SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Fit chinese wwm to new datasets (#9887)

Browse Source 
* MOD: fit chinese wwm to new datasets

* MOD: move wwm to new folder

* MOD: formate code

* Styling

* MOD add param and recover trainer

Co-authored-by: Sylvain Gugger <sylvain.gugger@gmail.com>
This commit is contained in:
wlhgtc
2021-02-01 16:37:59 +08:00
committed by GitHub
parent 24881008a6
commit 1682804ebd
6 changed files with 249 additions and 67 deletions
Download Patch File Download Diff File Expand all files Collapse all files

92
examples/research_projects/mlm_wwm/README.md Normal file
View File

@@ -0,0 +1,92 @@
<!---
Copyright 2020 The HuggingFace Team. 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.
-->
## Whole Word Mask Language Model
These scripts leverage the 🤗 Datasets library and the Trainer API. You can easily customize them to your needs if you
need extra processing on your datasets.
The following examples, will run on a datasets hosted on our [hub](https://huggingface.co/datasets) or with your own
text files for training and validation. We give examples of both below.
The BERT authors released a new version of BERT using Whole Word Masking in May 2019. Instead of masking randomly
selected tokens (which may be part of words), they mask randomly selected words (masking all the tokens corresponding
to that word). This technique has been refined for Chinese in [this paper](https://arxiv.org/abs/1906.08101).
To fine-tune a model using whole word masking, use the following script:
```bash
python run_mlm_wwm.py \
--model_name_or_path roberta-base \
--dataset_name wikitext \
--dataset_config_name wikitext-2-raw-v1 \
--do_train \
--do_eval \
--output_dir /tmp/test-mlm-wwm
```
For Chinese models, we need to generate a reference files (which requires the ltp library), because it's tokenized at
the character level.
**Q :** Why a reference file?
**A :** Suppose we have a Chinese sentence like: `我喜欢你` The original Chinese-BERT will tokenize it as
`['我','喜','欢','你']` (character level). But `喜欢` is a whole word. For whole word masking proxy, we need a result
like `['我','喜','##欢','你']`, so we need a reference file to tell the model which position of the BERT original token
should be added `##`.
**Q :** Why LTP ?
**A :** Cause the best known Chinese WWM BERT is [Chinese-BERT-wwm](https://github.com/ymcui/Chinese-BERT-wwm) by HIT.
It works well on so many Chines Task like CLUE (Chinese GLUE). They use LTP, so if we want to fine-tune their model,
we need LTP.
You could run the following:
```bash
export TRAIN_FILE=/path/to/dataset/wiki.train.raw
export LTP_RESOURCE=/path/to/ltp/tokenizer
export BERT_RESOURCE=/path/to/bert/tokenizer
export SAVE_PATH=/path/to/data/ref.txt
python run_chinese_ref.py \
--file_name=path_to_train_or_eval_file \
--ltp=path_to_ltp_tokenizer \
--bert=path_to_bert_tokenizer \
--save_path=path_to_reference_file
```
Then you can run the script like this:
```bash
python run_mlm_wwm.py \
--model_name_or_path roberta-base \
--train_file path_to_train_file \
--validation_file path_to_validation_file \
--train_ref_file path_to_train_chinese_ref_file \
--validation_ref_file path_to_validation_chinese_ref_file \
--do_train \
--do_eval \
--output_dir /tmp/test-mlm-wwm
```
**Note1:** On TPU, you should the flag `--pad_to_max_length` to make sure all your batches have the same length.
**Note2:** And if you have any questions or something goes wrong when runing this code, don't hesitate to pin @wlhgtc.

4
examples/research_projects/mlm_wwm/requirements.txt Normal file
View File

@@ -0,0 +1,4 @@
datasets >= 1.1.3
sentencepiece != 0.1.92
protobuf
ltp

147
examples/research_projects/mlm_wwm/run_chinese_ref.py Normal file
View File

@@ -0,0 +1,147 @@
import argparse
import json
from typing import List
from ltp import LTP
from transformers.models.bert.tokenization_bert import BertTokenizer
def _is_chinese_char(cp):
"""Checks whether CP is the codepoint of a CJK character."""
# This defines a "chinese character" as anything in the CJK Unicode block:
# https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
#
# Note that the CJK Unicode block is NOT all Japanese and Korean characters,
# despite its name. The modern Korean Hangul alphabet is a different block,
# as is Japanese Hiragana and Katakana. Those alphabets are used to write
# space-separated words, so they are not treated specially and handled
# like the all of the other languages.
if (
(cp >= 0x4E00 and cp <= 0x9FFF)
or (cp >= 0x3400 and cp <= 0x4DBF) #
or (cp >= 0x20000 and cp <= 0x2A6DF) #
or (cp >= 0x2A700 and cp <= 0x2B73F) #
or (cp >= 0x2B740 and cp <= 0x2B81F) #
or (cp >= 0x2B820 and cp <= 0x2CEAF) #
or (cp >= 0xF900 and cp <= 0xFAFF)
or (cp >= 0x2F800 and cp <= 0x2FA1F) #
): #
return True
return False
def is_chinese(word: str):
# word like '180' or '身高' or '神'
for char in word:
char = ord(char)
if not _is_chinese_char(char):
return 0
return 1
def get_chinese_word(tokens: List[str]):
word_set = set()
for token in tokens:
chinese_word = len(token) > 1 and is_chinese(token)
if chinese_word:
word_set.add(token)
word_list = list(word_set)
return word_list
def add_sub_symbol(bert_tokens: List[str], chinese_word_set: set()):
if not chinese_word_set:
return bert_tokens
max_word_len = max([len(w) for w in chinese_word_set])
bert_word = bert_tokens
start, end = 0, len(bert_word)
while start < end:
single_word = True
if is_chinese(bert_word[start]):
l = min(end - start, max_word_len)
for i in range(l, 1, -1):
whole_word = "".join(bert_word[start : start + i])
if whole_word in chinese_word_set:
for j in range(start + 1, start + i):
bert_word[j] = "##" + bert_word[j]
start = start + i
single_word = False
break
if single_word:
start += 1
return bert_word
def prepare_ref(lines: List[str], ltp_tokenizer: LTP, bert_tokenizer: BertTokenizer):
ltp_res = []
for i in range(0, len(lines), 100):
res = ltp_tokenizer.seg(lines[i : i + 100])[0]
res = [get_chinese_word(r) for r in res]
ltp_res.extend(res)
assert len(ltp_res) == len(lines)
bert_res = []
for i in range(0, len(lines), 100):
res = bert_tokenizer(lines[i : i + 100], add_special_tokens=True, truncation=True, max_length=512)
bert_res.extend(res["input_ids"])
assert len(bert_res) == len(lines)
ref_ids = []
for input_ids, chinese_word in zip(bert_res, ltp_res):
input_tokens = []
for id in input_ids:
token = bert_tokenizer._convert_id_to_token(id)
input_tokens.append(token)
input_tokens = add_sub_symbol(input_tokens, chinese_word)
ref_id = []
# We only save pos of chinese subwords start with ##, which mean is part of a whole word.
for i, token in enumerate(input_tokens):
if token[:2] == "##":
clean_token = token[2:]
# save chinese tokens' pos
if len(clean_token) == 1 and _is_chinese_char(ord(clean_token)):
ref_id.append(i)
ref_ids.append(ref_id)
assert len(ref_ids) == len(bert_res)
return ref_ids
def main(args):
# For Chinese (Ro)Bert, the best result is from : RoBERTa-wwm-ext (https://github.com/ymcui/Chinese-BERT-wwm)
# If we want to fine-tune these model, we have to use same tokenizer : LTP (https://github.com/HIT-SCIR/ltp)
with open(args.file_name, "r", encoding="utf-8") as f:
data = f.readlines()
data = [line.strip() for line in data if len(line) > 0 and not line.isspace()] # avoid delimiter like '\u2029'
ltp_tokenizer = LTP(args.ltp) # faster in GPU device
bert_tokenizer = BertTokenizer.from_pretrained(args.bert)
ref_ids = prepare_ref(data, ltp_tokenizer, bert_tokenizer)
with open(args.save_path, "w", encoding="utf-8") as f:
data = [json.dumps(ref) + "\n" for ref in ref_ids]
f.writelines(data)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="prepare_chinese_ref")
parser.add_argument(
"--file_name",
type=str,
default="./resources/chinese-demo.txt",
help="file need process, same as training data in lm",
)
parser.add_argument(
"--ltp", type=str, default="./resources/ltp", help="resources for LTP tokenizer, usually a path"
)
parser.add_argument("--bert", type=str, default="./resources/robert", help="resources for Bert tokenizer")
parser.add_argument("--save_path", type=str, default="./resources/ref.txt", help="path to save res")
args = parser.parse_args()
main(args)

408
examples/research_projects/mlm_wwm/run_mlm_wwm.py Normal file
View File

@@ -0,0 +1,408 @@
# coding=utf-8
# Copyright 2020 The HuggingFace Team 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 masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
text file or a dataset.
Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
https://huggingface.co/models?filter=masked-lm
"""
# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
import json
import logging
import math
import os
import sys
from dataclasses import dataclass, field
from typing import Optional
from datasets import Dataset, load_dataset
import transformers
from transformers import (
CONFIG_MAPPING,
MODEL_FOR_MASKED_LM_MAPPING,
AutoConfig,
AutoModelForMaskedLM,
AutoTokenizer,
DataCollatorForWholeWordMask,
HfArgumentParser,
Trainer,
TrainingArguments,
set_seed,
)
from transformers.trainer_utils import get_last_checkpoint, is_main_process
logger = logging.getLogger(__name__)
MODEL_CONFIG_CLASSES = list(MODEL_FOR_MASKED_LM_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."
"Don't set 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 huggingface.co"},
)
use_fast_tokenizer: bool = field(
default=True,
metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
)
model_revision: str = field(
default="main",
metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
)
use_auth_token: bool = field(
default=False,
metadata={
"help": "Will use the token generated when running `transformers-cli login` (necessary to use this script "
"with private models)."
},
)
@dataclass
class DataTrainingArguments:
"""
Arguments pertaining to what data we are going to input our model for training and eval.
"""
dataset_name: Optional[str] = field(
default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
)
dataset_config_name: Optional[str] = field(
default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
)
train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
validation_file: Optional[str] = field(
default=None,
metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
)
train_ref_file: Optional[str] = field(
default=None,
metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
)
validation_ref_file: Optional[str] = field(
default=None,
metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
)
overwrite_cache: bool = field(
default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
)
validation_split_percentage: Optional[int] = field(
default=5,
metadata={
"help": "The percentage of the train set used as validation set in case there's no validation split"
},
)
max_seq_length: Optional[int] = field(
default=None,
metadata={
"help": "The maximum total input sequence length after tokenization. Sequences longer "
"than this will be truncated. Default to the max input length of the model."
},
)
preprocessing_num_workers: Optional[int] = field(
default=None,
metadata={"help": "The number of processes to use for the preprocessing."},
)
mlm_probability: float = field(
default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
)
pad_to_max_length: bool = field(
default=False,
metadata={
"help": "Whether to pad all samples to `max_seq_length`. "
"If False, will pad the samples dynamically when batching to the maximum length in the batch."
},
)
def __post_init__(self):
if self.train_file is not None:
extension = self.train_file.split(".")[-1]
assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
if self.validation_file is not None:
extension = self.validation_file.split(".")[-1]
assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
def add_chinese_references(dataset, ref_file):
with open(ref_file, "r", encoding="utf-8") as f:
refs = [json.loads(line) for line in f.read().splitlines() if (len(line) > 0 and not line.isspace())]
assert len(dataset) == len(refs)
dataset_dict = {c: dataset[c] for c in dataset.column_names}
dataset_dict["chinese_ref"] = refs
return Dataset.from_dict(dataset_dict)
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))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome."
)
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
if "validation" not in datasets.keys():
datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
)
datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
if extension == "txt":
extension = "text"
datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config_kwargs = {
"cache_dir": model_args.cache_dir,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning("You are instantiating a new config instance from scratch.")
tokenizer_kwargs = {
"cache_dir": model_args.cache_dir,
"use_fast": model_args.use_fast_tokenizer,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, **tokenizer_kwargs)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, **tokenizer_kwargs)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"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 = AutoModelForMaskedLM.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,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForMaskedLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
padding = "max_length" if data_args.pad_to_max_length else False
def tokenize_function(examples):
# Remove empty lines
examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
return tokenizer(examples["text"], padding=padding, truncation=True, max_length=data_args.max_seq_length)
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=[text_column_name],
load_from_cache_file=not data_args.overwrite_cache,
)
# Add the chinese references if provided
if data_args.train_ref_file is not None:
tokenized_datasets["train"] = add_chinese_references(tokenized_datasets["train"], data_args.train_ref_file)
if data_args.validation_ref_file is not None:
tokenized_datasets["validation"] = add_chinese_references(
tokenized_datasets["validation"], data_args.validation_ref_file
)
# If we have ref files, need to avoid it removed by trainer
has_ref = data_args.train_ref_file or data_args.validation_ref_file
if has_ref:
training_args.remove_unused_columns = False
# Data collator
# This one will take care of randomly masking the tokens.
data_collator = DataCollatorForWholeWordMask(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_datasets["train"] if training_args.do_train else None,
eval_dataset=tokenized_datasets["validation"] if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
output_train_file = os.path.join(training_args.output_dir, "train_results.txt")
if trainer.is_world_process_zero():
with open(output_train_file, "w") as writer:
logger.info("***** Train results *****")
for key, value in sorted(train_result.metrics.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
# Need to save the state, since Trainer.save_model saves only the tokenizer with the model
trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
results["perplexity"] = perplexity
output_eval_file = os.path.join(training_args.output_dir, "eval_results_mlm_wwm.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in sorted(results.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def _mp_fn(index):
# For xla_spawn (TPUs)
main()
if __name__ == "__main__":
main()