Pytorch 中文语言模型（Bert/Roberta）进一步预训练（further pretrain）
1.Motivation
Bert是在大规模的语料下进行MLM训练得到的结果。然而，在具体的任务下，再在自己新的数据集进行finetune的效果并不会特别好。因此，需要利用domain内的语料对Bert预训练模型进行进一步的MLM训练，也就是further pretrain/ repretrain，与此相关的论文有ACL2020论文《Don’t Stop Pretraining 》，也有一篇将主题模型与bert相结合做semantic matching的工作，它们的目的都是提升Bert在target domain的效果。
2.相关链接
主要参考以下链接
按照链接1上的说明进行操作，我没有将代码跑通，主要是版本问题和Out of memory 问题。于是，尝试链接3下的官方链接，但是也失败了，数据预处理过程就出现了问题，因为服务器不能访问dataset 库里面提供的text.py外网链接，后来通过copy的方式将text.py 文件保存在本地解决了该问题，但是后面又出现了0 sample的问题、index出界的问题，根源还是数据的预处理过程，这样搞了一天，我也服了我自己。最后，还是根据链接1，重新debug，至少它的数据处理没问题，将代码跑通。
3. 具体步骤
3.1 依赖项
务必安装以下版本的transformer，其他版本不保证，至少最新的我试过不行。
transformers==3.0.2
torch = 1.4.0
3.2 数据格式
我的数据格式如下：
每一行就是一条数据，数据之间没有空一行，注意此数据格式对应的参数--line_by_line应该设置为True
若数据格式为：
每一行就是一条数据，数据之间有空一行，注意此数据格式对应的参数--line_by_line应该设置为False。这种数据我没有尝试过，我猜的是设置为False，哈哈哈哈
3.3 代码运行
run_language_model_bert.py 具体代码见附录
运行命令：
python run_language_model_bert.py --output_dir=further_pretrain_sentiment --model_type=roberta--model_name_or_path=robert_pretrain_model --do_train --train_data_file=train_sentiment_500k.txt --mlm --per_device_train_batch_size=32 --line_by_line --overwrite_output_dir --block_size=128
参数说明：
-
output_dir:输出路径，训练结束后，该文件夹下会保存训练之后的文件，包括pytorch_model.bin、vocab.txt、config.json等等 -
model_type：设置为bert或者roberta -
model_name_or_path：原始的Bert/Roberta模型路径，我这里是robert_pretrain_model -
do_train：训练的Flag -
train_data_file：自己的训练语料路径，我的文件是与代码在同一路径下的train_sentiment_500k.txt -
do_eval：验证的Flag，为了节省时间，我没有进行验证，可以自己选择 -
eval_data_file：自己的验证语料路径，在验证flag为True的情况下 -
mlm：我们要做的就是MLM训练，因此设置为True -
per_device_train_batch_size: 训练时，每个gpu上运行的batch 大小，根据自己情况进行设置，我这里GPU共2张卡，每张卡16GB，我设置的大小为32 -
line_by_line: 根据自己的语料格式，自行选择，若每两条数据之间没有空一行，务必设置为True -
overwrite_output_dir:是否覆盖output_dir下已有的文件。若不设置为True，输出路径下有文件时，代码会停止运行，给人提示。自己依情况选择 -
block_size:若设置为-1，则会取语料中句子的最大长度与512之间的较小值。一般语料句子最大长度都比较长，因此会很大，则会造成out of memory现象，即使再减小batch_size也无济于事，之前就是因为这个问题，解决了很久，最后看代码、问师兄才解决，我一看out of memory，就减小batch_size，然而并没有什么用。因此，务必自己设定一个句子长度最大值，我这里是128.
4. 结果
4.1 完整的目录结构
其中，cached_lm_BertTokenizer_1000000000000000019884624838654_train.txt等类似文件是代码运行数据预处理之后产生的。
4.2 训练过程
原始代码会自动调用多GPU，我设置了让它只调用0号GPU.
先运行：export CUDA_VISIBLE_DEVICES=0
再运行 3.3命令行
下图显示的是--per_device_train_batch_size=128 、--block_size=64 0 号 GPU使用情况：
4.3 训练结果
训练结果保存在指定的 output_dir下，训练结束后，该文件夹下生成以下文件：
若设置了do_eval 验证的Flag，该文件夹下还应有验证的结果保存文件eval_results_lm.txt
5 .附录
run_language_model_bert.py代码如下：
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Fine-tuning the library models for language modeling on a text file (GPT, GPT-2, CTRL, BERT, RoBERTa, XLNet).
GPT, GPT-2 and CTRL are fine-tuned using a causal language modeling (CLM) loss. BERT and RoBERTa are fine-tuned
using a masked language modeling (MLM) loss. XLNet is fine-tuned using a permutation language modeling (PLM) loss.
"""
import logging
import math
import os
import torch
from dataclasses import dataclass, field
from typing import Optional
from transformers import (
CONFIG_MAPPING,
MODEL_WITH_LM_HEAD_MAPPING,
AutoConfig,
AutoModelWithLMHead,
AutoTokenizer,
DataCollatorForLanguageModeling,
HfArgumentParser,
LineByLineTextDataset,
PreTrainedTokenizer,
TextDataset,
Trainer,
Trainer,
TrainingArguments,
set_seed,
)
logger = logging.getLogger(__name__)
MODEL_CONFIG_CLASSES = list(MODEL_WITH_LM_HEAD_MAPPING.keys())
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
@dataclass
class ModelArguments:
"""
Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
"""
model_name_or_path: Optional[str] = field(
default=None,
metadata={
"help": "The model checkpoint for weights initialization. Leave None if you want to train a model from scratch."
},
)
model_type: Optional[str] = field(
default=None,
metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
)
config_name: Optional[str] = field(
default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
)
tokenizer_name: Optional[str] = field(
default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
)
cache_dir: Optional[str] = field(
default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
)
@dataclass
class DataTrainingArguments:
"""
Arguments pertaining to what data we are going to input our model for training and eval.
"""
train_data_file: Optional[str] = field(
default=None, metadata={"help": "The input training data file (a text file)."}
)
eval_data_file: Optional[str] = field(
default=None,
metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
)
line_by_line: bool = field(
default=False,
metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
)
mlm: bool = field(
default=False, metadata={"help": "Train with masked-language modeling loss instead of language modeling."}
)
mlm_probability: float = field(
default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
)
plm_probability: float = field(
default=1 / 6,
metadata={
"help": "Ratio of length of a span of masked tokens to surrounding context length for permutation language modeling."
},
)
max_span_length: int = field(
default=5, metadata={"help": "Maximum length of a span of masked tokens for permutation language modeling."}
)
block_size: int = field(
default=128,
metadata={
"help": "Optional input sequence length after tokenization."
"The training dataset will be truncated in block of this size for training."
"Default to the model max input length for single sentence inputs (take into account special tokens)."
},
)
overwrite_cache: bool = field(
default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
)
def get_dataset(args: DataTrainingArguments, tokenizer: PreTrainedTokenizer, evaluate=False):
file_path = args.eval_data_file if evaluate else args.train_data_file
if args.line_by_line:
return LineByLineTextDataset(tokenizer=tokenizer, file_path=file_path, block_size=args.block_size)
else:
return TextDataset(
tokenizer=tokenizer, file_path=file_path, block_size=args.block_size, overwrite_cache=args.overwrite_cache
)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if data_args.eval_data_file is None and training_args.do_eval:
raise ValueError(
"Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
"or remove the --do_eval argument."
)
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning("You are instantiating a new config instance from scratch.")
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported, but you can do it from another script, save it,"
"and load it from here, using --tokenizer_name"
)
if model_args.model_name_or_path:
model = AutoModelWithLMHead.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
else:
logger.info("Training new model from scratch")
model = AutoModelWithLMHead.from_config(config)
model.resize_token_embeddings(len(tokenizer))
if config.model_type in ["bert", "roberta", "distilbert", "camembert"] and not data_args.mlm:
raise ValueError(
"BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the"
"--mlm flag (masked language modeling)."
)
if data_args.block_size <= 0:
data_args.block_size = tokenizer.max_len
# Our input block size will be the max possible for the model
else:
data_args.block_size = min(data_args.block_size, tokenizer.max_len)
# Get datasets
train_dataset = get_dataset(data_args, tokenizer=tokenizer) if training_args.do_train else None
eval_dataset = get_dataset(data_args, tokenizer=tokenizer, evaluate=True) if training_args.do_eval else None
data_collator = DataCollatorForLanguageModeling(
tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
data_collator=data_collator,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
prediction_loss_only=True,
)
# Training
if training_args.do_train:
model_path = (
model_args.model_name_or_path
if model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path)
else None
)
trainer.train(model_path=model_path)
trainer.save_model()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_master():
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
result = {"perplexity": perplexity}
output_eval_file = os.path.join(training_args.output_dir, "eval_results_lm.txt")
if trainer.is_world_master():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
results.update(result)
return results
def _mp_fn(index):
# For xla_spawn (TPUs)
main()
if __name__ == "__main__":
# torch.device('cuda:1')
main()