def prepare_model(args, device):
# Prepare model
config = BertConfig.from_json_file(args.bert_config_path)
# Padding for divisibility by 8
if config.vocab_size % 8 != 0:
config.vocab_size += 8 - (config.vocab_size % 8)
print('padded vocab size to: {}'.format(config.vocab_size))
# Set some options that the config file is expected to have (but don't need to be set properly
# at this point)
config.pad = False
config.unpad = False
config.dense_seq_output = False
config.fused_mha = False
config.fused_gelu_bias = False
config.fuse_qkv = False
config.fuse_scale = False
config.fuse_mask = False
config.fuse_dropout = False
config.apex_softmax = False
config.enable_stream = False
if config.fuse_mask == True: config.apex_softmax = True
if config.pad == False: config.enable_stream = True
if config.unpad == True: config.fused_mha = False
#Load from TF checkpoint
model = BertForPreTraining.from_pretrained(args.tf_checkpoint,
from_tf=True,
config=config)
return model
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--train_file",
default=None,
type=str,
required=True,
help="The input train corpus.")
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese."
)
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
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("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=3e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument(
"--on_memory",
action='store_true',
help="Whether to load train samples into memory or use disk")
parser.add_argument(
"--do_lower_case",
action='store_true',
help=
"Whether to lower case the input text. True for uncased models, False for cased models."
)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumualte before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
args = parser.parse_args()
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train 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):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
#train_examples = None
num_train_optimization_steps = None
if args.do_train:
print("Loading Train Dataset", args.train_file)
train_dataset = BERTDataset(args.train_file,
tokenizer,
seq_len=args.max_seq_length,
corpus_lines=None,
on_memory=args.on_memory)
num_train_optimization_steps = int(
len(train_dataset) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
model = BertForPreTraining.from_pretrained(args.bert_model)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
#TODO: check if this works with current data generator from disk that relies on next(file)
# (it doesn't return item back by index)
train_sampler = DistributedSampler(train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids, is_next = batch
loss = model(input_ids, segment_ids, input_mask, lm_label_ids,
is_next)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
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
# Save a trained model
logger.info("** ** * Saving fine - tuned model ** ** * ")
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
if args.do_train:
torch.save(model_to_save.state_dict(), output_model_file)
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(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--bert_model",
default='bert-base-multilingual-cased',
type=str,
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(
"--max_seq_length",
default=384,
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("--train_batch_size",
default=2,
type=int,
help="Total batch size for training.")
# parser.add_argument("--eval_batch_size",
# default=2,
# type=int,
# help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=3e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on GPUs")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumualte before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument('--visdom',
action='store_true',
help='Use visdom for loss visualization')
parser.add_argument('--check_saved_model',
action='store_true',
help='Use visdom for loss visualization')
parser.add_argument('--last_final_epoch',
type=int,
default=-1,
help="저번에 이미 최종 학습을 했고, 이에 이어서 트레이닝을 원할때 사용,\n"
"기존에 train_epoch를 3으로 세팅했다면, 2가 아닌 3을 입력하세요.")
args = parser.parse_args()
print(args)
if args.visdom:
import visdom
viz = visdom.Visdom()
# visdom을 통해서 loss를 시각화
os.makedirs(args.output_dir, exist_ok=True)
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train:
raise ValueError(
"Training is currently the only implemented execution option. Please set `do_train`."
)
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=False)
processor = DataProcessor()
label_list = processor.get_labels()
num_train_optimization_steps = None
if args.do_train:
print("Loading Train Dataset", args.data_dir)
train_examples = processor.get_train_examples(args.data_dir)
train_dataset = LazyDataset(train_examples, args.max_seq_length,
tokenizer)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
train_sampler = DistributedSampler(train_dataset)
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
loaded_epoch = -1
saved_model_path = -1
if args.last_final_epoch != -1:
last_model = os.path.join(args.output_dir, WEIGHTS_NAME)
if os.path.exists(last_model):
saved_model_path = last_model
loaded_epoch = args.last_final_epoch - 1
elif args.check_saved_model:
for epoch in range(int(args.num_train_epochs)):
tmp = os.path.join(args.output_dir,
(f"weight_on_ep{epoch}_" + WEIGHTS_NAME))
if os.path.exists(tmp):
saved_model_path = tmp
loaded_epoch = epoch
if saved_model_path != -1:
logger.info(f"Loading on saved model {saved_model_path}")
config_file = os.path.join(args.output_dir, CONFIG_NAME)
config = BertConfig(config_file)
logger.info("Model config {}".format(config))
model = BertForPreTraining(config)
model.load_state_dict(torch.load(saved_model_path))
else:
loaded_epoch = -1
model = BertForPreTraining.from_pretrained(args.bert_model)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
if args.visdom:
# 일단 visdom 기본 figure를 정의
vis_title = f'Baseline on {len(train_dataset)} dataset'
vis_legend = ['LM Loss', 'Click Loss', 'Total Loss']
iter_plot = create_vis_plot(viz, 'Iteration', 'Loss', vis_title,
vis_legend)
epoch_plot = create_vis_plot(viz, 'Epoch', 'Loss', vis_title,
vis_legend)
# if args.do_eval:
# eval_examples = processor.get_dev_examples(args.data_dir)
#
# logger.info("***** Running evaluation *****")
# logger.info(" Num examples = %d", len(eval_examples))
# logger.info(" Batch size = %d", args.eval_batch_size)
#
# eval_data = LazyDatasetClassifier(eval_examples, label_list, args.max_seq_length, tokenizer)
# # Run prediction for full data
# """
# cur_tensors = (torch.tensor(f.input_ids),
# torch.tensor(f.input_mask),
# torch.tensor(f.segment_ids),
# torch.tensor(f.lm_label_ids),
# torch.tensor(f.label))
# """
# eval_sampler = SequentialSampler(eval_data)
# eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
# save_eval_loss = []
global_step = 0
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
"""
cur_tensors = (torch.tensor(f.input_ids),
torch.tensor(f.input_mask),
torch.tensor(f.segment_ids),
torch.tensor(f.lm_label_ids),
torch.tensor(f.label))
"""
save_loss = []
save_epoch_loss = []
save_step = int(len(train_dataloader) // 5)
for epoch in trange((loaded_epoch + 1),
int(args.num_train_epochs),
desc="Epoch"):
# if args.do_eval and loaded_epoch != -1:
# model.eval()
# eval_loss, eval_accuracy = 0, 0
# nb_eval_steps, nb_eval_examples = 0, 0
#
# for batch in tqdm(eval_dataloader, desc="Evaluating"):
# batch = tuple(t.to(device) for t in batch)
# input_ids, input_mask, segment_ids, label_ids = batch
#
# with torch.no_grad():
# tmp_eval_loss = model(input_ids, segment_ids, input_mask, None, label_ids)
# prediction_scores, logits = model(input_ids, segment_ids, input_mask)
#
# if n_gpu > 1:
# tmp_eval_loss = tmp_eval_loss.mean() # mean() to average on multi-gpu.
#
# 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,
# 'global_step': global_step}
#
# save_eval_loss.append(eval_loss)
#
# output_eval_file = os.path.join(args.output_dir, f"Epoch_{epoch}_eval_results.txt")
# with open(output_eval_file, "w") as writer:
# logger.info(f"***** Eval results on Epoch {epoch} *****")
# for key in sorted(result.keys()):
# logger.info(" %s = %s", key, str(result[key]))
# writer.write("%s = %s\n" % (key, str(result[key])))
model.train()
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
tr_loss_ml = 0
tr_loss_click = 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, label = batch
# if global_step == 0:
# print(input_ids.shape, input_mask.shape, segment_ids.shape, lm_label_ids.shape, label.shape)
loss, loss_ml, loss_click = model(input_ids, segment_ids,
input_mask, lm_label_ids,
label)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
loss_ml = loss_ml.mean()
loss_click = loss_click.mean()
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss_ml = loss_ml / args.gradient_accumulation_steps
loss_click = loss_click / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
tr_loss_ml += loss_ml.item()
tr_loss_click += loss_click.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
if global_step != 0 and global_step % save_step == 0:
# 한 에포치당 5번 저장
logger.info(f'Saving state, iter: {global_step}')
model_to_save = model.module if hasattr(
model, 'module') else model
# Only save the model it-self
model_name = f"weight_on_{global_step}_" + WEIGHTS_NAME
output_model_file = os.path.join(args.output_dir,
model_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())
print("Loss at ", global_step, loss_ml.item(),
loss_click.item(), loss.item())
save_loss.append(
[loss_ml.item(),
loss_click.item(),
loss.item()])
if args.visdom:
update_vis_plot(viz, global_step, loss_ml.item(),
loss_click.item(), iter_plot, epoch_plot,
'append')
if epoch != (int(args.num_train_epochs) - 1):
# 각 에포치가 끝날때 마다 저장
logger.info(f'Saving state, epoch: {epoch}')
model_to_save = model.module if hasattr(model,
'module') else model
# Only save the model it-self
model_name = f"weight_on_ep{epoch}_" + WEIGHTS_NAME
output_model_file = os.path.join(args.output_dir, model_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())
print("Loss at epoch", epoch, tr_loss_ml, tr_loss_click,
tr_loss)
save_epoch_loss.append([tr_loss_ml, tr_loss_click, tr_loss])
if args.visdom:
update_vis_plot(viz, epoch, tr_loss_ml, tr_loss_click,
epoch_plot, None, 'append',
len(train_dataset) // args.train_batch_size)
# if args.do_eval and loaded_epoch == -1:
#
# model.eval()
# eval_loss, eval_accuracy = 0, 0
# nb_eval_steps, nb_eval_examples = 0, 0
#
# for batch in tqdm(eval_dataloader, desc="Evaluating"):
# batch = tuple(t.to(device) for t in batch)
# input_ids, input_mask, segment_ids, label_ids = batch
#
# with torch.no_grad():
# tmp_eval_loss = model(input_ids, segment_ids, input_mask, None, label_ids)
# prediction_scores, logits = model(input_ids, segment_ids, input_mask)
#
# if n_gpu > 1:
# tmp_eval_loss = tmp_eval_loss.mean() # mean() to average on multi-gpu.
#
# 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,
# 'global_step': global_step}
#
# save_eval_loss.append(eval_loss)
#
# output_eval_file = os.path.join(args.output_dir, f"Epoch_{epoch}_eval_results.txt")
# with open(output_eval_file, "w") as writer:
# logger.info(f"***** Eval results on Epoch {epoch} *****")
# 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
logger.info("** ** * Saving fine - tuned model ** ** * ")
# model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
# output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
# if args.do_train:
# torch.save(model_to_save.state_dict(), output_model_file)
save_loss = np.array(save_loss)
save_epoch_loss = np.array(save_epoch_loss)
np.save(os.path.join(args.output_dir, "save_loss.npy"), save_loss)
np.save(os.path.join(args.output_dir, "save_epoch_loss.npy"),
save_epoch_loss)
# if args.do_eval:
# save_eval_loss = np.array(save_eval_loss)
# np.save(os.path.join(args.output_dir, "save_eval_loss.npy"), save_eval_loss)
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())
corpus = load_lm_data(args.entity_dict, args.data, args.output_dir,
args.dataset, tokenizer)
## Training Dataset
train_iter = corpus.get_iterator('train',
args.batch_size,
args.max_seq_length,
args.max_doc_length,
device=device)
## total batch numbers and optim updating steps
total_train_steps = int(train_iter.batch_steps * args.num_train_epochs)
########################################################################################################################
# Building the model
########################################################################################################################
model = BertForPreTraining.from_pretrained(args.bert_model,
entity_num=train_iter.entity_num)
args.n_all_param = sum([p.nelement() for p in model.bert.parameters()])
args.n_nonemb_param = sum(
[p.nelement() for p in model.bert.encoder.parameters()])
logger.info('=' * 100)
for k, v in args.__dict__.items():
logger.info(' - {} : {}'.format(k, v))
logger.info('=' * 100)
logger.info('#params = {}'.format(args.n_all_param))
logger.info('#non emb params = {}'.format(args.n_nonemb_param))
if args.fp16:
model = model.half()
def prepare_model_and_optimizer(args, device):
global_step = 0
args.resume_step = 0
checkpoint = None
config = BertConfig.from_json_file(args.bert_config_path)
config.fused_mha = args.fused_mha
config.fused_gelu_bias = args.fused_gelu_bias
config.dense_seq_output = args.dense_seq_output
config.unpad = args.unpad
config.pad = args.pad
config.fuse_qkv = not args.disable_fuse_qkv
config.fuse_scale = not args.disable_fuse_scale
config.fuse_mask = not args.disable_fuse_mask
config.fuse_dropout = args.enable_fuse_dropout
config.apex_softmax = not args.disable_apex_softmax
config.enable_stream = args.enable_stream
if config.fuse_mask == True: config.apex_softmax = True
if config.pad == False: config.enable_stream = True
if config.unpad == True: config.fused_mha = False
# Padding for divisibility by 8
if config.vocab_size % 8 != 0:
config.vocab_size += 8 - (config.vocab_size % 8)
# Load from Pyt checkpoint - either given as init_checkpoint, or picked up from output_dir if found
if args.init_checkpoint is not None or found_resume_checkpoint(args):
# Prepare model
model = BertForPreTraining(config)
if args.init_checkpoint is None: # finding checkpoint in output_dir
checkpoint_str = "phase2_ckpt_*.pt" if args.phase2 else "phase1_ckpt_*.pt"
model_names = [f for f in glob.glob(os.path.join(args.output_dir, checkpoint_str))]
global_step = max([int(x.split('.pt')[0].split('_')[-1].strip()) for x in model_names])
args.resume_step = global_step #used for throughput computation
resume_init_checkpoint = os.path.join(args.output_dir, checkpoint_str.replace("*", str(global_step)))
print("Setting init checkpoint to %s - which is the latest in %s" %(resume_init_checkpoint, args.output_dir))
checkpoint=torch.load(resume_init_checkpoint, map_location="cpu")
else:
checkpoint=torch.load(args.init_checkpoint, map_location="cpu")["model"]
# Fused MHA requires a remapping of checkpoint parameters
if config.fused_mha:
checkpoint_remapped = remap_attn_parameters(checkpoint)
model.load_state_dict(checkpoint_remapped, strict=False)
else:
model.load_state_dict(checkpoint, strict=True)
else: #Load from TF Checkpoint
model = BertForPreTraining.from_pretrained(args.init_tf_checkpoint, from_tf=True, config=config)
model.to(device)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta', 'LayerNorm']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay_rate},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}]
mlperf_logger.log_event(key=mlperf_logger.constants.OPT_BASE_LR,
value=args.learning_rate, sync=False)
optimizer = FusedLAMB(optimizer_grouped_parameters,
lr=args.learning_rate,
betas=(args.opt_lamb_beta_1, args.opt_lamb_beta_2))
mlperf_logger.log_event(key='opt_epsilon', value=optimizer.defaults['eps'],
sync=False)
b1, b2 = optimizer.defaults['betas']
mlperf_logger.log_event(key='opt_lamb_beta_1', value=b1, sync=False)
mlperf_logger.log_event(key='opt_lamb_beta_2', value=b2, sync=False)
mlperf_logger.log_event(key='opt_lamb_weight_decay_rate',
value=optimizer.defaults['weight_decay'],
sync=False)
if args.warmup_steps == 0:
warmup_steps = int(args.max_steps * args.warmup_proportion)
warmup_start = 0
else:
warmup_steps = args.warmup_steps
warmup_start = args.start_warmup_step
lr_scheduler = LinearWarmupPolyDecayScheduler(optimizer, start_warmup_steps=warmup_start, warmup_steps=warmup_steps,
total_steps=args.max_steps, end_learning_rate=0.0, degree=1.0)
if args.fp16:
if args.loss_scale == 0:
model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale="dynamic")
else:
model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale=args.loss_scale)
amp._amp_state.loss_scalers[0]._loss_scale = float(os.getenv("INIT_LOSS_SCALE", 2**20))
if found_resume_checkpoint(args):
optimizer.load_state_dict(checkpoint['optimizer']) #restores m,v states (only if resuming checkpoint, not for init_checkpoint and init_tf_checkpoint for now)
# Restore AMP master parameters
if args.fp16:
optimizer._lazy_init_maybe_master_weights()
optimizer._amp_stash.lazy_init_called = True
optimizer.load_state_dict(checkpoint['optimizer'])
for param, saved_param in zip(amp.master_params(optimizer), checkpoint['master params']):
param.data.copy_(saved_param.data)
if args.local_rank != -1:
if not args.allreduce_post_accumulation:
model = DDP(model, message_size=250000000, gradient_predivide_factor=torch.distributed.get_world_size())
else:
flat_dist_call([param.data for param in model.parameters()], torch.distributed.broadcast, (0,) )
return model, optimizer, lr_scheduler, checkpoint, global_step
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)