Spaces:
Runtime error
Runtime error
File size: 9,337 Bytes
f4fac26 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 |
{
"cells": [
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"import pyarrow.parquet as pq \n",
"from transformers import AutoTokenizer, PreTrainedTokenizerFast\n",
"from rich import progress"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"#### 训练集数据训练tokenizer，小于16G内存的机器容易OOM"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"pq_file = '../data/my_dataset.shuffle.parquet'\n",
"pf = pq.read_table(pq_file)\n",
"\n",
"def get_training_corpus():\n",
" buffer = []\n",
" for prompt, response in progress.track(zip(pf['prompt'], pf['response']), total=pf.num_rows):\n",
"\n",
" buffer.append(\n",
" f\"{prompt.as_py()}\\n{response.as_py()}\"\n",
" )\n",
"\n",
" if len(buffer) >= 1000:\n",
" yield buffer\n",
" buffer = []\n",
"\n",
" if buffer: yield buffer\n",
"iter_training_corpus = get_training_corpus()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## step 1： 加载T5模型自带的tokenizer"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"old_tokenizer = AutoTokenizer.from_pretrained('t5-base')"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## step 2: 加载Wiki中文语料，1.6GB\n",
"备注: 全量预训练语料文本大小约7GB"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"lines = []\n",
"with open('../data/raw_data/wiki.simple.txt', 'r', encoding='utf-8') as f:\n",
" lines = f.readlines()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"len(lines)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## step 3 定义一个语料的迭代生成器\n",
"一个文本块（段落）的最小长度为2048，迭代一次返回1000个文本块"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"def get_training_corpus():\n",
" buffer = []\n",
" i = 0 \n",
" txt = []\n",
" len_cnt = 0\n",
" for line in progress.track(lines):\n",
" \n",
" len_cnt += len(line)\n",
" txt.append(line)\n",
" if len_cnt >= 2048:\n",
" buffer.append(\n",
" ''.join(txt)\n",
" )\n",
" txt = []\n",
" len_cnt = 0\n",
" \n",
" if len(buffer) >= 1000:\n",
" yield buffer\n",
" buffer = []\n",
" i += 1\n",
"\n",
" # yield last buffer\n",
" if len(buffer) > 0:\n",
" yield buffer\n",
"\n",
"iter_training_corpus = get_training_corpus()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"for i in get_training_corpus():\n",
" print(len(i))\n",
" print([len(t) for t in i][0:20])\n",
" break\n",
"## 1000\n",
"## [2104, 2053, 2176, 2224, 2172, 2068, 2054, 2258, 2058, 2085, 2142, 2274, 2184, 2246, 2144, 2223, 2075, 2058, 2164, 2178]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## step 4: 训练tokenizer"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"tokenizer = old_tokenizer.train_new_from_iterator(iter_training_corpus, vocab_size=40960)\n",
"\n",
"# cpu计算密集型任务 13600K大概需要1个小时，最大内存占用20G"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## step 5: 保存训练好的tokenizer"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"tokenizer.save_pretrained('../model_save/my_tokenizer_wiki')"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# 补充内容： 自定义模型、及特殊字符训练"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from transformers import PreTrainedTokenizerFast\n",
"from tokenizers.pre_tokenizers import Whitespace, Punctuation, Digits, ByteLevel, Metaspace\n",
"from tokenizers.normalizers import NFKC\n",
"from tokenizers import Tokenizer, decoders\n",
"from tokenizers.models import BPE\n",
"import tokenizers"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 字符级别的 BPE toeknizer"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"model = BPE(unk_token=\"[UNK]\")\n",
"tokenizer = Tokenizer(model)\n",
"\n",
"# 用兼容等价分解合并对utf编码进行等价组合，比如全角A转换为半角A\n",
"tokenizer.normalizer = tokenizers.normalizers.Sequence([NFKC()])\n",
"\n",
"# 标点符号，数字，及Metaspace预分割（否则decode出来没有空格）\n",
"tokenizer.pre_tokenizer = tokenizers.pre_tokenizers.Sequence(\n",
" [Punctuation(), Digits(individual_digits=True), Metaspace()])\n",
"\n",
"tokenizer.add_special_tokens([\"[PAD]\",\"[EOS]\",\"[SEP]\",\"[BOS]\", \"[CLS]\", \"[MASK]\", \"[UNK]\"])\n",
"tokenizer.decoder = decoders.Metaspace()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 字节级别（ByteLevel） BPE toeknizer"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# byte BPE n不需要unk_token\n",
"model = BPE() \n",
"tokenizer = Tokenizer(model)\n",
"\n",
"tokenizer.pre_tokenizer = tokenizers.pre_tokenizers.ByteLevel(add_prefix_space=False)\n",
"\n",
"tokenizer.add_special_tokens([\"[PAD]\",\"[EOS]\",\"[SEP]\",\"[BOS]\", \"[CLS]\", \"[MASK]\", \"[UNK]\"])\n",
"tokenizer.decoder = decoders.ByteLevel(add_prefix_space=True, use_regex=True)\n",
"tokenizer.post_processor = tokenizers.processors.ByteLevel(trim_offsets=False)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# PreTrainedTokenizerFast类无法从 tokenizer 对象推断出哪个标记是掩码标记、[CLS] 标记等，需要手动指定\n",
"# 上文的通过from_pretrained('t5-base')定义的old_tokenizer，自带了特殊标记，不用指定\n",
"# 到这一步和上文 step 4 一致了\n",
"old_tokenizer = PreTrainedTokenizerFast(\n",
" tokenizer_object=tokenizer,\n",
" unk_token=\"[UNK]\",\n",
" pad_token=\"[PAD]\",\n",
" cls_token=\"[CLS]\",\n",
" sep_token=\"[SEP]\",\n",
" mask_token=\"[MASK]\",\n",
" bos_token='[BOS]',\n",
" eos_token='[EOS]', \n",
")\n",
"tokenizer = old_tokenizer.train_new_from_iterator(iter_training_corpus, vocab_size=40960)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# add \\t \\n if char level tokenizer\n",
"# if '\\t' not in tokenizer.vcoab:\n",
"# tokenizer.add_tokens(['\\t'])\n",
"# if '\\n' not in tokenizer.vcoab:\n",
"# tokenizer.add_tokens(['\\n'])"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"tokenizer.save_pretrained('../model_save/my_tokenizer_wiki')"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"txt = '这是一段中英混输的句子, （chinese and English, here are words.）'\n",
"# toeknize\n",
"tokens = tokenizer.tokenize(txt)\n",
"print(tokens)\n",
"# 字级别输出：\n",
"# ['▁这是', '一段', '中英', '混', '输', '的', '句子', '▁,', '▁(', '▁ch', 'inese', '▁and', '▁Eng', 'lish', '▁,', '▁h', 'ere', '▁', 'are', '▁w', 'ord', 's', '▁.', '▁)']\n",
"\n",
"# Byte级别输出\n",
"# ['Ġè¿Ļæĺ¯', 'ä¸Ģ段', 'ä¸Ńèĭ±', 'æ··', 'è¾ĵ', 'çļĦ', 'åı¥åŃIJ', 'Ġ,', 'Ġ(', 'Ġch', 'inese', 'Ġand', 'ĠEng', 'lish', 'Ġ,', 'Ġh', 'ere', 'Ġare', 'Ġw', 'ord', 's', 'Ġ.', 'Ġ)']\n",
"\n",
"# decode\n",
"ids = tokenizer.encode(txt)\n",
"tokenizer.decode(ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "py310",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.10.12"
}
},
"nbformat": 4,
"nbformat_minor": 2
}
|