def configure_nccl_settings_from_env():
# Start by collecting parallel env info
global_size = get_global_size()
local_size = get_local_size()
print("local_size = {}".format(local_size))
print("global_size = {}".format(global_size))
set_environment_variables_for_nccl_backend(True, master_port=6105)
#global_size = 1
#local_size = 1
# TODO use logger
print('local_rank = {}'.format(local_rank))
print('global_size = {}'.format(global_size))
print('local_size = {}'.format(local_size))
import socket
os.environ['MASTER_ADDR'] = socket.gethostname()
os.environ['MASTER_PORT'] = '54965'
dist.init_process_group(
backend='nccl',
#init_method='/gpfs/gpfs0/groups/mozafari/ruixliu/tmp/misc/sharedfile',
rank=local_rank,
world_size=local_size)
set_environment_variables_for_nccl_backend(local_size == global_size)
# Prepare Logger
job_id = rutils.get_current_time()
logger = rutils.FileLogging('%s_bert_pretrain_%d' % (job_id, local_rank))
#logger = Logger(cuda=torch.cuda.is_available())
logger.info('job id: %s' % job_id)
logger.info(rutils.parser_args_to_dict(args))
# # Extact config file from blob storage
job_config = BertJobConfiguration(
config_file_path=os.path.join(args.config_file_path, config_file))
logger.info(job_config.config)
job_name = job_config.get_name()
# Setting the distributed variables
seed = args.seed
loss_scale = args.loss_scale
load_training_checkpoint = args.load_training_checkpoint
max_seq_length = args.max_seq_length
max_predictions_per_seq = args.max_predictions_per_seq
masked_lm_prob = args.masked_lm_prob
addr = args.master_addr
master_port = args.master_port
local_rank = args.local_rank
nprocs = args.nprocs = torch.cuda.device_count()
world_size = args.nprocs * args.nodes
train_batch_size = args.train_batch_size
if nprocs > train_batch_size:
train_batch_size = nprocs
set_environment_variables_for_nccl_backend(addr, master_port, local_rank,
world_size)
# Prepare Logger
logger = Logger(cuda=torch.cuda.is_available())
# # Extact config file from blob storage
job_config = BertJobConfiguration(
config_file_path=os.path.join(args.config_file_path, config_file))
job_name = job_config.get_name()
# Setting the distributed variables
# local_rank = 0
if not use_multigpu_with_single_device_per_process:
device = torch.device("cuda")
n_gpu = torch.cuda.device_count()
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."
)
parser.add_argument(
"--model_file_location",
default=None,
type=str,
required=True,
help=
"The directory in the blob storage which contains data and config files."
)
## 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(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument("--model_file",
type=str,
default="0",
help="Path to the Pretrained BERT Encoder File.")
args = parser.parse_args()
global_size = get_global_size()
local_size = get_local_size()
set_environment_variables_for_nccl_backend(local_size == global_size)
local_rank = get_local_rank()
is_distributed = bool(local_rank != -1)
args.model_file = os.path.join(args.model_file_location, args.model_file)
if not is_distributed 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(local_rank)
device = torch.device("cuda", 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(is_distributed), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_predict:
raise ValueError(
"At least one of `do_train` or `do_predict` must be True.")
if args.do_train:
if not args.train_file:
raise ValueError(
"If `do_train` is True, then `train_file` must be specified.")
if args.do_predict:
if not args.predict_file:
raise ValueError(
"If `do_predict` is True, then `predict_file` must be specified."
)
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
raise ValueError(
"Output directory () already exists and is not empty.")
os.makedirs(args.output_dir, exist_ok=True)
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 = 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.from_pretrained(
args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(local_rank))
if args.model_file is not "0":
logger.info(f"Loading Pretrained Bert Encoder from: {args.model_file}")
bert_state_dict = torch.load(args.model_file,
map_location=torch.device("cpu"))
model.bert.load_state_dict(bert_state_dict)
logger.info(f"Pretrained Bert Encoder Loaded from: {args.model_file}")
if args.fp16:
model.half()
model.to(device)
if is_distributed:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
# hack to remove pooler, which is not used
# thus it produce None grad that break apex
param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
t_total = num_train_steps
if args.do_train and local_rank != -1:
t_total = t_total // get_world_size()
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
global_step = 0
if args.do_train:
if (not is_distributed or torch.distributed.get_rank() == 0):
run.log('lr', np.float(args.learning_rate))
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))
train_features = None
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
except:
train_features = convert_examples_to_features(
examples=train_examples,
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 local_rank == -1 or torch.distributed.get_rank() == 0:
logger.info(" Saving train features into cached file %s",
cached_train_features_file)
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
logger.info("***** Running training *****")
logger.info(" Num orig examples = %d", len(train_examples))
logger.info(" Num split examples = %d", len(train_features))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_start_positions = torch.tensor(
[f.start_position for f in train_features], dtype=torch.long)
all_end_positions = torch.tensor(
[f.end_position for f in train_features], dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_start_positions,
all_end_positions)
if local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration", smoothing=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:
# modify learning rate with special warm up BERT uses
lr_this_step = args.learning_rate * warmup_linear(
global_step / t_total, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
# Save a trained model
#model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
#output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
#if args.do_train:
# torch.save(model_to_save.state_dict(), output_model_file)
# Load a trained model that you have fine-tuned
#model_state_dict = torch.load(output_model_file)
#model = BertForQuestionAnswering.from_pretrained(args.bert_model, state_dict=model_state_dict)
#model.to(device)
if args.do_predict and (local_rank == -1
or torch.distributed.get_rank() == 0):
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)
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"):
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")
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)
with open(args.predict_file) as predict_file:
dataset_json = json.load(predict_file)
dataset = dataset_json['data']
with open(output_prediction_file) as prediction_file:
predictions = json.load(prediction_file)
result = evaluate(dataset, predictions)
for key in result.keys():
logger.info(" %s = %s", key, str(result[key]))
run.log(key, result[key])
max_seq_length = args.max_seq_length
max_predictions_per_seq = args.max_predictions_per_seq
masked_lm_prob = args.masked_lm_prob
master_port = args.master_port
local_rank = -1
local_rank = get_local_rank()
global_size = get_global_size()
local_size = get_local_size()
# TODO use logger
print('local_rank = {}'.format(local_rank))
print('global_size = {}'.format(global_size))
print('local_size = {}'.format(local_size))
set_environment_variables_for_nccl_backend(local_size == global_size,
master_port)
# Prepare Logger
logger = Logger(cuda=torch.cuda.is_available())
# # Extact config file from blob storage
job_config = BertJobConfiguration(
config_file_path=os.path.join(args.config_file_path, config_file))
job_name = job_config.get_name()
# Setting the distributed variables
run = Run.get_context()
if not use_multigpu_with_single_device_per_process:
device = torch.device("cuda")
