def __init__(self, job_config, use_pretrain, tokenizer, cache_dir, device,
write_log, summary_writer):
self.job_config = job_config
if not use_pretrain:
model_config = self.job_config.get_model_config()
bert_config = BertConfig(**model_config)
bert_config.vocab_size = len(tokenizer.vocab)
self.bert_encoder = BertModel(bert_config)
# Use pretrained bert weights
else:
self.bert_encoder = BertModel.from_pretrained(
self.job_config.get_model_file_type(), cache_dir=cache_dir)
bert_config = self.bert_encoder.config
self.network = MTLRouting(self.bert_encoder,
write_log=write_log,
summary_writer=summary_writer)
#config_data=self.config['data']
# Pretrain Dataset
self.network.register_batch(BatchType.PRETRAIN_BATCH,
"pretrain_dataset",
loss_calculation=BertPretrainingLoss(
self.bert_encoder, bert_config))
self.device = device
def main():
parser = make_arg_parser()
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 = {
"cli": CLIProcessor,
}
num_labels_task = {
"cli": 7,
}
# Check whether bert_model_or_config_file is a file or directory
if os.path.isdir(args.bert_model_or_config_file):
pretrained = True
targets = [WEIGHTS_NAME, CONFIG_NAME, "tokenizer.pkl"]
for t in targets:
path = os.path.join(args.bert_model_or_config_file, t)
if not os.path.exists(path):
msg = "File '{}' not found".format(path)
raise ValueError(msg)
fp = os.path.join(args.bert_model_or_config_file, CONFIG_NAME)
config = BertConfig(fp)
else:
pretrained = False
config = BertConfig(args.bert_model_or_config_file)
# What GPUs do we use?
if args.num_gpus == -1:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
device_ids = None
else:
device = torch.device("cuda" if torch.cuda.is_available()
and args.num_gpus > 0 else "cpu")
n_gpu = args.num_gpus
if n_gpu > 1:
device_ids = list(range(n_gpu))
if args.local_rank != -1:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
# Check some other args
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
if not args.do_train and not args.do_eval and not args.do_predict:
raise ValueError(
"At least one of `do_train`, `do_eval` or `do_predict` must be True."
)
# Seed RNGs
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
# Prepare output directory
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) 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]()
num_labels = num_labels_task[task_name]
label_list = processor.get_labels()
# Get training data
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(
)
# Make tokenizer
if pretrained:
fp = os.path.join(args.bert_model_or_config_file, "tokenizer.pkl")
with open(fp, "rb") as f:
tokenizer = pickle.load(f)
else:
tokenizer = CuneiformCharTokenizer(
training_data=[x.text_a for x in train_examples])
tokenizer.trim_vocab(config.min_freq)
# Adapt vocab size in config
config.vocab_size = len(tokenizer.vocab)
print("Size of vocab: {}".format(len(tokenizer.vocab)))
# Prepare model
if pretrained:
model = BertForSequenceClassification.from_pretrained(
args.bert_model_or_config_file, num_labels=num_labels)
else:
model = BertForSequenceClassification(config, num_labels=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(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model, device_ids=device_ids)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
# Get dev data
if args.do_eval:
eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
# Prepare for training
global_step = 0
nb_tr_steps = 0
total_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_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)
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)
# Prepare log file
output_log_file = os.path.join(args.output_dir, "training_log.txt")
with open(output_log_file, "w") as f:
if args.do_eval:
f.write("Steps\tTrainLoss\tValLoss\tValAccuracy\tValFScore\n")
else:
f.write("Steps\tTrainLoss\n")
best_val_score = float("-inf")
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:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
avg_loss = tr_loss / nb_tr_examples
total_tr_steps += nb_tr_steps
log_data = [str(total_tr_steps), "{:.5f}".format(avg_loss)]
# Validate
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
predictions, eval_loss, eval_accuracy, fscore = evaluate(
model, eval_dataloader, device)
log_data.append("{:.5f}".format(eval_loss))
log_data.append("{:.5f}".format(eval_accuracy))
log_data.append("{:.5f}".format(fscore))
# Check if score has improved
if fscore > best_val_score:
best_val_score = fscore
save_model(model, tokenizer, args.output_dir)
else:
# If we can't validate, we save model at each epoch
save_model(model, tokenizer, args.output_dir)
# Log
with open(output_log_file, "a") as f:
f.write("\t".join(log_data) + "\n")
# Load model
if args.do_train:
# Load model we just fine-tuned
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
output_tokenizer_file = os.path.join(args.output_dir, "tokenizer.pkl")
config = BertConfig(output_config_file)
model = BertForSequenceClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
with open(output_tokenizer_file, "rb") as f:
tokenizer = pickle.load(f)
else:
# Load a model you fine-tuned previously
model = BertForSequenceClassification.from_pretrained(
args.bert_model_or_config_file, num_labels=num_labels)
model.to(device)
# Evaluate model on validation data
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
predictions, eval_loss, eval_accuracy, fscore = evaluate(
model, eval_dataloader, device)
loss = avg_loss if args.do_train else None
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'eval_fscore': fscore,
'global_step': global_step,
'loss': loss
}
# Write evaluation results
output_eval_file = os.path.join(args.output_dir, "dev_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])))
# Write predictions
output_pred_file = os.path.join(args.output_dir, "dev_pred.txt")
with open(output_pred_file, "w", encoding="utf-8") as writer:
for label_id in predictions:
label = label_list[label_id]
writer.write(label + "\n")
# Predict labels of test set
if args.do_predict:
test_examples = processor.get_test_examples(args.data_dir)
test_features = convert_examples_to_features(test_examples, label_list,
args.max_seq_length,
tokenizer)
all_input_ids = torch.tensor([f.input_ids for f in test_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in test_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in test_features],
dtype=torch.long)
test_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids)
test_sampler = SequentialSampler(test_data)
test_dataloader = DataLoader(test_data,
sampler=test_sampler,
batch_size=args.eval_batch_size)
logger.info("***** Running prediction *****")
logger.info(" Num examples = %d", len(test_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
predictions = predict(model, test_dataloader, device)
# Write predictions
output_pred_file = os.path.join(args.output_dir, "test_pred.txt")
with open(output_pred_file, "w", encoding="utf-8") as writer:
for label_id in predictions:
label = label_list[label_id]
writer.write(label + "\n")
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--model",
default="bert",
type=str,
required=True,
help="The model used for pretraining. Currently support bert or electra"
)
parser.add_argument(
"--config_file",
"--cf",
help="pointer to the configuration file of the experiment",
type=str,
required=True)
parser.add_argument(
"--config_file_path",
default=None,
type=str,
required=True,
help="The blob storage directory where config file is located.")
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("--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 checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--checkpoint_file",
default=None,
type=str,
help=
"The path to checkpoint file which will be used to initializ the model 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",
default=False,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
default=False,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
default=False,
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",
default=False,
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(
'--optimize_on_cpu',
default=False,
action='store_true',
help=
"Whether to perform optimization and keep the optimizer averages on CPU"
)
parser.add_argument(
'--fp16',
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=128,
help=
'Loss scaling, positive power of 2 values can improve fp16 convergence.'
)
parser.add_argument('--step_per_log',
type=int,
default=5,
help='Number of updates steps to log metrics.')
parser.add_argument(
"--process_count_per_node",
default=1,
type=int,
help="Total number of process count to launch per node.")
args = parser.parse_args()
#run = Run.get_context()
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mrpc": MrpcProcessor,
"qqp": QQPProcessor,
"qnli": QNLIProcessor,
"sst2": SST2Processor,
"stsb": STSBProcessor,
"rte": RTEProcessor,
}
comm = DistributedCommunicator(
accumulation_step=args.gradient_accumulation_steps)
rank = comm.rank
local_rank = comm.local_rank
world_size = comm.world_size
is_master = rank == 0
# Prepare logger
job_id = rutils.get_current_time()
logger = rutils.FileLogging('%s_bert_fine_tune_%d' % (job_id, local_rank))
logger.info("job id: %s" % job_id)
logger.info(rutils.parser_args_to_dict(args))
logger.info(
"world size: {}, local rank: {}, global rank: {}, fp16: {}".format(
world_size, local_rank, rank, args.fp16))
torch.cuda.set_device(local_rank)
device = torch.device("cuda", local_rank)
hostname = socket.gethostname()
n_gpu = torch.cuda.device_count()
logger.info("host: {}, device: {}, n_gpu: {}".format(
hostname, device, n_gpu))
# extract config
job_config = BertJobConfiguration(
config_file_path=os.path.join(args.config_file_path, args.config_file))
#if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
# raise ValueError("Output directory () already exists and is not empty.")
#os.makedirs(args.output_dir, exist_ok=True)
output_model_file = os.path.join(args.output_dir,
job_id + "_pytorch_model_fine_tune.bin")
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
if local_rank == -1:
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 not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
task_name = args.task_name.lower()
is_master = (local_rank == -1 or rank == 0)
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
label_list = processor.get_labels()
tokenizer = BertTokenizer.from_pretrained(job_config.get_token_file_type(),
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)
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
num_labels = len(processor.get_labels())
# Prepare model
model_name = args.model
model_config = job_config.get_model_config()
if model_name == 'bert':
config = BertConfig(**model_config)
config.vocab_size = len(tokenizer.vocab)
model = BertForSequenceClassification(config, num_labels=num_labels)
elif model_name == 'electra':
config = ElectraConfig(**model_config)
config.vocab_size = len(tokenizer.vocab)
model = ElectraForSequenceClassification(config, num_labels=num_labels)
#model = BertForSequenceClassification.from_pretrained(args.bert_model,
# cache_dir=PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.format(local_rank), num_labels=num_labels)
# Load checkpoint if specified
#import pdb;pdb.set_trace()
if os.path.exists(str(args.checkpoint_file)):
state_dict = torch.load(args.checkpoint_file)
if model_name == 'bert':
model.bert.load_state_dict(state_dict)
elif model_name == 'electra':
model.electra.load_state_dict(state_dict)
logger.info("Set the model parameter from the checkpoint %s" %
args.checkpoint_file)
if args.fp16:
model.half()
model.to(device)
comm.register_model(model, args.fp16)
if args.do_train:
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 // 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 run this."
)
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)
if is_master:
logger.info('lr: {}'.format(np.float(args.learning_rate)))
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer, logger)
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 local_rank != -1 and world_size > 1:
train_sampler = DistributedSampler(train_data)
else:
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
global_step, tr_loss = 0, 0
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
for _, 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)
loss = loss / args.gradient_accumulation_steps
loss.backward()
global_step += 1
tr_loss += loss.item()
if comm.synchronize():
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()
model.zero_grad()
if is_master and (global_step + 1) % args.step_per_log == 0:
logger.info('train_loss: {}'.format(
np.float(tr_loss / args.step_per_log)))
tr_loss = 0
if is_master:
# Save a trained model
torch.save(model.state_dict(), output_model_file)
logger.info('model checkpoint saved at %s' % output_model_file)
if args.do_eval and is_master:
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)
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 eval_dataloader:
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
result = {'eval_loss': eval_loss, 'eval_accuracy': eval_accuracy}
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s" % (key, str(result[key])))
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--bert_model_or_config_file",
default=None,
type=str,
required=True,
help=
"Directory containing pre-trained BERT model or path of configuration file (if no pre-training)."
)
parser.add_argument("--train_file",
default=None,
type=str,
required=True,
help="The input train corpus.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--learning_rate",
default=3e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument(
"--on_memory",
action='store_true',
help="Whether to load train samples into memory or use disk")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument(
"--num_gpus",
type=int,
default=-1,
help="Num GPUs to use for training (0 for none, -1 for all available)")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumualte before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
args = parser.parse_args()
# Check whether bert_model_or_config_file is a file or directory
if os.path.isdir(args.bert_model_or_config_file):
pretrained = True
targets = [WEIGHTS_NAME, CONFIG_NAME, "tokenizer.pkl"]
for t in targets:
path = os.path.join(args.bert_model_or_config_file, t)
if not os.path.exists(path):
msg = "File '{}' not found".format(path)
raise ValueError(msg)
fp = os.path.join(args.bert_model_or_config_file, CONFIG_NAME)
config = BertConfig(fp)
else:
pretrained = False
config = BertConfig(args.bert_model_or_config_file)
# What GPUs do we use?
if args.num_gpus == -1:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
device_ids = None
else:
device = torch.device("cuda" if torch.cuda.is_available()
and args.num_gpus > 0 else "cpu")
n_gpu = args.num_gpus
if n_gpu > 1:
device_ids = list(range(n_gpu))
if args.local_rank != -1:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
# Check some other args
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
if not args.do_train:
raise ValueError(
"Training is currently the only implemented execution option. Please set `do_train`."
)
# Seed RNGs
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
# Prepare output directory
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
# Make tokenizer
if pretrained:
fp = os.path.join(args.bert_model_or_config_file, "tokenizer.pkl")
with open(fp, "rb") as f:
tokenizer = pickle.load(f)
else:
training_data = [
line.strip() for line in open(args.train_file).readlines()
]
tokenizer = CuneiformCharTokenizer(training_data=training_data)
tokenizer.trim_vocab(config.min_freq)
# Adapt vocab size in config
config.vocab_size = len(tokenizer.vocab)
print("Size of vocab: {}".format(len(tokenizer.vocab)))
# Get training data
num_train_optimization_steps = None
if args.do_train:
print("Loading Train Dataset", args.train_file)
train_dataset = BERTDataset(args.train_file,
tokenizer,
seq_len=args.max_seq_length,
corpus_lines=None,
on_memory=args.on_memory)
num_train_optimization_steps = int(
len(train_dataset) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
if pretrained:
model = BertForMaskedLM.from_pretrained(args.bert_model_or_config_file)
else:
model = BertForMaskedLM(config)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model, device_ids=device_ids)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
# Prepare training log
output_log_file = os.path.join(args.output_dir, "training_log.txt")
with open(output_log_file, "w") as f:
f.write("Steps\tTrainLoss\n")
# Start training
global_step = 0
total_tr_steps = 0
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
#TODO: check if this works with current data generator from disk that relies on next(file)
# (it doesn't return item back by index)
train_sampler = DistributedSampler(train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids = batch
loss = model(input_ids, segment_ids, input_mask, lm_label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
avg_loss = tr_loss / nb_tr_examples
# Update training log
total_tr_steps += nb_tr_steps
log_data = [str(total_tr_steps), "{:.5f}".format(avg_loss)]
with open(output_log_file, "a") as f:
f.write("\t".join(log_data) + "\n")
# Save model
logger.info("** ** * Saving model ** ** * ")
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
fn = os.path.join(args.output_dir, "tokenizer.pkl")
with open(fn, "wb") as f:
pickle.dump(tokenizer, f)
