Taizo Kaneko
commited on
Commit
•
97c46f0
1
Parent(s):
3b2e896
commit files to HF hub
Browse files- .gitattributes +1 -0
- config.json +27 -0
- fasttext_fsc.py +56 -0
- fasttext_jp_embedding.py +48 -0
- fasttext_jp_tokenizer.py +143 -0
- mecab_tokenizer.py +92 -0
- pytorch_model.bin +3 -0
- special_tokens_map.json +4 -0
- tokenizer_config.json +12 -0
- vocab.txt +3 -0
.gitattributes
CHANGED
@@ -32,3 +32,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
32 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
33 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
34 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
32 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
33 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
34 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
35 |
+
vocab.txt filter=lfs diff=lfs merge=lfs -text
|
config.json
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architectures": [
|
3 |
+
"FastTextForSeuqenceClassification"
|
4 |
+
],
|
5 |
+
"auto_map": {
|
6 |
+
"AutoConfig": "fasttext_jp_embedding.FastTextJpConfig",
|
7 |
+
"AutoModel": "fasttext_fsc.FastTextForSeuqenceClassification"
|
8 |
+
},
|
9 |
+
"hidden_size": 300,
|
10 |
+
"id2label": {
|
11 |
+
"0": "entailment",
|
12 |
+
"1": "neutral",
|
13 |
+
"2": "contradiction"
|
14 |
+
},
|
15 |
+
"label2id": {
|
16 |
+
"contradiction": 2,
|
17 |
+
"entailment": 0,
|
18 |
+
"neutral": 1
|
19 |
+
},
|
20 |
+
"max_length": 128,
|
21 |
+
"model_type": "fasttext_jp",
|
22 |
+
"tokenizerI_class": "FastTextJpTokenizer",
|
23 |
+
"tokenizer_class": "FastTextJpTokenizer",
|
24 |
+
"torch_dtype": "float32",
|
25 |
+
"transformers_version": "4.23.1",
|
26 |
+
"vocab_size": 500
|
27 |
+
}
|
fasttext_fsc.py
ADDED
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from __future__ import annotations
|
2 |
+
from transformers import PretrainedConfig
|
3 |
+
from torch import nn
|
4 |
+
import torch
|
5 |
+
from torchtyping import TensorType
|
6 |
+
from .fasttext_jp_embedding import FastTextJpModel, FastTextJpConfig
|
7 |
+
from transformers.modeling_outputs import SequenceClassifierOutput
|
8 |
+
|
9 |
+
|
10 |
+
class FastTextForSeuqenceClassification(FastTextJpModel):
|
11 |
+
"""FastTextのベクトルをベースとした分類を行います。
|
12 |
+
"""
|
13 |
+
|
14 |
+
def __init__(self, config: FastTextJpConfig):
|
15 |
+
super().__init__(config)
|
16 |
+
|
17 |
+
def forward(self, **inputs) -> SequenceClassifierOutput:
|
18 |
+
"""embeddingを行います。
|
19 |
+
|
20 |
+
Returns:
|
21 |
+
TensorType["batch", "word", "vectors"]: 単語ごとにベクトルを返します。
|
22 |
+
"""
|
23 |
+
input_ids = inputs["input_ids"]
|
24 |
+
outputs = self.word_embeddings(input_ids)
|
25 |
+
sentence = outputs[torch.logical_and(inputs["attention_mask"] == 1,
|
26 |
+
inputs["token_type_ids"] == 0)]
|
27 |
+
candidate_label = outputs[torch.logical_and(
|
28 |
+
inputs["attention_mask"] == 1, inputs["token_type_ids"] == 1)]
|
29 |
+
|
30 |
+
sentence_mean = torch.mean(sentence, dim=-2, keepdim=True)
|
31 |
+
candidate_label_mean = torch.mean(candidate_label,
|
32 |
+
dim=-2,
|
33 |
+
keepdim=True)
|
34 |
+
if sentence_mean.dim() == 2:
|
35 |
+
p = torch.nn.functional.cosine_similarity(sentence_mean,
|
36 |
+
candidate_label_mean,
|
37 |
+
dim=1)
|
38 |
+
logits = [[torch.log(p), -torch.inf, torch.log(1 - p)]]
|
39 |
+
else:
|
40 |
+
logits = []
|
41 |
+
# batch
|
42 |
+
for sm, clm in zip(sentence_mean, candidate_label_mean):
|
43 |
+
p = torch.nn.functional.cosine_similarity(sm, clm, dim=1)
|
44 |
+
logits.append([[torch.log(p), -torch.inf, torch.log(1 - p)]])
|
45 |
+
logits = torch.FloatTensor(logits)
|
46 |
+
return SequenceClassifierOutput(
|
47 |
+
loss=None,
|
48 |
+
logits=logits, # type: ignore
|
49 |
+
hidden_states=None,
|
50 |
+
attentions=None,
|
51 |
+
)
|
52 |
+
|
53 |
+
|
54 |
+
# AutoModelに登録が必要だが、いろいろやり方が変わっているようで定まっていない。(2022/11/6)
|
55 |
+
# https://huggingface.co/docs/transformers/custom_models#sending-the-code-to-the-hub
|
56 |
+
FastTextForSeuqenceClassification.register_for_auto_class("AutoModel")
|
fasttext_jp_embedding.py
ADDED
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from __future__ import annotations
|
2 |
+
from transformers import PretrainedConfig
|
3 |
+
from transformers import PreTrainedModel
|
4 |
+
from torch import nn
|
5 |
+
import torch
|
6 |
+
from torchtyping import TensorType
|
7 |
+
|
8 |
+
|
9 |
+
class FastTextJpConfig(PretrainedConfig):
|
10 |
+
"""FastTextJpModelのConfig
|
11 |
+
"""
|
12 |
+
model_type = "fasttext_jp"
|
13 |
+
|
14 |
+
def __init__(self, tokenizer_class="FastTextJpTokenizer", **kwargs):
|
15 |
+
"""初期化処理
|
16 |
+
|
17 |
+
Args:
|
18 |
+
tokenizer_class (str, optional):
|
19 |
+
tokenizer_classを指定しないと、pipelineから読み込まれません。
|
20 |
+
config.jsonに記載されます。
|
21 |
+
"""
|
22 |
+
kwargs["tokenizer_class"] = tokenizer_class
|
23 |
+
super().__init__(**kwargs)
|
24 |
+
|
25 |
+
|
26 |
+
class FastTextJpModel(PreTrainedModel):
|
27 |
+
"""FastTextのEmbeddingを行います。
|
28 |
+
"""
|
29 |
+
config_class = FastTextJpConfig
|
30 |
+
|
31 |
+
def __init__(self, config: FastTextJpConfig):
|
32 |
+
super().__init__(config)
|
33 |
+
self.word_embeddings = nn.Embedding(config.vocab_size,
|
34 |
+
config.hidden_size)
|
35 |
+
|
36 |
+
def forward(self, **inputs) -> TensorType["batch", "word", "vectors"]:
|
37 |
+
"""embeddingを行います。
|
38 |
+
|
39 |
+
Returns:
|
40 |
+
TensorType["batch", "word", "vectors"]: 単語ごとにベクトルを返します。
|
41 |
+
"""
|
42 |
+
return self.word_embeddings(torch.Tensor(inputs["input_ids"]))
|
43 |
+
|
44 |
+
|
45 |
+
# AutoModelに登録が必要だが、いろいろやり方が変わっているようで定まっていない。(2022/11/6)
|
46 |
+
# https://huggingface.co/docs/transformers/custom_models#sending-the-code-to-the-hub
|
47 |
+
FastTextJpConfig.register_for_auto_class()
|
48 |
+
FastTextJpModel.register_for_auto_class("AutoModel")
|
fasttext_jp_tokenizer.py
ADDED
@@ -0,0 +1,143 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from __future__ import annotations
|
2 |
+
from .mecab_tokenizer import MeCabTokenizer
|
3 |
+
import os
|
4 |
+
|
5 |
+
VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt"}
|
6 |
+
|
7 |
+
|
8 |
+
def save_stoi(stoi: dict[str, int], vocab_file: str):
|
9 |
+
"""単語IDの辞書を配列にしてvocab_fileに保存します。
|
10 |
+
|
11 |
+
Args:
|
12 |
+
stoi (dict[str, int]): 単語IDのマッピング
|
13 |
+
vocab_file (str): 保存するパス
|
14 |
+
|
15 |
+
Raises:
|
16 |
+
ValueError: IDが途切れているとエラーを起こします。
|
17 |
+
"""
|
18 |
+
|
19 |
+
with open(vocab_file, "w", encoding="utf-8") as writer:
|
20 |
+
index = 0
|
21 |
+
for token, token_index in sorted(stoi.items(), key=lambda kv: kv[1]):
|
22 |
+
if index != token_index:
|
23 |
+
raise ValueError(
|
24 |
+
"Saving vocabulary to {vocab_file}: vocabulary indices are not consecutive."
|
25 |
+
" Please check that the vocabulary is not corrupted!")
|
26 |
+
writer.write(token + "\n")
|
27 |
+
index += 1
|
28 |
+
|
29 |
+
|
30 |
+
def load_stoi(vocab_file: str) -> dict[str, int]:
|
31 |
+
"""ファイルから単語IDの辞書をロードします。
|
32 |
+
|
33 |
+
Args:
|
34 |
+
vocab_file (str): ファイルのパス
|
35 |
+
|
36 |
+
Returns:
|
37 |
+
dict[str, int]: 単語IDのマッピング
|
38 |
+
"""
|
39 |
+
|
40 |
+
stoi: dict[str, int] = {}
|
41 |
+
# ファイルから読み出し
|
42 |
+
with open(vocab_file, "r", encoding="utf-8") as reader:
|
43 |
+
tokens = reader.readlines()
|
44 |
+
|
45 |
+
# 単語IDのマッピングを生成します。
|
46 |
+
for index, token in enumerate(tokens):
|
47 |
+
token = token.rstrip("\n")
|
48 |
+
stoi[token] = index
|
49 |
+
return stoi
|
50 |
+
|
51 |
+
|
52 |
+
class FastTextJpTokenizer(MeCabTokenizer):
|
53 |
+
|
54 |
+
# Configが認識するのに必要です。
|
55 |
+
# https://huggingface.co/docs/transformers/custom_models#writing-a-custom-configuration
|
56 |
+
model_type = "fasttext_jp"
|
57 |
+
|
58 |
+
# vocab.txtを認識するのにおそらく必要。
|
59 |
+
vocab_files_names = VOCAB_FILES_NAMES
|
60 |
+
|
61 |
+
def __init__(self,
|
62 |
+
vocab_file: str,
|
63 |
+
hinshi: list[str] | None = None,
|
64 |
+
mecab_dicdir: str | None = None,
|
65 |
+
**kwargs):
|
66 |
+
"""初期化処理
|
67 |
+
|
68 |
+
Args:
|
69 |
+
vocab_file (str): vocab_fileのpath
|
70 |
+
hinshi (list[str] | None, optional): 抽出する品詞
|
71 |
+
mecab_dicdir (str | None, optional): dicrcのあるディレクトリ
|
72 |
+
"""
|
73 |
+
super().__init__(hinshi, mecab_dicdir, **kwargs)
|
74 |
+
|
75 |
+
if not os.path.isfile(vocab_file):
|
76 |
+
raise ValueError(
|
77 |
+
f"Can't find a vocabulary file at path '{vocab_file}'. To load the vocabulary from a Google pretrained"
|
78 |
+
" model use `tokenizer = BertTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`"
|
79 |
+
)
|
80 |
+
self.stoi = load_stoi(vocab_file)
|
81 |
+
self.itos = dict([(ids, tok) for tok, ids in self.stoi.items()])
|
82 |
+
|
83 |
+
@property
|
84 |
+
def vocab_size(self) -> int:
|
85 |
+
"""ボキャブラリのサイズ
|
86 |
+
※PreTrainedTokenizerで実装すべき必須の関数。
|
87 |
+
|
88 |
+
Returns:
|
89 |
+
int: ボキャブラリのサイズ
|
90 |
+
"""
|
91 |
+
return len(self.stoi)
|
92 |
+
|
93 |
+
def _convert_token_to_id(self, token: str) -> int:
|
94 |
+
"""単語からID
|
95 |
+
※PreTrainedTokenizerで実装すべき必須の関数。
|
96 |
+
|
97 |
+
Args:
|
98 |
+
token (str): 単語
|
99 |
+
|
100 |
+
Returns:
|
101 |
+
int: ID
|
102 |
+
"""
|
103 |
+
return self.stoi[token]
|
104 |
+
|
105 |
+
def _convert_id_to_token(self, index: int) -> str:
|
106 |
+
"""IDから単語
|
107 |
+
※PreTrainedTokenizerで実装すべき必須の関数。
|
108 |
+
|
109 |
+
Args:
|
110 |
+
index (int): ID
|
111 |
+
|
112 |
+
Returns:
|
113 |
+
str: 単語
|
114 |
+
"""
|
115 |
+
return self.itos[index]
|
116 |
+
|
117 |
+
def save_vocabulary(self,
|
118 |
+
save_directory: str,
|
119 |
+
filename_prefix: str | None = None) -> tuple[str]:
|
120 |
+
"""ボキャブラリの保存
|
121 |
+
|
122 |
+
Args:
|
123 |
+
save_directory (str): 保存するディレクトリ。ファイル名はvocab.txtに固定
|
124 |
+
filename_prefix (str | None, optional): ファイルのprefix
|
125 |
+
|
126 |
+
Returns:
|
127 |
+
tuple[str]: ファイル名を返す。
|
128 |
+
"""
|
129 |
+
if os.path.isdir(save_directory):
|
130 |
+
vocab_file = os.path.join(
|
131 |
+
save_directory,
|
132 |
+
(filename_prefix + "-" if filename_prefix else "") +
|
133 |
+
VOCAB_FILES_NAMES["vocab_file"])
|
134 |
+
else:
|
135 |
+
vocab_file = (filename_prefix +
|
136 |
+
"-" if filename_prefix else "") + save_directory
|
137 |
+
save_stoi(self.stoi, vocab_file)
|
138 |
+
return (vocab_file, )
|
139 |
+
|
140 |
+
|
141 |
+
# AutoTokenizerに登録が必要だが、いろいろやり方が変わっているようで定まっていない。(2022/11/6)
|
142 |
+
# https://huggingface.co/docs/transformers/custom_models#sending-the-code-to-the-hub
|
143 |
+
FastTextJpTokenizer.register_for_auto_class("AutoTokenizer")
|
mecab_tokenizer.py
ADDED
@@ -0,0 +1,92 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from __future__ import annotations
|
2 |
+
from typing import NamedTuple
|
3 |
+
import MeCab
|
4 |
+
from transformers import PreTrainedTokenizer
|
5 |
+
|
6 |
+
|
7 |
+
class MeCabResult(NamedTuple):
|
8 |
+
"""MeCab解析結果の型
|
9 |
+
"""
|
10 |
+
hyosokei: str
|
11 |
+
hinshi: str
|
12 |
+
hinshi_saibunrui_1: str
|
13 |
+
hinshi_saibunrui_2: str
|
14 |
+
hinshi_saibunrui_3: str
|
15 |
+
katsuyokei_1: str
|
16 |
+
katsuyokei_2: str
|
17 |
+
genkei: str
|
18 |
+
yomi: str
|
19 |
+
hatsuon: str
|
20 |
+
|
21 |
+
|
22 |
+
class MeCabTokenizer(PreTrainedTokenizer):
|
23 |
+
|
24 |
+
def __init__(self,
|
25 |
+
hinshi: list[str] | None = None,
|
26 |
+
mecab_dicdir: str | None = None,
|
27 |
+
**kwargs):
|
28 |
+
"""初期化処理
|
29 |
+
|
30 |
+
Args:
|
31 |
+
hinshi (list[str] | None): 抽出する品詞
|
32 |
+
mecab_dicdir (str | None, optional): dicrcのあるディレクトリ
|
33 |
+
"""
|
34 |
+
|
35 |
+
self.target_hinshi = hinshi
|
36 |
+
if mecab_dicdir is not None:
|
37 |
+
self.mecab = MeCab.Tagger(f"-d {mecab_dicdir}")
|
38 |
+
else:
|
39 |
+
self.mecab = MeCab.Tagger()
|
40 |
+
|
41 |
+
super().__init__(**kwargs)
|
42 |
+
|
43 |
+
def _tokenize(self, text: str) -> list[str]:
|
44 |
+
"""文章から特定の品詞の単語を返します。
|
45 |
+
|
46 |
+
Args:
|
47 |
+
text (str): 文章
|
48 |
+
|
49 |
+
Returns:
|
50 |
+
list[str]: 特定の品詞の単語
|
51 |
+
"""
|
52 |
+
|
53 |
+
out = []
|
54 |
+
# Mecabで分析します。
|
55 |
+
result_words = self.mecab_analyze(text)
|
56 |
+
for result_word in result_words:
|
57 |
+
# 最初と最後は空文字
|
58 |
+
if result_word.hyosokei == "":
|
59 |
+
continue
|
60 |
+
if self.target_hinshi is not None:
|
61 |
+
if result_word.hinshi in self.target_hinshi:
|
62 |
+
# 特定の品詞のみ返します。
|
63 |
+
out.append(result_word.hyosokei)
|
64 |
+
else:
|
65 |
+
continue
|
66 |
+
else:
|
67 |
+
out.append(result_word.hyosokei)
|
68 |
+
return out
|
69 |
+
|
70 |
+
def mecab_analyze(self, text: str) -> list[MeCabResult]:
|
71 |
+
"""文章をMecabで分析します。
|
72 |
+
|
73 |
+
Args:
|
74 |
+
text (str): 文章
|
75 |
+
|
76 |
+
Returns:
|
77 |
+
list[MeCabResult]: MeCabの解析結果
|
78 |
+
"""
|
79 |
+
node = self.mecab.parseToNode(text)
|
80 |
+
#形態素1つ1つを処理
|
81 |
+
out = []
|
82 |
+
while node:
|
83 |
+
args = []
|
84 |
+
args.append(node.surface)
|
85 |
+
feature = node.feature.split(",")
|
86 |
+
args.extend(feature)
|
87 |
+
mecab_result = MeCabResult(args[0], args[1], args[2], args[3],
|
88 |
+
args[4], args[5], args[6], args[7],
|
89 |
+
args[8], args[9])
|
90 |
+
out.append(mecab_result)
|
91 |
+
node = node.next # 最後のEOSを省く
|
92 |
+
return out
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:6bd689ac294c1623aa08045af576207197ec480898bbb9f4057b062f63cfdf4f
|
3 |
+
size 600829
|
special_tokens_map.json
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"pad_token": "*",
|
3 |
+
"unk_token": "*"
|
4 |
+
}
|
tokenizer_config.json
ADDED
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"auto_map": {
|
3 |
+
"AutoTokenizer": [
|
4 |
+
"fasttext_jp_tokenizer.FastTextJpTokenizer",
|
5 |
+
null
|
6 |
+
]
|
7 |
+
},
|
8 |
+
"model_max_length": 128,
|
9 |
+
"pad_token": "*",
|
10 |
+
"tokenizer_class": "FastTextJpTokenizer",
|
11 |
+
"unk_token": "*"
|
12 |
+
}
|
vocab.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:3a1770ed0a47f44e882afc3f56271a16bc8dba675f18dd61e2cffac276b49acc
|
3 |
+
size 29910902
|