def load_model(bison_args, device, data_handler, output_model_file=None):
"""
Load a model.
:param bison_args: instance of :py:class:BisonArguments
:param device: the device to move the model to
:param data_handler: the dataset handler, an instance of :py:class:BitextHandler or a subclass
:param output_model_file: the location of the model to load
:return: the loaded model
"""
model_state_dict = None
if output_model_file is not None:
model_state_dict = torch.load(output_model_file)
if bison_args.bert_model == 'bert-vanilla':
# randomly initialises BERT weights instead of using a pre-trained model
model = BertForMaskedLM(BertConfig.from_default_settings())
else:
model = BertForMaskedLM.from_pretrained(bison_args.bert_model,
state_dict=model_state_dict)
model.to(device)
return model
def main():
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
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
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.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer(vocab_file=args.vocab_file)
train_examples = None
num_train_optimization_steps = None
vocab_list = []
with open(args.vocab_file, 'r') as fr:
for line in fr:
vocab_list.append(line.strip("\n"))
if args.do_train:
train_examples = create_examples(
data_path=args.pretrain_train_path,
max_seq_length=args.max_seq_length,
masked_lm_prob=args.masked_lm_prob,
max_predictions_per_seq=args.max_predictions_per_seq,
vocab_list=vocab_list)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
model = BertForMaskedLM(
config=BertConfig.from_json_file(args.bert_config_json))
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
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)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
best_loss = 100000
if args.do_train:
train_features = convert_examples_to_features(train_examples,
args.max_seq_length,
tokenizer)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
for e in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
# masked_lm_loss
loss = model(input_ids, segment_ids, input_mask, label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
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 that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
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 nb_tr_steps > 0 and nb_tr_steps % 100 == 0:
logger.info(
"===================== -epoch %d -train_step %d -train_loss %.4f\n"
% (e, nb_tr_steps, tr_loss / nb_tr_steps))
if nb_tr_steps > 0 and nb_tr_steps % 2000 == 0:
eval_examples = create_examples(
data_path=args.pretrain_dev_path,
max_seq_length=args.max_seq_length,
masked_lm_prob=args.masked_lm_prob,
max_predictions_per_seq=args.max_predictions_per_seq,
vocab_list=vocab_list)
eval_features = convert_examples_to_features(
eval_examples, args.max_seq_length, tokenizer)
all_input_ids = torch.tensor(
[f.input_ids for f in eval_features], dtype=torch.long)
all_input_mask = torch.tensor(
[f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor(
[f.segment_ids for f in eval_features], dtype=torch.long)
all_label_ids = torch.tensor(
[f.label_id for f in eval_features], dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss = 0
nb_eval_steps = 0
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
loss = model(input_ids, segment_ids, input_mask,
label_ids)
eval_loss += loss.item()
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
if eval_loss < best_loss:
# 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)
torch.save(model_to_save.state_dict(), output_model_file)
best_loss = eval_loss
logger.info(
"============================ -epoch %d -train_loss %.4f -eval_loss %.4f\n"
% (e, tr_loss / nb_tr_steps, eval_loss))
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--bert_model_or_config_file",
default=None,
type=str,
required=True,
help=
"Directory containing pre-trained BERT model or path of configuration file (if no pre-training)."
)
parser.add_argument("--train_file",
default=None,
type=str,
required=True,
help="The input train corpus.")
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(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
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(
"--on_memory",
action='store_true',
help="Whether to load train samples into memory or use disk")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument(
"--num_gpus",
type=int,
default=-1,
help="Num GPUs to use for training (0 for none, -1 for all 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 accumualte before performing a backward/update pass."
)
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")
args = parser.parse_args()
# Check whether bert_model_or_config_file is a file or directory
if os.path.isdir(args.bert_model_or_config_file):
pretrained = True
targets = [WEIGHTS_NAME, CONFIG_NAME, "tokenizer.pkl"]
for t in targets:
path = os.path.join(args.bert_model_or_config_file, t)
if not os.path.exists(path):
msg = "File '{}' not found".format(path)
raise ValueError(msg)
fp = os.path.join(args.bert_model_or_config_file, CONFIG_NAME)
config = BertConfig(fp)
else:
pretrained = False
config = BertConfig(args.bert_model_or_config_file)
# What GPUs do we use?
if args.num_gpus == -1:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
device_ids = None
else:
device = torch.device("cuda" if torch.cuda.is_available()
and args.num_gpus > 0 else "cpu")
n_gpu = args.num_gpus
if n_gpu > 1:
device_ids = list(range(n_gpu))
if args.local_rank != -1:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
# Check some other args
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
if not args.do_train:
raise ValueError(
"Training is currently the only implemented execution option. Please set `do_train`."
)
# Seed RNGs
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)
# Prepare output directory
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
# Make tokenizer
if pretrained:
fp = os.path.join(args.bert_model_or_config_file, "tokenizer.pkl")
with open(fp, "rb") as f:
tokenizer = pickle.load(f)
else:
training_data = [
line.strip() for line in open(args.train_file).readlines()
]
tokenizer = CuneiformCharTokenizer(training_data=training_data)
tokenizer.trim_vocab(config.min_freq)
# Adapt vocab size in config
config.vocab_size = len(tokenizer.vocab)
print("Size of vocab: {}".format(len(tokenizer.vocab)))
# Get training data
num_train_optimization_steps = None
if args.do_train:
print("Loading Train Dataset", args.train_file)
train_dataset = BERTDataset(args.train_file,
tokenizer,
seq_len=args.max_seq_length,
corpus_lines=None,
on_memory=args.on_memory)
num_train_optimization_steps = int(
len(train_dataset) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
if pretrained:
model = BertForMaskedLM.from_pretrained(args.bert_model_or_config_file)
else:
model = BertForMaskedLM(config)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model, device_ids=device_ids)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
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=num_train_optimization_steps)
# Prepare training log
output_log_file = os.path.join(args.output_dir, "training_log.txt")
with open(output_log_file, "w") as f:
f.write("Steps\tTrainLoss\n")
# Start training
global_step = 0
total_tr_steps = 0
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
#TODO: check if this works with current data generator from disk that relies on next(file)
# (it doesn't return item back by index)
train_sampler = DistributedSampler(train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids = batch
loss = model(input_ids, segment_ids, input_mask, lm_label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
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 that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
avg_loss = tr_loss / nb_tr_examples
# Update training log
total_tr_steps += nb_tr_steps
log_data = [str(total_tr_steps), "{:.5f}".format(avg_loss)]
with open(output_log_file, "a") as f:
f.write("\t".join(log_data) + "\n")
# Save model
logger.info("** ** * Saving 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, WEIGHTS_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
fn = os.path.join(args.output_dir, "tokenizer.pkl")
with open(fn, "wb") as f:
pickle.dump(tokenizer, f)
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--train_file",
default=None,
type=str,
required=True,
help="The input train corpus.")
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese."
)
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
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("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument(
"--on_memory",
action='store_true',
help="Whether to load train samples into memory or use disk")
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('--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(
'--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("--hybrid_attention",
action='store_true',
help="Whether to use hybrid attention")
parser.add_argument("--continue_training",
action='store_true',
help="Continue training from a checkpoint")
parser.add_argument("--no_pretrain",
default="",
action='store_true',
help="Whether not to use pretrained model")
parser.add_argument(
"--config_path",
default="",
type=str,
help="Where to load the config file when not using pretrained model")
args = parser.parse_args()
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = 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:
raise ValueError(
"Training is currently the only implemented execution option. Please set `do_train`."
)
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and not args.continue_training:
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
#train_examples = None
num_train_optimization_steps = None
if args.do_train:
print("Loading Train Dataset", args.train_file)
train_dataset = BERTDataset(args.train_file,
tokenizer,
seq_len=args.max_seq_length,
corpus_lines=None,
on_memory=args.on_memory)
num_train_optimization_steps = len(
train_dataset
) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps / torch.distributed.get_world_size(
)
num_train_optimization_steps = math.ceil(num_train_optimization_steps)
if args.no_pretrain:
if not args.config_path:
raise ValueError(
"Config file is needed when not using the pretrained model")
config = BertConfig(args.config_path)
model = BertForMaskedLM(config)
else:
# Prepare model
model = BertForMaskedLM.from_pretrained(args.bert_model)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
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)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
if args.hybrid_attention:
max_seq_length = args.max_seq_length
attention_mask = torch.ones(12,
max_seq_length,
max_seq_length,
dtype=torch.long)
# left attention
attention_mask[:2, :, :] = torch.tril(
torch.ones(max_seq_length, max_seq_length, dtype=torch.long))
# right attention
attention_mask[2:4, :, :] = torch.triu(
torch.ones(max_seq_length, max_seq_length, dtype=torch.long))
# local attention, window size = 3
attention_mask[4:6, :, :] = torch.triu(
torch.tril(
torch.ones(max_seq_length, max_seq_length, dtype=torch.long),
1), -1)
attention_mask = torch.cat(
[attention_mask.unsqueeze(0) for _ in range(8)])
attention_mask = attention_mask.to(device)
else:
attention_mask = None
global_step = 0
epoch_start = 0
if args.do_train:
if args.continue_training:
# if checkpoint file exists, find the last checkpoint
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
all_cp = os.listdir(args.output_dir)
steps = [
int(re.search('_\d+', cp).group()[1:]) for cp in all_cp
if re.search('_\d+', cp)
]
if len(steps) == 0:
raise ValueError(
"No existing checkpoint. Please do not use --continue_training."
)
max_step = max(steps)
# load checkpoint
checkpoint = torch.load(
os.path.join(args.output_dir,
'checkpoints_' + str(max_step) + '.pt'))
logger.info("***** Loading checkpoint *****")
logger.info(" Num steps = %d", checkpoint['global_step'])
logger.info(" Num epoch = %d", checkpoint['epoch'])
logger.info(" Loss = %d, %d", checkpoint['loss'],
checkpoint['loss_now'])
model.module.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
global_step = checkpoint['global_step']
epoch_start = checkpoint['epoch']
del checkpoint
else:
raise ValueError(
"No existing checkpoint. Please do not use --continue_training."
)
writer = SummaryWriter(log_dir=os.environ['HOME'])
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
#TODO: check if this works with current data generator from disk that relies on next(file)
# (it doesn't return item back by index)
train_sampler = DistributedSampler(train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
tr_loss_1000 = 0
for ep in trange(epoch_start, int(args.num_train_epochs),
desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids = batch
loss = model(input_ids,
segment_ids,
input_mask,
lm_label_ids,
hybrid_mask=attention_mask)
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()
tr_loss += loss.item()
tr_loss_1000 += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
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 that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
# log the training loss for every 1000 steps
if global_step % 1000 == 999:
writer.add_scalar('data/loss', tr_loss_1000 / 1000,
global_step)
logger.info("training steps: %s", global_step)
logger.info("training loss per 1000: %s",
tr_loss_1000 / 1000)
tr_loss_1000 = 0
# save the checkpoint for every 10000 steps
if global_step % 10000 == 0:
model_to_save = model.module if hasattr(
model,
'module') else model # Only save the model it-self
output_file = os.path.join(
args.output_dir,
"checkpoints_" + str(global_step) + ".pt")
checkpoint = {
'model': model_to_save.state_dict(),
'optimizer': optimizer.state_dict(),
'epoch': ep,
'global_step': global_step,
'loss': tr_loss / nb_tr_steps,
'loss_now': tr_loss_1000
}
if args.do_train:
torch.save(checkpoint, output_file)
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_" + str(ep))
if args.do_train:
torch.save(model_to_save.state_dict(), output_model_file)
logger.info("training loss: %s", tr_loss / nb_tr_steps)
# Save a trained model
logger.info("** ** * Saving fine - tuned 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)
eval_examples = create_examples(data_path=args.pretrain_dev_path,
max_seq_length=args.max_seq_length,
masked_lm_prob=args.masked_lm_prob,
max_predictions_per_seq=args.max_predictions_per_seq,
vocab_list=vocab_list)
eval_features = convert_examples_to_features(
eval_examples, args.max_seq_length, tokenizer)
all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
model.eval()
eval_loss = 0
nb_eval_steps = 0
# device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
device = torch.device("cpu")
model.to(device)
for input_ids, input_mask, segment_ids, label_ids in tqdm(eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
loss = model(input_ids, segment_ids, input_mask, label_ids)
print(loss)