def __init__(self, args):
self.config = args.config
if not args.use_pretrain:
if args.progressive_layer_drop:
print("BertConfigPreLnLayerDrop")
from nvidia.modelingpreln_layerdrop import BertForPreTrainingPreLN, BertForMaskedLM, BertConfig
else:
from nvidia.modelingpreln import BertForPreTrainingPreLN, BertForMaskedLM, BertConfig
bert_config = BertConfig(**self.config["bert_model_config"])
bert_config.vocab_size = len(args.tokenizer.vocab)
# Padding for divisibility by 8
if bert_config.vocab_size % 8 != 0:
bert_config.vocab_size += 8 - (bert_config.vocab_size % 8)
print("VOCAB SIZE:", bert_config.vocab_size)
self.network = BertForPreTrainingPreLN(bert_config, args)
# self.network = BertForMaskedLM(bert_config) # something else should be changes for this to work
# Use pretrained bert weights
else:
self.bert_encoder = BertModel.from_pretrained(
self.config['bert_model_file'],
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(args.local_rank))
bert_config = self.bert_encoder.config
self.device = None
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
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("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
# Other parameters
parser.add_argument(
"--cache_dir",
default="",
type=str,
help=
"Where do you want to store the pre-trained models downloaded from s3")
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("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
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("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--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(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--deepscale',
default=False,
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=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('--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("--model_file",
type=str,
default="0",
help="Path to the Pretrained BERT Encoder File.")
parser.add_argument('--random',
default=False,
action='store_true',
help="Whether to fientune for random initialization")
parser.add_argument('--focal',
default=False,
action='store_true',
help="Whether to use Focal Loss for finetuning.")
parser.add_argument('--gamma',
type=float,
default=0.5,
help="Gamma parameter to be used in focal loss.")
parser.add_argument('--deepspeed_sparse_attention',
default=False,
action='store_true',
help='Use DeepSpeed sparse self attention.')
parser.add_argument('--deepspeed_transformer_kernel',
default=False,
action='store_true',
help='Use DeepSpeed transformer kernel to accelerate.')
parser.add_argument(
'--progressive_layer_drop',
default=False,
action='store_true',
help="Whether to enable progressive layer dropping or not")
args = parser.parse_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()
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mnli-mm": MnliMismatchedProcessor,
"mrpc": MrpcProcessor,
"sst-2": Sst2Processor,
"sts-b": StsbProcessor,
"qqp": QqpProcessor,
"qnli": QnliProcessor,
"rte": RteProcessor,
"wnli": WnliProcessor,
}
output_modes = {
"cola": "classification",
"mnli": "classification",
"mrpc": "classification",
"sst-2": "classification",
"sts-b": "regression",
"qqp": "classification",
"qnli": "classification",
"rte": "classification",
"wnli": "classification",
}
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
args.seed = random.randint(1, 1000)
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_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if (args.local_rank == -1 or torch.distributed.get_rank() == 0):
# 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)
if args.local_rank != -1:
torch.distributed.barrier()
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
output_mode = output_modes[task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
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:
train_examples = processor.get_train_examples(args.data_dir)
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(
)
# Prepare model
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(
args.local_rank))
bert_base_model_config = {
"vocab_size_or_config_json_file": 119547,
"hidden_size": 768,
"num_hidden_layers": 12,
"num_attention_heads": 12,
"intermediate_size": 3072,
"hidden_act": "gelu",
"hidden_dropout_prob": 0.1,
"attention_probs_dropout_prob": 0.1,
"max_position_embeddings": 512,
"type_vocab_size": 2,
"initializer_range": 0.02
}
if args.progressive_layer_drop:
print("BertBaseConfigPreLnLayerDrop")
from nvidia.modelingpreln_layerdrop import BertForSequenceClassification, BertConfig
else:
from nvidia.modelingpreln import BertForSequenceClassification, BertConfig
bert_config = BertConfig(**bert_base_model_config)
bert_config.vocab_size = len(tokenizer.vocab)
# Padding for divisibility by 8
if bert_config.vocab_size % 8 != 0:
bert_config.vocab_size += 8 - (bert_config.vocab_size % 8)
model = BertForSequenceClassification(args,
bert_config,
num_labels=num_labels)
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)
# model.bert.load_state_dict(bert_state_dict)
checkpoint_state_dict = torch.load(args.model_file,
map_location=torch.device("cpu"))
if 'module' in checkpoint_state_dict:
logger.info('Loading DeepSpeed v2.0 style checkpoint')
model.load_state_dict(checkpoint_state_dict['module'],
strict=False)
elif 'model_state_dict' in checkpoint_state_dict:
model.load_state_dict(checkpoint_state_dict['model_state_dict'],
strict=False)
else:
raise ValueError("Unable to find model state in checkpoint")
logger.info(f"Pretrained Bert Encoder Loaded from: {args.model_file}")
if args.random:
logger.info("USING RANDOM INITIALISATION FOR FINETUNING")
model.apply(model.init_bert_weights)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
if args.deepscale:
print("Enabling DeepScale")
from deepscale.distributed_apex import DistributedDataParallel as DDP
else:
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
nb_tr_steps = 0
tr_loss = 0
if args.do_train:
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer, output_mode)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_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)
if output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
elif output_mode == "regression":
if args.fp16:
all_label_ids = torch.tensor(
[f.label_id for f in train_features], dtype=torch.half)
else:
all_label_ids = torch.tensor(
[f.label_id for f in train_features], dtype=torch.float)
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 _ 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
# define a new function to compute loss values for both output_modes
logits = model(input_ids, segment_ids, input_mask, labels=None)
if output_mode == "classification":
if args.focal:
loss_fct = FocalLoss(class_num=num_labels,
gamma=args.gamma)
else:
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
elif output_mode == "regression":
loss_fct = MSELoss()
loss = loss_fct(logits.view(-1), label_ids.view(-1))
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.deepscale and args.local_rank != -1:
model.disable_need_reduction()
if (step + 1) % args.gradient_accumulation_steps == 0:
model.enable_need_reduction()
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 args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer, output_mode)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_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)
if output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.float)
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
preds = []
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():
logits = model(input_ids, segment_ids, input_mask, labels=None)
# create eval loss and other metric required by the task
if output_mode == "classification":
if args.focal:
loss_fct = FocalLoss(class_num=num_labels,
gamma=args.gamma)
else:
loss_fct = CrossEntropyLoss()
tmp_eval_loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
elif output_mode == "regression":
loss_fct = MSELoss()
print(logits.type())
print(label_ids.type())
if task_name == "sts-b":
tmp_eval_loss = loss_fct(logits.float().view(-1),
label_ids.view(-1))
else:
tmp_eval_loss = loss_fct(logits.view(-1),
label_ids.view(-1))
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if len(preds) == 0:
preds.append(logits.detach().cpu().numpy())
else:
preds[0] = np.append(preds[0],
logits.detach().cpu().numpy(),
axis=0)
eval_loss = eval_loss / nb_eval_steps
preds = preds[0]
if output_mode == "classification":
preds = np.argmax(preds, axis=1)
elif output_mode == "regression":
preds = np.squeeze(preds)
result = compute_metrics(task_name, preds, all_label_ids.numpy())
loss = tr_loss / nb_tr_steps if args.do_train else None
result['eval_loss'] = eval_loss
result['global_step'] = global_step
result['loss'] = loss
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
# hack for MNLI-MM
if task_name == "mnli":
task_name = "mnli-mm"
processor = processors[task_name]()
if os.path.exists(args.output_dir +
'-MM') and os.listdir(args.output_dir +
'-MM') and args.do_train:
raise ValueError(
"Output directory ({}{}) already exists and is not empty.".
format(args.output_dir, '-MM'))
if not os.path.exists(args.output_dir + '-MM'):
os.makedirs(args.output_dir + '-MM')
eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(
eval_examples, label_list, args.max_seq_length, tokenizer,
output_mode)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_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_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
preds = []
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():
logits = model(input_ids,
segment_ids,
input_mask,
labels=None)
if args.focal:
loss_fct = FocalLoss(class_num=num_labels,
gamma=args.gamma)
else:
loss_fct = CrossEntropyLoss()
tmp_eval_loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if len(preds) == 0:
preds.append(logits.detach().cpu().numpy())
else:
preds[0] = np.append(preds[0],
logits.detach().cpu().numpy(),
axis=0)
eval_loss = eval_loss / nb_eval_steps
preds = preds[0]
preds = np.argmax(preds, axis=1)
result = compute_metrics(task_name, preds, all_label_ids.numpy())
loss = tr_loss / nb_tr_steps if args.do_train else None
result['eval_loss'] = eval_loss
result['global_step'] = global_step
result['loss'] = loss
output_eval_file = os.path.join(args.output_dir + '-MM',
"eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))