output_mode)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
if output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.float)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss = 0
nb_eval_steps = 0
preds = []
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
logits = model(input_ids, segment_ids, input_mask, labels=None)
# create eval loss and other metric required by the task
if output_mode == "classification":
loss_fct = CrossEntropyLoss()
tmp_eval_loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
elif output_mode == "regression":
loss_fct = MSELoss()
tmp_eval_loss = loss_fct(logits.view(-1), label_ids.view(-1))
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if len(preds) == 0:
preds.append(logits.detach().cpu().numpy())
else:
preds[0] = np.append(preds[0],
logits.detach().cpu().numpy(),
axis=0)
eval_loss = eval_loss / nb_eval_steps
preds = preds[0]
if output_mode == "classification":
preds = np.argmax(preds, axis=1)
elif output_mode == "regression":
preds = np.squeeze(preds)
result = compute_metrics(task_name, preds, all_label_ids.numpy())
loss = tr_loss / nb_tr_steps if args.do_train else None
result['eval_loss'] = eval_loss
result['global_step'] = global_step
result['loss'] = loss
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
# test
eval_examples = processor.get_test_examples(args.data_dir)
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer, output_mode)
logger.info("***** Running test *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
if output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.float)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss = 0
nb_eval_steps = 0
preds = []
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
logits = model(input_ids, segment_ids, input_mask, labels=None)
# create eval loss and other metric required by the task
if output_mode == "classification":
loss_fct = CrossEntropyLoss()
tmp_eval_loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
elif output_mode == "regression":
loss_fct = MSELoss()
tmp_eval_loss = loss_fct(logits.view(-1), label_ids.view(-1))
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if len(preds) == 0:
preds.append(logits.detach().cpu().numpy())
else:
preds[0] = np.append(preds[0],
logits.detach().cpu().numpy(),
axis=0)
eval_loss = eval_loss / nb_eval_steps
preds = preds[0]
if output_mode == "classification":
preds = np.argmax(preds, axis=1)
elif output_mode == "regression":
preds = np.squeeze(preds)
result = compute_metrics(task_name, preds, all_label_ids.numpy())
loss = tr_loss / nb_tr_steps if args.do_train else None
result['eval_loss'] = eval_loss
result['global_step'] = global_step
result['loss'] = loss
output_eval_file = os.path.join(args.output_dir, "test_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Test results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
def train(args, train_dataset, model, tokenizer):
""" Train the model """
if args.local_rank in [-1, 0]:
tb_writer = SummaryWriter()
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = RandomSampler(
train_dataset) if args.local_rank == -1 else DistributedSampler(
train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (
len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ['bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
'weight_decay':
args.weight_decay
}, {
'params': [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0
}]
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, T_max=args.num_train_epochs, eta_min=0)
# scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
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)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1))
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
train_iterator = trange(int(args.num_train_epochs),
desc="Epoch",
disable=args.local_rank not in [-1, 0])
set_seed(
args) # Added here for reproductibility (even between python 2 and 3)
train_step = 0
train_loss = 0
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])
scheduler.step()
for step, batch in enumerate(epoch_iterator):
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {
'input_ids':
batch[0],
'attention_mask':
batch[1],
'token_type_ids':
batch[2] if args.model_type in ['bert', 'xlnet'] else
None, # XLM don't use segment_ids
'labels':
batch[3]
}
ouputs = model(**inputs)
loss = ouputs[
0] # model outputs are always tuple in pytorch-transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
train_step += 1
train_loss += loss.item()
if train_step % 10000 == 0:
print(train_loss * 1. / train_step)
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer),
args.max_grad_norm)
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
# scheduler.step() # Update learning rate schedule
optimizer.step()
model.zero_grad()
global_step += 1
if args.local_rank in [
-1, 0
] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
# Log metrics
if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well
results = evaluate(args, model, tokenizer)
for key, value in results.items():
tb_writer.add_scalar('eval_{}'.format(key), value,
global_step)
tb_writer.add_scalar('lr',
scheduler.get_lr()[0], global_step)
tb_writer.add_scalar('loss', (tr_loss - logging_loss) /
args.logging_steps, global_step)
logging_loss = tr_loss
if args.local_rank in [
-1, 0
] and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(
args.output_dir, 'checkpoint-{}'.format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
torch.save({'state_dict': model.state_dict()},
os.path.join(output_dir, 'training_args.bin'))
# model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training
# model_to_save.save_pretrained(output_dir)
# torch.save(args, os.path.join(output_dir, 'training_args.bin'))
logger.info("Saving model checkpoint to %s", output_dir)
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
if args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
if args.local_rank in [-1, 0]:
tb_writer.close()
return global_step, tr_loss / global_step
def main():
cfg = load_config(CONFIG)
# distributed settings
parser = argparse.ArgumentParser()
parser.add_argument('--local_rank', type=int, default=0)
parser.add_argument('--ngpu', type=int, default=4)
args = parser.parse_args()
if cfg['training']['data_parallel'] == 'distributed':
distributed = True
local_rank = args.local_rank
torch.cuda.set_device(local_rank)
world_size = args.ngpu
torch.distributed.init_process_group(
'nccl',
init_method="tcp://localhost:16534",
world_size=world_size,
rank=local_rank)
else:
distributed = False
local_rank = None
world_size = None
# settings
bs_train, bs_test, n_workers = cfg['training']['batch_size_train'], cfg[
'training']['batch_size_test'], cfg['training']['n_workers']
n_epochs = cfg['training']['n_epochs']
transforms_train, transforms_test = load_transforms(cfg)
# data
ds_train = E32Dataset(cfg,
cfg['paths']['data_train'],
'train',
transforms=transforms_train)
ds_test = E32Dataset(cfg,
cfg['paths']['data_test'],
'test',
transforms=transforms_test)
sampler_train = DistributedSampler(
ds_train, num_replicas=world_size,
rank=local_rank) if distributed else None
sampler_test = DistributedSampler(ds_test,
num_replicas=world_size,
rank=local_rank) if distributed else None
dl_train = DataLoader(ds_train,
bs_train,
shuffle=False if distributed else True,
num_workers=n_workers,
pin_memory=False,
sampler=sampler_train)
dl_test = DataLoader(ds_test,
bs_test,
shuffle=False,
num_workers=n_workers,
pin_memory=False,
sampler=sampler_test)
# model
model, starting_epoch, state = load_model(cfg, local_rank)
optimizer, scheduler = load_optimizer(model,
cfg,
state,
steps_per_epoch=(len(dl_train)))
train_criterion, test_criterion = load_criterion(cfg)
# WandB
if not local_rank:
wandb.init(project="a4c3d",
config=cfg,
notes=cfg.get("description", None))
wandb.save("*.mp4") # Write MP4 files immediately to WandB
wandb.watch(model)
# training
best_loss, best_path, last_save_path = 1e10, None, None
for epoch in range(starting_epoch, n_epochs + 1):
print(f"\nEpoch {epoch} of {n_epochs}")
sampler_train.set_epoch(epoch)
sampler_test.set_epoch(epoch)
# Cycle
train_loss = cycle('train',
model,
dl_train,
epoch,
train_criterion,
optimizer,
cfg,
scheduler,
local_rank=local_rank)
test_loss = cycle('test',
model,
dl_test,
epoch,
test_criterion,
optimizer,
cfg,
scheduler,
local_rank=local_rank)
# Save state if required
if local_rank == 0:
model_weights = model.module.state_dict(
) if cfg['training']['data_parallel'] else model.state_dict()
state = {
'epoch': epoch + 1,
'model': model_weights,
'optimizer': optimizer.state_dict(),
'scheduler': scheduler
}
save_name = f"{epoch}_{test_loss:.05f}.pt"
best_loss, last_save_path = save_state(state,
save_name,
test_loss,
best_loss,
cfg,
last_save_path,
lowest_best=True)
# Vis seg
vis_mse(ds_test, model, epoch, cfg)
if local_rank == 0:
save_name = f"FINAL_{epoch}_{test_loss:.05f}.pt"
save_state(state,
save_name,
test_loss,
best_loss,
cfg,
last_save_path,
force=True)
def LoadDatasets(args, task_cfg, ids, split="trainval"):
tokenizer = BertTokenizer.from_pretrained(
args.bert_model, do_lower_case=args.do_lower_case
)
task_feature_reader1 = {}
task_feature_reader2 = {}
for i, task_id in enumerate(ids):
task = "TASK" + task_id
if task_cfg[task]["features_h5path1"] not in task_feature_reader1:
task_feature_reader1[task_cfg[task]["features_h5path1"]] = None
if task_cfg[task]["features_h5path2"] not in task_feature_reader2:
task_feature_reader2[task_cfg[task]["features_h5path2"]] = None
# initilzie the feature reader
for features_h5path in task_feature_reader1.keys():
if features_h5path != "":
task_feature_reader1[features_h5path] = ImageFeaturesH5Reader(
features_h5path, args.in_memory
)
for features_h5path in task_feature_reader2.keys():
if features_h5path != "":
task_feature_reader2[features_h5path] = ImageFeaturesH5Reader(
features_h5path, args.in_memory
)
task_datasets_train = {}
task_datasets_val = {}
task_dataloader_train = {}
task_dataloader_val = {}
task_ids = []
task_batch_size = {}
task_num_iters = {}
for i, task_id in enumerate(ids):
task = "TASK" + task_id
task_name = task_cfg[task]["name"]
task_ids.append(task)
batch_size = task_cfg[task]["batch_size"] // args.gradient_accumulation_steps
num_workers = args.num_workers
if args.local_rank != -1:
batch_size = int(batch_size / dist.get_world_size())
num_workers = int(num_workers / dist.get_world_size())
# num_workers = int(num_workers / len(ids))
logger.info(
"Loading %s Dataset with batch size %d"
% (task_cfg[task]["name"], batch_size)
)
task_datasets_train[task] = None
if "train" in split:
task_datasets_train[task] = DatasetMapTrain[task_name](
task=task_cfg[task]["name"],
dataroot=task_cfg[task]["dataroot"],
annotations_jsonpath=task_cfg[task]["train_annotations_jsonpath"],
split=task_cfg[task]["train_split"],
image_features_reader=task_feature_reader1[
task_cfg[task]["features_h5path1"]
],
gt_image_features_reader=task_feature_reader2[
task_cfg[task]["features_h5path2"]
],
tokenizer=tokenizer,
bert_model=args.bert_model,
clean_datasets=args.clean_train_sets,
padding_index=0,
max_seq_length=task_cfg[task]["max_seq_length"],
max_region_num=task_cfg[task]["max_region_num"],
)
task_datasets_val[task] = None
if "val" in split:
task_datasets_val[task] = DatasetMapTrain[task_name](
task=task_cfg[task]["name"],
dataroot=task_cfg[task]["dataroot"],
annotations_jsonpath=task_cfg[task]["val_annotations_jsonpath"],
split=task_cfg[task]["val_split"],
image_features_reader=task_feature_reader1[
task_cfg[task]["features_h5path1"]
],
gt_image_features_reader=task_feature_reader2[
task_cfg[task]["features_h5path2"]
],
tokenizer=tokenizer,
bert_model=args.bert_model,
clean_datasets=args.clean_train_sets,
padding_index=0,
max_seq_length=task_cfg[task]["max_seq_length"],
max_region_num=task_cfg[task]["max_region_num"],
)
task_num_iters[task] = 0
task_batch_size[task] = 0
if "train" in split:
if args.local_rank == -1:
train_sampler = RandomSampler(task_datasets_train[task])
else:
# TODO: check if this works with current data generator from disk that relies on next(file)
# (it doesn't return item back by index)
train_sampler = DistributedSampler(task_datasets_train[task])
task_dataloader_train[task] = DataLoader(
task_datasets_train[task],
sampler=train_sampler,
batch_size=batch_size,
num_workers=num_workers,
pin_memory=True,
)
task_num_iters[task] = len(task_dataloader_train[task])
task_batch_size[task] = batch_size
if "val" in split:
task_dataloader_val[task] = DataLoader(
task_datasets_val[task],
shuffle=False,
batch_size=batch_size,
num_workers=2,
pin_memory=True,
)
return (
task_batch_size,
task_num_iters,
task_ids,
task_datasets_train,
task_datasets_val,
task_dataloader_train,
task_dataloader_val,
)
def run(rank, world_size):
os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '12355'
dist.init_process_group('nccl', rank=rank, world_size=world_size)
dataset = PygGraphPropPredDataset('ogbg-molhiv',
'../../data/OGB',
pre_transform=T.ToSparseTensor())
split_idx = dataset.get_idx_split()
evaluator = Evaluator('ogbg-molhiv')
train_dataset = dataset[split_idx['train']]
train_sampler = DistributedSampler(train_dataset,
num_replicas=world_size,
rank=rank)
train_loader = DataLoader(train_dataset,
batch_size=128,
sampler=train_sampler)
torch.manual_seed(12345)
model = GIN(128, dataset.num_tasks, num_layers=3, dropout=0.5).to(rank)
model = DistributedDataParallel(model, device_ids=[rank])
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
criterion = torch.nn.BCEWithLogitsLoss()
if rank == 0:
val_loader = DataLoader(dataset[split_idx['valid']], batch_size=256)
test_loader = DataLoader(dataset[split_idx['test']], batch_size=256)
for epoch in range(1, 51):
model.train()
total_loss = 0
for data in train_loader:
data = data.to(rank)
optimizer.zero_grad()
logits = model(data.x, data.adj_t, data.batch)
loss = criterion(logits, data.y.to(torch.float))
loss.backward()
optimizer.step()
total_loss += float(loss) * logits.size(0)
loss = total_loss / len(train_loader.dataset)
dist.barrier()
if rank == 0: # We evaluate on a single GPU for now.
model.eval()
y_pred, y_true = [], []
for data in val_loader:
data = data.to(rank)
with torch.no_grad():
y_pred.append(model.module(data.x, data.adj_t, data.batch))
y_true.append(data.y)
val_rocauc = evaluator.eval({
'y_pred': torch.cat(y_pred, dim=0),
'y_true': torch.cat(y_true, dim=0),
})['rocauc']
y_pred, y_true = [], []
for data in test_loader:
data = data.to(rank)
with torch.no_grad():
y_pred.append(model.module(data.x, data.adj_t, data.batch))
y_true.append(data.y)
test_rocauc = evaluator.eval({
'y_pred': torch.cat(y_pred, dim=0),
'y_true': torch.cat(y_true, dim=0),
})['rocauc']
print(f'Epoch: {epoch:03d}, Loss: {loss:.4f}, '
f'Val: {val_rocauc:.4f}, Test: {test_rocauc:.4f}')
dist.barrier()
dist.destroy_process_group()
