def train(config):
net = BertForMaskedLM.from_pretrained(config.model)
lossFunc = KLDivLoss(config)
if torch.cuda.is_available():
net = net.cuda()
lossFunc = lossFunc.cuda()
if config.dataParallel:
net = DataParallelModel(net)
lossFunc = DataParallelCriterion(lossFunc)
options = optionsLoader(LOG, config.optionFrames, disp=False)
Tokenizer = BertTokenizer.from_pretrained(config.model)
prepareFunc = prepare_data
trainSet = Dataset('train', config.batch_size,
lambda x: len(x[0]) + len(x[1]), prepareFunc, Tokenizer,
options['dataset'], LOG, 'train')
validSet = Dataset('valid', config.batch_size,
lambda x: len(x[0]) + len(x[1]), prepareFunc, Tokenizer,
options['dataset'], LOG, 'valid')
print(trainSet.__len__())
Q = []
best_vloss = 1e99
counter = 0
lRate = config.lRate
prob_src = config.prob_src
prob_tgt = config.prob_tgt
num_train_optimization_steps = trainSet.__len__(
) * options['training']['stopConditions']['max_epoch']
param_optimizer = list(net.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 = BertAdam(optimizer_grouped_parameters,
lr=lRate,
e=1e-9,
t_total=num_train_optimization_steps,
warmup=0.0)
for epoch_idx in range(options['training']['stopConditions']['max_epoch']):
total_seen = 0
total_similar = 0
total_unseen = 0
total_source = 0
trainSet.setConfig(config, prob_src, prob_tgt)
trainLoader = data.DataLoader(dataset=trainSet,
batch_size=1,
shuffle=True,
num_workers=config.dataLoader_workers,
pin_memory=True)
validSet.setConfig(config, 0.0, prob_tgt)
validLoader = data.DataLoader(dataset=validSet,
batch_size=1,
shuffle=False,
num_workers=config.dataLoader_workers,
pin_memory=True)
for batch_idx, batch_data in enumerate(trainLoader):
if (batch_idx + 1) % 10000 == 0:
gc.collect()
start_time = time.time()
net.train()
inputs, positions, token_types, labels, masks, batch_seen, batch_similar, batch_unseen, batch_source = batch_data
inputs = inputs[0].cuda()
positions = positions[0].cuda()
token_types = token_types[0].cuda()
labels = labels[0].cuda()
masks = masks[0].cuda()
total_seen += batch_seen
total_similar += batch_similar
total_unseen += batch_unseen
total_source += batch_source
n_token = int((labels.data != 0).data.sum())
predicts = net(inputs, positions, token_types, masks)
loss = lossFunc(predicts, labels, n_token).sum()
Q.append(float(loss))
if len(Q) > 200:
Q.pop(0)
loss_avg = sum(Q) / len(Q)
optimizer.zero_grad()
loss.backward()
optimizer.step()
LOG.log(
'Epoch %2d, Batch %6d, Loss %9.6f, Average Loss %9.6f, Time %9.6f'
% (epoch_idx + 1, batch_idx + 1, loss, loss_avg,
time.time() - start_time))
# Checkpoints
idx = epoch_idx * trainSet.__len__() + batch_idx + 1
if (idx >= options['training']['checkingPoints']['checkMin']) and (
idx % options['training']['checkingPoints']['checkFreq']
== 0):
if config.do_eval:
vloss = 0
total_tokens = 0
for bid, batch_data in enumerate(validLoader):
inputs, positions, token_types, labels, masks, batch_seen, batch_similar, batch_unseen, batch_source = batch_data
inputs = inputs[0].cuda()
positions = positions[0].cuda()
token_types = token_types[0].cuda()
labels = labels[0].cuda()
masks = masks[0].cuda()
n_token = int((labels.data != config.PAD).data.sum())
with torch.no_grad():
net.eval()
predicts = net(inputs, positions, token_types,
masks)
vloss += float(lossFunc(predicts, labels).sum())
total_tokens += n_token
vloss /= total_tokens
is_best = vloss < best_vloss
best_vloss = min(vloss, best_vloss)
LOG.log(
'CheckPoint: Validation Loss %11.8f, Best Loss %11.8f'
% (vloss, best_vloss))
if is_best:
LOG.log('Best Model Updated')
save_check_point(
{
'epoch': epoch_idx + 1,
'batch': batch_idx + 1,
'options': options,
'config': config,
'state_dict': net.state_dict(),
'best_vloss': best_vloss
},
is_best,
path=config.save_path,
fileName='latest.pth.tar')
counter = 0
else:
counter += options['training']['checkingPoints'][
'checkFreq']
if counter >= options['training']['stopConditions'][
'rateReduce_bound']:
counter = 0
for param_group in optimizer.param_groups:
lr_ = param_group['lr']
param_group['lr'] *= 0.55
_lr = param_group['lr']
LOG.log(
'Reduce Learning Rate from %11.8f to %11.8f' %
(lr_, _lr))
LOG.log('Current Counter = %d' % (counter))
else:
save_check_point(
{
'epoch': epoch_idx + 1,
'batch': batch_idx + 1,
'options': options,
'config': config,
'state_dict': net.state_dict(),
'best_vloss': 1e99
},
False,
path=config.save_path,
fileName='checkpoint_Epoch' + str(epoch_idx + 1) +
'_Batch' + str(batch_idx + 1) + '.pth.tar')
LOG.log('CheckPoint Saved!')
if options['training']['checkingPoints']['everyEpoch']:
save_check_point(
{
'epoch': epoch_idx + 1,
'batch': batch_idx + 1,
'options': options,
'config': config,
'state_dict': net.state_dict(),
'best_vloss': 1e99
},
False,
path=config.save_path,
fileName='checkpoint_Epoch' + str(epoch_idx + 1) + '.pth.tar')
LOG.log('Epoch Finished.')
LOG.log(
'Total Seen: %d, Total Unseen: %d, Total Similar: %d, Total Source: %d.'
% (total_seen, total_unseen, total_similar, total_source))
gc.collect()
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)
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
args.warmup_steps = t_total // 100
# Prepare optimizer and schedule (linear warmup and decay)
optimizer_grouped_parameters = get_param_groups(args, model)
optimizer = RAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
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)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
# model = torch.nn.DataParallel(model)
model = DataParallelModel(model)
# 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)
args.logging_steps = len(train_dataloader) // 1
args.save_steps = args.logging_steps
global_step = 0
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
train_iterator = trange(int(args.num_train_epochs), desc="Epoch")
set_seed(args)
for _ in train_iterator:
args.current_epoch = _
epoch_iterator = tqdm(train_dataloader, desc="Iteration")
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 and RoBERTa don't use segment_ids
# 'labels': batch[3]}
outputs = model(**inputs)
outputs = [outputs[i][0] for i in range(len(outputs))]
loss_fct = CrossEntropyLoss()
loss_fct = DataParallelCriterion(loss_fct)
loss = loss_fct(outputs, batch[3])
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()
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:
optimizer.step()
scheduler.step()
model.zero_grad()
global_step += 1
if 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.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)
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
class Trainer:
"""
trainer class
"""
def __init__(self, cfg: Namespace, data: Dataset):
"""
Args:
cfg: configuration
data: train dataset
"""
self.cfg = cfg
self.train, self.valid = data.split(0.8)
RATING_FIELD.build_vocab(self.train)
self.device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu') # pylint: disable=no-member
self.batch_size = cfg.batch_size
if torch.cuda.is_available():
self.batch_size *= torch.cuda.device_count()
self.trn_itr = BucketIterator(
self.train,
device=self.device,
batch_size=self.batch_size,
shuffle=True,
train=True,
sort_within_batch=True,
sort_key=lambda exam: -len(exam.comment_text))
self.vld_itr = BucketIterator(
self.valid,
device=self.device,
batch_size=self.batch_size,
shuffle=False,
train=False,
sort_within_batch=True,
sort_key=lambda exam: -len(exam.comment_text))
self.log_step = 1000
if len(self.vld_itr) < 100:
self.log_step = 10
elif len(self.vld_itr) < 1000:
self.log_step = 100
bert_path = cfg.bert_path if cfg.bert_path else 'bert-base-cased'
self.model = BertForSequenceClassification.from_pretrained(
bert_path, num_labels=2)
pos_weight = (
len([exam for exam in self.train.examples if exam.target < 0.5]) /
len([exam for exam in self.train.examples if exam.target >= 0.5]))
pos_wgt_tensor = torch.tensor([1.0, pos_weight], device=self.device) # pylint: disable=not-callable
self.criterion = nn.CrossEntropyLoss(weight=pos_wgt_tensor)
if torch.cuda.is_available():
self.model = DataParallelModel(self.model.cuda())
self.criterion = DataParallelCriterion(self.criterion)
self.optimizer = optim.Adam(self.model.parameters(), cfg.learning_rate)
def run(self):
"""
do train
"""
max_f_score = -9e10
max_epoch = -1
for epoch in range(self.cfg.epoch):
train_loss = self._train_epoch(epoch)
metrics = self._evaluate(epoch)
max_f_score_str = f' < {max_f_score:.2f}'
if metrics['f_score'] > max_f_score:
max_f_score_str = ' is max'
max_f_score = metrics['f_score']
max_epoch = epoch
torch.save(self.model.state_dict(), self.cfg.model_path)
logging.info('EPOCH[%d]: train loss: %.6f, valid loss: %.6f, acc: %.2f,' \
' F: %.2f%s', epoch, train_loss, metrics['loss'],
metrics['accuracy'], metrics['f_score'], max_f_score_str)
if (epoch - max_epoch) >= self.cfg.patience:
logging.info('early stopping...')
break
logging.info('epoch: %d, f-score: %.2f', max_epoch, max_f_score)
def _train_epoch(self, epoch: int) -> float:
"""
train single epoch
Args:
epoch: epoch number
Returns:
average loss
"""
self.model.train()
progress = tqdm(self.trn_itr,
f'EPOCH[{epoch}]',
mininterval=1,
ncols=100)
losses = []
for step, batch in enumerate(progress, start=1):
outputs = self.model(batch.comment_text)
# output of model wrapped with DataParallelModel is a list of outputs from each GPU
# make input of DataParallelCriterion as a list of tuples
if isinstance(self.model, DataParallelModel):
loss = self.criterion([(output, ) for output in outputs],
batch.target)
else:
loss = self.criterion(outputs, batch.target)
losses.append(loss.item())
if step % self.log_step == 0:
avg_loss = sum(losses) / len(losses)
progress.set_description(f'EPOCH[{epoch}] ({avg_loss:.6f})')
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()
return sum(losses) / len(losses)
def _evaluate(self, epoch: int) -> Dict[str, float]:
"""
evaluate on validation data
Args:
epoch: epoch number
Returns:
metrics
"""
self.model.eval()
progress = tqdm(self.vld_itr,
f' EVAL[{epoch}]',
mininterval=1,
ncols=100)
losses = []
preds = []
golds = []
for step, batch in enumerate(progress, start=1):
with torch.no_grad():
outputs = self.model(batch.comment_text)
if isinstance(self.model, DataParallelModel):
loss = self.criterion([(output, ) for output in outputs],
batch.target)
for output in outputs:
preds.extend([(0 if o[0] < o[1] else 1)
for o in output])
else:
loss = self.criterion(outputs, batch.target)
preds.extend([(0 if output[0] < output[1] else 1)
for output in outputs])
losses.append(loss.item())
golds.extend([gold.item() for gold in batch.target])
if step % self.log_step == 0:
avg_loss = sum(losses) / len(losses)
progress.set_description(
f' EVAL[{epoch}] ({avg_loss:.6f})')
metrics = self._get_metrics(preds, golds)
metrics['loss'] = sum(losses) / len(losses)
return metrics
@classmethod
def _get_metrics(cls, preds: List[float],
golds: List[float]) -> Dict[str, float]:
"""
get metric values
Args:
preds: predictions
golds: gold standards
Returns:
metric
"""
assert len(preds) == len(golds)
true_pos = 0
false_pos = 0
false_neg = 0
true_neg = 0
for pred, gold in zip(preds, golds):
if pred >= 0.5:
if gold >= 0.5:
true_pos += 1
else:
false_pos += 1
else:
if gold >= 0.5:
false_neg += 1
else:
true_neg += 1
accuracy = (true_pos + true_neg) / (true_pos + false_pos + false_neg +
true_neg)
precision = 0.0
if (true_pos + false_pos) > 0:
precision = true_pos / (true_pos + false_pos)
recall = 0.0
if (true_pos + false_neg) > 0:
recall = true_pos / (true_pos + false_neg)
f_score = 0.0
if (precision + recall) > 0.0:
f_score = 2.0 * precision * recall / (precision + recall)
return {
'accuracy': 100.0 * accuracy,
'precision': 100.0 * precision,
'recall': 100.0 * recall,
'f_score': 100.0 * f_score,
}