def BertSquad(file="",
mode='predict',
bert_model="bert-base-uncased",
output='./output'):
parser = {}
parser["bert_model"] = bert_model
parser["output_dir"] = output
parser["train_file"] = file
parser["predict_file"] = file
parser["max_seq_length"] = 384
parser["doc_stride"] = 128
parser["max_query_length"] = 64
parser["do_train"] = mode == 'train'
parser["do_predict"] = mode == 'predict'
parser["train_batch_size"] = 32
parser["predict_batch_size"] = 8
parser["learning_rate"] = 5e-5
parser["num_train_epochs"] = 3.0
parser["warmup_proportion"] = 0.1
parser["n_best_size"] = 20
parser["max_answer_length"] = 30
parser["verbose_logging"] = False
parser["no_cuda"] = False
parser['seed'] = 42
parser['gradient_accumulation_steps'] = 1
parser["do_lower_case"] = ('uncased' in bert_model)
parser["local_rank"] = -1
parser['fp16'] = False
parser['overwrite_output_dir'] = False
parser['loss_scale'] = 0
parser['version_2_with_negative'] = False
parser['null_score_diff_threshold'] = 0.0
parser['server_ip'] = ''
parser['server_port'] = ''
args = AttrDict.AttrDict(parser)
print(args)
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_predict:
raise ValueError(
"At least one of `do_train` or `do_predict` must be True.")
if args.do_train:
if not args.train_file:
raise ValueError(
"If `do_train` is True, then `train_file` must be specified.")
if args.do_predict:
if not args.predict_file:
raise ValueError(
"If `do_predict` is True, then `predict_file` must be specified."
)
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir
) and args.do_train and not args.overwrite_output_dir:
raise ValueError(
"Output directory () already exists and is not empty.")
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
if args.local_rank not in [-1, 0]:
torch.distributed.barrier(
) # Make sure only the first process in distributed training will download model & vocab
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
model = BertForQuestionAnswering.from_pretrained(args.bert_model)
if args.local_rank == 0:
torch.distributed.barrier()
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,
find_unused_parameters=True)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
if args.do_train:
if args.local_rank in [-1, 0]:
tb_writer = SummaryWriter()
# Prepare data loader
train_examples = read_squad_examples(
input_file=args.train_file,
is_training=True,
version_2_with_negative=args.version_2_with_negative)
cached_train_features_file = args.train_file + '_{0}_{1}_{2}_{3}'.format(
list(filter(None, args.bert_model.split('/'))).pop(),
str(args.max_seq_length), str(args.doc_stride),
str(args.max_query_length))
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
except:
train_features = convert_examples_to_features(
examples=train_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=True)
if args.local_rank == -1 or torch.distributed.get_rank() == 0:
logger.info(" Saving train features into cached file %s",
cached_train_features_file)
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
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)
num_train_optimization_steps = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
# if args.local_rank != -1:
# num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
# Prepare optimizer
param_optimizer = list(model.named_parameters())
# hack to remove pooler, which is not used
# thus it produce None grad that break apex
param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
warmup_linear = WarmupLinearSchedule(
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
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_optimization_steps)
model.train()
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
for step, batch in enumerate(
tqdm(train_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])):
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.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used and handles this automatically
lr_this_step = args.learning_rate * warmup_linear.get_lr(
global_step, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
if args.local_rank in [-1, 0]:
tb_writer.add_scalar('lr',
optimizer.get_lr()[0],
global_step)
tb_writer.add_scalar('loss', loss.item(), global_step)
if args.do_train and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
# Save a trained model, configuration and tokenizer
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
# If we save using the predefined names, we can load using `from_pretrained`
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
model_to_save.config.to_json_file(output_config_file)
tokenizer.save_vocabulary(args.output_dir)
# Load a trained model and vocabulary that you have fine-tuned
model = BertForQuestionAnswering.from_pretrained(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(
args.output_dir, do_lower_case=args.do_lower_case)
# Good practice: save your training arguments together with the trained model
output_args_file = os.path.join(args.output_dir, 'training_args.bin')
torch.save(args, output_args_file)
else:
# Load a trained model and vocabulary that you have fine-tuned
model = BertForQuestionAnswering.from_pretrained(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(
args.output_dir, do_lower_case=args.do_lower_case)
model.to(device)
if args.do_predict and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = read_squad_examples(
input_file=args.predict_file,
is_training=False,
version_2_with_negative=args.version_2_with_negative)
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)
logger.info("***** Running predictions *****")
logger.info(" Num orig examples = %d", len(eval_examples))
logger.info(" Num split examples = %d", len(eval_features))
logger.info(" Batch size = %d", args.predict_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_example_index = torch.arange(all_input_ids.size(0),
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_example_index)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.predict_batch_size)
model.eval()
all_results = []
logger.info("Start evaluating")
for input_ids, input_mask, segment_ids, example_indices in tqdm(
eval_dataloader,
desc="Evaluating",
disable=args.local_rank not in [-1, 0]):
if len(all_results) % 1000 == 0:
logger.info("Processing example: %d" % (len(all_results)))
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
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")
output_nbest_file = os.path.join(args.output_dir,
"nbest_predictions.json")
output_null_log_odds_file = os.path.join(args.output_dir,
"null_odds.json")
write_predictions(eval_examples, eval_features, all_results,
args.n_best_size, args.max_answer_length,
args.do_lower_case, output_prediction_file,
output_nbest_file, output_null_log_odds_file,
args.verbose_logging, args.version_2_with_negative,
args.null_score_diff_threshold)
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-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--output_dir", default=None, type=str, required=True,
help="The output directory where the model checkpoints and 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("--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", action='store_true', help="Whether to run training.")
parser.add_argument("--do_predict", 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("--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", 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",
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('--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",
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',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--overwrite_output_dir',
action='store_true',
help="Overwrite the content of the output directory")
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('--version_2_with_negative',
action='store_true',
help='If true, the SQuAD examples contain some that do not have an answer.')
parser.add_argument('--null_score_diff_threshold',
type=float, default=0.0,
help="If null_score - best_non_null is greater than the threshold predict null.")
parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
args = parser.parse_args()
print(args)
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
ptvsd.wait_for_attach()
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt = '%m/%d/%Y %H:%M:%S',
level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_predict:
raise ValueError("At least one of `do_train` or `do_predict` must be True.")
if args.do_train:
if not args.train_file:
raise ValueError(
"If `do_train` is True, then `train_file` must be specified.")
if args.do_predict:
if not args.predict_file:
raise ValueError(
"If `do_predict` is True, then `predict_file` must be specified.")
if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
raise ValueError("Output directory () already exists and is not empty.")
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
if args.local_rank not in [-1, 0]:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
model = BertForQuestionAnswering.from_pretrained(args.bert_model)
if args.local_rank == 0:
torch.distributed.barrier()
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,
find_unused_parameters=True)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
if args.do_train:
if args.local_rank in [-1, 0]:
tb_writer = SummaryWriter()
# Prepare data loader
train_examples = read_squad_examples(
input_file=args.train_file, is_training=True, version_2_with_negative=args.version_2_with_negative)
cached_train_features_file = args.train_file+'_{0}_{1}_{2}_{3}'.format(
list(filter(None, args.bert_model.split('/'))).pop(), str(args.max_seq_length), str(args.doc_stride), str(args.max_query_length))
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
except:
train_features = convert_examples_to_features(
examples=train_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=True)
if args.local_rank == -1 or torch.distributed.get_rank() == 0:
logger.info(" Saving train features into cached file %s", cached_train_features_file)
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
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)
num_train_optimization_steps = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
# if args.local_rank != -1:
# num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
# Prepare optimizer
param_optimizer = list(model.named_parameters())
# hack to remove pooler, which is not used
# thus it produce None grad that break apex
param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
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_optimization_steps)
model.train()
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])):
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.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used and handles this automatically
lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
if args.local_rank in [-1, 0]:
tb_writer.add_scalar('lr', optimizer.get_lr()[0], global_step)
tb_writer.add_scalar('loss', loss.item(), global_step)
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
# Save a trained model, configuration and tokenizer
model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
# If we save using the predefined names, we can load using `from_pretrained`
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
model_to_save.config.to_json_file(output_config_file)
tokenizer.save_vocabulary(args.output_dir)
# Load a trained model and vocabulary that you have fine-tuned
model = BertForQuestionAnswering.from_pretrained(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
# Good practice: save your training arguments together with the trained model
output_args_file = os.path.join(args.output_dir, 'training_args.bin')
torch.save(args, output_args_file)
else:
model = BertForQuestionAnswering.from_pretrained(args.bert_model)
model.to(device)
if args.do_predict and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
eval_examples = read_squad_examples(
input_file=args.predict_file, is_training=False, version_2_with_negative=args.version_2_with_negative)
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)
logger.info("***** Running predictions *****")
logger.info(" Num orig examples = %d", len(eval_examples))
logger.info(" Num split examples = %d", len(eval_features))
logger.info(" Batch size = %d", args.predict_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.predict_batch_size)
model.eval()
all_results = []
logger.info("Start evaluating")
for input_ids, input_mask, segment_ids, example_indices in tqdm(eval_dataloader, desc="Evaluating", disable=args.local_rank not in [-1, 0]):
if len(all_results) % 1000 == 0:
logger.info("Processing example: %d" % (len(all_results)))
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
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")
output_nbest_file = os.path.join(args.output_dir, "nbest_predictions.json")
output_null_log_odds_file = os.path.join(args.output_dir, "null_odds.json")
write_predictions(eval_examples, eval_features, all_results,
args.n_best_size, args.max_answer_length,
args.do_lower_case, output_prediction_file,
output_nbest_file, output_null_log_odds_file, args.verbose_logging,
args.version_2_with_negative, args.null_score_diff_threshold)
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-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model checkpoints and 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(
"--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",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_predict",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--do_evaluate",
action='store_true',
help="Whether to eval after training.")
parser.add_argument("--train_batch_size",
default=8,
type=int,
help="Total batch size for training.")
parser.add_argument("--predict_batch_size",
default=32,
type=int,
help="Total batch size for predictions.")
parser.add_argument("--learning_rate",
default=3e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. E.g., 0.1 = 10%% "
"of training.")
parser.add_argument(
"--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",
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("--num_train_samples",
default=-1,
type=int,
help="Total number of training samples used.")
parser.add_argument("--no_cuda",
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(
'--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",
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',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--overwrite_output_dir',
action='store_true',
help="Overwrite the content of the output directory")
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(
'--version_2_with_negative',
action='store_true',
help=
'If true, the SQuAD examples contain some that do not have an answer.')
parser.add_argument(
'--null_score_diff_threshold',
type=float,
default=0.0,
help=
"If null_score - best_non_null is greater than the threshold predict null."
)
parser.add_argument('--server_ip',
type=str,
default='',
help="Can be used for distant debugging.")
parser.add_argument('--server_port',
type=str,
default='',
help="Can be used for distant debugging.")
parser.add_argument('--eval_period', type=int, default=2000)
parser.add_argument('--wait_step', type=int, default=7)
parser.add_argument('--load_from_cache',
action='store_true',
help="Load train features from cache.")
parser.add_argument('--indiv_digits',
action='store_true',
help="Tokenize numbers into indiv digits.")
parser.add_argument('--use_segment_ids',
action='store_true',
help="Use segment ids.")
args = parser.parse_args()
#print(args)
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if 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:
make_output_dir(
args, scripts_to_save=[sys.argv[0], 'run_squad_dataset_utils.py'])
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."
)
with open(args.predict_file, encoding='utf-8') as pf:
dev_data = json.load(pf)["data"]
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir
) and args.do_train and not args.overwrite_output_dir:
raise ValueError(
"Output directory () already exists and is not empty.")
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
if args.local_rank not in [-1, 0]:
torch.distributed.barrier(
) # Make sure only the first process in distributed training will download model & vocab
if args.do_train:
tokenizer = BertTokenizer.from_pretrained(
args.bert_model, do_lower_case=args.do_lower_case)
model = BertForQuestionAnswering.from_pretrained(args.bert_model)
elif args.do_evaluate:
# Load a trained model and vocabulary that you have fine-tuned
model = BertForQuestionAnswering.from_pretrained(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(
args.output_dir, do_lower_case=args.do_lower_case)
if args.local_rank == 0:
torch.distributed.barrier()
model.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
if args.do_train:
if args.local_rank in [-1, 0]:
tb_writer = SummaryWriter(os.path.join(args.output_dir,
'log')) # tensorboard
# Prepare data loader
train_examples = read_squad_examples(
input_file=args.train_file,
is_training=True,
version_2_with_negative=args.version_2_with_negative,
num_train_samples=args.num_train_samples)
cached_train_features_file = args.train_file + '_{0}_{1}_{2}_{3}_{4}'.format(
list(filter(None, args.bert_model.split('/'))).pop(),
str(args.max_seq_length), str(args.doc_stride),
str(args.max_query_length), str(args.num_train_samples))
try:
if args.load_from_cache:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
else:
raise Exception
except:
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,
indiv_digits=args.indiv_digits)
if args.local_rank == -1:
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)
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)
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 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
}]
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 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_optimization_steps)
if args.do_predict and args.local_rank == -1:
eval_examples = read_squad_examples(
input_file=args.predict_file,
is_training=False,
version_2_with_negative=args.version_2_with_negative)
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,
indiv_digits=args.indiv_digits)
logger.info("***** Running predictions *****")
logger.info(" Num orig examples = %d", len(eval_examples))
logger.info(" Num split examples = %d", len(eval_features))
logger.info(" Batch size = %d", args.predict_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_example_index = torch.arange(all_input_ids.size(0),
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_example_index)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.predict_batch_size)
if args.do_train:
best_f1 = 0
wait_step = 0
global_step = 0
stop_training = False
do_eval = False
model.train()
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
if stop_training:
break
for step, batch in enumerate(
tqdm(train_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])):
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
segment_ids = segment_ids if args.use_segment_ids else None
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.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
global_step += 1
if global_step % args.eval_period == 0:
do_eval = True
if do_eval:
do_eval = False
model.eval()
scores = predict(args, model, eval_examples, eval_features,
eval_dataloader, dev_data, device)
em, f1 = scores['exact_match'], scores['f1']
logger.info("f1: %.3f, em: %.3f on epoch=%d" %
(f1, em, epoch))
print("f1: %.3f, em: %.3f on epoch=%d" % (f1, em, epoch))
if best_f1 < f1:
logger.info("Saving model with best f1: %.3f -> %.3f on epoch=%d" % \
(best_f1, f1, epoch))
save_model(args, model, tokenizer)
model.to(device)
best_f1 = f1
wait_step = 0
stop_training = False
else:
wait_step += 1
if best_f1 > 86 and wait_step == args.wait_step:
stop_training = True
break
model.train()
# end of epoch
do_eval = True
elif args.do_evaluate:
model.eval()
scores = predict(args, model, eval_examples, eval_features,
eval_dataloader, dev_data, device, True)
em, f1 = scores['exact_match'], scores['f1']
print("f1: %.3f, em: %.3f" % (f1, em))