# Add whole word mask support for lm fine-tune (#7925)

* ADD: add whole word mask proxy for both eng and chinese

* MOD: adjust format

* MOD: reformat code

* MOD: update import

* MOD: fix bug

* MOD: add import

* MOD: fix bug

* MOD: decouple code and update readme

* MOD: reformat code

* Update examples/language-modeling/README.md

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update examples/language-modeling/README.md

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update examples/language-modeling/run_language_modeling.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update examples/language-modeling/run_language_modeling.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update examples/language-modeling/run_language_modeling.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update examples/language-modeling/run_language_modeling.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* change wwm to whole_word_mask

* reformat code

* reformat

* format

* Code quality

* ADD: update chinese ref readme

* MOD: small changes

* MOD: small changes2

* update readme

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <sylvain.gugger@gmail.com>

examples/language-modeling/README.md

@@ -45,6 +45,8 @@ slightly slower (over-fitting takes more epochs).
 
 We use the `--mlm` flag so that the script may change its loss function.
 
+If using whole-word masking, use both the`--mlm` and `--wwm` flags.
+
 ```bash
 export TRAIN_FILE=/path/to/dataset/wiki.train.raw
 export TEST_FILE=/path/to/dataset/wiki.test.raw
@@ -57,7 +59,55 @@ python run_language_modeling.py \
     --train_data_file=$TRAIN_FILE \
     --do_eval \
     --eval_data_file=$TEST_FILE \
-    --mlm
+    --mlm \
+    --wwm
+```
+
+For Chinese models, it's same with English model with only --mlm`. If using whole-word masking, we need to generate a reference files, case it's char level.
+
+**Q :** Why ref 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']` in char level.
+Actually, `喜欢` is a whole word. For whole word mask proxy, We need res like `['我','喜','##欢','你']`.
+So we need a ref file to tell model which pos of 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.
+
+```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 chinese_ref.py \
+    --file_name=$TRAIN_FILE \
+    --ltp=$LTP_RESOURCE
+    --bert=$BERT_RESOURCE \
+    --save_path=$SAVE_PATH 
+```
+Now Chinese Ref is only supported by `LineByLineWithRefDataset` Class, so we need add `line_by_line` flag: 
+
+
+```bash
+export TRAIN_FILE=/path/to/dataset/wiki.train.raw
+export TEST_FILE=/path/to/dataset/wiki.test.raw
+export REF_FILE=/path/to/ref.txt
+
+python run_language_modeling.py \
+    --output_dir=output \
+    --model_type=roberta \
+    --model_name_or_path=roberta-base \
+    --do_train \
+    --train_data_file=$TRAIN_FILE \
+    --chinese_ref_file=$REF_FILE \
+    --do_eval \
+    --eval_data_file=$TEST_FILE \
+    --mlm \
+    --line_by_line \
+    --wwm
 ```
 
 ### XLNet and permutation language modeling

examples/language-modeling/chinese_ref.py

@@ -0,0 +1,147 @@
+import argparse
+import json
+from typing import List
+
+from ltp import LTP
+from transformers.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()
+
+    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)

examples/language-modeling/run_language_modeling.py

@@ -37,8 +37,10 @@ from transformers import (
     AutoTokenizer,
     DataCollatorForLanguageModeling,
     DataCollatorForPermutationLanguageModeling,
+    DataCollatorForWholeWordMask,
     HfArgumentParser,
     LineByLineTextDataset,
+    LineByLineWithRefDataset,
     PreTrainedTokenizer,
     TextDataset,
     Trainer,
@@ -101,6 +103,10 @@ class DataTrainingArguments:
         default=None,
         metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
     )
+    chinese_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input ref data file for whole word mask in Chinees."},
+    )
     line_by_line: bool = field(
         default=False,
         metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
@@ -109,6 +115,7 @@ class DataTrainingArguments:
     mlm: bool = field(
         default=False, metadata={"help": "Train with masked-language modeling loss instead of language modeling."}
     )
+    whole_word_mask: bool = field(default=False, metadata={"help": "Whether ot not to use whole word mask."})
     mlm_probability: float = field(
         default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
     )
@@ -143,6 +150,16 @@ def get_dataset(
 ):
     def _dataset(file_path):
         if args.line_by_line:
+            if args.chinese_ref_file is not None:
+                if not args.whole_word_mask or not args.mlm:
+                    raise ValueError("You need to set world whole masking and mlm to True for Chinese Whole Word Mask")
+                return LineByLineWithRefDataset(
+                    tokenizer=tokenizer,
+                    file_path=file_path,
+                    block_size=args.block_size,
+                    ref_path=args.chinese_ref_file,
+                )
+
             return LineByLineTextDataset(tokenizer=tokenizer, file_path=file_path, block_size=args.block_size)
         else:
             return TextDataset(
@@ -174,7 +191,6 @@ def main():
             "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)
@@ -269,6 +285,11 @@ def main():
             plm_probability=data_args.plm_probability,
             max_span_length=data_args.max_span_length,
         )
+    else:
+        if data_args.mlm and data_args.whole_word_mask:
+            data_collator = DataCollatorForWholeWordMask(
+                tokenizer=tokenizer, mlm_probability=data_args.mlm_probability
+            )
         else:
             data_collator = DataCollatorForLanguageModeling(
                 tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability

src/transformers/__init__.py

@@ -284,6 +284,7 @@ if is_torch_available():
         DataCollatorForNextSentencePrediction,
         DataCollatorForPermutationLanguageModeling,
         DataCollatorForSOP,
+        DataCollatorForWholeWordMask,
         DataCollatorWithPadding,
         default_data_collator,
     )
@@ -291,6 +292,7 @@ if is_torch_available():
         GlueDataset,
         GlueDataTrainingArguments,
         LineByLineTextDataset,
+        LineByLineWithRefDataset,
         LineByLineWithSOPTextDataset,
         SquadDataset,
         SquadDataTrainingArguments,

src/transformers/data/data_collator.py

@@ -1,3 +1,4 @@
+import random
 from dataclasses import dataclass
 from typing import Any, Callable, Dict, List, NewType, Optional, Tuple, Union
 
@@ -195,6 +196,124 @@ class DataCollatorForLanguageModeling:
         return inputs, labels
 
 
+@dataclass
+class DataCollatorForWholeWordMask(DataCollatorForLanguageModeling):
+    """
+    Data collator used for language modeling.
+    - collates batches of tensors, honoring their tokenizer's pad_token
+    - preprocesses batches for masked language modeling
+    """
+
+    def __call__(
+        self, examples: List[Union[List[int], torch.Tensor, Dict[str, torch.Tensor]]]
+    ) -> Dict[str, torch.Tensor]:
+        if isinstance(examples[0], (dict, BatchEncoding)):
+            input_ids = [e["input_ids"] for e in examples]
+        else:
+            input_ids = examples
+            examples = [{"input_ids": e} for e in examples]
+
+        batch_input = self._tensorize_batch(input_ids)
+
+        mask_labels = []
+        for e in examples:
+            ref_tokens = []
+            for id in e["input_ids"].tolist():
+                token = self.tokenizer._convert_id_to_token(id)
+                ref_tokens.append(token)
+
+            # For Chinese tokens, we need extra inf to mark sub-word, e.g [喜,欢]-> [喜，##欢]
+            if "chinese_ref" in e:
+                ref_pos = e["chinese_ref"].tolist()
+                len_seq = e["input_ids"].size(0)
+                for i in range(len_seq):
+                    if i in ref_pos:
+                        ref_tokens[i] = "##" + ref_tokens[i]
+            mask_labels.append(self._whole_word_mask(ref_tokens))
+        batch_mask = self._tensorize_batch(mask_labels)
+        inputs, labels = self.mask_tokens(batch_input, batch_mask)
+        return {"input_ids": inputs, "labels": labels}
+
+    def _whole_word_mask(self, input_tokens: List[str], max_predictions=512):
+        """
+        Get 0/1 labels for masked tokens with whole word mask proxy
+        """
+
+        cand_indexes = []
+        for (i, token) in enumerate(input_tokens):
+            if token == "[CLS]" or token == "[SEP]":
+                continue
+
+            if len(cand_indexes) >= 1 and token.startswith("##"):
+                cand_indexes[-1].append(i)
+            else:
+                cand_indexes.append([i])
+
+        random.shuffle(cand_indexes)
+        num_to_predict = min(max_predictions, max(1, int(round(len(input_tokens) * self.mlm_probability))))
+        masked_lms = []
+        covered_indexes = set()
+        for index_set in cand_indexes:
+            if len(masked_lms) >= num_to_predict:
+                break
+            # If adding a whole-word mask would exceed the maximum number of
+            # predictions, then just skip this candidate.
+            if len(masked_lms) + len(index_set) > num_to_predict:
+                continue
+            is_any_index_covered = False
+            for index in index_set:
+                if index in covered_indexes:
+                    is_any_index_covered = True
+                    break
+            if is_any_index_covered:
+                continue
+            for index in index_set:
+                covered_indexes.add(index)
+                masked_lms.append(index)
+
+        assert len(covered_indexes) == len(masked_lms)
+        mask_labels = [1 if i in covered_indexes else 0 for i in range(len(input_tokens))]
+        return mask_labels
+
+    def mask_tokens(self, inputs: torch.Tensor, mask_labels: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
+        Set 'mask_labels' means we use whole word mask (wwm), we directly mask idxs according to it's ref.
+        """
+
+        if self.tokenizer.mask_token is None:
+            raise ValueError(
+                "This tokenizer does not have a mask token which is necessary for masked language modeling. Remove the --mlm flag if you want to use this tokenizer."
+            )
+        labels = inputs.clone()
+        # We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
+
+        probability_matrix = mask_labels
+
+        special_tokens_mask = [
+            self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()
+        ]
+        probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
+        if self.tokenizer._pad_token is not None:
+            padding_mask = labels.eq(self.tokenizer.pad_token_id)
+            probability_matrix.masked_fill_(padding_mask, value=0.0)
+
+        masked_indices = probability_matrix.bool()
+        labels[~masked_indices] = -100  # We only compute loss on masked tokens
+
+        # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
+        indices_replaced = torch.bernoulli(torch.full(labels.shape, 0.8)).bool() & masked_indices
+        inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
+
+        # 10% of the time, we replace masked input tokens with random word
+        indices_random = torch.bernoulli(torch.full(labels.shape, 0.5)).bool() & masked_indices & ~indices_replaced
+        random_words = torch.randint(len(self.tokenizer), labels.shape, dtype=torch.long)
+        inputs[indices_random] = random_words[indices_random]
+
+        # The rest of the time (10% of the time) we keep the masked input tokens unchanged
+        return inputs, labels
+
+
 @dataclass
 class DataCollatorForSOP(DataCollatorForLanguageModeling):
     """

src/transformers/data/datasets/__init__.py

@@ -5,6 +5,7 @@
 from .glue import GlueDataset, GlueDataTrainingArguments
 from .language_modeling import (
     LineByLineTextDataset,
+    LineByLineWithRefDataset,
     LineByLineWithSOPTextDataset,
     TextDataset,
     TextDatasetForNextSentencePrediction,

src/transformers/data/datasets/language_modeling.py

@@ -1,3 +1,4 @@
+import json
 import os
 import pickle
 import random
@@ -106,12 +107,48 @@ class LineByLineTextDataset(Dataset):
 
         batch_encoding = tokenizer(lines, add_special_tokens=True, truncation=True, max_length=block_size)
         self.examples = batch_encoding["input_ids"]
+        self.examples = [{"input_ids": torch.tensor(e, dtype=torch.long)} for e in self.examples]
 
     def __len__(self):
         return len(self.examples)
 
-    def __getitem__(self, i) -> torch.Tensor:
+    def __getitem__(self, i) -> Dict[str, torch.tensor]:
-        return torch.tensor(self.examples[i], dtype=torch.long)
+        return self.examples[i]
+
+
+class LineByLineWithRefDataset(Dataset):
+    """
+    This will be superseded by a framework-agnostic approach
+    soon.
+    """
+
+    def __init__(self, tokenizer: PreTrainedTokenizer, file_path: str, block_size: int, ref_path: str):
+        assert os.path.isfile(file_path), f"Input file path {file_path} not found"
+        assert os.path.isfile(ref_path), f"Ref file path {file_path} not found"
+        # Here, we do not cache the features, operating under the assumption
+        # that we will soon use fast multithreaded tokenizers from the
+        # `tokenizers` repo everywhere =)
+        logger.info("Creating features from dataset file at %s", file_path)
+        logger.info("Use ref segment results at %s", ref_path)
+        with open(file_path, encoding="utf-8") as f:
+            data = [line for line in f.read().splitlines() if (len(line) > 0 and not line.isspace())]
+        batch_encoding = tokenizer(data, add_special_tokens=True, truncation=True, max_length=block_size)
+        self.examples = batch_encoding["input_ids"]
+        self.examples = [{"input_ids": torch.tensor(e, dtype=torch.long)} for e in self.examples]
+
+        # Get ref inf from file
+        with open(ref_path, encoding="utf-8") as f:
+            ref = [json.loads(line) for line in f.read().splitlines() if (len(line) > 0 and not line.isspace())]
+        assert len(data) == len(ref)
+        n = len(self.examples)
+        for i in range(n):
+            self.examples[i]["chinese_ref"] = torch.tensor(ref[i], dtype=torch.long)
+
+    def __len__(self):
+        return len(self.examples)
+
+    def __getitem__(self, i) -> Dict[str, torch.tensor]:
+        return self.examples[i]
 
 
 class LineByLineWithSOPTextDataset(Dataset):

src/transformers/utils/dummy_pt_objects.py

@@ -45,6 +45,11 @@ class DataCollatorForSOP:
         requires_pytorch(self)
 
 
+class DataCollatorForWholeWordMask:
+    def __init__(self, *args, **kwargs):
+        requires_pytorch(self)
+
+
 class DataCollatorWithPadding:
     def __init__(self, *args, **kwargs):
         requires_pytorch(self)
@@ -69,6 +74,11 @@ class LineByLineTextDataset:
         requires_pytorch(self)
 
 
+class LineByLineWithRefDataset:
+    def __init__(self, *args, **kwargs):
+        requires_pytorch(self)
+
+
 class LineByLineWithSOPTextDataset:
     def __init__(self, *args, **kwargs):
         requires_pytorch(self)