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("--src_file",
default=None,
type=str,
help="The input data file name.")
parser.add_argument("--tgt_file",
default=None,
type=str,
help="The output data file name.")
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("--config_path",
default=None,
type=str,
help="Bert config file path.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
parser.add_argument(
"--log_dir",
default='',
type=str,
required=True,
help="The output directory where the log will be written.")
parser.add_argument("--model_recover_path",
default=None,
type=str,
required=True,
help="The file of fine-tuned pretraining model.")
parser.add_argument("--optim_recover_path",
default=None,
type=str,
help="The file of pretraining optimizer.")
# 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("--local_debug",
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=64,
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("--label_smoothing",
default=0,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay",
default=0.01,
type=float,
help="The weight decay rate for Adam.")
parser.add_argument("--finetune_decay",
action='store_true',
help="Weight decay to the original weights.")
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("--hidden_dropout_prob",
default=0.1,
type=float,
help="Dropout rate for hidden states.")
parser.add_argument("--attention_probs_dropout_prob",
default=0.1,
type=float,
help="Dropout rate for attention probabilities.")
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(
'--fp32_embedding',
action='store_true',
help=
"Whether to use 32-bit float precision instead of 16-bit for embeddings"
)
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('--amp',
action='store_true',
help="Whether to use amp for fp16")
parser.add_argument(
'--from_scratch',
action='store_true',
help=
"Initialize parameters with random values (i.e., training from scratch)."
)
parser.add_argument('--new_segment_ids',
action='store_true',
help="Use new segment ids for bi-uni-directional LM.")
parser.add_argument('--new_pos_ids',
action='store_true',
help="Use new position ids for LMs.")
parser.add_argument('--tokenized_input',
action='store_true',
help="Whether the input is tokenized.")
parser.add_argument('--max_len_a',
type=int,
default=0,
help="Truncate_config: maximum length of segment A.")
parser.add_argument('--max_len_b',
type=int,
default=0,
help="Truncate_config: maximum length of segment B.")
parser.add_argument(
'--trunc_seg',
default='',
help="Truncate_config: first truncate segment A/B (option: a, b).")
parser.add_argument(
'--always_truncate_tail',
action='store_true',
help="Truncate_config: Whether we should always truncate tail.")
parser.add_argument(
"--mask_prob",
default=0.15,
type=float,
help=
"Number of prediction is sometimes less than max_pred when sequence is short."
)
parser.add_argument(
"--mask_prob_eos",
default=0,
type=float,
help=
"Number of prediction is sometimes less than max_pred when sequence is short."
)
parser.add_argument('--max_pred',
type=int,
default=20,
help="Max tokens of prediction.")
parser.add_argument("--num_workers",
default=0,
type=int,
help="Number of workers for the data loader.")
parser.add_argument('--mask_source_words',
action='store_true',
help="Whether to mask source words for training")
parser.add_argument('--skipgram_prb',
type=float,
default=0.0,
help='prob of ngram mask')
parser.add_argument('--skipgram_size',
type=int,
default=1,
help='the max size of ngram mask')
parser.add_argument('--mask_whole_word',
action='store_true',
help="Whether masking a whole word.")
parser.add_argument('--do_l2r_training',
action='store_true',
help="Whether to do left to right training")
parser.add_argument(
'--has_sentence_oracle',
action='store_true',
help="Whether to have sentence level oracle for training. "
"Only useful for summary generation")
parser.add_argument('--max_position_embeddings',
type=int,
default=None,
help="max position embeddings")
parser.add_argument('--relax_projection',
action='store_true',
help="Use different projection layers for tasks.")
parser.add_argument('--ffn_type',
default=0,
type=int,
help="0: default mlp; 1: W((Wx+b) elem_prod x);")
parser.add_argument('--num_qkv',
default=0,
type=int,
help="Number of different <Q,K,V>.")
parser.add_argument('--seg_emb',
action='store_true',
help="Using segment embedding for self-attention.")
parser.add_argument(
'--s2s_special_token',
action='store_true',
help="New special tokens ([S2S_SEP]/[S2S_CLS]) of S2S.")
parser.add_argument('--s2s_add_segment',
action='store_true',
help="Additional segmental for the encoder of S2S.")
parser.add_argument(
'--s2s_share_segment',
action='store_true',
help=
"Sharing segment embeddings for the encoder of S2S (used with --s2s_add_segment)."
)
parser.add_argument('--pos_shift',
action='store_true',
help="Using position shift for fine-tuning.")
args = parser.parse_args()
assert Path(
args.model_recover_path).exists(), "--model_recover_path doesn't exist"
args.output_dir = args.output_dir.replace('[PT_OUTPUT_DIR]',
os.getenv('PT_OUTPUT_DIR', ''))
args.log_dir = args.log_dir.replace('[PT_OUTPUT_DIR]',
os.getenv('PT_OUTPUT_DIR', ''))
os.makedirs(args.output_dir, exist_ok=True)
os.makedirs(args.log_dir, exist_ok=True)
json.dump(args.__dict__,
open(os.path.join(args.output_dir, 'opt.json'), 'w'),
sort_keys=True,
indent=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
dist.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 args.local_rank not in (-1, 0):
# Make sure only the first process in distributed training will download model & vocab
dist.barrier()
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
if args.max_position_embeddings:
tokenizer.max_len = args.max_position_embeddings
data_tokenizer = WhitespaceTokenizer(
) if args.tokenized_input else tokenizer
if args.local_rank == 0:
dist.barrier()
if args.do_train:
print("Loading Train Dataset", args.data_dir)
bi_uni_pipeline = [
seq2seq_loader.Preprocess4Seq2seq(
args.max_pred,
args.mask_prob,
list(tokenizer.vocab.keys()),
tokenizer.convert_tokens_to_ids,
args.max_seq_length,
new_segment_ids=args.new_segment_ids,
truncate_config={
'max_len_a': args.max_len_a,
'max_len_b': args.max_len_b,
'trunc_seg': args.trunc_seg,
'always_truncate_tail': args.always_truncate_tail
},
mask_source_words=args.mask_source_words,
skipgram_prb=args.skipgram_prb,
skipgram_size=args.skipgram_size,
mask_whole_word=args.mask_whole_word,
mode="s2s",
has_oracle=args.has_sentence_oracle,
num_qkv=args.num_qkv,
s2s_special_token=args.s2s_special_token,
s2s_add_segment=args.s2s_add_segment,
s2s_share_segment=args.s2s_share_segment,
pos_shift=args.pos_shift)
]
file_oracle = None
if args.has_sentence_oracle:
file_oracle = os.path.join(args.data_dir, 'train.oracle')
fn_src = os.path.join(args.data_dir,
args.src_file if args.src_file else 'train.src')
fn_tgt = os.path.join(args.data_dir,
args.tgt_file if args.tgt_file else 'train.tgt')
train_dataset = seq2seq_loader.Seq2SeqDataset(
fn_src,
fn_tgt,
args.train_batch_size,
data_tokenizer,
args.max_seq_length,
file_oracle=file_oracle,
bi_uni_pipeline=bi_uni_pipeline,
corpus_preprocessors=corpus_preprocessors)
train_dataset.initial()
print(len(train_dataset.ex_list))
print(train_dataset.batch_size)
# assert 1==0
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset, replacement=False)
_batch_size = args.train_batch_size
else:
train_sampler = DistributedSampler(train_dataset)
_batch_size = args.train_batch_size // dist.get_world_size()
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=_batch_size,
sampler=train_sampler,
num_workers=args.num_workers,
collate_fn=seq2seq_loader.batch_list_to_batch_tensors,
pin_memory=False)
# c = 0
# for i_epoch in trange(0, int(args.num_train_epochs)+1, desc="Epoch", disable=args.local_rank not in (-1, 0)):
# if args.local_rank != -1:
# train_sampler.set_epoch(i_epoch)
# iter_bar = tqdm(train_dataloader, desc='Iter (loss=X.XXX)',
# disable=args.local_rank not in (-1, 0))
# for step, batch in enumerate(iter_bar):
# batch = [
# t.to(device) if t is not None else None for t in batch]
# if args.has_sentence_oracle:
# input_ids, segment_ids, input_mask, mask_qkv, lm_label_ids, masked_pos, masked_weights, is_next, task_idx, sop_label, oracle_pos, oracle_weights, oracle_labels = batch
# else:
# input_ids, segment_ids, input_mask, mask_qkv, lm_label_ids, masked_pos, masked_weights, is_next, task_idx, sop_label = batch
# oracle_pos, oracle_weights, oracle_labels = None, None, None
# c += input_ids.shape[0]
#
# # print(input_ids)
# #
# # print(input_ids.shape)
# # print(segment_ids)
# # print(segment_ids.shape)
# # print(is_next)
# # print(task_idx)
# # print(sop_label)
# # print(task_idx.shape)
# # for i in range(input_mask.shape[0]):
# # print(input_mask[i])
# print(c)
# print(train_dataset.c)
# note: args.train_batch_size has been changed to (/= args.gradient_accumulation_steps)
# t_total = int(math.ceil(len(train_dataset.ex_list) / args.train_batch_size)
t_total = int(
len(train_dataloader) * args.num_train_epochs /
args.gradient_accumulation_steps)
amp_handle = None
if args.fp16 and args.amp:
from apex import amp
amp_handle = amp.init(enable_caching=True)
logger.info("enable fp16 with amp")
# Prepare model
recover_step = _get_max_epoch_model(args.output_dir)
cls_num_labels = 2
type_vocab_size = 6 + \
(1 if args.s2s_add_segment else 0) if args.new_segment_ids else 2
num_sentlvl_labels = 2 if args.has_sentence_oracle else 0
relax_projection = 4 if args.relax_projection else 0
if args.local_rank not in (-1, 0):
# Make sure only the first process in distributed training will download model & vocab
dist.barrier()
if (recover_step is None) and (args.model_recover_path is None):
# if _state_dict == {}, the parameters are randomly initialized
# if _state_dict == None, the parameters are initialized with bert-init
_state_dict = {} if args.from_scratch else None
model = BertForPreTrainingLossMask.from_pretrained(
args.bert_model,
state_dict=_state_dict,
num_labels=cls_num_labels,
num_rel=0,
type_vocab_size=type_vocab_size,
config_path=args.config_path,
task_idx=3,
num_sentlvl_labels=num_sentlvl_labels,
max_position_embeddings=args.max_position_embeddings,
label_smoothing=args.label_smoothing,
fp32_embedding=args.fp32_embedding,
relax_projection=relax_projection,
new_pos_ids=args.new_pos_ids,
ffn_type=args.ffn_type,
hidden_dropout_prob=args.hidden_dropout_prob,
attention_probs_dropout_prob=args.attention_probs_dropout_prob,
num_qkv=args.num_qkv,
seg_emb=args.seg_emb,
local_debug=args.local_debug)
global_step = 0
else:
if recover_step:
logger.info("***** Recover model: %d *****", recover_step)
model_recover = torch.load(os.path.join(
args.output_dir, "model.{0}.bin".format(recover_step)),
map_location='cpu')
# recover_step == number of epochs
global_step = math.floor(recover_step * t_total /
args.num_train_epochs)
elif args.model_recover_path:
logger.info("***** Recover model: %s *****",
args.model_recover_path)
model_recover = torch.load(args.model_recover_path,
map_location='cpu')
global_step = 0
model = BertForPreTrainingLossMask.from_pretrained(
args.bert_model,
state_dict=model_recover,
num_labels=cls_num_labels,
num_rel=0,
type_vocab_size=type_vocab_size,
config_path=args.config_path,
task_idx=3,
num_sentlvl_labels=num_sentlvl_labels,
max_position_embeddings=args.max_position_embeddings,
label_smoothing=args.label_smoothing,
fp32_embedding=args.fp32_embedding,
relax_projection=relax_projection,
new_pos_ids=args.new_pos_ids,
ffn_type=args.ffn_type,
hidden_dropout_prob=args.hidden_dropout_prob,
attention_probs_dropout_prob=args.attention_probs_dropout_prob,
num_qkv=args.num_qkv,
seg_emb=args.seg_emb,
local_debug=args.local_debug)
if args.local_rank == 0:
dist.barrier()
if args.fp16:
model.half()
if args.fp32_embedding:
model.bert.embeddings.word_embeddings.float()
model.bert.embeddings.position_embeddings.float()
model.bert.embeddings.token_type_embeddings.float()
model.to(device)
if args.local_rank != -1:
try:
from torch.nn.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError("DistributedDataParallel")
model = DDP(model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
elif n_gpu > 1:
# model = torch.nn.DataParallel(model)
model = DataParallelImbalance(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 pytorch_pretrained_bert.optimization_fp16 import FP16_Optimizer_State
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_State(optimizer,
dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer_State(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 recover_step:
logger.info("***** Recover optimizer: %d *****", recover_step)
optim_recover = torch.load(os.path.join(
args.output_dir, "optim.{0}.bin".format(recover_step)),
map_location='cpu')
if hasattr(optim_recover, 'state_dict'):
optim_recover = optim_recover.state_dict()
optimizer.load_state_dict(optim_recover)
if args.loss_scale == 0:
logger.info("***** Recover optimizer: dynamic_loss_scale *****")
optimizer.dynamic_loss_scale = True
logger.info("***** CUDA.empty_cache() *****")
torch.cuda.empty_cache()
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", t_total)
model.train()
if recover_step:
start_epoch = recover_step + 1
else:
start_epoch = 1
for i_epoch in trange(start_epoch,
int(args.num_train_epochs) + 1,
desc="Epoch",
disable=args.local_rank not in (-1, 0)):
if args.local_rank != -1:
train_sampler.set_epoch(i_epoch)
iter_bar = tqdm(train_dataloader,
desc='Iter (loss=X.XXX)',
disable=args.local_rank not in (-1, 0))
for step, batch in enumerate(iter_bar):
batch = [
t.to(device) if t is not None else None for t in batch
]
if args.has_sentence_oracle:
input_ids, segment_ids, input_mask, mask_qkv, lm_label_ids, masked_pos, masked_weights, is_next, task_idx, sop_label, oracle_pos, oracle_weights, oracle_labels = batch
else:
input_ids, segment_ids, input_mask, mask_qkv, lm_label_ids, masked_pos, masked_weights, is_next, task_idx, sop_label = batch
print(sop_label)
print(task_idx)
oracle_pos, oracle_weights, oracle_labels = None, None, None
# loss_tuple = model(input_ids, segment_ids, input_mask, lm_label_ids, is_next,
# masked_pos=masked_pos, masked_weights=masked_weights, task_idx=task_idx,
# masked_pos_2=oracle_pos, masked_weights_2=oracle_weights,
# masked_labels_2=oracle_labels, mask_qkv=mask_qkv)
loss_tuple = model(input_ids,
segment_ids,
input_mask,
lm_label_ids,
sop_label,
masked_pos=masked_pos,
masked_weights=masked_weights,
task_idx=task_idx,
masked_pos_2=oracle_pos,
masked_weights_2=oracle_weights,
masked_labels_2=oracle_labels,
mask_qkv=mask_qkv)
masked_lm_loss, next_sentence_loss = loss_tuple
if n_gpu > 1: # mean() to average on multi-gpu.
# loss = loss.mean()
masked_lm_loss = masked_lm_loss.mean()
next_sentence_loss = next_sentence_loss.mean()
print('mask_lm_loss {}'.format(masked_lm_loss))
print('next_sentence_loss {}'.format(next_sentence_loss))
print('----------------------------------------------')
loss = masked_lm_loss + next_sentence_loss
# logging for each step (i.e., before normalization by args.gradient_accumulation_steps)
iter_bar.set_description('Iter (loss=%5.3f)' % loss.item())
# ensure that accumlated gradients are normalized
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
if amp_handle:
amp_handle._clear_cache()
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
lr_this_step = args.learning_rate * \
warmup_linear(global_step/t_total,
args.warmup_proportion)
if args.fp16:
# modify learning rate with special warm up BERT uses
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
if (args.local_rank == -1 or torch.distributed.get_rank() == 0):
logger.info(
"** ** * Saving fine-tuned model and optimizer ** ** * ")
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, "model.{0}.bin".format(i_epoch))
torch.save(model_to_save.state_dict(), output_model_file)
output_optim_file = os.path.join(
args.output_dir, "optim.{0}.bin".format(i_epoch))
torch.save(optimizer.state_dict(), output_optim_file)
logger.info("***** CUDA.empty_cache() *****")
torch.cuda.empty_cache()
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("--src_file",
default=None,
type=str,
help="The input data file name.")
parser.add_argument("--topic_model_recover_path",
default=None,
type=str,
help="The file of fine-tuned pretraining topic model.")
parser.add_argument("--topic_model_dict_path",
default=None,
type=str,
help="The file of fine-tuned pretraining topic model.")
parser.add_argument("--tgt_file",
default=None,
type=str,
help="The output data file name.")
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("--config_path",
default=None,
type=str,
help="Bert config file path.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
parser.add_argument(
"--log_dir",
default='',
type=str,
required=True,
help="The output directory where the log will be written.")
parser.add_argument("--model_recover_path",
default=None,
type=str,
required=True,
help="The file of fine-tuned pretraining model.")
parser.add_argument("--optim_recover_path",
default=None,
type=str,
help="The file of pretraining optimizer.")
parser.add_argument('--topic_mode',
default=1,
type=float,
help="1:idea1 1.1:idea1_wo_theta 2:idea2 ")
parser.add_argument('--topic_model',
default=False,
type=bool,
help="if only use topic model")
# Other parameters
parser.add_argument(
"--max_seq_length",
default=192,
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.") #batch_size = batch_size/n_gpus
parser.add_argument("--eval_batch_size",
default=16,
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("--label_smoothing",
default=0,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay",
default=0.01,
type=float,
help="The weight decay rate for Adam.")
parser.add_argument("--finetune_decay",
action='store_true',
help="Weight decay to the original weights.")
parser.add_argument("--num_train_epochs",
default=30,
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("--hidden_dropout_prob",
default=0.1,
type=float,
help="Dropout rate for hidden states.")
parser.add_argument("--attention_probs_dropout_prob",
default=0.1,
type=float,
help="Dropout rate for attention probabilities.")
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(
'--fp32_embedding',
action='store_true',
help=
"Whether to use 32-bit float precision instead of 16-bit for embeddings"
)
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('--amp',
action='store_true',
help="Whether to use amp for fp16")
parser.add_argument(
'--from_scratch',
action='store_true',
help=
"Initialize parameters with random values (i.e., training from scratch)."
)
parser.add_argument('--new_segment_ids',
action='store_true',
help="Use new segment ids for bi-uni-directional LM.")
parser.add_argument('--new_pos_ids',
action='store_true',
help="Use new position ids for LMs.")
parser.add_argument('--tokenized_input',
action='store_true',
help="Whether the input is tokenized.")
parser.add_argument('--max_len_a',
type=int,
default=0,
help="Truncate_config: maximum length of segment A.")
parser.add_argument('--max_len_b',
type=int,
default=0,
help="Truncate_config: maximum length of segment B.")
parser.add_argument(
'--trunc_seg',
default='',
help="Truncate_config: first truncate segment A/B (option: a, b).")
parser.add_argument(
'--always_truncate_tail',
action='store_true',
help="Truncate_config: Whether we should always truncate tail.")
parser.add_argument(
"--mask_prob",
default=0.15,
type=float,
help=
"Number of prediction is sometimes less than max_pred when sequence is short."
)
parser.add_argument(
"--mask_prob_eos",
default=0,
type=float,
help=
"Number of prediction is sometimes less than max_pred when sequence is short."
)
parser.add_argument('--max_pred',
type=int,
default=20,
help="Max tokens of prediction.")
parser.add_argument("--num_workers",
default=0,
type=int,
help="Number of workers for the data loader.")
parser.add_argument('--mask_source_words',
action='store_true',
help="Whether to mask source words for training")
parser.add_argument('--skipgram_prb',
type=float,
default=0.0,
help='prob of ngram mask')
parser.add_argument('--skipgram_size',
type=int,
default=1,
help='the max size of ngram mask')
parser.add_argument('--mask_whole_word',
action='store_true',
help="Whether masking a whole word.")
parser.add_argument('--do_l2r_training',
action='store_true',
help="Whether to do left to right training")
parser.add_argument(
'--has_sentence_oracle',
action='store_true',
help="Whether to have sentence level oracle for training. "
"Only useful for summary generation")
parser.add_argument('--max_position_embeddings',
type=int,
default=None,
help="max position embeddings")
parser.add_argument('--relax_projection',
action='store_true',
help="Use different projection layers for tasks.")
parser.add_argument('--ffn_type',
default=0,
type=int,
help="0: default mlp; 1: W((Wx+b) elem_prod x);")
parser.add_argument('--num_qkv',
default=0,
type=int,
help="Number of different <Q,K,V>.")
parser.add_argument('--seg_emb',
action='store_true',
help="Using segment embedding for self-attention.")
parser.add_argument(
'--s2s_special_token',
action='store_true',
help="New special tokens ([S2S_SEP]/[S2S_CLS]) of S2S.")
parser.add_argument('--s2s_add_segment',
action='store_true',
help="Additional segmental for the encoder of S2S.")
parser.add_argument(
'--s2s_share_segment',
action='store_true',
help=
"Sharing segment embeddings for the encoder of S2S (used with --s2s_add_segment)."
)
parser.add_argument('--pos_shift',
action='store_true',
help="Using position shift for fine-tuning.")
args = parser.parse_args()
assert Path(
args.model_recover_path).exists(), "--model_recover_path doesn't exist"
args.output_dir = args.output_dir.replace('[PT_OUTPUT_DIR]',
os.getenv('PT_OUTPUT_DIR', ''))
args.log_dir = args.log_dir.replace('[PT_OUTPUT_DIR]',
os.getenv('PT_OUTPUT_DIR', ''))
os.makedirs(args.output_dir, exist_ok=True)
os.makedirs(args.log_dir, exist_ok=True)
json.dump(args.__dict__,
open(os.path.join(args.output_dir, 'opt.json'), 'w'),
sort_keys=True,
indent=2)
print("args.local_rank", args.local_rank)
print("args.no_cuda", args.no_cuda)
if args.local_rank == -1 or args.no_cuda: #-1 False
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu") #device = cuda
n_gpu = torch.cuda.device_count()
print("n_gpu_1", n_gpu)
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
dist.init_process_group(backend='nccl')
print("n_gpu_1", n_gpu)
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 args.local_rank not in (-1, 0):
# Make sure only the first process in distributed training will download model & vocab
dist.barrier()
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
if args.max_position_embeddings:
tokenizer.max_len = args.max_position_embeddings
data_tokenizer = WhitespaceTokenizer(
) if args.tokenized_input else tokenizer
if args.local_rank == 0:
dist.barrier()
if args.do_train:
bi_uni_pipeline = [
seq2seq_loader.Preprocess4Seq2seq(
args.max_pred,
args.mask_prob,
list(tokenizer.vocab.keys()),
tokenizer.convert_tokens_to_ids,
args.max_seq_length,
new_segment_ids=args.new_segment_ids,
truncate_config={
'max_len_a': args.max_len_a,
'max_len_b': args.max_len_b,
'trunc_seg': args.trunc_seg,
'always_truncate_tail': args.always_truncate_tail
},
mask_source_words=args.mask_source_words,
skipgram_prb=args.skipgram_prb,
skipgram_size=args.skipgram_size,
mask_whole_word=args.mask_whole_word,
mode="s2s",
has_oracle=args.has_sentence_oracle,
num_qkv=args.num_qkv,
s2s_special_token=args.s2s_special_token,
s2s_add_segment=args.s2s_add_segment,
s2s_share_segment=args.s2s_share_segment,
pos_shift=args.pos_shift)
]
file_oracle = None
if args.has_sentence_oracle:
file_oracle = os.path.join(args.data_dir, 'train.oracle')
fn_src = os.path.join(args.data_dir,
args.src_file if args.src_file else 'train.src')
fn_tgt = os.path.join(args.data_dir,
args.tgt_file if args.tgt_file else 'train.tgt')
train_dataset = seq2seq_loader.Seq2SeqDataset(
fn_src,
fn_tgt,
args.data_dir,
args.topic_model_dict_path,
args.train_batch_size,
data_tokenizer,
args.max_seq_length,
file_oracle=file_oracle,
bi_uni_pipeline=bi_uni_pipeline)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset, replacement=False)
_batch_size = args.train_batch_size
else:
train_sampler = DistributedSampler(train_dataset)
_batch_size = args.train_batch_size // dist.get_world_size()
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=_batch_size,
sampler=train_sampler,
num_workers=args.num_workers,
collate_fn=seq2seq_loader.batch_list_to_batch_tensors,
pin_memory=False)
# note: args.train_batch_size has been changed to (/= args.gradient_accumulation_steps)
# t_total = int(math.ceil(len(train_dataset.ex_list) / args.train_batch_size)
t_total = int(
len(train_dataloader) * args.num_train_epochs /
args.gradient_accumulation_steps)
amp_handle = None
if args.fp16 and args.amp:
from apex import amp
amp_handle = amp.init(enable_caching=True)
logger.info("enable fp16 with amp")
# Prepare model
recover_step = _get_max_epoch_model(args.output_dir)
cls_num_labels = 2
type_vocab_size = 6 + \
(1 if args.s2s_add_segment else 0) if args.new_segment_ids else 2 ### type_vocab_size=6
num_sentlvl_labels = 2 if args.has_sentence_oracle else 0
relax_projection = 4 if args.relax_projection else 0
if args.local_rank not in (-1, 0):
# Make sure only the first process in distributed training will download model & vocab
dist.barrier()
if (recover_step is None) and (args.model_recover_path is None):
# if _state_dict == {}, the parameters are randomly initialized
# if _state_dict == None, the parameters are initialized with bert-init
_state_dict = {} if args.from_scratch else None
unilm = BertForPreTrainingLossMask.from_pretrained(
args.bert_model,
state_dict=_state_dict,
num_labels=cls_num_labels,
num_rel=0,
type_vocab_size=type_vocab_size,
config_path=args.config_path,
task_idx=3,
num_sentlvl_labels=num_sentlvl_labels,
max_position_embeddings=args.max_position_embeddings,
label_smoothing=args.label_smoothing,
fp32_embedding=args.fp32_embedding,
relax_projection=relax_projection,
new_pos_ids=args.new_pos_ids,
ffn_type=args.ffn_type,
hidden_dropout_prob=args.hidden_dropout_prob,
attention_probs_dropout_prob=args.attention_probs_dropout_prob,
num_qkv=args.num_qkv,
seg_emb=args.seg_emb)
global_step = 0
else:
if recover_step:
logger.info("***** Recover model: %d *****", recover_step)
model_recover = torch.load(os.path.join(
args.output_dir, "model.{0}.bin".format(recover_step)),
map_location='cpu')
# recover_step == number of epochs
global_step = math.floor(recover_step * t_total /
args.num_train_epochs)
elif args.model_recover_path: # here is the entrance
logger.info("***** Recover model: %s *****",
args.model_recover_path)
model_recover = torch.load(args.model_recover_path,
map_location='cpu')
global_step = 0
unilm = BertForPreTrainingLossMask.from_pretrained(
args.bert_model,
state_dict=model_recover,
num_labels=cls_num_labels,
num_rel=0,
type_vocab_size=type_vocab_size,
config_path=args.config_path,
task_idx=3,
num_sentlvl_labels=num_sentlvl_labels,
max_position_embeddings=args.max_position_embeddings,
label_smoothing=args.label_smoothing,
fp32_embedding=args.fp32_embedding,
relax_projection=relax_projection,
new_pos_ids=args.new_pos_ids,
ffn_type=args.ffn_type,
hidden_dropout_prob=args.hidden_dropout_prob,
attention_probs_dropout_prob=args.attention_probs_dropout_prob,
num_qkv=args.num_qkv,
seg_emb=args.seg_emb)
#1. 模型初始化，入口定义好
gsm = GSM(train_dataset.vocabsize)
gsm_checkpoint = torch.load(args.topic_model_recover_path)
gsm.load_state_dict(gsm_checkpoint["net"])
if args.local_rank == 0:
dist.barrier()
if args.fp16:
unilm.half()
gsm.half()
if args.fp32_embedding:
unilm.bert.embeddings.word_embeddings.float()
unilm.bert.embeddings.position_embeddings.float()
unilm.bert.embeddings.token_type_embeddings.float()
unilm.to(device)
gsm.to(device)
if args.local_rank != -1:
try:
from torch.nn.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError("DistributedDataParallel")
unilm = DDP(unilm,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
elif n_gpu > 1:
# model = torch.nn.DataParallel(model)
unilm = DataParallelImbalance(unilm)
gsm = DataParallelImbalance(gsm)
# Prepare optimizer
total = 0
param_optimizer = list(unilm.named_parameters())
param_optimizer_topic = list(gsm.named_parameters())
for name, parameters in unilm.named_parameters():
if "idea" in name:
if "11" in name and "idea2" in name:
total += np.prod(parameters.size())
# print(name, ':', parameters.size())
else:
total += np.prod(parameters.size())
# print(name, ':', parameters.size())
print("gsm have {} paramerters in total".format(
sum(x.numel() for x in gsm.parameters())))
print("Number of parameter: %.6fM" % (total / 1e6))
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
if not args.topic_model:
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,
'topic':
False
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0,
'topic':
False
}, {
'params': [p for n, p in param_optimizer_topic],
'weight_decay':
0.0,
'lr':
1e-3,
'topic':
True
}]
else:
optimizer_grouped_parameters = [{
'params': [p for n, p in param_optimizer_topic],
'weight_decay':
0.0,
'lr':
1e-3,
'topic':
True
}]
#一部分是有weight的，一部分是没有weight_dacay的
# print("optimizer_grouped_parameters", optimizer_grouped_parameters)
if args.fp16:
try:
# from apex.optimizers import FP16_Optimizer
from pytorch_pretrained_bert.optimization_fp16 import FP16_Optimizer_State
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_State(optimizer,
dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer_State(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 recover_step:
logger.info("***** Recover optimizer: %d *****", recover_step)
optim_recover = torch.load(os.path.join(
args.output_dir, "optim.{0}.bin".format(recover_step)),
map_location='cpu')
if hasattr(optim_recover, 'state_dict'):
optim_recover = optim_recover.state_dict()
optimizer.load_state_dict(optim_recover)
if args.loss_scale == 0:
logger.info("***** Recover optimizer: dynamic_loss_scale *****")
optimizer.dynamic_loss_scale = True
logger.info("***** CUDA.empty_cache() *****")
torch.cuda.empty_cache()
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", t_total)
unilm.train()
gsm.train()
if recover_step:
start_epoch = recover_step + 1
else:
start_epoch = 1
print("000000", args.local_rank, start_epoch,
int(args.num_train_epochs) + 1)
topicloss = []
unilmloss = []
topicloss_lst = []
unilmloss_lst = []
for i_epoch in trange(start_epoch,
int(args.num_train_epochs) + 1,
desc="Epoch",
disable=args.local_rank not in (-1, 0)):
loss_sum = 0.0
ppx_sum = 0.0
word_count = 0.0
doc_count = 0.0
if args.local_rank != -1:
train_sampler.set_epoch(i_epoch)
iter_bar = tqdm(train_dataloader,
desc='Iter (loss=X.XXX)',
disable=args.local_rank not in (-1, 0))
for step, batch in enumerate(iter_bar):
batch = [
t.to(device) if t is not None else None for t in batch
]
if args.has_sentence_oracle: #false
input_ids, segment_ids, input_mask, mask_qkv, lm_label_ids, masked_pos, masked_weights, is_next, task_idx, oracle_pos, oracle_weights, oracle_labels = batch
else: #这里加了bows
input_ids, segment_ids, input_mask, mask_qkv, lm_label_ids, masked_pos, masked_weights, is_next, task_idx, bows = batch
oracle_pos, oracle_weights, oracle_labels = None, None, None
p_x, mus, log_vars, theta, beta, topic_embedding = gsm(bows)
if not args.topic_model:
loss_tuple = unilm(input_ids,
theta,
beta,
topic_embedding,
args.topic_mode,
segment_ids,
input_mask,
lm_label_ids,
is_next,
masked_pos=masked_pos,
masked_weights=masked_weights,
task_idx=task_idx,
masked_pos_2=oracle_pos,
masked_weights_2=oracle_weights,
masked_labels_2=oracle_labels,
mask_qkv=mask_qkv)
masked_lm_loss, next_sentence_loss = loss_tuple
## topic loss
logsoftmax = torch.log(p_x + 1e-10)
rec_loss = -1.0 * torch.sum(
bows * logsoftmax
) #bows*logsoftmax = [batch_size, |V|], 其中torch.sum 把所有的loss全部加起来了，也可以只用加某一维度。
rec_loss_per = -1.0 * torch.sum(bows * logsoftmax, dim=1)
rec_loss_per = rec_loss_per.cpu().detach().numpy()
kl_div = -0.5 * torch.sum(1 + log_vars - mus.pow(2) -
log_vars.exp())
loss_topic = rec_loss + kl_div
if n_gpu > 1: # mean() to average on multi-gpu.
loss_topic = loss_topic.mean()
if not args.topic_model:
masked_lm_loss = masked_lm_loss.mean()
next_sentence_loss = next_sentence_loss.mean()
if not args.topic_model:
loss_unilm = masked_lm_loss + next_sentence_loss
# cal perplexity
word_count_list = []
loss_sum += loss_topic.item()
for bow in bows:
word_num = torch.sum(bow).cpu().numpy()
word_count_list.append(word_num)
word_count += word_num
word_count_np = np.array(word_count_list)
doc_count += len(bows)
ppx_sum += np.sum(np.true_divide(rec_loss_per, word_count_np))
topicloss_lst.append(loss_topic.item() / len(bows))
if not args.topic_model:
unilmloss_lst.append(loss_unilm.item())
#topic_loss end
if not args.topic_model:
loss = loss_unilm + loss_topic
else:
loss = loss_topic
# ensure that accumlated gradients are normalized
if args.gradient_accumulation_steps > 1: # =1
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
if amp_handle:
amp_handle._clear_cache()
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
lr_this_step = args.learning_rate * \
warmup_linear(global_step/t_total,
args.warmup_proportion)
if args.fp16:
# modify learning rate with special warm up BERT uses
for param_group in optimizer.param_groups:
if not param_group['topic']:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
if not args.topic_model:
iter_bar.set_description(
'Iter (loss_unilm=%5.3f),Iter (ppl=%5.3f)' %
(loss_unilm.item(),
np.sum(np.true_divide(rec_loss_per, word_count_np))))
else:
iter_bar.set_description(
'Iter (loss_topic=%5.3f), (ppl=%5.3f)' %
(loss_topic.item(),
np.sum(np.true_divide(rec_loss_per, word_count_np))))
#Save a trained model
ppx_word = np.exp(loss_sum / word_count)
ppx_document = np.exp(ppx_sum / doc_count)
print("********")
print("word_count", word_count)
print("ppx_word", ppx_word)
print("ppx_document", ppx_document)
if (args.local_rank == -1 or torch.distributed.get_rank() == 0):
#save unilm model
logger.info(
"** ** * Saving fine-tuned model and optimizer ** ** * ")
unilm_model_to_save = unilm.module if hasattr(
unilm, 'module') else unilm # Only save the model it-self
output_unilm_model_file = os.path.join(
args.output_dir, "unilm.{0}.bin".format(i_epoch))
torch.save(unilm_model_to_save.state_dict(),
output_unilm_model_file)
#save topic model
logger.info(
"** ** * Saving topic model and optimizer ** ** * ")
topic_model_to_save = gsm.module if hasattr(
gsm, 'module') else gsm # Only save the model it-self
output_topic_model_file = os.path.join(
args.output_dir, "topic.{0}.ckpt".format(i_epoch))
torch.save(topic_model_to_save.state_dict(),
output_topic_model_file)
logger.info("***** CUDA.empty_cache() *****")
torch.cuda.empty_cache()
smth_pts = smooth_curve(topicloss_lst)
# plt.plot(range(len(topicloss_lst)), topicloss_lst)
plt.plot(range(len(smth_pts)), smth_pts)
plt.xlabel('epochs')
plt.title('Topic Model Train Loss')
plt.savefig(args.output_dir + '/topic_loss.png')
plt.cla()
plt.plot(range(len(unilmloss_lst)), unilmloss_lst)
plt.xlabel('epochs')
plt.title('Unilm Train Loss')
plt.savefig(args.output_dir + '/unilm_loss.png')