raise ImportError(
"Please install NVIDIA Apex for distributed and FP16 training")
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.lr,
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.lr,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
trainer = BertTrainer(model, optimizer, processor, args)
if not args.trained_model:
trainer.train()
model = torch.load(trainer.snapshot_path)
else:
model = model = HierarchicalBert(args.model)
model_ = torch.load(args, map_location=lambda storage, loc: storage)
state = {}
for key in model_.state_dict().keys():
new_key = key.replace("module.", "")
state[new_key] = model_.state_dict()[key]
model.load_state_dict(state)
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--train_corpus",
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")
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):
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_corpus)
train_dataset = BERTDataset(args.train_corpus, 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
model = BertForPreTraining.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
if args.do_train:
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)
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
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, is_next = batch
loss = model(input_ids, segment_ids, input_mask, lm_label_ids, is_next)
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.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
# 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, WEIGHTS_NAME)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
if args.do_train:
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)
def do_main():
# Set default configuration in args.py
args = get_args()
if args.local_rank == -1 or not args.cuda:
device = torch.device(
"cuda" if torch.cuda.is_available() and args.cuda else "cpu")
n_gpu = torch.cuda.device_count()
torch.cuda.set_device(args.gpu)
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')
print('Device:', str(device).upper())
print('Number of GPUs:', n_gpu)
print('Distributed training:', bool(args.local_rank != -1))
print('FP16:', args.fp16)
# Set random seed for reproducibility
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)
dataset_map = {
'SST-2': SST2Processor,
'Reuters': ReutersProcessor,
'IMDB': IMDBProcessor,
'AAPD': AAPDProcessor,
'AGNews': AGNewsProcessor,
'Yelp2014': Yelp2014Processor,
'Sogou': SogouProcessor,
'Personality': PersonalityProcessor,
'News_art': News_artProcessor,
'News': News_Processor,
'UCI_yelp': UCI_yelpProcessor,
'Procon': ProconProcessor,
'Style': StyleProcessor,
'ProconDual': ProconDualProcessor,
'Pan15': Pan15_Processor,
'Pan14E': Pan14E_Processor,
'Pan14N': Pan14N_Processor,
'Perspectrum': PerspectrumProcessor
}
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
if args.dataset not in dataset_map:
raise ValueError('Unrecognized dataset')
args.batch_size = args.batch_size // args.gradient_accumulation_steps
args.device = device
args.n_gpu = n_gpu
args.num_labels = dataset_map[args.dataset].NUM_CLASSES
args.is_multilabel = dataset_map[args.dataset].IS_MULTILABEL
if not args.trained_model:
save_path = os.path.join(args.save_path,
dataset_map[args.dataset].NAME)
os.makedirs(save_path, exist_ok=True)
processor = dataset_map[args.dataset]()
args.is_lowercase = 'uncased' in args.model
args.is_hierarchical = False
tokenizer = BertTokenizer.from_pretrained(args.model,
is_lowercase=args.is_lowercase)
train_examples = None
num_train_optimization_steps = None
if not args.trained_model:
train_examples = processor.get_train_examples(args.data_dir,
args.train_name)
num_train_optimization_steps = int(
len(train_examples) / args.batch_size /
args.gradient_accumulation_steps) * args.epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(
args.local_rank))
model = BertForSequenceClassification.from_pretrained(
args.model, cache_dir=cache_dir, num_labels=args.num_labels)
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(
"Install NVIDIA Apex to use distributed and FP16 training.")
model = DDP(model)
'''elif n_gpu > 1: changed by marjan
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 NVIDIA Apex for distributed and FP16 training")
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.lr,
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.lr,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
trainer = BertTrainer(model, optimizer, processor, args)
if not args.trained_model:
trainer.train()
model = torch.load(trainer.snapshot_path)
else:
model = BertForSequenceClassification.from_pretrained(
args.model, num_labels=args.num_labels)
model_ = torch.load(args.trained_model,
map_location=lambda storage, loc: storage)
state = {}
for key in model_.state_dict().keys():
new_key = key.replace("module.", "")
state[new_key] = model_.state_dict()[key]
model.load_state_dict(state)
model = model.to(device)
evaluate_split(model, processor, args, split=args.dev_name)
evaluate_split(model, processor, args, split=args.test_name)
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learn_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=total_train_steps)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learn_rate,
warmup=args.warmup_proportion,
t_total=total_train_steps)
########################################################################################################################
# Training
########################################################################################################################
logger.info("** ** ** Running Training ** ** **")
logger.info(" Num documents = %d", train_iter.num_docs)
logger.info(" Num sentences = %d", train_iter.batch_steps * args.batch_size)
logger.info(" Num steps = %d ", total_train_steps)
# train
def train():
global train_step, train_loss, train_entity_pred, train_entity_right, log_start_time, \
def do_main():
# Set default configuration in args.py
args = get_args()
if args.local_rank == -1 or not args.cuda:
device = torch.device(
"cuda" if torch.cuda.is_available() and args.cuda else "cpu")
n_gpu = torch.cuda.device_count()
torch.cuda.set_device(args.gpu)
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')
print('Device:', str(device).upper())
print('Number of GPUs:', n_gpu)
print('Distributed training:', bool(args.local_rank != -1))
print('FP16:', args.fp16)
# Set random seed for reproducibility
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)
dataset_map = {'News_art': News_artProcessor, 'News': News_Processor}
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
if args.dataset not in dataset_map:
raise ValueError('Unrecognized dataset')
args.batch_size = args.batch_size // args.gradient_accumulation_steps
args.device = device
args.n_gpu = n_gpu
args.num_labels = dataset_map[args.dataset].NUM_CLASSES
args.is_multilabel = dataset_map[args.dataset].IS_MULTILABEL
if not args.trained_model:
save_path = os.path.join(args.save_path,
dataset_map[args.dataset].NAME)
os.makedirs(save_path, exist_ok=True)
processor = dataset_map[args.dataset]()
args.is_lowercase = 'uncased' in args.model
args.is_hierarchical = False
tokenizer = BertTokenizer.from_pretrained(args.model,
is_lowercase=args.is_lowercase)
train_examples = None
num_train_optimization_steps = None
if args.trained_model:
train_examples = processor.get_train_examples(args.data_dir,
args.train_name)
num_train_optimization_steps = int(
math.ceil(len(train_examples) / args.batch_size) /
args.gradient_accumulation_steps) * args.epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(
args.local_rank))
model = BertForSequenceClassification.from_pretrained(
args.model, num_labels=2) # creating news model!
#model = BertForSequenceClassification.from_pretrained(args.model, cache_dir=cache_dir, num_labels=args.num_labels)
if args.fp16:
model.half()
model.to(device)
#model = BertForSequenceClassification.from_pretrained(args.model, num_labels=args.num_labels)
model_ = torch.load(
args.trained_model,
map_location=lambda storage, loc: storage) # load personality model
state = {}
for key in model_.state_dict().keys():
new_key = key.replace("module.", "")
state[new_key] = model_.state_dict()[key]
del state['classifier.weight'] # removing personality classifier!
del state['classifier.bias']
model.load_state_dict(state, strict=False)
model = model.to(device)
# 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
}]
print('t_total :', num_train_optimization_steps)
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.lr,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
args.freez_bert = False
trainer = BertTrainer(model, optimizer, processor, args)
trainer.train()
model = torch.load(trainer.snapshot_path)
evaluate_split(model, processor, args, split=args.dev_name)
evaluate_split(model, processor, args, split=args.test_name)