class OneCycleOptimizer(BertAdam):
def __init__(self,
params,
start,
stop,
warmup=-1,
t_total=-1,
schedule='warmup_linear',
b1=0.9,
b2=0.999,
e=1e-6,
weight_decay_rate=0.01,
steps=200,
max_grad_norm=1.0):
defaults = dict(lr=start,
schedule=schedule,
warmup=warmup,
t_total=t_total,
b1=b1,
b2=b2,
e=e,
weight_decay_rate=weight_decay_rate,
max_grad_norm=max_grad_norm)
self.bertadam = BertAdam(params, **defaults)
self.dir = 1
self.start = start
self.stop = stop
self.curr = start
self.incr = 2 * (stop - start) / steps
super(OneCycleOptimizer, self).__init__(params, **defaults)
def get_lr(self):
lr_list = self.bertadam.get_lr()
if self.dir == 1:
self.curr += self.incr
if self.curr > self.stop:
self.dir = -1
elif self.dir == -1:
self.curr -= self.incr
if self.curr < self.start:
self.dir = 1
return lr_list / self.curr
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--examples_n_features_dir",
default='data/examples_n_features/',
type=str,
help="Dir containing drop examples and features.")
parser.add_argument("--mlm_dir",
default='../data/MLM_train/',
type=Path,
help="The data dir with MLM taks. Should contain the .jsonl files (or other data files) for the task.")
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-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--output_dir",
default='./out_drop_finetune',
type=str,
help="The output directory where the model checkpoints will be written.")
parser.add_argument("--model", default="bert-base-uncased", type=str)
parser.add_argument("--init_weights_dir",
default='',
type=str,
help="The directory where init model wegihts an config are stored.")
parser.add_argument("--max_seq_length",
default=-1,
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_inference",
action='store_true',
help="Whether to run inference 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("--mlm_batch_size",
default=-1,
type=int,
help="Total batch size for mlm train data.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_samples",
default=-1,
type=int,
help="Total number of training samples used.")
parser.add_argument("--num_train_epochs",
default=6.0,
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('--freeze_encoder',
action='store_true',
help="Whether to freeze the bert encoder, embeddings.")
# parser.add_argument('--indiv_digits',
# action='store_true',
# help="Whether to tokenize numbers as digits.")
parser.add_argument('--rand_init',
action='store_true',
help="Whether to use random init instead of BERT.")
parser.add_argument('--random_shift',
action='store_true',
help="Whether to randomly shift position ids of encoder input.")
parser.add_argument('--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--mlm_scale',
type=float, default=1.0,
help="mlm loss scaling factor.")
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()
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))
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 args.do_train:
make_output_dir(args, scripts_to_save=[sys.argv[0], 'modeling.py', 'create_examples_n_features.py'])
#tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True)
if args.init_weights_dir:
model = BertTransformer.from_pretrained(args.init_weights_dir)
else:
# prepare model
model = BertTransformer.from_pretrained(args.model,
cache_dir=os.path.join(str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(args.local_rank)))
model.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
eval_data = DropDataset(args, 'eval')
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_data))
logger.info(" Batch size = %d", args.eval_batch_size)
if args.do_eval and args.do_inference:
inference(args, model, eval_dataloader, device, tokenizer)
exit()
if args.do_train:
# Prepare data loader
train_data = DropDataset(args, 'train')
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
num_train_optimization_steps = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
# 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]
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=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_data))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
'''
------------------------------------------------------------------------------
TODO: check training resume, fp16, use --random_shift for short inputs
------------------------------------------------------------------------------
'''
# using fp16
fp16 = False
# try:
# from apex import amp
# model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
# fp16 = True
# except ImportError:
# logger.info("Not using 16-bit training due to apex import error.")
# if n_gpu > 1:
# model = torch.nn.DataParallel(model)
tb_writer = SummaryWriter(os.path.join(args.output_dir, 'log')) # tensorboard
# masked LM data
do_mlm_task = False
if args.mlm_batch_size > 0:
mlm_dataset = MLMDataset(training_path=args.mlm_dir, tokenizer=tokenizer)
mlm_dataloader = DataLoader(mlm_dataset, sampler=RandomSampler(mlm_dataset), batch_size=args.mlm_batch_size)
mlm_iter = iter(mlm_dataloader)
do_mlm_task = True
model.train()
(global_step, all_losses, all_errors, all_dec_losses, all_dec_errors, eval_errors,
best, best_mlm, t_prev, do_eval) = 0, [], [], [], [], [], 1000, 1000, time.time(), False
mlm_losses, mlm_errors, all_span_losses, all_span_errors = [], [], [], []
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
# grads wrt to train data
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids, head_type, q_spans, p_spans = batch
losses = model(input_ids, segment_ids, input_mask, random_shift=args.random_shift, target_ids=label_ids,
target_mask=None, answer_as_question_spans=q_spans, answer_as_passage_spans=p_spans,
head_type=head_type)
loss, errs, dec_loss, dec_errors, span_loss, span_errors, type_loss, type_errors, type_preds = losses
# aggregate on multi-gpu
take_mean = lambda x: x.mean() if x is not None and sum(x.size()) > 1 else x
take_sum = lambda x: x.sum() if x is not None and sum(x.size()) > 1 else x
[loss, dec_loss, span_loss, type_loss] = list(map(take_mean, [loss, dec_loss, span_loss, type_loss]))
[errs, dec_errors, span_errors, type_errors] = list(map(take_sum, [errs, dec_errors, span_errors, type_errors]))
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
all_losses.append(loss.item()); all_dec_losses.append(dec_loss.item());
all_errors.append(errs.item() / input_ids.size(0))
all_dec_errors.append(dec_errors.item() / input_ids.size(0))
all_span_losses.append(span_loss.item()); all_span_errors.append(span_errors.item() / input_ids.size(0))
if fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
if do_mlm_task:
# grads wrt to mlm data
while True:
try:
batch = next(mlm_iter) # sample next mlm batch
break
except StopIteration: # end of epoch: reset and shuffle
mlm_iter = iter(mlm_dataloader)
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, label_ids = batch
loss, errs = model(input_ids, None, input_mask, target_ids=label_ids,
ignore_idx=IGNORE_IDX, task='mlm')
loss, err_sum = take_mean(loss), take_sum(errs) # for multi-gpu
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss = args.mlm_scale * loss
mlm_losses.append(loss.item()); mlm_errors.append(err_sum.item() / input_ids.size(0))
if fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
else:
mlm_losses.append(-1); mlm_errors.append(-1)
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step() # update step
optimizer.zero_grad()
global_step += 1
train_result = {'trn_loss': all_losses[-1], 'trn_dec_loss': all_dec_losses[-1],
'trn_err': all_errors[-1], 'trn_dec_err': all_dec_errors[-1],
'lr': optimizer.get_lr()[0], 'trn_span_loss': all_span_losses[-1],
'trn_span_err': all_span_errors[-1], 'epoch': epoch}
mlm_result = {'trn_mlm_loss': mlm_losses[-1], 'trn_mlm_err': mlm_errors[-1]}
tb_writer.add_scalars('train', train_result, len(all_losses))
tb_writer.add_scalars('mlm', mlm_result, len(all_losses)) if do_mlm_task else None
if time.time() - t_prev > 60*60: # evaluate every hr
do_eval = True
t_prev = time.time()
if do_eval:
do_eval = False
eval_result = evaluate(args, model, eval_dataloader, device, len(train_data))
eval_err = eval_result['eval_err']
if eval_err < best or (eval_err < best + 0.005 and np.mean(mlm_errors[-1000:]) < best_mlm):
# if eval err is in range of best, look at MLM err
train_state = {'global_step': global_step, 'optimizer_state_dict': optimizer.state_dict()}
train_state.update(train_result)
save(args, model, tokenizer, train_state)
best_mlm = min(best_mlm, np.mean(mlm_errors[-1000:]))
best = min(best, eval_err)
eval_errors.append((len(all_losses), eval_err))
model.train()
tb_writer.add_scalars('eval', eval_result, len(all_losses))
# for name, param in model.named_parameters():
# tb_writer.add_histogram(name, param.clone().cpu().data.numpy(), len(all_losses))
# end of epoch
do_eval = True
# training complete
tb_writer.export_scalars_to_json(os.path.join(args.output_dir, 'training_scalars.json'))
tb_writer.close()
def do_training(train_fs, train_exs):
"""Runs BERT fine-tuning."""
# Allows to write to enclosed variables global_step
nonlocal global_step
# Create the batched training data out of the features.
train_data = create_tensor_dataset(train_fs)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
# Calculate the number of optimization steps.
num_train_optimization_steps = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
# 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 = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
# Log some information about the training.
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_exs))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
# Set the model to training mode and train for X epochs.
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
# Iterate over all batches.
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
# Get the Logits and calculate the loss.
logits = model(input_ids, token_type_ids=segment_ids, attention_mask=input_mask)
loss = CrossEntropyLoss()(logits.view(-1, num_labels), label_ids.view(-1))
# Scale the loss in gradient accumulation mode.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
# Calculate the gradients.
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
# Update the weights every gradient_accumulation_steps steps.
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
global_step += 1
tb_writer.add_scalar('lr', optimizer.get_lr()[0], global_step)
tb_writer.add_scalar('loss', loss.item(), global_step)
class BertTrainer:
def __init__(self, hypers: Hypers, model_name, checkpoint,
**extra_model_args):
"""
initialize the BertOptimizer, with common logic for setting weight_decay_rate, doing gradient accumulation and
tracking loss
:param hypers: the core hyperparameters for the bert model
:param model_name: the fully qualified name of the bert model we will train
like pytorch_pretrained_bert.modeling.BertForQuestionAnswering
:param checkpoint: if resuming training,
this is the checkpoint that contains the optimizer state as checkpoint['optimizer']
"""
self.init_time = time.time()
self.model = self.get_model(hypers, model_name, checkpoint,
**extra_model_args)
self.step = 0
self.hypers = hypers
self.train_stats = TrainStats(hypers)
self.model.train()
logger.info('configured model for training')
# show parameter names
# logger.info(str([n for (n, p) in self.model.named_parameters()]))
# Prepare optimizer
if hasattr(hypers, 'exclude_pooler') and hypers.exclude_pooler:
# module.bert.pooler.dense.weight, module.bert.pooler.dense.bias
# see https://github.com/NVIDIA/apex/issues/131
self.param_optimizer = [
(n, p) for (n, p) in self.model.named_parameters()
if '.pooler.' not in n
]
else:
self.param_optimizer = list(self.model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in self.param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay_rate':
0.01
}, {
'params': [
p for n, p in self.param_optimizer
if any(nd in n for nd in no_decay)
],
'weight_decay_rate':
0.0
}]
self.t_total = hypers.num_train_steps
self.global_step = hypers.global_step
if hypers.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=hypers.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if hypers.loss_scale == 0:
self.optimizer = FP16_Optimizer(
optimizer,
dynamic_loss_scale=True,
verbose=(hypers.global_rank == 0))
else:
self.optimizer = FP16_Optimizer(
optimizer,
static_loss_scale=hypers.loss_scale,
verbose=(hypers.global_rank == 0))
else:
self.optimizer = BertAdam(optimizer_grouped_parameters,
lr=hypers.learning_rate,
warmup=hypers.warmup_proportion,
t_total=self.t_total)
logger.info('created optimizer')
if checkpoint and type(
checkpoint) is dict and 'optimizer' in checkpoint:
self.optimizer.load_state_dict(checkpoint['optimizer'])
if hypers.fp16:
pass
else:
# if we load this state, we need to set the t_total to what we passed, not what was saved
self.optimizer.set_t_total(self.t_total)
# show state of optimizer
lrs = self.optimizer.get_lr()
logger.info('Min and max learn rate: %s',
str([min(lrs), max(lrs)]))
logger.info('Min and max step in state: %s',
str(self.optimizer.get_steps()))
instances_per_step = hypers.train_batch_size * hypers.gradient_accumulation_steps * hypers.world_size
if 'seen_instances' in checkpoint:
self.global_step = int(checkpoint['seen_instances'] /
instances_per_step)
self.train_stats.previous_instances = checkpoint[
'seen_instances']
logger.info('got global step from checkpoint = %i',
self.global_step)
logger.info('Loaded optimizer state:')
logger.info(repr(self.optimizer))
def reset(self):
"""
reset any gradient accumulation
:return:
"""
self.model.zero_grad()
self.step = 0
def should_continue(self):
"""
:return: True if training should continue
"""
if self.global_step >= self.t_total:
logger.info(
'stopping due to train step %i >= target train steps %i',
self.global_step, self.t_total)
return False
if 0 < self.hypers.time_limit <= (time.time() - self.init_time):
logger.info('stopping due to time out %i seconds',
self.hypers.time_limit)
return False
return True
def save_simple(self, filename):
if self.hypers.global_rank != 0:
logger.info('skipping save in %i', torch.distributed.get_rank())
return
model_to_save = self.model.module if hasattr(
self.model, 'module') else self.model # Only save the model itself
torch.save(model_to_save.state_dict(), filename)
logger.info(f'saved model only to {filename}')
def save(self, filename, **extra_checkpoint_info):
"""
save a checkpoint with the model parameters, the optimizer state and any additional checkpoint info
:param filename:
:param extra_checkpoint_info:
:return:
"""
# only local_rank 0, in fact only global rank 0
if self.hypers.global_rank != 0:
logger.info('skipping save in %i', torch.distributed.get_rank())
return
start_time = time.time()
checkpoint = extra_checkpoint_info
model_to_save = self.model.module if hasattr(
self.model, 'module') else self.model # Only save the model itself
os.makedirs(os.path.dirname(filename), exist_ok=True)
# also save the optimizer state, since we will likely resume from partial pre-training
checkpoint['state_dict'] = model_to_save.state_dict()
checkpoint['optimizer'] = self.optimizer.state_dict()
# include world size in instances_per_step calculation
instances_per_step = self.hypers.train_batch_size * \
self.hypers.gradient_accumulation_steps * \
self.hypers.world_size
checkpoint['seen_instances'] = self.global_step * instances_per_step
checkpoint['num_instances'] = self.t_total * instances_per_step
# CONSIDER: also save hypers?
torch.save(checkpoint, filename)
logger.info(
f'saved model to {filename} in {time.time()-start_time} seconds')
def get_instance_count(self):
instances_per_step = self.hypers.train_batch_size * \
self.hypers.gradient_accumulation_steps * \
self.hypers.world_size
return self.global_step * instances_per_step
def step_loss(self, loss):
"""
accumulates the gradient, tracks the loss and applies the gradient to the model
:param loss: the loss from evaluating the model
"""
if self.global_step == 0:
logger.info('first step_loss')
if self.hypers.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
self.train_stats.note_loss(loss.item())
if self.hypers.gradient_accumulation_steps > 1:
loss = loss / self.hypers.gradient_accumulation_steps
if self.hypers.fp16:
self.optimizer.backward(loss)
else:
loss.backward()
if (self.step + 1) % self.hypers.gradient_accumulation_steps == 0:
lr_this_step = self.hypers.learning_rate * warmup_linear(
self.global_step / self.t_total, self.hypers.warmup_proportion)
for param_group in self.optimizer.param_groups:
param_group['lr'] = lr_this_step
self.optimizer.step()
self.model.zero_grad()
self.global_step += 1
self.step += 1
@classmethod
def get_files(cls, train_file, completed_files):
logger.info('completed files = %s, count = %i',
str(completed_files[:min(5, len(completed_files))]),
len(completed_files))
# multiple train files
if not os.path.isdir(train_file):
train_files = [train_file]
else:
if not train_file.endswith('/'):
train_file = train_file + '/'
train_files = glob.glob(train_file + '**', recursive=True)
train_files = [f for f in train_files if not os.path.isdir(f)]
# exclude completed files
if not set(train_files) == set(completed_files):
train_files = [f for f in train_files if f not in completed_files]
else:
completed_files = [] # new epoch
logger.info('train files = %s, count = %i',
str(train_files[:min(5, len(train_files))]),
len(train_files))
return train_files, completed_files
@classmethod
def get_model(cls, hypers, model_name, checkpoint, **extra_model_args):
override_state_dict = None
if checkpoint:
if type(checkpoint) is dict and 'state_dict' in checkpoint:
logger.info('loading from multi-part checkpoint')
override_state_dict = checkpoint['state_dict']
else:
logger.info('loading from saved model parameters')
override_state_dict = checkpoint
# create the model object by name
# https://stackoverflow.com/questions/4821104/python-dynamic-instantiation-from-string-name-of-a-class-in-dynamically-imported
import importlib
clsdot = model_name.rfind('.')
class_ = getattr(importlib.import_module(model_name[0:clsdot]),
model_name[clsdot + 1:])
model_args = {
'state_dict': override_state_dict,
'cache_dir': PYTORCH_PRETRAINED_BERT_CACHE
}
model_args.update(extra_model_args)
# logger.info(pprint.pformat(extra_model_args, indent=4))
model = class_.from_pretrained(hypers.bert_model, **model_args)
logger.info('built model')
# configure model for fp16, multi-gpu and/or distributed training
if hypers.fp16:
model.half()
logger.info('model halved')
logger.info('sending model to %s', str(hypers.device))
model.to(hypers.device)
logger.info('sent model to %s', str(hypers.device))
if hypers.local_rank != -1:
if not hypers.no_apex:
try:
from apex.parallel import DistributedDataParallel as DDP
model = DDP(model)
except ImportError:
raise ImportError("Please install apex")
else:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[hypers.local_rank],
output_device=hypers.local_rank)
logger.info('using DistributedDataParallel for world size %i',
hypers.world_size)
elif hypers.n_gpu > 1:
model = torch.nn.DataParallel(model)
return model
@classmethod
def get_base_parser(cls):
parser = argparse.ArgumentParser()
# Required parameters
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."
)
# Other parameters
parser.add_argument(
"--num_instances",
default=-1,
type=int,
help="Total number of training instances to train over.")
parser.add_argument(
"--seen_instances",
default=-1,
type=int,
help=
"When resuming training, the number of instances we have already trained over."
)
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
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("--no_apex",
default=False,
action='store_true',
help="Whether not to use apex when available")
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(
'--optimize_on_cpu',
default=False,
action='store_true',
help=
"Whether to perform optimization and keep the optimizer averages on CPU"
)
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, positive power of 2 values can improve fp16 convergence. '
'Leave at zero to use dynamic loss scaling')
return parser
def main():
parser = argparse.ArgumentParser()
# Required parameters
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("--vocab_file",
default='bert-base-uncased-vocab.txt',
type=str,
required=True)
parser.add_argument("--model_file",
default='bert-base-uncased.tar.gz',
type=str,
required=True)
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model checkpoints and predictions will be written."
)
parser.add_argument(
"--predict_dir",
default=None,
type=str,
required=True,
help="The output directory where the predictions will be written.")
# Other parameters
parser.add_argument("--train_file",
default=None,
type=str,
help="SQuAD json for training. E.g., train-v1.1.json")
parser.add_argument(
"--predict_file",
default=None,
type=str,
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json"
)
parser.add_argument("--test_file", default=None, type=str)
parser.add_argument(
"--max_seq_length",
default=384,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. Sequences "
"longer than this will be truncated, and sequences shorter than this will be padded."
)
parser.add_argument(
"--doc_stride",
default=128,
type=int,
help=
"When splitting up a long document into chunks, how much stride to take between chunks."
)
parser.add_argument(
"--max_query_length",
default=64,
type=int,
help=
"The maximum number of tokens for the question. Questions longer than this will "
"be truncated to this length.")
parser.add_argument("--do_train",
default=False,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_predict",
default=False,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--predict_batch_size",
default=8,
type=int,
help="Total batch size for predictions.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=2.0,
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(
"--n_best_size",
default=20,
type=int,
help=
"The total number of n-best predictions to generate in the nbest_predictions.json "
"output file.")
parser.add_argument(
"--max_answer_length",
default=30,
type=int,
help=
"The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.")
parser.add_argument(
"--verbose_logging",
default=False,
action='store_true',
help=
"If true, all of the warnings related to data processing will be printed. "
"A number of warnings are expected for a normal SQuAD evaluation.")
parser.add_argument("--no_cuda",
default=False,
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument('--view_id',
type=int,
default=1,
help="view id of multi-view co-training(two-view)")
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(
"--do_lower_case",
default=True,
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(
'--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('--save_all', default=False, action='store_true')
# Base setting
parser.add_argument('--pretrain', type=str, default=None)
parser.add_argument('--max_ctx', type=int, default=2)
parser.add_argument('--task_name', type=str, default='race')
parser.add_argument('--bert_name', type=str, default='pool-race')
parser.add_argument('--reader_name', type=str, default='race')
parser.add_argument('--per_eval_step', type=int, default=10000000)
# model parameters
parser.add_argument('--evidence_lambda', type=float, default=0.8)
# Parameters for running labeling model
parser.add_argument('--do_label', default=False, action='store_true')
parser.add_argument('--sentence_id_file', nargs='*')
parser.add_argument('--weight_threshold', type=float, default=0.0)
parser.add_argument('--only_correct', default=False, action='store_true')
parser.add_argument('--label_threshold', type=float, default=0.0)
parser.add_argument('--multi_evidence', default=False, action='store_true')
parser.add_argument('--metric', default='accuracy', type=str)
parser.add_argument('--num_evidence', default=1, type=int)
parser.add_argument('--power_length', default=1., type=float)
parser.add_argument('--num_choices', default=4, type=int)
parser.add_argument('--split_type', default=0, type=int)
parser.add_argument('--use_gumbel', default=False, action='store_true')
parser.add_argument('--sample_steps', type=int, default=10)
parser.add_argument('--reward_func', type=int, default=0)
parser.add_argument('--freeze_bert', default=False, action='store_true')
args = parser.parse_args()
logger = setting_logger(args.output_dir)
logger.info('================== Program start. ========================')
logger.info(
f'================== Running with seed {args.seed} =========================='
)
model_params = prepare_model_params(args)
read_params = prepare_read_params(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_predict and not args.do_label:
raise ValueError(
"At least one of `do_train` or `do_predict` or `do_label` 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 args.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.")
os.makedirs(args.output_dir, exist_ok=True)
if args.do_predict or args.do_label:
os.makedirs(args.predict_dir, exist_ok=True)
tokenizer = BertTokenizer.from_pretrained(args.vocab_file)
data_reader = initialize_reader(args.reader_name)
num_train_steps = None
if args.do_train or args.do_label:
train_examples = data_reader.read(input_file=args.train_file,
**read_params)
cached_train_features_file = args.train_file + '_{0}_{1}_{2}_{3}_{4}_{5}'.format(
args.bert_model, str(args.max_seq_length), str(args.doc_stride),
str(args.max_query_length), str(args.max_ctx), str(args.task_name))
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
except FileNotFoundError:
train_features = data_reader.convert_examples_to_features(
examples=train_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length)
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)
num_train_steps = int(
len(train_features) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
if args.pretrain is not None:
logger.info('Load pretrained model from {}'.format(args.pretrain))
model_state_dict = torch.load(args.pretrain, map_location='cuda:0')
model = initialize_model(args.bert_name,
args.model_file,
state_dict=model_state_dict,
**model_params)
else:
model = initialize_model(args.bert_name, args.model_file,
**model_params)
# 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())
# Remove frozen parameters
param_optimizer = [n for n in param_optimizer if n[1].requires_grad]
# 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 num_train_steps is not None else -1
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)
# warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion, t_total=t_total)
# logger.info(f"warm up linear: warmup = {warmup_linear.warmup}, t_total = {warmup_linear.t_total}.")
# else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
if args.fp16:
from apex import amp
model, optimizer = amp.initialize(model, optimizer, opt_level='O2')
# Prepare data
eval_examples = data_reader.read(input_file=args.predict_file,
**read_params)
eval_features = data_reader.convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length)
eval_tensors = data_reader.data_to_tensors(eval_features)
eval_data = TensorDataset(*eval_tensors)
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.predict_batch_size)
if args.do_train:
if args.do_label:
logger.info('Training in State Wise.')
sentence_label_file = args.sentence_id_file
if sentence_label_file is not None:
for file in sentence_label_file:
train_features = data_reader.generate_features_sentence_ids(
train_features, file)
else:
logger.info('No sentence id supervision is found.')
else:
logger.info('Training in traditional way.')
logger.info("***** Running training *****")
logger.info(" Num orig examples = %d", len(train_examples))
logger.info(" Num split examples = %d", len(train_features))
logger.info(" Num train total optimization steps = %d", t_total)
logger.info(" Batch size = %d", args.predict_batch_size)
train_loss = AverageMeter()
best_acc = 0.0
best_loss = 1000000
summary_writer = SummaryWriter(log_dir=args.output_dir)
global_step = 0
eval_loss = AverageMeter()
eval_accuracy = CategoricalAccuracy()
eval_epoch = 0
train_tensors = data_reader.data_to_tensors(train_features)
train_data = TensorDataset(*train_tensors)
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)
for epoch in range(int(args.num_train_epochs)):
logger.info(f'Running at Epoch {epoch}')
# Train
for step, batch in enumerate(
tqdm(train_dataloader,
desc="Iteration",
dynamic_ncols=True)):
model.train()
if n_gpu == 1:
batch = batch_to_device(
batch, device) # multi-gpu does scattering it-self
inputs = data_reader.generate_inputs(
batch, train_features, model_state=ModelState.Train)
model_output = model(**inputs)
loss = model_output['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)
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used and handles this automatically
# if args.fp16:
# lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step)
# for param_group in optimizer.param_groups:
# param_group['lr'] = lr_this_step
# summary_writer.add_scalar('lr', lr_this_step, global_step)
# else:
summary_writer.add_scalar('lr',
optimizer.get_lr()[0],
global_step)
optimizer.step()
optimizer.zero_grad()
global_step += 1
train_loss.update(loss.item(), 1)
summary_writer.add_scalar('train_loss', train_loss.avg,
global_step)
# logger.info(f'Train loss: {train_loss.avg}')
if (step + 1) % args.per_eval_step == 0 or step == len(
train_dataloader) - 1:
# Evaluation
model.eval()
logger.info("Start evaluating")
for _, eval_batch in enumerate(
tqdm(eval_dataloader,
desc="Evaluating",
dynamic_ncols=True)):
if n_gpu == 1:
eval_batch = batch_to_device(
eval_batch,
device) # multi-gpu does scattering it-self
inputs = data_reader.generate_inputs(
eval_batch,
eval_features,
model_state=ModelState.Evaluate)
with torch.no_grad():
output_dict = model(**inputs)
loss, choice_logits = output_dict[
'loss'], output_dict['choice_logits']
eval_loss.update(loss.item(), 1)
eval_accuracy(choice_logits, inputs["labels"])
eval_epoch_loss = eval_loss.avg
summary_writer.add_scalar('eval_loss', eval_epoch_loss,
eval_epoch)
eval_loss.reset()
current_acc = eval_accuracy.get_metric(reset=True)
summary_writer.add_scalar('eval_acc', current_acc,
eval_epoch)
torch.cuda.empty_cache()
if args.save_all:
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, f"pytorch_model_{eval_epoch}.bin")
torch.save(model_to_save.state_dict(),
output_model_file)
if current_acc > best_acc:
best_acc = current_acc
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)
if eval_epoch_loss < best_loss:
best_loss = eval_epoch_loss
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_loss_model.bin")
torch.save(model_to_save.state_dict(),
output_model_file)
logger.info(
'Eval Epoch: %d, Accuracy: %.4f (Best Accuracy: %.4f)'
% (eval_epoch, current_acc, best_acc))
eval_epoch += 1
logger.info(
f'Epoch {epoch}: Accuracy: {best_acc}, Train Loss: {train_loss.avg}'
)
summary_writer.close()
for output_model_name in ["pytorch_model.bin", "pytorch_loss_model.bin"]:
# Loading trained model
output_model_file = os.path.join(args.output_dir, output_model_name)
model_state_dict = torch.load(output_model_file, map_location='cuda:0')
model = initialize_model(args.bert_name,
args.model_file,
state_dict=model_state_dict,
**model_params)
model.to(device)
# Write Yes/No predictions
if args.do_predict and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
test_examples = data_reader.read(args.test_file)
test_features = data_reader.convert_examples_to_features(
test_examples, tokenizer, args.max_seq_length)
test_tensors = data_reader.data_to_tensors(test_features)
test_data = TensorDataset(*test_tensors)
test_sampler = SequentialSampler(test_data)
test_dataloader = DataLoader(test_data,
sampler=test_sampler,
batch_size=args.predict_batch_size)
logger.info("***** Running predictions *****")
logger.info(" Num orig examples = %d", len(test_examples))
logger.info(" Num split examples = %d", len(test_features))
logger.info(" Batch size = %d", args.predict_batch_size)
model.eval()
all_results = []
test_acc = CategoricalAccuracy()
logger.info("Start predicting yes/no on Dev set.")
for batch in tqdm(test_dataloader, desc="Testing"):
if n_gpu == 1:
batch = batch_to_device(
batch, device) # multi-gpu does scattering it-self
inputs = data_reader.generate_inputs(
batch, test_features, model_state=ModelState.Evaluate)
with torch.no_grad():
batch_choice_logits = model(**inputs)['choice_logits']
test_acc(batch_choice_logits, inputs['labels'])
example_indices = batch[-1]
for i, example_index in enumerate(example_indices):
choice_logits = batch_choice_logits[i].detach().cpu(
).tolist()
test_feature = test_features[example_index.item()]
unique_id = int(test_feature.unique_id)
all_results.append(
RawResultChoice(unique_id=unique_id,
choice_logits=choice_logits))
if "loss" in output_model_name:
logger.info(
'Predicting question choice on test set using model with lowest loss on validation set.'
)
output_prediction_file = os.path.join(args.predict_dir,
'loss_predictions.json')
else:
logger.info(
'Predicting question choice on test set using model with best accuracy on validation set,'
)
output_prediction_file = os.path.join(args.predict_dir,
'predictions.json')
data_reader.write_predictions(test_examples, test_features,
all_results, output_prediction_file)
logger.info(
f"Accuracy on Test set: {test_acc.get_metric(reset=True)}")
# Loading trained model.
if args.metric == 'accuracy':
logger.info("Load model with best accuracy on validation set.")
output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
elif args.metric == 'loss':
logger.info("Load model with lowest loss on validation set.")
output_model_file = os.path.join(args.output_dir,
"pytorch_loss_model.bin")
else:
raise RuntimeError(
f"Wrong metric type for {args.metric}, which must be in ['accuracy', 'loss']."
)
model_state_dict = torch.load(output_model_file, map_location='cuda:0')
model = initialize_model(args.bert_name,
args.model_file,
state_dict=model_state_dict,
**model_params)
model.to(device)
# Labeling sentence id.
if args.do_label and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
f = open('debug_log.txt', 'w')
def softmax(x):
"""Compute softmax values for each sets of scores in x."""
e_x = np.exp(x - np.max(x))
return e_x / e_x.sum()
def topk(sentence_sim):
"""
:param sentence_sim: numpy
:return:
"""
max_length = min(args.num_evidence, len(sentence_sim))
sorted_scores = np.array(sorted(sentence_sim, reverse=True))
scores = []
for idx in range(max_length):
scores.append(np.log(softmax(sorted_scores[idx:])[0]))
scores = [np.mean(scores[:(j + 1)]) for j in range(max_length)]
top_k = int(np.argmax(scores) + 1)
sorted_scores = sorted(enumerate(sentence_sim),
key=lambda x: x[1],
reverse=True)
evidence_ids = [x[0] for x in sorted_scores[:top_k]]
sentence = {
'sentences': evidence_ids,
'value': float(np.exp(scores[top_k - 1]))
}
return sentence
def batch_topk(sentence_sim, sentence_mask):
batch_size = sentence_sim.size(0)
num_choices = sentence_sim.size(1)
sentence_sim = sentence_sim.numpy() + 1e-15
sentence_mask = sentence_mask.numpy()
sentence_ids = []
for b in range(batch_size):
choice_sentence_ids = [
topk(_sim[:int(sum(_mask))])
for _sim, _mask in zip(sentence_sim[b], sentence_mask[b])
]
assert len(choice_sentence_ids) == num_choices
sentence_ids.append(choice_sentence_ids)
return sentence_ids
test_examples = train_examples
test_features = train_features
test_tensors = data_reader.data_to_tensors(test_features)
test_data = TensorDataset(*test_tensors)
test_sampler = SequentialSampler(test_data)
test_dataloader = DataLoader(test_data,
sampler=test_sampler,
batch_size=args.predict_batch_size)
logger.info("***** Running labeling *****")
logger.info(" Num orig examples = %d", len(test_examples))
logger.info(" Num split examples = %d", len(test_features))
logger.info(" Batch size = %d", args.predict_batch_size)
model.eval()
all_results = []
logger.info("Start labeling.")
for batch in tqdm(test_dataloader, desc="Testing"):
if n_gpu == 1:
batch = batch_to_device(batch, device)
inputs = data_reader.generate_inputs(batch,
test_features,
model_state=ModelState.Test)
with torch.no_grad():
output_dict = model(**inputs)
batch_choice_logits, batch_sentence_logits = output_dict[
"choice_logits"], output_dict["sentence_logits"]
batch_sentence_mask = output_dict["sentence_mask"]
example_indices = batch[-1]
# batch_beam_results = batch_choice_beam_search(batch_sentence_logits, batch_sentence_mask)
batch_topk_results = batch_topk(batch_sentence_logits,
batch_sentence_mask)
for i, example_index in enumerate(example_indices):
choice_logits = batch_choice_logits[i].detach().cpu()
evidence_list = batch_topk_results[i]
test_feature = test_features[example_index.item()]
unique_id = int(test_feature.unique_id)
all_results.append(
RawOutput(unique_id=unique_id,
model_output={
"choice_logits": choice_logits,
"evidence_list": evidence_list
}))
output_prediction_file = os.path.join(args.predict_dir,
'sentence_id_file.json')
data_reader.predict_sentence_ids(
test_examples,
test_features,
all_results,
output_prediction_file,
weight_threshold=args.weight_threshold,
only_correct=args.only_correct,
label_threshold=args.label_threshold)
print('Start training!')
writer = SummaryWriter(dir+'models/logs/%s/'%LOG_NAME)
loss_fct = torch.nn.CrossEntropyLoss()
model.train()
for epoch in range(current_epoch, EPOCH):
loss_train = 0
for idx, batch in enumerate(tqdm(train_loader), 1):
loss, correct = train(batch, model, loss_fct, optimizer, idx, gradient_accumulation_steps)
writer.add_scalar('loss_train_batch/loss_train_batch', loss, global_batch_counter_train)
writer.add_scalar('accuracy_train_batch/accuracy_train_batch', correct/batch[0].shape[0], global_batch_counter_train)
writer.add_scalar('epoch/lr', optimizer.get_lr()[-1], global_batch_counter_train)
global_batch_counter_train += 1
loss_train += loss
if (test_per_n_batch_one_epoch and (idx % test_per_n_batch_one_epoch == 0)) or (DEBUG and idx == 4 and test_per_n_batch_one_epoch):
print('Start testing...')
loss_test = 0
total = 0
correct = 0
y_pred = []
y_true = []
model.eval()
with torch.no_grad():
for test_idx, batch in enumerate(tqdm(test_loader), 1):
def train(**kwargs):
config = Config()
config.update(**kwargs)
print('当前设置为:\n', config)
if args.use_cuda:
torch.cuda.set_device(config.gpu)
print('loading corpus')
vocab = load_vocab(args.vocab_file)
label_dic = load_vocab(config.label_file)
index2label={v:k for k,v in label_dic.items()}
tagset_size = len(label_dic)
train_data,_ = read_corpus(args.pretrain_train_path, max_length=args.max_seq_length, label_dic=label_dic, vocab=vocab)
dev_data,dev_len = read_corpus(args.pretrain_dev_path, max_length=args.max_seq_length, label_dic=label_dic, vocab=vocab)
num_train_optimization_steps = int(
len(train_data) / 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()
train_ids = torch.LongTensor([temp.input_id for temp in train_data])
train_masks = torch.LongTensor([temp.input_mask for temp in train_data])
train_tags = torch.LongTensor([temp.label_id for temp in train_data])
train_dataset = TensorDataset(train_ids, train_masks, train_tags)
train_loader = DataLoader(train_dataset, shuffle=True, batch_size=args.train_batch_size)
dev_ids = torch.LongTensor([temp.input_id for temp in dev_data])
dev_masks = torch.LongTensor([temp.input_mask for temp in dev_data])
dev_tags = torch.LongTensor([temp.label_id for temp in dev_data])
dev_dataset = TensorDataset(dev_ids, dev_masks, dev_tags)
dev_loader = DataLoader(dev_dataset, shuffle=False, batch_size=args.eval_batch_size)
model = BERT_LSTM_CRF(args, tagset_size, config.bert_embedding, config.rnn_hidden, config.rnn_layer, dropout_ratio=config.dropout_ratio, dropout1=config.dropout1, use_cuda=config.use_cuda)
if config.use_cuda:
model=model.cuda()
if config.load_model:
if config.flag=='submit':
assert config.load_path is not None
test_data, test_len = read_corpus(args.submit_test_path, max_length=args.max_seq_length, label_dic=label_dic,
vocab=vocab,flag='submit')
test_ids = torch.LongTensor([temp.input_id for temp in test_data])
test_masks = torch.LongTensor([temp.input_mask for temp in test_data])
test_dataset = TensorDataset(test_ids, test_masks)
test_loader = DataLoader(test_dataset, shuffle=False, batch_size=args.eval_batch_size)
model = load_model(model, name=None)
test(model, test_loader, config, index2label, test_len)
# dev(model, test_loader, None, config)
if config.flag=='test':
assert config.load_path is not None
test_data, test_len = read_corpus(args.pretrain_test_path, max_length=args.max_seq_length, label_dic=label_dic,
vocab=vocab)
test_ids = torch.LongTensor([temp.input_id for temp in test_data])
test_masks = torch.LongTensor([temp.input_mask for temp in test_data])
test_tags = torch.LongTensor([temp.label_id for temp in test_data])
test_dataset = TensorDataset(test_ids, test_masks, test_tags)
test_loader = DataLoader(test_dataset, shuffle=False, batch_size=args.eval_batch_size)
model = load_model(model, name=None)
# test(model, test_loader, config, index2label, test_len)
dev(model, test_loader, 0, config, index2label, dev_len)
else:
# print(model)
model.train()
bert_param_optimizer = list(model.word_embeds.named_parameters())
lstm_param_optimizer = list(model.lstm.named_parameters())
liner_param_optimizer = list(model.liner.named_parameters())
crf_param_optimizer = list(model.crf.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
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': 0.01,'lr':config.lr},
{'params': [p for n, p in bert_param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0,'lr':config.lr},
{'params': [p for n, p in lstm_param_optimizer if not any(nd in n for nd in no_decay)],'weight_decay': 0.001, 'lr': config.lr*5},
{'params': [p for n, p in lstm_param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0,'lr': config.lr*5},
{'params': [p for n, p in liner_param_optimizer if not any(nd in n for nd in no_decay)],'weight_decay': 0.001, 'lr': config.lr * 2},
{'params': [p for n, p in liner_param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0, 'lr':config.lr * 2},
{'params': [p for n, p in crf_param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.001, 'lr': config.lr * 3},
{'params': [p for n, p in crf_param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0,'lr': config.lr * 3},
]
# print(optimizer_grouped_parameters)
optimizer = BertAdam(optimizer_grouped_parameters,
# lr=config.lr,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
# optimizer = optimizer(model.parameters(), lr=config.lr, weight_decay=config.weight_decay)
eval_f1 = 0.0
for epoch in range(config.base_epoch):
print(optimizer.get_lr())
step = 0
for i, batch in enumerate(train_loader):
step += 1
model.zero_grad()
inputs, masks, tags = batch
inputs, masks, tags = Variable(inputs), Variable(masks), Variable(tags)
if config.use_cuda:
inputs, masks, tags = inputs.cuda(), masks.cuda(), tags.cuda()
feats = model(inputs, masks)
loss = model.loss(feats, masks,tags)
loss.backward()
optimizer.step()
if step % 50 == 0:
print('step: {} | epoch: {}| loss: {}'.format(step, epoch, loss.item()))
f_measure = dev(model, dev_loader, epoch, config,index2label,dev_len)
if eval_f1 < f_measure:
eval_f1=f_measure
save_model(model,epoch,f_measure)
def train(params,
datasets,
student,
val_iterator,
cuda_device=-1,
teachers=None):
""" Train the model """
tb_writer = SummaryWriter()
t_total = params.params.get('trainer').get('optimizer').get("t_total")
no_decay = ['bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in student.named_parameters()
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params': [
p for n, p in student.named_parameters()
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0
}]
# BertAdam already has a scheduler
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon,
t_total=t_total)
# parameters = [[n, p] for n, p in student.named_parameters() if p.requires_grad]
# optimizer = Optimizer.from_params(parameters, params.get("trainer").pop("optimizer"))
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(list(datasets["train"])))
logger.info(" Num Epochs = %d", num_epochs)
logger.info(" Instantaneous batch size per GPU = %d",
params.get("iterator").get("batch_size"))
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
total_training_loss = 0.0
logging_loss = 0.0
best_f1 = 0.0
set_seed(
args) # Added here for reproductibility (even between python 2 and 3)
iterator = DataIterator.from_params(params.pop("iterator"))
iterator.index_with(student.vocab)
for epoch in range(num_epochs):
tr_loss = 0.0
student.zero_grad()
batches_this_epoch = 0
epoch_start_time = time.time()
logger.info("Training")
logger.info("Epoch %d/%d", epoch, num_epochs - 1)
# Get tqdm for the training batches
train_generator = iterator(datasets["train"],
num_epochs=1,
shuffle=True)
num_training_batches = math.ceil(
iterator.get_num_batches(datasets["train"]))
train_generator_tqdm = Tqdm.tqdm(train_generator,
total=num_training_batches)
for batch_group in train_generator_tqdm:
batches_this_epoch += 1
set_name = batch_group["metadata"][0]["dataset"]
if teachers != {}:
assert set_name in teachers.keys()
teacher = teachers.get(set_name, None)
student.train()
if teacher is not None:
teacher.eval()
batch_group = nn_util.move_to_device(batch_group, cuda_device)
output_dict = student(**batch_group)
start_logits_stu = output_dict["span_start_logits"]
end_logits_stu = output_dict["span_end_logits"]
loss = output_dict["loss"] # pure student loss, gold loss
# Distillation loss
if teacher is not None:
with torch.no_grad():
teacher_output_dict = teacher(**batch_group)
start_logits_tea = teacher_output_dict["span_start_logits"]
end_logits_tea = teacher_output_dict["span_end_logits"]
assert start_logits_stu.size() == start_logits_tea.size()
assert end_logits_stu.size() == end_logits_tea.size()
bias_weights = get_bias_weight(args.bias_type, batch_group,
set_name)
# confidence regularization method
if args.method == "CR":
loss_start = probability_scaling(start_logits_stu,
bias_weights[:, 0],
start_logits_tea)
loss_end = probability_scaling(end_logits_stu,
bias_weights[:, 1],
end_logits_tea)
else:
# the WL method
loss_start = loss_reweighting(start_logits_stu,
bias_weights[:, 0],
start_logits_tea)
loss_end = loss_reweighting(end_logits_stu,
bias_weights[:, 1],
end_logits_tea)
loss = loss_start + loss_end
loss.backward()
tr_loss += loss.item()
optimizer.step()
student.zero_grad()
global_step += 1
metrics = training_util.get_metrics(student, tr_loss,
batches_this_epoch)
description = training_util.description_from_metrics(
metrics) + "\n"
train_generator_tqdm.set_description(description, refresh=False)
training_util.get_metrics(student,
tr_loss,
batches_this_epoch,
reset=True)
# evaluate on the validation dataset
with torch.no_grad():
logging.info("validation student")
metrics = evaluate(student,
datasets["validation"],
val_iterator,
cuda_device,
batch_weight_key="")
current_f1 = metrics["f1"]
for key, value in metrics.items():
tb_writer.add_scalar("eval_{}".format(key), value, global_step)
tb_writer.add_scalar("lr", optimizer.get_lr()[0], global_step)
tb_writer.add_scalar("loss", (tr_loss - logging_loss) /
args.logging_steps, global_step)
logging_loss = tr_loss
logger.info("{'Epoch %d/%d, student exact_match': %s, 'f1': %s}",
epoch, num_epochs - 1, metrics["EM"], metrics["f1"])
# Save model checkpoint
model_path = os.path.join(args.output_dir,
"model_state_epoch_{}.th".format(epoch))
best_path = os.path.join(args.output_dir, "best.th")
torch.save(student.state_dict(), model_path)
if current_f1 > best_f1:
torch.save(student.state_dict(), best_path)
best_f1 = current_f1
logger.info("Saving model checkpoint to %s", args.output_dir)
epoch_elapsed_time = time.time() - epoch_start_time
logger.info("Epoch duration: %s",
datetime.timedelta(seconds=epoch_elapsed_time))
total_training_loss += tr_loss
student.get_metrics(reset=True)
tb_writer.close()
return global_step, total_training_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-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(args):
logging = config.get_logging(args.log_name)
logging.info("##"*20)
logging.info("##"*20)
logging.info("##"*20)
logging.info(args)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
logging.info("| question first :: {}".format(args.question_first))
logging.info("| gpu count : {}".format(n_gpu))
logging.info("| train batch size in each gpu : {}".format(args.train_batch_size))
logging.info("| biuid tokenizer and model in : {}".format(args.pre_dir))
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
tokenizer = BertTokenizer.build_tokenizer(args)
train_data_iter = MSmarco_iterator(args, tokenizer, batch_size=args.train_batch_size, world_size=n_gpu, accumulation_steps=args.gradient_accumulation_steps, name="msmarco_train.pk")
dev_data_iter = MSmarco_iterator(args, tokenizer, batch_size=args.valid_batch_size, world_size=n_gpu, name="msmarco_dev.pk")
data_size = len(train_data_iter)
gradient_accumulation_steps = args.gradient_accumulation_steps
num_train_steps = args.num_train_epochs*data_size//gradient_accumulation_steps
# logging.info("| load dataset {}".format(data_size))
logging.info("| train data size {}".format(len(train_data_iter)*n_gpu*args.train_batch_size))
logging.info("| dev data size {}".format(len(dev_data_iter)*n_gpu*args.valid_batch_size))
logging.info("| train batch data size {}".format(len(train_data_iter)))
logging.info("| dev batch data size {}".format(len(dev_data_iter)))
logging.info("| update in each train data {}".format(data_size//gradient_accumulation_steps))
logging.info("| total update {}".format(num_train_steps))
# num_train_steps = (96032//2//2)+(data_size-96032)//2
model = MSmarco.build_model(args)
model.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta', 'layer_norm']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.0}
]
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.lr,
warmup=args.warmup_proportion,
t_total=num_train_steps)
logging.info("| init lr is {}".format(optimizer.get_lr()))
global_update = 0
for epochs in range(args.num_train_epochs):
total_loss = 0
merge_batch = []
# count = 0
for step, batch in enumerate(tqdm(train_data_iter, desc="Train Iteration")):
model.train()
# if step < 96032:
# merge_batch.append(batch)
# if len(merge_batch) == 2:
# batch = merger_tensor(merge_batch)
# merge_batch = []
# else:
# continue
if n_gpu==1:
for key in batch.keys():
batch[key]=batch[key].to(device)
loss = model(**batch)
# count += 1
# pdb.set_trace()
if n_gpu > 1:
loss = loss.mean()
if args.gradient_accumulation_steps > 1:
loss = loss/args.gradient_accumulation_steps
loss.backward()
if (step+1) % args.gradient_accumulation_steps == 0:
optimizer.step()
model.zero_grad()
global_update += 1
if global_update % args.validate_updates==0:
validation(args, model, dev_data_iter, n_gpu, epochs, global_update, logging)
if (step+1) % args.loss_interval==0:
logging.info("TRAIN ::Epoch {} updates {}, train loss {}".format(epochs, global_update, loss.item()))
# save_checkpoint(args, model, epochs)
validation(args, model, dev_data_iter, n_gpu, epochs, global_update, logging)