def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--bert_config_file",
default=None,
type=str,
required=True,
help="The config json file corresponding to the pre-trained BERT model. "
"This specifies the model architecture.")
parser.add_argument(
"--vocab_file",
default=None,
type=str,
required=True,
help="The vocabulary file that the BERT model was trained on.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model checkpoints 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(
"--init_checkpoint",
default=None,
type=str,
help="Initial checkpoint (usually from a pre-trained BERT model).")
parser.add_argument(
"--do_lower_case",
default=True,
action='store_true',
help="Whether to lower case the input text. Should be True for uncased "
"models and False for cased models.")
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=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("--save_checkpoints_steps",
default=1000,
type=int,
help="How often to save the model checkpoint.")
parser.add_argument("--iterations_per_loop",
default=1000,
type=int,
help="How many steps to make in each estimator call.")
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("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument(
"--accumulate_gradients",
type=int,
default=1,
help=
"Number of steps to accumulate gradient on (divide the batch_size and accumulate)"
)
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=128,
help=
'Loss scaling, positive power of 2 values can improve fp16 convergence.'
)
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:
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')
if args.fp16:
logger.info(
"16-bits training currently not supported in distributed training"
)
args.fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496)
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits trainiing: {}"
.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."
)
bert_config = BertConfig.from_json_file(args.bert_config_file)
if args.max_seq_length > bert_config.max_position_embeddings:
raise ValueError(
"Cannot use sequence length %d because the BERT model "
"was only trained up to sequence length %d" %
(args.max_seq_length, bert_config.max_position_embeddings))
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)
tokenizer = tokenization.FullTokenizer(vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
if args.do_train:
train_examples = read_squad_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 = BertForQuestionAnswering(bert_config)
if args.init_checkpoint is not None:
model.bert.load_state_dict(
torch.load(args.init_checkpoint, map_location='cpu'))
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
if args.fp16:
param_optimizer = [(n, param.clone().detach().to('cpu').float().requires_grad_()) \
for n, param in model.named_parameters()]
elif args.optimize_on_cpu:
param_optimizer = [(n, param.clone().detach().to('cpu').requires_grad_()) \
for n, param in model.named_parameters()]
else:
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta']
optimizer_grouped_parameters = [{
'params': [p for n, p in param_optimizer if n not in no_decay],
'weight_decay_rate':
0.01
}, {
'params': [p for n, p in param_optimizer if n in no_decay],
'weight_decay_rate':
0.0
}]
optimizer = BERTAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_steps)
global_step = 0
if args.do_train:
train_features = convert_examples_to_features(
examples=train_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=True)
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)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_start_positions,
all_end_positions)
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)
eval_examples = read_squad_examples(input_file=args.predict_file,
is_training=False)
eval_features = convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=False)
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)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_example_index)
if args.local_rank == -1:
eval_sampler = SequentialSampler(eval_data)
else:
eval_sampler = DistributedSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.predict_batch_size)
model.train()
global_step = 0
for t_epoch in trange(int(args.num_train_epochs), desc="Epoch"):
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
global_step += 1
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 = batch
loss = model(input_ids, segment_ids, input_mask,
start_positions, end_positions)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.fp16 and args.loss_scale != 1.0:
# rescale loss for fp16 training
# see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html
loss = loss * args.loss_scale
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16 or args.optimize_on_cpu:
if args.fp16 and args.loss_scale != 1.0:
# scale down gradients for fp16 training
for param in model.parameters():
param.grad.data = param.grad.data / args.loss_scale
is_nan = set_optimizer_params_grad(
param_optimizer,
model.named_parameters(),
test_nan=True)
if is_nan:
logger.info(
"FP16 TRAINING: Nan in gradients, reducing loss scaling"
)
args.loss_scale = args.loss_scale / 2
model.zero_grad()
continue
optimizer.step()
copy_optimizer_params_to_model(
model.named_parameters(), param_optimizer)
else:
optimizer.step()
model.zero_grad()
global_step += 1
if args.do_predict and global_step % args.save_checkpoints_steps == 0:
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)
model.eval()
all_results = []
logger.info("Start evaluating")
for input_ids, input_mask, segment_ids, example_indices in tqdm(
eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
with torch.no_grad():
batch_start_logits, batch_end_logits = model(
input_ids, segment_ids, 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()
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))
output_prediction_file = os.path.join(
args.output_dir,
"predictions_{}_{}.json".format(t_epoch, global_step))
output_nbest_file = os.path.join(
args.output_dir, "nbest_predictions_{}_{}.json".format(
t_epoch, global_step))
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)
torch.save(
model.state_dict(),
os.path.join(args.output_dir,
'model' + global_step + '.pt'))
model.train()
def main():
parser = argparse.ArgumentParser()
BERT_DIR = "./model/uncased_L-12_H-768_A-12/"
## Required parameters
parser.add_argument("--bert_config_file", default=BERT_DIR+"bert_config.json", \
type=str, help="The config json file corresponding to the pre-trained BERT model. "
"This specifies the model architecture.")
parser.add_argument("--vocab_file", default=BERT_DIR+"vocab.txt", type=str, \
help="The vocabulary file that the BERT model was trained on.")
parser.add_argument("--output_dir", default="out", type=str, \
help="The output directory where the model checkpoints will be written.")
## Other parameters
parser.add_argument("--train_file", type=str, \
help="SQuAD json for training. E.g., train-v1.1.json", \
default="")
parser.add_argument("--predict_file", type=str,
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json", \
default="")
parser.add_argument("--init_checkpoint", type=str,
help="Initial checkpoint (usually from a pre-trained BERT model).", \
default=BERT_DIR+"pytorch_model.bin")
parser.add_argument(
"--do_lower_case",
default=True,
action='store_true',
help="Whether to lower case the input text. Should be True for uncased "
"models and False for cased models.")
parser.add_argument(
"--max_seq_length",
default=300,
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=128,
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=10.0,
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("--save_checkpoints_steps",
default=1000,
type=int,
help="How often to save the model checkpoint.")
parser.add_argument("--iterations_per_loop",
default=1000,
type=int,
help="How many steps to make in each estimator call.")
parser.add_argument(
"--n_best_size",
default=3,
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("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument(
"--accumulate_gradients",
type=int,
default=1,
help=
"Number of steps to accumulate gradient on (divide the batch_size and accumulate)"
)
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('--eval_period', type=int, default=2000)
parser.add_argument('--max_n_answers', type=int, default=5)
parser.add_argument('--merge_query', type=int, default=-1)
parser.add_argument('--reduce_layers', type=int, default=-1)
parser.add_argument('--reduce_layers_to_tune', type=int, default=-1)
parser.add_argument('--only_comp', action="store_true", default=False)
parser.add_argument('--train_subqueries_file', type=str, default="") #500
parser.add_argument('--predict_subqueries_file', type=str,
default="") #500
parser.add_argument('--prefix', type=str, default="") #500
parser.add_argument('--model', type=str, default="qa") #500
parser.add_argument('--pooling', type=str, default="max")
parser.add_argument('--debug', action="store_true", default=False)
parser.add_argument('--output_dropout_prob', type=float, default=0)
parser.add_argument('--wait_step', type=int, default=30)
parser.add_argument('--with_key', action="store_true", default=False)
parser.add_argument('--add_noise', action="store_true", default=False)
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:
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 %s n_gpu %d distributed training %r", device, n_gpu,
bool(args.local_rank != -1))
if args.accumulate_gradients < 1:
raise ValueError(
"Invalid accumulate_gradients parameter: {}, should be >= 1".
format(args.accumulate_gradients))
args.train_batch_size = int(args.train_batch_size /
args.accumulate_gradients)
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 not args.predict_file:
raise ValueError(
"If `do_train` is True, then `predict_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."
)
bert_config = BertConfig.from_json_file(args.bert_config_file)
if args.do_train and args.max_seq_length > bert_config.max_position_embeddings:
raise ValueError(
"Cannot use sequence length %d because the BERT model "
"was only trained up to sequence length %d" %
(args.max_seq_length, bert_config.max_position_embeddings))
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
logger.info("Output directory () already exists and is not empty.")
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir, exist_ok=True)
tokenizer = tokenization.FullTokenizer(vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
eval_dataloader, eval_examples, eval_features, _ = get_dataloader(
logger=logger,
args=args,
input_file=args.predict_file,
subqueries_file=args.predict_subqueries_file,
is_training=False,
batch_size=args.predict_batch_size,
num_epochs=1,
tokenizer=tokenizer)
if args.do_train:
train_dataloader, train_examples, _, num_train_steps = get_dataloader(
logger=logger, args=args, \
input_file=args.train_file, \
subqueries_file=args.train_subqueries_file, \
is_training=True,
batch_size=args.train_batch_size,
num_epochs=args.num_train_epochs,
tokenizer=tokenizer)
#a = input()
if args.model == 'qa':
model = BertForQuestionAnswering(bert_config, 4)
metric_name = "F1"
elif args.model == 'classifier':
if args.reduce_layers != -1:
bert_config.num_hidden_layers = args.reduce_layers
model = BertClassifier(bert_config, 2, args.pooling)
metric_name = "F1"
elif args.model == "span-predictor":
if args.reduce_layers != -1:
bert_config.num_hidden_layers = args.reduce_layers
if args.with_key:
Model = BertForQuestionAnsweringWithKeyword
else:
Model = BertForQuestionAnswering
model = Model(bert_config, 2)
metric_name = "Accuracy"
else:
raise NotImplementedError()
if args.init_checkpoint is not None and args.do_predict and \
len(args.init_checkpoint.split(','))>1:
assert args.model == "qa"
model = [model]
for i, checkpoint in enumerate(args.init_checkpoint.split(',')):
if i > 0:
model.append(BertForQuestionAnswering(bert_config, 4))
print("Loading from", checkpoint)
state_dict = torch.load(checkpoint, map_location='cpu')
filter = lambda x: x[7:] if x.startswith('module.') else x
state_dict = {filter(k): v for (k, v) in state_dict.items()}
model[-1].load_state_dict(state_dict)
model[-1].to(device)
else:
if args.init_checkpoint is not None:
print("Loading from", args.init_checkpoint)
state_dict = torch.load(args.init_checkpoint, map_location='cpu')
if args.reduce_layers != -1:
state_dict = {k:v for k, v in state_dict.items() \
if not '.'.join(k.split('.')[:3]) in \
['encoder.layer.{}'.format(i) for i in range(args.reduce_layers, 12)]}
if args.do_predict:
filter = lambda x: x[7:] if x.startswith('module.') else x
state_dict = {filter(k): v for (k, v) in state_dict.items()}
model.load_state_dict(state_dict)
else:
model.bert.load_state_dict(state_dict)
if args.reduce_layers_to_tune != -1:
model.bert.embeddings.required_grad = False
n_layers = 12 if args.reduce_layers == -1 else args.reduce_layers
for i in range(n_layers - args.reduce_layers_to_tune):
model.bert.encoder.layer[i].require_grad = False
model.to(device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
if args.do_train:
no_decay = ['bias', 'gamma', 'beta']
optimizer_parameters = [{
'params':
[p for n, p in model.named_parameters() if n not in no_decay],
'weight_decay_rate':
0.01
}, {
'params':
[p for n, p in model.named_parameters() if n in no_decay],
'weight_decay_rate':
0.0
}]
optimizer = BERTAdam(optimizer_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_steps)
global_step = 0
best_f1 = 0
wait_step = 0
model.train()
global_step = 0
stop_training = False
for epoch in range(int(args.num_train_epochs)):
for step, batch in tqdm(enumerate(train_dataloader)):
global_step += 1
batch = [t.to(device) for t in batch]
loss = model(batch, global_step)
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
loss.backward()
if global_step % args.gradient_accumulation_steps == 0:
optimizer.step() # We have accumulated enought gradients
model.zero_grad()
if global_step % args.eval_period == 0:
model.eval()
f1 = predict(args, model, eval_dataloader, eval_examples, eval_features, \
device, write_prediction=False)
logger.info("%s: %.3f on epoch=%d" %
(metric_name, f1 * 100.0, epoch))
if best_f1 < f1:
logger.info("Saving model with best %s: %.3f -> %.3f on epoch=%d" % \
(metric_name, best_f1*100.0, f1*100.0, epoch))
model_state_dict = {
k: v.cpu()
for (k, v) in model.state_dict().items()
}
torch.save(
model_state_dict,
os.path.join(args.output_dir, "best-model.pt"))
model = model.cuda()
best_f1 = f1
wait_step = 0
stop_training = False
else:
wait_step += 1
if best_f1 > 0.1 and wait_step == args.wait_step:
stop_training = True
model.train()
if stop_training:
break
elif args.do_predict:
if type(model) == list:
model = [m.eval() for m in model]
else:
model.eval()
f1 = predict(args, model, eval_dataloader, eval_examples,
eval_features, device)
logger.info("Final %s score: %.3f%%" % (metric_name, f1 * 100.0))
def main():
parser = argparse.ArgumentParser()
BERT_DIR = "uncased_L-12_H-768_A-12/"
## Required parameters
parser.add_argument("--bert_config_file", default=BERT_DIR+"bert_config.json", \
type=str, help="The config json file corresponding to the pre-trained BERT model. "
"This specifies the model architecture.")
parser.add_argument("--vocab_file", default=BERT_DIR+"vocab.txt", type=str, \
help="The vocabulary file that the BERT model was trained on.")
parser.add_argument("--output_dir", default="out", type=str, \
help="The output directory where the model checkpoints will be written.")
## Other parameters
parser.add_argument("--load", default=False, action='store_true')
parser.add_argument("--train_file", type=str, \
help="SQuAD json for training. E.g., train-v1.1.json", \
default="/home/sewon/data/squad/train-v1.1.json")
parser.add_argument("--predict_file", type=str,
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json", \
default="/home/sewon/data/squad/dev-v1.1.json")
parser.add_argument("--init_checkpoint", type=str,
help="Initial checkpoint (usually from a pre-trained BERT model).", \
default=BERT_DIR+"pytorch_model.bin")
parser.add_argument(
"--do_lower_case",
default=True,
action='store_true',
help="Whether to lower case the input text. Should be True for uncased "
"models and False for cased models.")
parser.add_argument(
"--max_seq_length",
default=300,
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=39,
type=int,
help="Total batch size for training.")
parser.add_argument("--predict_batch_size",
default=300,
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=1000.0,
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("--save_checkpoints_steps",
default=1000,
type=int,
help="How often to save the model checkpoint.")
parser.add_argument("--iterations_per_loop",
default=1000,
type=int,
help="How many steps to make in each estimator call.")
parser.add_argument(
"--n_best_size",
default=3,
type=int,
help=
"The total number of n-best predictions to generate in the nbest_predictions.json "
"output file.")
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("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument(
"--accumulate_gradients",
type=int,
default=1,
help=
"Number of steps to accumulate gradient on (divide the batch_size and accumulate)"
)
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('--eval_period', type=int, default=500)
parser.add_argument('--max_n_answers', type=int, default=20)
parser.add_argument('--n_paragraphs', type=str, default='40')
parser.add_argument('--verbose', action="store_true", default=False)
parser.add_argument('--wait_step', type=int, default=12)
# Learning method variation
parser.add_argument('--loss_type', type=str, default="mml")
parser.add_argument('--tau', type=float, default=12000.0)
# For evaluation
parser.add_argument('--prefix', type=str, default="") #500
parser.add_argument('--debug', action="store_true", default=False)
args = parser.parse_args()
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
print("Output directory () already exists and is not empty.")
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir, exist_ok=True)
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO,
handlers=[
logging.FileHandler(os.path.join(args.output_dir, "log.txt")),
logging.StreamHandler()
])
logger = logging.getLogger(__name__)
logger.info(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:
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 %s n_gpu %d distributed training %r", device, n_gpu,
bool(args.local_rank != -1))
if args.accumulate_gradients < 1:
raise ValueError(
"Invalid accumulate_gradients parameter: {}, should be >= 1".
format(args.accumulate_gradients))
args.train_batch_size = int(args.train_batch_size /
args.accumulate_gradients)
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 not args.predict_file:
raise ValueError(
"If `do_train` is True, then `predict_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."
)
bert_config = BertConfig.from_json_file(args.bert_config_file)
if args.do_train and args.max_seq_length > bert_config.max_position_embeddings:
raise ValueError(
"Cannot use sequence length %d because the BERT model "
"was only trained up to sequence length %d" %
(args.max_seq_length, bert_config.max_position_embeddings))
model = BertForQuestionAnswering(bert_config,
device,
4,
loss_type=args.loss_type,
tau=args.tau)
metric_name = "EM"
tokenizer = tokenization.FullTokenizer(vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
train_split = ',' in args.train_file
if train_split:
n_train_files = len(args.train_file.split(','))
eval_dataloader, eval_examples, eval_features, _ = get_dataloader(
logger=logger,
args=args,
input_file=args.predict_file,
is_training=False,
batch_size=args.predict_batch_size,
num_epochs=1,
tokenizer=tokenizer)
if args.do_train:
train_file = args.train_file
if train_split:
train_file = args.train_file.split(',')[0]
train_dataloader, _, _, num_train_steps = get_dataloader(
logger=logger, args=args, \
input_file=train_file, \
is_training=True,
batch_size=args.train_batch_size,
num_epochs=args.num_train_epochs,
tokenizer=tokenizer)
if args.init_checkpoint is not None:
logger.info("Loading from {}".format(args.init_checkpoint))
state_dict = torch.load(args.init_checkpoint, map_location='cpu')
if args.do_train and args.init_checkpoint.endswith(
'pytorch_model.bin'):
model.bert.load_state_dict(state_dict)
else:
filter = lambda x: x[7:] if x.startswith('module.') else x
state_dict = {filter(k): v for (k, v) in state_dict.items()}
model.load_state_dict(state_dict)
model.to(device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
if args.do_train:
no_decay = ['bias', 'gamma', 'beta']
optimizer_parameters = [{
'params':
[p for n, p in model.named_parameters() if n not in no_decay],
'weight_decay_rate':
0.01
}, {
'params':
[p for n, p in model.named_parameters() if n in no_decay],
'weight_decay_rate':
0.0
}]
optimizer = BERTAdam(optimizer_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_steps)
global_step = 0
best_f1 = (-1, -1)
wait_step = 0
model.train()
global_step = 0
stop_training = False
train_losses = []
for epoch in range(int(args.num_train_epochs)):
if epoch > 0 and train_split:
train_file = args.train_file.split(',')[epoch % n_train_files]
train_dataloader = get_dataloader(
logger=logger, args=args, \
input_file=train_file, \
is_training=True,
batch_size=args.train_batch_size,
num_epochs=args.num_train_epochs,
tokenizer=tokenizer)[0]
for step, batch in enumerate(train_dataloader):
global_step += 1
batch = [t.to(device) for t in batch]
loss = model(batch, global_step)
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
train_losses.append(loss.detach().cpu())
loss.backward()
if global_step % args.gradient_accumulation_steps == 0:
optimizer.step() # We have accumulated enought gradients
model.zero_grad()
if global_step % args.eval_period == 0:
model.eval()
f1 = predict(logger, args, model, eval_dataloader, eval_examples, eval_features, \
device, write_prediction=False)
logger.info(
"Step %d Train loss %.2f EM %.2f F1 %.2f on epoch=%d" %
(global_step, np.mean(train_losses), f1[0] * 100,
f1[1] * 100, epoch))
train_losses = []
if best_f1 < f1:
logger.info("Saving model with best %s: %.2f (F1 %.2f) -> %.2f (F1 %.2f) on epoch=%d" % \
(metric_name, best_f1[0]*100, best_f1[1]*100, f1[0]*100, f1[1]*100, epoch))
model_state_dict = {
k: v.cpu()
for (k, v) in model.state_dict().items()
}
torch.save(
model_state_dict,
os.path.join(args.output_dir, "best-model.pt"))
model = model.to(device)
best_f1 = f1
wait_step = 0
stop_training = False
else:
wait_step += 1
if wait_step == args.wait_step:
stop_training = True
model.train()
if stop_training:
break
logger.info("Training finished!")
elif args.do_predict:
if type(model) == list:
model = [m.eval() for m in model]
else:
model.eval()
f1 = predict(logger,
args,
model,
eval_dataloader,
eval_examples,
eval_features,
device,
varying_n_paragraphs=len(args.n_paragraphs) > 1)