def __init__(self, model_dir, top_rnns=False, vocab_size=None, entity_num=0, device='cpu', finetuning=False):
super(Net, self).__init__()
# print(entity_num)
self.bert = BertModel.from_pretrained(model_dir, entity_num=entity_num)
self.hidden_dim = self.bert.config.hidden_size
# print(self.hidden_dim)
self.top_rnns=top_rnns
if top_rnns:
self.rnn = nn.LSTM(bidirectional=True, num_layers=2, input_size=self.hidden_dim, hidden_size=self.hidden_dim//2, batch_first=True)
self.fc = nn.Linear(self.hidden_dim, vocab_size)
self.device = device
self.finetuning = finetuning
self._param_init(self.fc)
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=16,
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(
'--pretrain',
action='store_true',
help="Whether to load a pre-trained model for continuing training")
parser.add_argument('--pretrained_model_file',
type=str,
help="The path of the pretrained_model_file")
parser.add_argument(
"--percent",
default=100,
type=float,
help="The percentage of examples used in the training data.\n")
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 = 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)
train_examples = train_examples[:int(
len(train_examples) * args.percent / 100)]
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
model = ConditionalSemanticBertForQuestionAnswering.from_pretrained(
args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(args.local_rank))
model.QABert = BertModel.from_pretrained(
args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(args.local_rank))
if args.pretrain:
# Load a pre-trained model
print('load a pre-trained model from ' + args.pretrained_model_file)
pretrained_state_dict = torch.load(args.pretrained_model_file)
model_state_dict = model.state_dict()
print(
'QABert', model_state_dict[
'QABert.encoder.layer.4.attention.self.value.weight'])
print(
'bert', model_state_dict[
'bert.encoder.layer.4.attention.self.value.weight'])
print('pretrained_state_dict', pretrained_state_dict.keys())
print('model_state_dict', model_state_dict.keys())
pretrained_state = {
k: v
for k, v in pretrained_state_dict.items() if k in model_state_dict
and v.size() == model_state_dict[k].size()
}
print('stored pretrained dict', pretrained_state.keys())
model_state_dict.update(pretrained_state)
print('updated_state_dict', model_state_dict.keys())
model.load_state_dict(model_state_dict)
model.to(device)
# for param in model.bert.parameters():
# param.requires_grad = False
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())
# 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.local_rank != -1:
t_total = t_total // torch.distributed.get_world_size()
if t_total is None:
t_total = -1
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:
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:
with open('not load features', "rb") as reader:
train_features = pickle.load(reader)
print('not load stored features, but generate itself')
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)
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 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"):
total_loss = 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
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
total_loss += loss
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
print('loss', total_loss)
# 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 = ConditionalSemanticBertForQuestionAnswering.from_pretrained(
args.bert_model, state_dict=model_state_dict)
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)
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)
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(
"--percent",
default=100,
type=int,
help="The percentage of examples used in the training data.\n")
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(
'--pretrain',
action='store_true',
help="Whether to load a pre-trained model for continuing training")
parser.add_argument('--pretrained_model_file',
type=str,
help="The path of the pretrained_model_file")
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.")
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 = {"rc": RCProcessor}
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 not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
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)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[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)
train_examples = train_examples[:int(
len(train_examples) * args.percent / 100)]
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))
model = SemanticBertForREAttBiLSTM.from_pretrained(args.bert_model,
cache_dir=cache_dir,
num_labels=num_labels)
model.cls_bert = BertModel.from_pretrained(args.bert_model,
cache_dir=cache_dir)
if args.pretrain:
# Load a pre-trained model
print('load a pre-trained model from ' + args.pretrained_model_file)
pretrained_state_dict = torch.load(args.pretrained_model_file)
model_state_dict = model.state_dict()
# print('pretrained_state_dict', pretrained_state_dict.keys())
# print('model_state_dict', model_state_dict.keys())
print(
'cls_bert', model_state_dict[
'cls_bert.encoder.layer.4.attention.self.value.weight'])
print(
'bert', model_state_dict[
'bert.encoder.layer.4.attention.self.value.weight'])
pretrained_state = {
k: v
for k, v in pretrained_state_dict.items() if k in model_state_dict
and v.size() == model_state_dict[k].size()
}
print('pretrained_state_dict', pretrained_state.keys())
model_state_dict.update(pretrained_state)
# print('updated_state_dict', model_state_dict.keys())
model.load_state_dict(model_state_dict)
model.to(device)
for param in model.bert.parameters():
param.requires_grad = False
for param in model.hidden_ESRL.parameters():
param.requires_grad = False
for param in model.cls_bert.parameters():
param.requires_grad = False
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)
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:
if num_train_optimization_steps is None:
num_train_optimization_steps = 0
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:
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer)
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_positions_one = torch.tensor(
[f.positions_one for f in train_features], dtype=torch.long)
all_positions_two = torch.tensor(
[f.positions_two for f in train_features], dtype=torch.long)
all_index_one = torch.tensor([f.index_one for f in train_features],
dtype=torch.long)
all_index_two = torch.tensor([f.index_two 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_positions_one,
all_positions_two, all_index_one,
all_index_two, 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, positions_one, positions_two, index_one, index_two, input_mask, segment_ids, label_ids = batch
loss = model(input_ids, positions_one, positions_two,
index_one, index_two, 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
print('train loss', tr_loss)
# Save a trained model and the associated configuration
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())
label_map = {i: label for i, label in enumerate(label_list, 0)}
model_config = {
"bert_model": args.bert_model,
"do_lower": args.do_lower_case,
"max_seq_length": args.max_seq_length,
"num_labels": len(label_list) + 1,
"label_map": label_map
}
json.dump(
model_config,
open(os.path.join(args.output_dir, "model_config.json"), "w"))
# Load a trained model and config that you have fine-tuned
else:
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
model_state_dict = torch.load(output_model_file)
model = SemanticBertForREAttBiLSTM.from_pretrained(
args.bert_model,
state_dict=model_state_dict,
num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
model.to(device)
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_examples = eval_examples[:int(len(eval_examples) * args.percent / 100)]
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer)
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_positions_one = torch.tensor(
[f.positions_one for f in eval_features], dtype=torch.long)
all_positions_two = torch.tensor(
[f.positions_two for f in eval_features], dtype=torch.long)
all_index_one = torch.tensor([f.index_one for f in eval_features],
dtype=torch.long)
all_index_two = torch.tensor([f.index_two 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_positions_one,
all_positions_two, all_index_one,
all_index_two, 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()
y_true = []
y_pred = []
label_map = {i: label for i, label in enumerate(label_list, 0)}
for input_ids, positions_one, positions_two, index_one, index_two, input_mask, segment_ids, label_ids in \
tqdm(eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
positions_one = positions_one.to(device)
positions_two = positions_two.to(device)
index_one = index_one.to(device)
index_two = index_two.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, positions_one, positions_two,
index_one, index_two, segment_ids, input_mask)
# print(F.softmax(logits, dim=1))
logits = torch.argmax(F.softmax(logits, dim=1), dim=1)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
# print('logits', logits)
# print('label ids', label_ids)
for index in range(len(label_ids)):
y_true.append(label_map[label_ids[index]])
y_pred.append(label_map[logits[index]])
error_file = os.path.join(args.output_dir, 'errors.txt')
pred_file = os.path.join(args.output_dir, 'pred.txt')
gold_file = os.path.join(args.output_dir, 'gold.txt')
write_predictions(eval_examples, y_true, y_pred, pred_file, gold_file,
error_file)
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--input_file", default=None, type=str, required=True)
parser.add_argument("--output_file", default=None, type=str, required=True)
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.")
## Other parameters
parser.add_argument("--do_lower_case", action='store_true', help="Set this flag if you are using an uncased model.")
parser.add_argument("--layers", default="-1,-2,-3,-4", type=str)
parser.add_argument("--max_seq_length", default=128, 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("--batch_size", default=32, type=int, help="Batch size for predictions.")
parser.add_argument("--local_rank",
type=int,
default=-1,
help = "local_rank for distributed training on gpus")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
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:
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: {}".format(device, n_gpu, bool(args.local_rank != -1)))
layer_indexes = [int(x) for x in args.layers.split(",")]
tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
examples = read_examples(args.input_file)
features = convert_examples_to_features(
examples=examples, seq_length=args.max_seq_length, tokenizer=tokenizer)
unique_id_to_feature = {}
for feature in features:
unique_id_to_feature[feature.unique_id] = feature
model = BertModel.from_pretrained(args.bert_model)
model.to(device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
output_device=args.local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in 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_example_index)
if args.local_rank == -1:
eval_sampler = SequentialSampler(eval_data)
else:
eval_sampler = DistributedSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.batch_size)
model.eval()
with open(args.output_file, "w", encoding='utf-8') as writer:
for input_ids, input_mask, example_indices in eval_dataloader:
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
all_encoder_layers, _ = model(input_ids, token_type_ids=None, attention_mask=input_mask)
all_encoder_layers = all_encoder_layers
for b, example_index in enumerate(example_indices):
feature = features[example_index.item()]
unique_id = int(feature.unique_id)
# feature = unique_id_to_feature[unique_id]
output_json = collections.OrderedDict()
output_json["linex_index"] = unique_id
all_out_features = []
for (i, token) in enumerate(feature.tokens):
all_layers = []
for (j, layer_index) in enumerate(layer_indexes):
layer_output = all_encoder_layers[int(layer_index)].detach().cpu().numpy()
layer_output = layer_output[b]
layers = collections.OrderedDict()
layers["index"] = layer_index
layers["values"] = [
round(x.item(), 6) for x in layer_output[i]
]
all_layers.append(layers)
out_features = collections.OrderedDict()
out_features["token"] = token
out_features["layers"] = all_layers
all_out_features.append(out_features)
output_json["features"] = all_out_features
writer.write(json.dumps(output_json) + "\n")
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument("--input_dir",
type=str,
required=True)
parser.add_argument("--teacher_model",
default=None,
type=str,
required=True)
parser.add_argument("--student_model",
default=None,
type=str,
required=True)
parser.add_argument("--output_dir",
default=None,
type=str,
required=True)
parser.add_argument('--vocab_file',
type=str,
default=None,
required=True,
help="Vocabulary mapping/file BERT was pretrainined on")
# 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("--reduce_memory",
action="store_true",
help="Store training data as on-disc memmaps to massively reduce memory usage")
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('--weight_decay',
'--wd',
default=1e-4,
type=float, metavar='W',
help='weight decay')
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('--steps_per_epoch',
type=int,
default=-1,
help="Number of updates steps to in one epoch.")
parser.add_argument('--max_steps',
type=int,
default=-1,
help="Number of training steps.")
parser.add_argument('--amp',
action='store_true',
default=False,
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--continue_train',
action='store_true',
default=False,
help='Whether to train from checkpoints')
parser.add_argument('--disable_progress_bar',
default=False,
action='store_true',
help='Disable tqdm progress bar')
parser.add_argument('--max_grad_norm',
type=float,
default=1.,
help="Gradient Clipping threshold")
# Additional arguments
parser.add_argument('--eval_step',
type=int,
default=1000)
# This is used for running on Huawei Cloud.
parser.add_argument('--data_url',
type=str,
default="")
#Distillation specific
parser.add_argument('--value_state_loss',
action='store_true',
default=False)
parser.add_argument('--hidden_state_loss',
action='store_true',
default=False)
parser.add_argument('--use_last_layer',
action='store_true',
default=False)
parser.add_argument('--use_kld',
action='store_true',
default=False)
parser.add_argument('--use_cosine',
action='store_true',
default=False)
parser.add_argument('--distill_config',
default="distillation_config.json",
type=str,
help="path the distillation config")
parser.add_argument('--num_workers',
type=int,
default=4,
help='number of DataLoader worker processes per rank')
args = parser.parse_args()
logger.info('args:{}'.format(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')
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,
stream=sys.stdout)
logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
device, n_gpu, bool(args.local_rank != -1), args.amp))
if args.gradient_accumulation_steps < 1:
raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
args.gradient_accumulation_steps))
# Reference params
author_gbs = 256
author_steps_per_epoch = 22872
author_epochs = 3
author_max_steps = author_steps_per_epoch * author_epochs
# Compute present run params
if args.max_steps == -1 or args.steps_per_epoch == -1:
args.steps_per_epoch = author_steps_per_epoch * author_gbs // (args.train_batch_size * get_world_size() * args.gradient_accumulation_steps)
args.max_steps = author_max_steps * author_gbs // (args.train_batch_size * get_world_size() * args.gradient_accumulation_steps)
#Set seed
set_seed(args.seed, n_gpu)
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) and is_main_process():
os.makedirs(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.teacher_model, do_lower_case=args.do_lower_case)
teacher_model, teacher_config = BertModel.from_pretrained(args.teacher_model,
distill_config=args.distill_config)
# Required to make sure model's fwd doesn't return anything. required for DDP.
# fwd output not being used in loss computation crashes DDP
teacher_model.make_teacher()
if args.continue_train:
student_model, student_config = BertForPreTraining.from_pretrained(args.student_model,
distill_config=args.distill_config)
else:
student_model, student_config = BertForPreTraining.from_scratch(args.student_model,
distill_config=args.distill_config)
# We need a projection layer since teacher.hidden_size != student.hidden_size
use_projection = student_config.hidden_size != teacher_config.hidden_size
if use_projection:
project = Project(student_config, teacher_config)
if args.continue_train:
project_model_file = os.path.join(args.student_model, "project.bin")
project_ckpt = torch.load(project_model_file, map_location="cpu")
project.load_state_dict(project_ckpt)
distill_config = {"nn_module_names": []} #Empty list since we don't want to use nn module hooks here
distill_hooks_student, distill_hooks_teacher = DistillHooks(distill_config), DistillHooks(distill_config)
student_model.register_forward_hook(distill_hooks_student.child_to_main_hook)
teacher_model.register_forward_hook(distill_hooks_teacher.child_to_main_hook)
## Register hooks on nn.Modules
# student_fwd_pre_hook = student_model.register_forward_pre_hook(distill_hooks_student.register_nn_module_hook)
# teacher_fwd_pre_hook = teacher_model.register_forward_pre_hook(distill_hooks_teacher.register_nn_module_hook)
student_model.to(device)
teacher_model.to(device)
if use_projection:
project.to(device)
if args.local_rank != -1:
teacher_model = torch.nn.parallel.DistributedDataParallel(
teacher_model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=False
)
student_model = torch.nn.parallel.DistributedDataParallel(
student_model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=False
)
if use_projection:
project = torch.nn.parallel.DistributedDataParallel(
project, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=False
)
size = 0
for n, p in student_model.named_parameters():
logger.info('n: {}'.format(n))
logger.info('p: {}'.format(p.nelement()))
size += p.nelement()
logger.info('Total parameters: {}'.format(size))
# Prepare optimizer
param_optimizer = list(student_model.named_parameters())
if use_projection:
param_optimizer += list(project.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}
]
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False)
scheduler = LinearWarmUpScheduler(optimizer, warmup=args.warmup_proportion, total_steps=args.max_steps)
global_step = 0
logging.info("***** Running training *****")
logging.info(" Num examples = {}".format(args.train_batch_size * args.max_steps))
logging.info(" Batch size = %d", args.train_batch_size)
logging.info(" Num steps = %d", args.max_steps)
# Prepare the data loader.
if is_main_process():
tic = time.perf_counter()
train_dataloader = lddl.torch.get_bert_pretrain_data_loader(
args.input_dir,
local_rank=args.local_rank,
vocab_file=args.vocab_file,
data_loader_kwargs={
'batch_size': args.train_batch_size * n_gpu,
'num_workers': args.num_workers,
'pin_memory': True,
},
base_seed=args.seed,
log_dir=None if args.output_dir is None else os.path.join(args.output_dir, 'lddl_log'),
log_level=logging.WARNING,
start_epoch=0,
)
if is_main_process():
print('get_bert_pretrain_data_loader took {} s!'.format(time.perf_counter() - tic))
train_dataloader = tqdm(train_dataloader, desc="Iteration", disable=args.disable_progress_bar) if is_main_process() else train_dataloader
tr_loss, tr_att_loss, tr_rep_loss, tr_value_loss = 0., 0., 0., 0.
nb_tr_examples, local_step = 0, 0
student_model.train()
scaler = torch.cuda.amp.GradScaler()
transformer_losses = TransformerLosses(student_config, teacher_config, device, args)
iter_start = time.time()
while global_step < args.max_steps:
for batch in train_dataloader:
if global_step >= args.max_steps:
break
#remove forward_pre_hook after one forward pass
#the purpose of forward_pre_hook is to register
#forward_hooks on nn_module_names provided in config
# if idx == 1:
# student_fwd_pre_hook.remove()
# teacher_fwd_pre_hook.remove()
# # return
# Initialize loss metrics
if global_step % args.steps_per_epoch == 0:
tr_loss, tr_att_loss, tr_rep_loss, tr_value_loss = 0., 0., 0., 0.
mean_loss, mean_att_loss, mean_rep_loss, mean_value_loss = 0., 0., 0., 0.
batch = {k: v.to(device) for k, v in batch.items()}
input_ids, segment_ids, input_mask, lm_label_ids, is_next = batch['input_ids'], batch['token_type_ids'], batch['attention_mask'], batch['labels'], batch['next_sentence_labels']
att_loss = 0.
rep_loss = 0.
value_loss = 0.
with torch.cuda.amp.autocast(enabled=args.amp):
student_model(input_ids, segment_ids, input_mask, None)
# Gather student states extracted by hooks
temp_model = unwrap_ddp(student_model)
student_atts = flatten_states(temp_model.distill_states_dict, "attention_scores")
student_reps = flatten_states(temp_model.distill_states_dict, "hidden_states")
student_values = flatten_states(temp_model.distill_states_dict, "value_states")
student_embeddings = flatten_states(temp_model.distill_states_dict, "embedding_states")
bsz, attn_heads, seq_len, _ = student_atts[0].shape
#No gradient for teacher training
with torch.no_grad():
teacher_model(input_ids, segment_ids, input_mask)
# Gather teacher states extracted by hooks
temp_model = unwrap_ddp(teacher_model)
teacher_atts = [i.detach() for i in flatten_states(temp_model.distill_states_dict, "attention_scores")]
teacher_reps = [i.detach() for i in flatten_states(temp_model.distill_states_dict, "hidden_states")]
teacher_values = [i.detach() for i in flatten_states(temp_model.distill_states_dict, "value_states")]
teacher_embeddings = [i.detach() for i in flatten_states(temp_model.distill_states_dict, "embedding_states")]
teacher_layer_num = len(teacher_atts)
student_layer_num = len(student_atts)
#MiniLM
if student_config.distillation_config["student_teacher_layer_mapping"] == "last_layer":
if student_config.distillation_config["use_attention_scores"]:
student_atts = [student_atts[-1]]
new_teacher_atts = [teacher_atts[-1]]
if student_config.distillation_config["use_value_states"]:
student_values = [student_values[-1]]
new_teacher_values = [teacher_values[-1]]
if student_config.distillation_config["use_hidden_states"]:
new_teacher_reps = [teacher_reps[-1]]
new_student_reps = [student_reps[-1]]
else:
assert teacher_layer_num % student_layer_num == 0
layers_per_block = int(teacher_layer_num / student_layer_num)
if student_config.distillation_config["use_attention_scores"]:
new_teacher_atts = [teacher_atts[i * layers_per_block + layers_per_block - 1]
for i in range(student_layer_num)]
if student_config.distillation_config["use_value_states"]:
new_teacher_values = [teacher_values[i * layers_per_block + layers_per_block - 1]
for i in range(student_layer_num)]
if student_config.distillation_config["use_hidden_states"]:
new_teacher_reps = [teacher_reps[i * layers_per_block + layers_per_block - 1]
for i in range(student_layer_num)]
new_student_reps = student_reps
if student_config.distillation_config["use_attention_scores"]:
att_loss = transformer_losses.compute_loss(student_atts, new_teacher_atts, loss_name="attention_loss")
if student_config.distillation_config["use_hidden_states"]:
if use_projection:
rep_loss = transformer_losses.compute_loss(project(new_student_reps), new_teacher_reps, loss_name="hidden_state_loss")
else:
rep_loss = transformer_losses.compute_loss(new_student_reps, new_teacher_reps, loss_name="hidden_state_loss")
if student_config.distillation_config["use_embedding_states"]:
if use_projection:
rep_loss += transformer_losses.compute_loss(project(student_embeddings), teacher_embeddings, loss_name="embedding_state_loss")
else:
rep_loss += transformer_losses.compute_loss(student_embeddings, teacher_embeddings, loss_name="embedding_state_loss")
if student_config.distillation_config["use_value_states"]:
value_loss = transformer_losses.compute_loss(student_values, new_teacher_values, loss_name="value_state_loss")
loss = att_loss + rep_loss + value_loss
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
tr_att_loss += att_loss.item() / args.gradient_accumulation_steps
if student_config.distillation_config["use_hidden_states"]:
tr_rep_loss += rep_loss.item() / args.gradient_accumulation_steps
if student_config.distillation_config["use_value_states"]:
tr_value_loss += value_loss.item() / args.gradient_accumulation_steps
if args.amp:
scaler.scale(loss).backward()
scaler.unscale_(optimizer)
else:
loss.backward()
if use_projection:
torch.nn.utils.clip_grad_norm_(chain(student_model.parameters(), project.parameters()), args.max_grad_norm, error_if_nonfinite=False)
else:
torch.nn.utils.clip_grad_norm_(student_model.parameters(), args.max_grad_norm, error_if_nonfinite=False)
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
local_step += 1
if local_step % args.gradient_accumulation_steps == 0:
scheduler.step()
if args.amp:
scaler.step(optimizer)
scaler.update()
else:
optimizer.step()
optimizer.zero_grad()
global_step = optimizer.param_groups[0]["step"] if "step" in optimizer.param_groups[0] else 0
if (global_step % args.steps_per_epoch) > 0:
mean_loss = tr_loss / (global_step % args.steps_per_epoch)
mean_att_loss = tr_att_loss / (global_step % args.steps_per_epoch)
mean_rep_loss = tr_rep_loss / (global_step % args.steps_per_epoch)
value_loss = tr_value_loss / (global_step % args.steps_per_epoch)
if (global_step + 1) % args.eval_step == 0 and is_main_process():
result = {}
result['global_step'] = global_step
result['lr'] = optimizer.param_groups[0]["lr"]
result['loss'] = mean_loss
result['att_loss'] = mean_att_loss
result['rep_loss'] = mean_rep_loss
result['value_loss'] = value_loss
result['perf'] = (global_step + 1) * get_world_size() * args.train_batch_size * args.gradient_accumulation_steps / (time.time() - iter_start)
output_eval_file = os.path.join(args.output_dir, "log.txt")
if is_main_process():
with open(output_eval_file, "a") 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])))
# Save a trained model
model_name = "{}".format(WEIGHTS_NAME)
logging.info("** ** * Saving fine-tuned model ** ** * ")
# Only save the model it-self
model_to_save = student_model.module if hasattr(student_model, 'module') else student_model
if use_projection:
project_to_save = project.module if hasattr(project, 'module') else project
output_model_file = os.path.join(args.output_dir, model_name)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
output_project_file = os.path.join(args.output_dir, "project.bin")
torch.save(model_to_save.state_dict(), output_model_file)
if use_projection:
torch.save(project_to_save.state_dict(), output_project_file)
model_to_save.config.to_json_file(output_config_file)
tokenizer.save_vocabulary(args.output_dir)
if oncloud:
logging.info(mox.file.list_directory(args.output_dir, recursive=True))
logging.info(mox.file.list_directory('.', recursive=True))
mox.file.copy_parallel(args.output_dir, args.data_url)
mox.file.copy_parallel('.', args.data_url)
model_name = "{}".format(WEIGHTS_NAME)
logging.info("** ** * Saving fine-tuned model ** ** * ")
model_to_save = student_model.module if hasattr(student_model, 'module') else student_model
if use_projection:
project_to_save = project.module if hasattr(project, 'module') else project
output_project_file = os.path.join(args.output_dir, "project.bin")
if is_main_process():
torch.save(project_to_save.state_dict(), output_project_file)
output_model_file = os.path.join(args.output_dir, model_name)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
if is_main_process():
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)
if oncloud:
logging.info(mox.file.list_directory(args.output_dir, recursive=True))
logging.info(mox.file.list_directory('.', recursive=True))
mox.file.copy_parallel(args.output_dir, args.data_url)
mox.file.copy_parallel('.', args.data_url)
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(
"--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(
'--pretrain',
action='store_true',
help="Whether to load a pre-trained model for continuing training")
parser.add_argument('--pretrained_model_file',
type=str,
help="The path of the pretrained_model_file")
parser.add_argument(
"--percent",
default=100,
type=int,
help="The percentage of examples used in the training data.\n")
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()
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mrpc": MrpcProcessor,
"rte": RteProcessor,
"wnli": WnliProcessor,
'sst-2': Sst2Processor,
}
num_labels_task = {
"cola": 2,
"mnli": 3,
"mrpc": 2,
'rte': 2,
'wnli': 2,
"sst-2": 2,
}
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 = 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_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
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))
os.makedirs(args.output_dir, exist_ok=True)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
num_labels = num_labels_task[task_name]
label_list = processor.get_labels()
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 = processor.get_train_examples(args.data_dir)
train_examples = train_examples[:int(
len(train_examples) * args.percent / 100)]
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
model = ConditionalSemanticBertForSequenceClassification.from_pretrained(
args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(args.local_rank),
num_labels=num_labels)
model.cls_bert = BertModel.from_pretrained(
args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(args.local_rank))
if args.pretrain:
# Load a pre-trained model
print('load a pre-trained model from ' + args.pretrained_model_file)
pretrained_state_dict = torch.load(args.pretrained_model_file)
model_state_dict = model.state_dict()
print(
'cls_bert', model_state_dict[
'cls_bert.encoder.layer.4.attention.self.value.weight'])
print(
'bert', model_state_dict[
'bert.encoder.layer.4.attention.self.value.weight'])
print('pretrained_state_dict', pretrained_state_dict.keys())
print('model_state_dict', model_state_dict.keys())
pretrained_state = {
k: v
for k, v in pretrained_state_dict.items() if k in model_state_dict
and v.size() == model_state_dict[k].size()
}
model_state_dict.update(pretrained_state)
print('updated_state_dict', model_state_dict.keys())
model.load_state_dict(model_state_dict)
model.to(device)
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
}]
t_total = num_train_steps
if args.local_rank != -1:
t_total = t_total // torch.distributed.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
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)
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_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_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 _ 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
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:
# 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 = ConditionalSemanticBertForSequenceClassification.from_pretrained(
args.bert_model, state_dict=model_state_dict, num_labels=num_labels)
model.to(device)
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)
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, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 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():
tmp_eval_loss = model(input_ids, segment_ids, input_mask,
label_ids)
logits = model(input_ids, segment_ids, input_mask)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_eval_accuracy = accuracy(logits, label_ids)
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
loss = tr_loss / nb_tr_steps if args.do_train else None
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'global_step': global_step,
'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])))