├── Quora
    ├── A look at different embeddings.!.ipynb
    └── how-to-preprocessing-when-using-embeddings.ipynb
├── README.md
├── Toxic Comment Classification Challenge
    └── Do_Pretrained_Embeddings_Give_You_The_Extra_Edge？.ipynb
└── use_tutorial.md


/Quora/how-to-preprocessing-when-using-embeddings.ipynb:
--------------------------------------------------------------------------------
  1 | {
  2 |  "cells": [
  3 |   {
  4 |    "cell_type": "code",
  5 |    "execution_count": null,
  6 |    "metadata": {},
  7 |    "outputs": [],
  8 |    "source": [
  9 |     "#在这个内核中，我想说明在构建深度学习NLP模型时如何进行有意义的预处理。\n",
 10 |     "#我开始的两条黄金法则:\n",
 11 |     "#1.使用标准的预处理步骤不喜欢阻止或stopword切除时pre-trained嵌入\n",
 12 |     "#一些您可能使用标准的预处理步骤时基于字数等特征提取(例如TFIDF)删除stopwords,引发等。原因很简单:你宽松的有价值的信息,这将有助于你神经网络图的东西。\n",
 13 |     "#2.让你的词汇量尽可能接近嵌入\n",
 14 |     "#我将集中在这个笔记本，如何实现这一点。以GoogleNews预培训的嵌入式为例，这种选择没有更深层次的原因。"
 15 |    ]
 16 |   },
 17 |   {
 18 |    "cell_type": "code",
 19 |    "execution_count": 1,
 20 |    "metadata": {},
 21 |    "outputs": [],
 22 |    "source": [
 23 |     "import pandas as pd\n",
 24 |     "from tqdm import tqdm\n",
 25 |     "tqdm.pandas()"
 26 |    ]
 27 |   },
 28 |   {
 29 |    "cell_type": "code",
 30 |    "execution_count": 2,
 31 |    "metadata": {},
 32 |    "outputs": [
 33 |     {
 34 |      "name": "stdout",
 35 |      "output_type": "stream",
 36 |      "text": [
 37 |       "Train shape: (1306122, 3)\n",
 38 |       "Test shape: (56370, 2)\n"
 39 |      ]
 40 |     }
 41 |    ],
 42 |    "source": [
 43 |     "# 加载数据\n",
 44 |     "train = pd.read_csv('../input/train.csv')\n",
 45 |     "test = pd.read_csv('../input/test.csv')\n",
 46 |     "print('Train shape:', train.shape)\n",
 47 |     "print('Test shape:', test.shape)"
 48 |    ]
 49 |   },
 50 |   {
 51 |    "cell_type": "code",
 52 |    "execution_count": 3,
 53 |    "metadata": {},
 54 |    "outputs": [],
 55 |    "source": [
 56 |     "#我将使用下面的函数来跟踪我们的训练词汇，它将遍历我们的所有文本并计算包含的单词的出现次数。\n",
 57 |     "def build_vocab(sentences, verbose = True):\n",
 58 |     "    # 参数  sentences list of list of words，就是二维的\n",
 59 |     "    # 返回值 对应  词和词的次数 的字典\n",
 60 |     "    vocab = {}\n",
 61 |     "    for sentence in tqdm(sentences, disable = (not verbose)):\n",
 62 |     "        for word in sentence:\n",
 63 |     "            try:\n",
 64 |     "                vocab[word] += 1\n",
 65 |     "            except KeyError:\n",
 66 |     "                vocab[word] = 1\n",
 67 |     "    return vocab"
 68 |    ]
 69 |   },
 70 |   {
 71 |    "cell_type": "code",
 72 |    "execution_count": 4,
 73 |    "metadata": {},
 74 |    "outputs": [
 75 |     {
 76 |      "name": "stderr",
 77 |      "output_type": "stream",
 78 |      "text": [
 79 |       "100%|███| 1306122/1306122 [00:04<00:00, 266825.76it/s]\n",
 80 |       "100%|███| 1306122/1306122 [00:05<00:00, 254818.90it/s]\n"
 81 |      ]
 82 |     },
 83 |     {
 84 |      "name": "stdout",
 85 |      "output_type": "stream",
 86 |      "text": [
 87 |       "{'How': 261930, 'did': 33489, 'Quebec': 97, 'nationalists': 91, 'see': 9003}\n"
 88 |      ]
 89 |     }
 90 |    ],
 91 |    "source": [
 92 |     "#因此，让我们填充词汇表并显示前5个元素及其计数。注意，现在我们可以使用progess_apply查看进度条\n",
 93 |     "sentences = train['question_text'].progress_apply(lambda x: x.split()).values\n",
 94 |     "vocab = build_vocab(sentences)\n",
 95 |     "print({k: vocab[k] for k in list(vocab)[:5]})"
 96 |    ]
 97 |   },
 98 |   {
 99 |    "cell_type": "code",
100 |    "execution_count": 6,
101 |    "metadata": {},
102 |    "outputs": [],
103 |    "source": [
104 |     "#接下来，我们导入我们稍后要在模型中使用的Embedding。为了说明这一点，我在这里使用GoogleNews\n",
105 |     "from gensim.models import KeyedVectors\n",
106 |     "\n",
107 |     "news_path = '../input/embeddings/GoogleNews-vectors-negative300/GoogleNews-vectors-negative300.bin'\n",
108 |     "embedding_index = KeyedVectors.load_word2vec_format(news_path, binary=True)"
109 |    ]
110 |   },
111 |   {
112 |    "cell_type": "code",
113 |    "execution_count": 15,
114 |    "metadata": {},
115 |    "outputs": [],
116 |    "source": [
117 |     "#接下来，我定义一个函数来检查词汇表和嵌入之间的交集。它将输出一个out of vocabulary (oov)单词列表，我们可以使用它来改进我们的预处理\n",
118 |     "import operator\n",
119 |     "def check_coverage(vocab, embeddings_index):\n",
120 |     "    a = {}\n",
121 |     "    oov = {}\n",
122 |     "    k = 0\n",
123 |     "    i = 0\n",
124 |     "    for word in tqdm(vocab):\n",
125 |     "        try:\n",
126 |     "            a[word] = embeddings_index[word]\n",
127 |     "            k += vocab[word]\n",
128 |     "        except:\n",
129 |     "            oov[word] = vocab[word]\n",
130 |     "            i += vocab[word]\n",
131 |     "            pass\n",
132 |     "    \n",
133 |     "    print('Found embeddings for {:.2%} of vocab'.format(len(a) /  len(vocab)))\n",
134 |     "    print('Found embeddings for {:.2%} of all text'.format(k / (k + i)))\n",
135 |     "    sorted_x = sorted(oov.items(), key=operator.itemgetter(1))[::-1]#取axis=1维度进行排序，并换为逆序\n",
136 |     "    \n",
137 |     "    return sorted_x"
138 |    ]
139 |   },
140 |   {
141 |    "cell_type": "code",
142 |    "execution_count": 16,
143 |    "metadata": {},
144 |    "outputs": [
145 |     {
146 |      "name": "stderr",
147 |      "output_type": "stream",
148 |      "text": [
149 |       "100%|█████| 508823/508823 [00:01<00:00, 374703.65it/s]\n"
150 |      ]
151 |     },
152 |     {
153 |      "name": "stdout",
154 |      "output_type": "stream",
155 |      "text": [
156 |       "Found embeddings for 24.31% of vocab\n",
157 |       "Found embeddings for 78.75% of all text\n"
158 |      ]
159 |     }
160 |    ],
161 |    "source": [
162 |     "oov = check_coverage(vocab, embedding_index)"
163 |    ]
164 |   },
165 |   {
166 |    "cell_type": "code",
167 |    "execution_count": 17,
168 |    "metadata": {},
169 |    "outputs": [
170 |     {
171 |      "data": {
172 |       "text/plain": [
173 |        "[('to', 403183),\n",
174 |        " ('a', 402682),\n",
175 |        " ('of', 330825),\n",
176 |        " ('and', 251973),\n",
177 |        " ('India?', 16384),\n",
178 |        " ('it?', 12900),\n",
179 |        " ('do?', 8753),\n",
180 |        " ('life?', 7753),\n",
181 |        " ('you?', 6295),\n",
182 |        " ('me?', 6202)]"
183 |       ]
184 |      },
185 |      "execution_count": 17,
186 |      "metadata": {},
187 |      "output_type": "execute_result"
188 |     }
189 |    ],
190 |    "source": [
191 |     "#哎哟，只有24%的词汇表会有嵌入，这使得21%的数据或多或少是无用的。所以让我们来看看并开始改进。为此，我们可以很容易地看一看顶部的oov单词\n",
192 |     "oov[:10]"
193 |    ]
194 |   },
195 |   {
196 |    "cell_type": "code",
197 |    "execution_count": 18,
198 |    "metadata": {},
199 |    "outputs": [
200 |     {
201 |      "name": "stdout",
202 |      "output_type": "stream",
203 |      "text": [
204 |       "False\n",
205 |       "True\n"
206 |      ]
207 |     }
208 |    ],
209 |    "source": [
210 |     "#：首先是“to”。为什么?仅仅是因为“to”在训练GoogleNews嵌入时被删除了。稍后我们将对此进行修复，因为现在我们要注意标点符号的分割，因为这似乎也是一个问题。但是，我们该如何处理标点符号呢?我们是想删除标点符号，还是将其视为一种标记?我想说:这要看情况。如果标记有嵌入，保留它，如果没有，我们就不再需要它了。我们检查:\n",
211 |     "print('?' in embedding_index)\n",
212 |     "print('&' in embedding_index)"
213 |    ]
214 |   },
215 |   {
216 |    "cell_type": "code",
217 |    "execution_count": 19,
218 |    "metadata": {},
219 |    "outputs": [],
220 |    "source": [
221 |     "#有趣。虽然“&”出现在谷歌新闻的嵌入中，“?”却不是。因此，我们基本上定义了一个函数，它分割“&”并删除其他标点符号。\n",
222 |     "\n",
223 |     "def clean_text(x):\n",
224 |     "    x = str(x)\n",
225 |     "    for punct in \"/-'\":\n",
226 |     "        x = x.replace(punct, ' ')\n",
227 |     "    for punct in '&':\n",
228 |     "        x = x.replace(punct, f' {punct} ')#f''解释：你不再需要直接调用一个字符串的.format()方法，但是要简单地用前缀f来标记格式以及内联最终字符串中你想要包括的表达式，不然它们就会被期望着去提供如同你从.format()函数得到的相同功能。这些格式化字符串也在文档中被称为“f字符串”。\n",
229 |     "    for punct in '?!.,\"#$%\\'()*+-/:;<=>@[\\\\]^_`{|}~' + '“”’':\n",
230 |     "        x = x.replace(punct, '')\n",
231 |     "    return x"
232 |    ]
233 |   },
234 |   {
235 |    "cell_type": "code",
236 |    "execution_count": 21,
237 |    "metadata": {},
238 |    "outputs": [
239 |     {
240 |      "name": "stderr",
241 |      "output_type": "stream",
242 |      "text": [
243 |       "100%|████| 1306122/1306122 [00:16<00:00, 79190.25it/s]\n",
244 |       "100%|███| 1306122/1306122 [00:05<00:00, 218846.61it/s]\n"
245 |      ]
246 |     }
247 |    ],
248 |    "source": [
249 |     "train['question_text'] =  train['question_text'].progress_apply(lambda x: clean_text(x))\n",
250 |     "sentences = train['question_text'].apply(lambda x: x.split())\n",
251 |     "vocab = build_vocab(sentences)"
252 |    ]
253 |   },
254 |   {
255 |    "cell_type": "code",
256 |    "execution_count": 23,
257 |    "metadata": {},
258 |    "outputs": [
259 |     {
260 |      "name": "stderr",
261 |      "output_type": "stream",
262 |      "text": [
263 |       "100%|█████| 253623/253623 [00:00<00:00, 329145.72it/s]\n"
264 |      ]
265 |     },
266 |     {
267 |      "name": "stdout",
268 |      "output_type": "stream",
269 |      "text": [
270 |       "Found embeddings for 57.38% of vocab\n",
271 |       "Found embeddings for 89.99% of all text\n"
272 |      ]
273 |     }
274 |    ],
275 |    "source": [
276 |     "oov = check_coverage(vocab, embedding_index)"
277 |    ]
278 |   },
279 |   {
280 |    "cell_type": "code",
281 |    "execution_count": 25,
282 |    "metadata": {},
283 |    "outputs": [
284 |     {
285 |      "data": {
286 |       "text/plain": [
287 |        "[('to', 406298),\n",
288 |        " ('a', 403852),\n",
289 |        " ('of', 332964),\n",
290 |        " ('and', 254081),\n",
291 |        " ('2017', 8781),\n",
292 |        " ('2018', 7373),\n",
293 |        " ('10', 6642),\n",
294 |        " ('12', 3694),\n",
295 |        " ('20', 2942),\n",
296 |        " ('100', 2883)]"
297 |       ]
298 |      },
299 |      "execution_count": 25,
300 |      "metadata": {},
301 |      "output_type": "execute_result"
302 |     }
303 |    ],
304 |    "source": [
305 |     "#好了!我们能够增加我们的嵌入比从24%到57%仅仅通过处理穿刺。好的，让我们检查一下这些单词。\n",
306 |     "oov[:10]"
307 |    ]
308 |   },
309 |   {
310 |    "cell_type": "code",
311 |    "execution_count": 30,
312 |    "metadata": {},
313 |    "outputs": [
314 |     {
315 |      "ename": "AttributeError",
316 |      "evalue": "'KeyedVectors' object has no attribute 'index2entity'",
317 |      "output_type": "error",
318 |      "traceback": [
319 |       "\u001b[1;31m-------------------------------------------------------\u001b[0m",
320 |       "\u001b[1;31mAttributeError\u001b[0m        Traceback (most recent call last)",
321 |       "\u001b[1;32m<ipython-input-30-7aac54b0cb07>\u001b[0m in \u001b[0;36m<module>\u001b[1;34m()\u001b[0m\n\u001b[0;32m      1\u001b[0m \u001b[1;31m#嗯，似乎数字也是个问题。让我们检查一下前10个嵌入来获得线索。\u001b[0m\u001b[1;33m\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m      2\u001b[0m \u001b[1;32mfor\u001b[0m \u001b[0mi\u001b[0m \u001b[1;32min\u001b[0m \u001b[0mrange\u001b[0m\u001b[1;33m(\u001b[0m\u001b[1;36m10\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m:\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[1;32m----> 3\u001b[1;33m     \u001b[0mprint\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0membedding_index\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mindex2entity\u001b[0m\u001b[1;33m[\u001b[0m\u001b[0mi\u001b[0m\u001b[1;33m]\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0m",
322 |       "\u001b[1;31mAttributeError\u001b[0m: 'KeyedVectors' object has no attribute 'index2entity'"
323 |      ]
324 |     }
325 |    ],
326 |    "source": [
327 |     "#嗯，似乎数字也是个问题。让我们检查一下前10个嵌入来获得线索。\n",
328 |     "for i in range(10):\n",
329 |     "    print(embedding_index.index2entity[i])"
330 |    ]
331 |   },
332 |   {
333 |    "cell_type": "code",
334 |    "execution_count": 31,
335 |    "metadata": {},
336 |    "outputs": [],
337 |    "source": [
338 |     "#为什么里面有\"##\" ?原因很简单，因为作为一个再处理，所有大于9的数字都被hashs替换了。即成为# #,123变成# # #或15.80€变成# #,# #€。因此，让我们模拟这个预处理步骤来进一步改进我们的嵌入式覆盖率\n",
339 |     "import re\n",
340 |     "\n",
341 |     "def clean_numbers(x):\n",
342 |     "    x = re.sub('[0-9]{5,}', '#####', x)\n",
343 |     "    x = re.sub('[0-9]{4}', '####', x)\n",
344 |     "    x = re.sub('[0-9]{3}', '###', x)\n",
345 |     "    x = re.sub('[0-9]{2}', '##', x)\n",
346 |     "    return x"
347 |    ]
348 |   },
349 |   {
350 |    "cell_type": "code",
351 |    "execution_count": 33,
352 |    "metadata": {},
353 |    "outputs": [
354 |     {
355 |      "name": "stderr",
356 |      "output_type": "stream",
357 |      "text": [
358 |       "100%|████| 1306122/1306122 [00:16<00:00, 80227.57it/s]\n",
359 |       "100%|███| 1306122/1306122 [00:05<00:00, 249280.32it/s]\n",
360 |       "100%|███| 1306122/1306122 [00:05<00:00, 259681.59it/s]\n"
361 |      ]
362 |     }
363 |    ],
364 |    "source": [
365 |     "train['question_text'] = train['question_text'].progress_apply(lambda x: clean_numbers(x))\n",
366 |     "sentences = train['question_text'].progress_apply(lambda x: x.split())\n",
367 |     "vocab = build_vocab(sentences)"
368 |    ]
369 |   },
370 |   {
371 |    "cell_type": "code",
372 |    "execution_count": 37,
373 |    "metadata": {},
374 |    "outputs": [
375 |     {
376 |      "name": "stderr",
377 |      "output_type": "stream",
378 |      "text": [
379 |       "100%|█████| 242997/242997 [00:00<00:00, 319105.53it/s]\n"
380 |      ]
381 |     },
382 |     {
383 |      "name": "stdout",
384 |      "output_type": "stream",
385 |      "text": [
386 |       "Found embeddings for 60.41% of vocab\n",
387 |       "Found embeddings for 90.75% of all text\n"
388 |      ]
389 |     }
390 |    ],
391 |    "source": [
392 |     "oov = check_coverage(vocab,embedding_index)"
393 |    ]
394 |   },
395 |   {
396 |    "cell_type": "code",
397 |    "execution_count": 24,
398 |    "metadata": {},
399 |    "outputs": [
400 |     {
401 |      "data": {
402 |       "text/plain": [
403 |        "[('to', 406298),\n",
404 |        " ('a', 403852),\n",
405 |        " ('of', 332964),\n",
406 |        " ('and', 254081),\n",
407 |        " ('2017', 8781),\n",
408 |        " ('2018', 7373),\n",
409 |        " ('10', 6642),\n",
410 |        " ('12', 3694),\n",
411 |        " ('20', 2942),\n",
412 |        " ('100', 2883),\n",
413 |        " ('15', 2762),\n",
414 |        " ('12th', 2551),\n",
415 |        " ('11', 2356),\n",
416 |        " ('30', 2163),\n",
417 |        " ('18', 2066),\n",
418 |        " ('50', 1993),\n",
419 |        " ('16', 1589),\n",
420 |        " ('14', 1533),\n",
421 |        " ('17', 1505),\n",
422 |        " ('13', 1390)]"
423 |       ]
424 |      },
425 |      "execution_count": 24,
426 |      "metadata": {},
427 |      "output_type": "execute_result"
428 |     }
429 |    ],
430 |    "source": [
431 |     "#好了!另一个3%的增长。现在就像处理撞击一样，但是每一点都有帮助。让我们检查oov单词\n",
432 |     "oov[:20]"
433 |    ]
434 |   },
435 |   {
436 |    "cell_type": "code",
437 |    "execution_count": 38,
438 |    "metadata": {},
439 |    "outputs": [],
440 |    "source": [
441 |     "#好了，现在我们来处理一下在使用美式/英式vocab时常见的拼写错误，并将一些“现代”单词替换为“social media”。此外，我们将简单地删除“a”、“to”、“and”和“of”等词，因为在培训GoogleNews嵌入式时，这些词显然已被下采样。\n",
442 |     "def _get_mispell(mispell_dict):\n",
443 |     "    mispell_re = re.compile('(%s)' % '|'.join(mispell_dict.keys()))#编写一个正则式\n",
444 |     "    return mispell_dict, mispell_re\n",
445 |     "\n",
446 |     "\n",
447 |     "mispell_dict = {'colour':'color',\n",
448 |     "                'centre':'center',\n",
449 |     "                'didnt':'did not',\n",
450 |     "                'doesnt':'does not',\n",
451 |     "                'isnt':'is not',\n",
452 |     "                'shouldnt':'should not',\n",
453 |     "                'favourite':'favorite',\n",
454 |     "                'travelling':'traveling',\n",
455 |     "                'counselling':'counseling',\n",
456 |     "                'theatre':'theater',\n",
457 |     "                'cancelled':'canceled',\n",
458 |     "                'labour':'labor',\n",
459 |     "                'organisation':'organization',\n",
460 |     "                'wwii':'world war 2',\n",
461 |     "                'citicise':'criticize',\n",
462 |     "                'instagram': 'social medium',\n",
463 |     "                'whatsapp': 'social medium',\n",
464 |     "                'snapchat': 'social medium'\n",
465 |     "\n",
466 |     "                }\n",
467 |     "mispellings, mispellings_re = _get_mispell(mispell_dict)"
468 |    ]
469 |   },
470 |   {
471 |    "cell_type": "code",
472 |    "execution_count": 39,
473 |    "metadata": {},
474 |    "outputs": [],
475 |    "source": [
476 |     "def replace_typical_misspell(text):\n",
477 |     "    def replace(match):\n",
478 |     "        return mispellings[match.group(0)]\n",
479 |     "\n",
480 |     "    return mispellings_re.sub(replace, text)"
481 |    ]
482 |   },
483 |   {
484 |    "cell_type": "code",
485 |    "execution_count": 47,
486 |    "metadata": {},
487 |    "outputs": [
488 |     {
489 |      "name": "stderr",
490 |      "output_type": "stream",
491 |      "text": [
492 |       "\n",
493 |       "  0%|                     | 0/1306122 [00:00<?, ?it/s]\n",
494 |       "  1%|     | 11727/1306122 [00:00<00:11, 116416.34it/s]\n",
495 |       "  2%|     | 30571/1306122 [00:00<00:09, 131288.09it/s]\n",
496 |       "  4%|▏    | 49145/1306122 [00:00<00:08, 143715.22it/s]\n",
497 |       "  5%|▎    | 68121/1306122 [00:00<00:08, 154721.00it/s]\n",
498 |       "  7%|▎    | 87093/1306122 [00:00<00:07, 163476.85it/s]\n",
499 |       "  8%|▎   | 105626/1306122 [00:00<00:07, 169118.02it/s]\n",
500 |       " 10%|▍   | 124101/1306122 [00:00<00:06, 173166.11it/s]\n",
501 |       " 11%|▍   | 143467/1306122 [00:00<00:06, 178484.38it/s]\n",
502 |       " 12%|▍   | 163252/1306122 [00:00<00:06, 183508.04it/s]\n",
503 |       " 14%|▌   | 181392/1306122 [00:01<00:06, 181354.71it/s]\n",
504 |       " 15%|▌   | 199386/1306122 [00:01<00:06, 174507.78it/s]\n",
505 |       " 17%|▋   | 216793/1306122 [00:01<00:06, 163104.15it/s]\n",
506 |       " 18%|▋   | 234639/1306122 [00:01<00:06, 167386.47it/s]\n",
507 |       " 19%|▊   | 254011/1306122 [00:01<00:06, 174388.73it/s]\n",
508 |       " 21%|▊   | 273563/1306122 [00:01<00:05, 179869.00it/s]\n",
509 |       " 22%|▉   | 291702/1306122 [00:01<00:05, 177298.68it/s]\n",
510 |       " 24%|▉   | 311585/1306122 [00:01<00:05, 182867.13it/s]\n",
511 |       " 25%|█   | 331034/1306122 [00:01<00:05, 185825.79it/s]\n",
512 |       " 27%|█   | 350033/1306122 [00:01<00:05, 186660.03it/s]\n",
513 |       " 28%|█▏  | 369543/1306122 [00:02<00:04, 188713.76it/s]\n",
514 |       " 30%|█▏  | 389234/1306122 [00:02<00:04, 190676.99it/s]\n",
515 |       " 31%|█▎  | 408965/1306122 [00:02<00:04, 192208.42it/s]\n",
516 |       " 33%|█▎  | 428219/1306122 [00:02<00:04, 190366.21it/s]\n",
517 |       " 34%|█▎  | 448524/1306122 [00:02<00:04, 193598.85it/s]\n",
518 |       " 36%|█▍  | 468425/1306122 [00:02<00:04, 194771.21it/s]\n",
519 |       " 37%|█▍  | 487927/1306122 [00:02<00:04, 193842.92it/s]\n",
520 |       " 39%|█▌  | 507330/1306122 [00:02<00:04, 192497.96it/s]\n",
521 |       " 40%|█▌  | 526595/1306122 [00:02<00:04, 180553.95it/s]\n",
522 |       " 42%|█▋  | 544812/1306122 [00:02<00:04, 177735.67it/s]\n",
523 |       " 43%|█▋  | 564557/1306122 [00:03<00:04, 182867.43it/s]\n",
524 |       " 45%|█▊  | 583580/1306122 [00:03<00:03, 184620.78it/s]\n",
525 |       " 46%|█▊  | 602133/1306122 [00:03<00:03, 182321.18it/s]\n",
526 |       " 48%|█▉  | 620435/1306122 [00:03<00:03, 181590.79it/s]\n",
527 |       " 49%|█▉  | 639600/1306122 [00:03<00:03, 184108.92it/s]\n",
528 |       " 50%|██  | 659352/1306122 [00:03<00:03, 187544.12it/s]\n",
529 |       " 52%|██  | 678401/1306122 [00:03<00:03, 188011.57it/s]\n",
530 |       " 53%|██▏ | 697235/1306122 [00:03<00:03, 186829.94it/s]\n",
531 |       " 55%|██▏ | 715943/1306122 [00:03<00:03, 182150.63it/s]\n",
532 |       " 56%|██▏ | 734201/1306122 [00:04<00:03, 179195.96it/s]\n",
533 |       " 58%|██▎ | 752162/1306122 [00:04<00:03, 178397.31it/s]\n",
534 |       " 59%|██▎ | 772026/1306122 [00:04<00:02, 183901.93it/s]\n",
535 |       " 61%|██▍ | 791545/1306122 [00:04<00:02, 186741.14it/s]\n",
536 |       " 62%|██▍ | 811467/1306122 [00:04<00:02, 189932.76it/s]\n",
537 |       " 64%|██▌ | 831070/1306122 [00:04<00:02, 191314.13it/s]\n",
538 |       " 65%|██▌ | 850240/1306122 [00:04<00:02, 188186.96it/s]\n",
539 |       " 67%|██▋ | 869098/1306122 [00:04<00:02, 184605.77it/s]\n",
540 |       " 68%|██▋ | 888029/1306122 [00:04<00:02, 185596.66it/s]\n",
541 |       " 69%|██▊ | 906620/1306122 [00:04<00:02, 183642.24it/s]\n",
542 |       " 71%|██▊ | 925011/1306122 [00:05<00:02, 178529.28it/s]\n",
543 |       " 72%|██▉ | 944306/1306122 [00:05<00:01, 182233.19it/s]\n",
544 |       " 74%|██▉ | 963091/1306122 [00:05<00:01, 183501.89it/s]\n",
545 |       " 75%|███ | 982721/1306122 [00:05<00:01, 186755.99it/s]\n",
546 |       " 77%|██▎| 1001564/1306122 [00:05<00:01, 186865.01it/s]\n",
547 |       " 78%|██▎| 1020280/1306122 [00:05<00:01, 177617.31it/s]\n",
548 |       " 79%|██▍| 1038158/1306122 [00:05<00:01, 176425.81it/s]\n",
549 |       " 81%|██▍| 1055884/1306122 [00:05<00:01, 174762.00it/s]\n",
550 |       " 82%|██▍| 1075189/1306122 [00:05<00:01, 179724.63it/s]\n",
551 |       " 84%|██▌| 1095021/1306122 [00:05<00:01, 184819.18it/s]\n",
552 |       " 85%|██▌| 1114281/1306122 [00:06<00:01, 186688.45it/s]\n",
553 |       " 87%|██▌| 1133019/1306122 [00:06<00:00, 185934.67it/s]\n",
554 |       " 88%|██▋| 1151661/1306122 [00:06<00:00, 182411.78it/s]\n",
555 |       " 90%|██▋| 1169952/1306122 [00:06<00:00, 181081.75it/s]\n",
556 |       " 91%|██▋| 1188097/1306122 [00:06<00:00, 178666.83it/s]\n",
557 |       " 92%|██▊| 1205997/1306122 [00:06<00:00, 176779.09it/s]\n",
558 |       " 94%|██▊| 1224641/1306122 [00:06<00:00, 179208.16it/s]\n",
559 |       " 95%|██▊| 1242674/1306122 [00:06<00:00, 179153.02it/s]\n",
560 |       " 97%|██▉| 1261028/1306122 [00:06<00:00, 180058.52it/s]\n",
561 |       " 98%|██▉| 1279049/1306122 [00:07<00:00, 175520.64it/s]\n",
562 |       " 99%|██▉| 1297771/1306122 [00:07<00:00, 178485.67it/s]\n",
563 |       "100%|███| 1306122/1306122 [00:07<00:00, 179723.74it/s]\n",
564 |       "  0%|                     | 0/1306122 [00:00<?, ?it/s]\n",
565 |       "  2%|     | 19770/1306122 [00:00<00:06, 196276.65it/s]\n",
566 |       "  4%|▏    | 52114/1306122 [00:00<00:05, 222184.13it/s]\n",
567 |       "  7%|▎    | 84949/1306122 [00:00<00:04, 245651.77it/s]\n",
568 |       "  9%|▎   | 118742/1306122 [00:00<00:04, 267107.67it/s]\n",
569 |       " 12%|▍   | 151915/1306122 [00:00<00:04, 283119.76it/s]\n",
570 |       " 14%|▌   | 180806/1306122 [00:00<00:03, 284259.51it/s]\n",
571 |       " 16%|▋   | 211236/1306122 [00:00<00:03, 289384.22it/s]\n",
572 |       " 18%|▋   | 240005/1306122 [00:00<00:03, 288568.25it/s]\n",
573 |       " 21%|▊   | 271492/1306122 [00:00<00:03, 295543.88it/s]\n",
574 |       " 23%|▉   | 303744/1306122 [00:01<00:03, 302779.28it/s]\n",
575 |       " 26%|█   | 340001/1306122 [00:01<00:03, 317799.71it/s]\n",
576 |       " 28%|█▏  | 371814/1306122 [00:01<00:02, 314823.31it/s]\n",
577 |       " 31%|█▏  | 403730/1306122 [00:01<00:02, 315887.08it/s]\n",
578 |       " 34%|█▎  | 438336/1306122 [00:01<00:02, 323339.88it/s]\n",
579 |       " 36%|█▍  | 473779/1306122 [00:01<00:02, 331883.32it/s]\n",
580 |       " 39%|█▌  | 508572/1306122 [00:01<00:02, 335811.01it/s]\n",
581 |       " 42%|█▋  | 542233/1306122 [00:01<00:02, 334326.05it/s]\n",
582 |       " 44%|█▊  | 578425/1306122 [00:01<00:02, 341423.56it/s]\n",
583 |       " 47%|█▉  | 614085/1306122 [00:01<00:02, 345139.39it/s]\n",
584 |       " 50%|█▉  | 649810/1306122 [00:02<00:01, 347909.57it/s]\n",
585 |       " 53%|██  | 686681/1306122 [00:02<00:01, 353150.70it/s]\n",
586 |       " 55%|██▏ | 722486/1306122 [00:02<00:01, 353876.73it/s]\n",
587 |       " 58%|██▎ | 757987/1306122 [00:02<00:01, 352354.75it/s]\n",
588 |       " 61%|██▍ | 793251/1306122 [00:02<00:01, 350625.20it/s]\n",
589 |       " 63%|██▌ | 828957/1306122 [00:02<00:01, 351737.15it/s]\n",
590 |       " 66%|██▋ | 864147/1306122 [00:02<00:01, 351049.13it/s]\n",
591 |       " 69%|██▊ | 899264/1306122 [00:02<00:01, 345133.67it/s]\n",
592 |       " 72%|██▊ | 934728/1306122 [00:02<00:01, 347218.06it/s]\n",
593 |       " 74%|██▉ | 969475/1306122 [00:02<00:01, 313765.19it/s]\n",
594 |       " 77%|██▎| 1001955/1306122 [00:03<00:00, 316319.39it/s]\n",
595 |       " 79%|██▍| 1037449/1306122 [00:03<00:00, 326335.61it/s]\n",
596 |       " 82%|██▍| 1072289/1306122 [00:03<00:00, 331961.04it/s]\n",
597 |       " 85%|██▌| 1108408/1306122 [00:03<00:00, 339518.49it/s]\n",
598 |       " 88%|██▋| 1144206/1306122 [00:03<00:00, 344100.47it/s]\n",
599 |       " 90%|██▋| 1179285/1306122 [00:03<00:00, 345329.94it/s]\n",
600 |       " 93%|██▊| 1215513/1306122 [00:03<00:00, 349531.92it/s]\n",
601 |       " 96%|██▊| 1251657/1306122 [00:03<00:00, 352232.09it/s]\n",
602 |       " 99%|██▉| 1286960/1306122 [00:03<00:00, 351702.14it/s]\n",
603 |       "100%|███| 1306122/1306122 [00:03<00:00, 330858.24it/s]\n",
604 |       "  0%|                     | 0/1306122 [00:00<?, ?it/s]\n",
605 |       "  2%|     | 31915/1306122 [00:00<00:04, 316844.02it/s]\n",
606 |       "  5%|▎    | 67049/1306122 [00:00<00:03, 326350.31it/s]\n",
607 |       "  8%|▎   | 102439/1306122 [00:00<00:03, 333927.07it/s]\n",
608 |       " 10%|▍   | 134545/1306122 [00:00<00:03, 329216.61it/s]\n",
609 |       " 13%|▌   | 170813/1306122 [00:00<00:03, 337869.40it/s]\n",
610 |       " 16%|▋   | 204571/1306122 [00:00<00:03, 337044.50it/s]\n",
611 |       " 18%|▋   | 234628/1306122 [00:00<00:03, 316264.65it/s]\n",
612 |       " 21%|▊   | 268246/1306122 [00:00<00:03, 321338.80it/s]\n",
613 |       " 23%|▉   | 303627/1306122 [00:00<00:03, 329762.43it/s]\n",
614 |       " 26%|█   | 336436/1306122 [00:01<00:02, 328498.33it/s]\n",
615 |       " 28%|█▏  | 370884/1306122 [00:01<00:02, 332466.05it/s]\n",
616 |       " 29%|█▋    | 376737/1306122 [00:14<11:02, 1402.62it/s]\n",
617 |       " 29%|█▋    | 376737/1306122 [00:14<11:02, 1402.62it/s]\n",
618 |       " 31%|█▊    | 405073/1306122 [00:15<07:30, 1999.47it/s]\n",
619 |       " 33%|█▉    | 432022/1306122 [00:15<05:06, 2847.26it/s]\n",
620 |       " 35%|██    | 461210/1306122 [00:15<03:28, 4050.46it/s]\n",
621 |       " 37%|██▏   | 486790/1306122 [00:15<02:22, 5747.09it/s]\n",
622 |       " 40%|██▍   | 517722/1306122 [00:15<01:36, 8144.81it/s]\n",
623 |       " 42%|██   | 544550/1306122 [00:15<01:06, 11484.91it/s]\n",
624 |       " 44%|██▏  | 569550/1306122 [00:15<00:45, 16084.86it/s]\n",
625 |       " 46%|██▎  | 602398/1306122 [00:15<00:31, 22502.65it/s]\n",
626 |       " 49%|██▍  | 633690/1306122 [00:15<00:21, 31182.80it/s]\n",
627 |       " 51%|██▌  | 666592/1306122 [00:15<00:14, 42803.32it/s]\n",
628 |       " 53%|██▋  | 697482/1306122 [00:16<00:10, 57712.69it/s]\n",
629 |       " 56%|██▊  | 736071/1306122 [00:16<00:07, 77446.46it/s]\n",
630 |       " 59%|██▎ | 772892/1306122 [00:16<00:05, 101424.97it/s]\n",
631 |       " 62%|██▍ | 809861/1306122 [00:16<00:03, 129550.82it/s]\n",
632 |       " 65%|██▌ | 845751/1306122 [00:16<00:02, 160126.19it/s]\n",
633 |       " 67%|██▋ | 880703/1306122 [00:16<00:02, 190980.92it/s]\n",
634 |       " 70%|██▊ | 918610/1306122 [00:16<00:01, 224090.51it/s]\n",
635 |       " 73%|██▉ | 956306/1306122 [00:16<00:01, 254750.82it/s]\n",
636 |       " 76%|███ | 992510/1306122 [00:16<00:01, 275633.53it/s]\n",
637 |       " 79%|██▎| 1030892/1306122 [00:17<00:00, 299852.39it/s]\n",
638 |       " 82%|██▍| 1068546/1306122 [00:17<00:00, 318799.35it/s]\n",
639 |       " 85%|██▌| 1105113/1306122 [00:17<00:00, 325572.08it/s]\n",
640 |       " 87%|██▌| 1140987/1306122 [00:17<00:00, 324266.87it/s]\n",
641 |       " 90%|██▋| 1176971/1306122 [00:17<00:00, 334056.67it/s]\n",
642 |       " 93%|██▊| 1213478/1306122 [00:17<00:00, 342554.21it/s]\n",
643 |       " 96%|██▊| 1248977/1306122 [00:17<00:00, 343088.68it/s]\n",
644 |       " 99%|██▉| 1287269/1306122 [00:17<00:00, 353169.13it/s]\n",
645 |       "100%|████| 1306122/1306122 [00:17<00:00, 73469.50it/s]\n",
646 |       "  0%|                     | 0/1306122 [00:00<?, ?it/s]\n",
647 |       "  1%|       | 9014/1306122 [00:00<00:14, 89484.26it/s]\n",
648 |       "  3%|▏    | 37631/1306122 [00:00<00:11, 112632.78it/s]\n",
649 |       "  5%|▎    | 65334/1306122 [00:00<00:09, 136878.90it/s]\n",
650 |       "  7%|▎    | 94288/1306122 [00:00<00:07, 162398.72it/s]\n",
651 |       "  9%|▎   | 121752/1306122 [00:00<00:06, 184818.65it/s]\n",
652 |       " 11%|▍   | 149610/1306122 [00:00<00:05, 205243.45it/s]\n",
653 |       " 13%|▌   | 175531/1306122 [00:00<00:05, 218512.62it/s]\n",
654 |       " 16%|▌   | 203885/1306122 [00:00<00:04, 234462.67it/s]\n",
655 |       " 18%|▋   | 230317/1306122 [00:00<00:04, 242221.55it/s]\n",
656 |       " 20%|▊   | 256091/1306122 [00:01<00:04, 246144.19it/s]\n",
657 |       " 22%|▊   | 281741/1306122 [00:01<00:04, 248496.20it/s]\n",
658 |       " 24%|▉   | 307426/1306122 [00:01<00:03, 250429.41it/s]\n",
659 |       " 26%|█   | 334872/1306122 [00:01<00:03, 257010.57it/s]\n",
660 |       " 28%|█   | 361756/1306122 [00:01<00:03, 259899.24it/s]\n",
661 |       " 30%|█▏  | 390010/1306122 [00:01<00:03, 265737.38it/s]\n",
662 |       " 32%|█▎  | 418433/1306122 [00:01<00:03, 270476.52it/s]\n",
663 |       " 34%|█▎  | 446407/1306122 [00:01<00:03, 272607.33it/s]\n",
664 |       " 36%|█▍  | 475012/1306122 [00:01<00:03, 275921.51it/s]\n",
665 |       " 39%|█▌  | 503859/1306122 [00:01<00:02, 278977.46it/s]\n",
666 |       " 41%|█▋  | 531838/1306122 [00:02<00:02, 268969.65it/s]\n",
667 |       " 43%|█▋  | 558869/1306122 [00:02<00:02, 268804.84it/s]\n",
668 |       " 45%|█▊  | 586881/1306122 [00:02<00:02, 271487.01it/s]\n",
669 |       " 47%|█▉  | 614478/1306122 [00:02<00:02, 272255.08it/s]\n",
670 |       " 49%|█▉  | 644078/1306122 [00:02<00:02, 278369.78it/s]\n",
671 |       " 51%|██  | 671985/1306122 [00:02<00:02, 277131.92it/s]\n",
672 |       " 54%|██▏ | 701364/1306122 [00:02<00:02, 281339.06it/s]\n",
673 |       " 56%|██▏ | 729549/1306122 [00:02<00:02, 275099.92it/s]\n",
674 |       " 58%|██▎ | 757850/1306122 [00:02<00:01, 276865.77it/s]\n",
675 |       " 60%|██▍ | 785586/1306122 [00:02<00:01, 276405.58it/s]\n",
676 |       " 62%|██▍ | 813261/1306122 [00:03<00:01, 264825.91it/s]\n",
677 |       " 64%|██▌ | 840550/1306122 [00:03<00:01, 266614.22it/s]\n",
678 |       " 66%|██▋ | 868417/1306122 [00:03<00:01, 269553.43it/s]\n",
679 |       " 69%|██▋ | 897360/1306122 [00:03<00:01, 274666.34it/s]\n",
680 |       " 71%|██▊ | 925014/1306122 [00:03<00:01, 275032.21it/s]\n",
681 |       " 73%|██▉ | 954231/1306122 [00:03<00:01, 279370.55it/s]\n",
682 |       " 75%|███ | 983186/1306122 [00:03<00:01, 281655.84it/s]\n",
683 |       " 77%|██▎| 1011394/1306122 [00:03<00:01, 276201.48it/s]\n",
684 |       " 80%|██▍| 1039067/1306122 [00:03<00:00, 270934.72it/s]\n",
685 |       " 82%|██▍| 1067708/1306122 [00:03<00:00, 274760.56it/s]\n",
686 |       " 84%|██▌| 1095239/1306122 [00:04<00:00, 271084.52it/s]\n",
687 |       " 86%|██▌| 1123082/1306122 [00:04<00:00, 272666.14it/s]\n",
688 |       " 88%|██▋| 1150386/1306122 [00:04<00:00, 265063.00it/s]\n",
689 |       " 90%|██▋| 1179019/1306122 [00:04<00:00, 270571.96it/s]\n",
690 |       " 92%|██▊| 1207102/1306122 [00:04<00:00, 272984.63it/s]\n",
691 |       " 95%|██▊| 1234499/1306122 [00:04<00:00, 272685.23it/s]\n",
692 |       " 97%|██▉| 1263786/1306122 [00:04<00:00, 277863.66it/s]\n",
693 |       " 99%|██▉| 1291626/1306122 [00:04<00:00, 274958.03it/s]\n",
694 |       "100%|███| 1306122/1306122 [00:04<00:00, 269247.95it/s]"
695 |      ]
696 |     }
697 |    ],
698 |    "source": [
699 |     "train[\"question_text\"] = train[\"question_text\"].progress_apply(lambda x: replace_typical_misspell(x))\n",
700 |     "sentences = train[\"question_text\"].progress_apply(lambda x: x.split())\n",
701 |     "to_remove = ['a','to','of','and']\n",
702 |     "sentences = [[word for word in sentence if not word in to_remove] for sentence in tqdm(sentences)]\n",
703 |     "vocab = build_vocab(sentences)"
704 |    ]
705 |   },
706 |   {
707 |    "cell_type": "code",
708 |    "execution_count": 48,
709 |    "metadata": {},
710 |    "outputs": [
711 |     {
712 |      "name": "stderr",
713 |      "output_type": "stream",
714 |      "text": [
715 |       "\n",
716 |       "  0%|                      | 0/242935 [00:00<?, ?it/s]\n",
717 |       "  0%|          | 160/242935 [00:00<02:32, 1588.38it/s]\n",
718 |       "  0%|          | 548/242935 [00:00<02:05, 1928.33it/s]\n",
719 |       "  1%|         | 1552/242935 [00:00<01:34, 2543.84it/s]\n",
720 |       "  2%|▏        | 4969/242935 [00:00<01:07, 3520.90it/s]\n",
721 |       "  4%|▎        | 9448/242935 [00:00<00:47, 4864.76it/s]\n",
722 |       "  6%|▍       | 15121/242935 [00:00<00:33, 6701.57it/s]\n",
723 |       "  9%|▊       | 22964/242935 [00:00<00:23, 9233.08it/s]\n",
724 |       " 14%|▉      | 34243/242935 [00:00<00:16, 12739.90it/s]\n",
725 |       " 20%|█▍     | 48984/242935 [00:00<00:11, 17545.09it/s]\n",
726 |       " 27%|█▊     | 64970/242935 [00:01<00:07, 23930.84it/s]\n",
727 |       " 34%|██▎    | 82019/242935 [00:01<00:04, 32242.48it/s]\n",
728 |       " 42%|██▍   | 100951/242935 [00:01<00:03, 42906.98it/s]\n",
729 |       " 50%|██▉   | 120275/242935 [00:01<00:02, 55932.64it/s]\n",
730 |       " 59%|███▌  | 143094/242935 [00:01<00:01, 72259.87it/s]\n",
731 |       " 66%|███▉  | 161090/242935 [00:01<00:01, 80879.94it/s]\n",
732 |       " 76%|███▊ | 184088/242935 [00:01<00:00, 100312.85it/s]\n",
733 |       " 83%|████▏| 202109/242935 [00:01<00:00, 104330.03it/s]\n",
734 |       " 93%|████▋| 225743/242935 [00:01<00:00, 125188.00it/s]\n",
735 |       "100%|█████| 242935/242935 [00:02<00:00, 121215.29it/s]"
736 |      ]
737 |     },
738 |     {
739 |      "name": "stdout",
740 |      "output_type": "stream",
741 |      "text": [
742 |       "Found embeddings for 60.43% of vocab\n",
743 |       "Found embeddings for 98.96% of all text\n"
744 |      ]
745 |     }
746 |    ],
747 |    "source": [
748 |     "oov = check_coverage(vocab, embedding_index)"
749 |    ]
750 |   },
751 |   {
752 |    "cell_type": "code",
753 |    "execution_count": 49,
754 |    "metadata": {},
755 |    "outputs": [
756 |     {
757 |      "data": {
758 |       "text/plain": [
759 |        "[('bitcoin', 987),\n",
760 |        " ('Quorans', 858),\n",
761 |        " ('cryptocurrency', 822),\n",
762 |        " ('Snapchat', 807),\n",
763 |        " ('btech', 632),\n",
764 |        " ('Brexit', 493),\n",
765 |        " ('cryptocurrencies', 481),\n",
766 |        " ('blockchain', 474),\n",
767 |        " ('behaviour', 468),\n",
768 |        " ('upvotes', 432),\n",
769 |        " ('programme', 402),\n",
770 |        " ('Redmi', 379),\n",
771 |        " ('realise', 371),\n",
772 |        " ('defence', 364),\n",
773 |        " ('KVPY', 349),\n",
774 |        " ('Paytm', 334),\n",
775 |        " ('grey', 299),\n",
776 |        " ('mtech', 281),\n",
777 |        " ('Btech', 262),\n",
778 |        " ('bitcoins', 254)]"
779 |       ]
780 |      },
781 |      "execution_count": 49,
782 |      "metadata": {},
783 |      "output_type": "execute_result"
784 |     }
785 |    ],
786 |    "source": [
787 |     "#我们发现，尽管我们改进了所有文本的嵌入量，从89%提高到99%。让我们再检查一遍oov单词\n",
788 |     "oov[:20]"
789 |    ]
790 |   },
791 |   {
792 |    "cell_type": "code",
793 |    "execution_count": 50,
794 |    "metadata": {},
795 |    "outputs": [],
796 |    "source": [
797 |     "#看起来不错。没有明显的oov词，我们可以快速修复。谢谢你的阅读和快乐的kaggling"
798 |    ]
799 |   },
800 |   {
801 |    "cell_type": "code",
802 |    "execution_count": 51,
803 |    "metadata": {},
804 |    "outputs": [],
805 |    "source": [
806 |     "train.to_csv('process_train.csv')"
807 |    ]
808 |   },
809 |   {
810 |    "cell_type": "code",
811 |    "execution_count": 54,
812 |    "metadata": {},
813 |    "outputs": [],
814 |    "source": [
815 |     "import pickle\n",
816 |     "\n",
817 |     "with open('vocab.pkl', 'wb') as f:\n",
818 |     "    pickle.dump(vocab, f, pickle.HIGHEST_PROTOCOL)\n",
819 |     "    \n",
820 |     "# def load_obj(name ):\n",
821 |     "#     with open('obj/' + name + '.pkl', 'rb') as f:\n",
822 |     "#         return pickle.load(f)"
823 |    ]
824 |   },
825 |   {
826 |    "cell_type": "code",
827 |    "execution_count": 56,
828 |    "metadata": {},
829 |    "outputs": [],
830 |    "source": [
831 |     "f = open('vocab.txt','w', encoding='utf-8')\n",
832 |     "f.write(str(vocab))\n",
833 |     "f.close()"
834 |    ]
835 |   },
836 |   {
837 |    "cell_type": "code",
838 |    "execution_count": null,
839 |    "metadata": {},
840 |    "outputs": [],
841 |    "source": []
842 |   }
843 |  ],
844 |  "metadata": {
845 |   "kernelspec": {
846 |    "display_name": "tensorflow-gpu",
847 |    "language": "python",
848 |    "name": "tensorflow-gpu"
849 |   },
850 |   "language_info": {
851 |    "codemirror_mode": {
852 |     "name": "ipython",
853 |     "version": 3
854 |    },
855 |    "file_extension": ".py",
856 |    "mimetype": "text/x-python",
857 |    "name": "python",
858 |    "nbconvert_exporter": "python",
859 |    "pygments_lexer": "ipython3",
860 |    "version": "3.6.2"
861 |   }
862 |  },
863 |  "nbformat": 4,
864 |  "nbformat_minor": 2
865 | }
866 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | # pretrained_embedding_tutorial
2 | 1.16-1.20小白我对于预训练词向量的使用的一些经验总结~
3 | 


--------------------------------------------------------------------------------
/Toxic Comment Classification Challenge/Do_Pretrained_Embeddings_Give_You_The_Extra_Edge？.ipynb:
--------------------------------------------------------------------------------
  1 | {
  2 |  "cells": [
  3 |   {
  4 |    "cell_type": "code",
  5 |    "execution_count": 3,
  6 |    "metadata": {},
  7 |    "outputs": [
  8 |     {
  9 |      "name": "stderr",
 10 |      "output_type": "stream",
 11 |      "text": [
 12 |       "Using TensorFlow backend.\n",
 13 |       "F:\\Anaconda\\envs\\tensorflow-gpu\\lib\\site-packages\\gensim\\utils.py:865: UserWarning: detected Windows; aliasing chunkize to chunkize_serial\n",
 14 |       "  warnings.warn(\"detected Windows; aliasing chunkize to chunkize_serial\")\n"
 15 |      ]
 16 |     }
 17 |    ],
 18 |    "source": [
 19 |     "import sys, os, re, csv, codecs, numpy as np, pandas as pd\n",
 20 |     "import matplotlib.pyplot as plt\n",
 21 |     "%matplotlib inline\n",
 22 |     "from keras.preprocessing.text import Tokenizer\n",
 23 |     "from keras.preprocessing.sequence import pad_sequences\n",
 24 |     "from keras.layers import Dense, Input, LSTM, Embedding, Dropout, Activation\n",
 25 |     "from keras.layers import Bidirectional, GlobalMaxPool1D,Bidirectional\n",
 26 |     "from keras.models import Model\n",
 27 |     "from keras import initializers, regularizers, constraints, optimizers, layers\n",
 28 |     "import matplotlib.pyplot as plt\n",
 29 |     "%matplotlib inline\n",
 30 |     "import gensim.models.keyedvectors as word2vec\n",
 31 |     "import gc"
 32 |    ]
 33 |   },
 34 |   {
 35 |    "cell_type": "code",
 36 |    "execution_count": 4,
 37 |    "metadata": {},
 38 |    "outputs": [],
 39 |    "source": [
 40 |     "train = pd.read_csv('../input/train.csv')\n",
 41 |     "test = pd.read_csv('../input/test.csv')\n",
 42 |     "embed_size = 0"
 43 |    ]
 44 |   },
 45 |   {
 46 |    "cell_type": "code",
 47 |    "execution_count": 5,
 48 |    "metadata": {},
 49 |    "outputs": [],
 50 |    "source": [
 51 |     "list_classes = ['toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate']\n",
 52 |     "y = train[list_classes].values\n",
 53 |     "list_sentences_train = train['comment_text']\n",
 54 |     "list_sentences_test = test['comment_text']"
 55 |    ]
 56 |   },
 57 |   {
 58 |    "cell_type": "code",
 59 |    "execution_count": 6,
 60 |    "metadata": {},
 61 |    "outputs": [],
 62 |    "source": [
 63 |     "max_features = 20000\n",
 64 |     "tokenizer = Tokenizer(num_words=max_features)\n",
 65 |     "tokenizer.fit_on_texts(list(list_sentences_train))\n",
 66 |     "list_tokenized_train = tokenizer.texts_to_sequences(list_sentences_train)\n",
 67 |     "list_tokenized_test = tokenizer.texts_to_sequences(list_sentences_test)"
 68 |    ]
 69 |   },
 70 |   {
 71 |    "cell_type": "code",
 72 |    "execution_count": 7,
 73 |    "metadata": {},
 74 |    "outputs": [],
 75 |    "source": [
 76 |     "maxlen = 200\n",
 77 |     "X_t = pad_sequences(list_tokenized_train, maxlen=maxlen)\n",
 78 |     "X_te = pad_sequences(list_tokenized_test, maxlen=maxlen)"
 79 |    ]
 80 |   },
 81 |   {
 82 |    "cell_type": "code",
 83 |    "execution_count": 12,
 84 |    "metadata": {},
 85 |    "outputs": [],
 86 |    "source": [
 87 |     "#由于我们将计算一些单词嵌入，让我们定义一个函数，以便能够正确地运行我们的实验。我将在下面的函数中添加一些注释，以便更好地理解。\n",
 88 |     "# 请注意，Kaggle数据集中嵌入了很多手套，我觉得使用基于Twitter文本进行训练的手套更合适。由于我们数据集中的注释由随意的、用户生成的短消息组成，因此使用的语义可能非常相似。因此，我们可能能够捕捉本质，并使用它来产生一个良好的准确的分数。\n",
 89 |     "# 同样，我也使用了经过谷歌负面新闻文本语料库训练的Word2Vec嵌入，希望它的负面词汇能在我们的“有毒”上下文中更好地发挥作用。"
 90 |    ]
 91 |   },
 92 |   {
 93 |    "cell_type": "code",
 94 |    "execution_count": 27,
 95 |    "metadata": {},
 96 |    "outputs": [],
 97 |    "source": [
 98 |     "def loadEmbeddingMatrix(typeToLoad):\n",
 99 |     "    #根据Embedding的不同，从Kaggle加载不同的嵌入文件\n",
100 |     "    #我们要实验的矩阵\n",
101 |     "    if(typeToLoad=='glove'):\n",
102 |     "        EMBEDDING_FILE = '../input/glove.twitter.27B.25d.txt'\n",
103 |     "        embed_size = 25\n",
104 |     "    elif(typeToLoad=='word2vec'):\n",
105 |     "        word2vecDict = word2vec.KeyedVectors.load_word2vec_format('../../Quora/input/embeddings/GoogleNews-vectors-negative300/GoogleNews-vectors-negative300.bin', binary=True)\n",
106 |     "        embed_size = 300\n",
107 |     "    elif(typeToLoad=='fasttext'):\n",
108 |     "        EMBEDDING_FILE='../input/wiki.simple.vec'\n",
109 |     "        embed_size = 300\n",
110 |     "    \n",
111 |     "    if(typeToLoad=='glove' or typeToLoad=='fasttext'):\n",
112 |     "        embeddings_index = dict()\n",
113 |     "        #通过遍历文件的每一行，将Embedding权重转移到字典中。\n",
114 |     "        f = open(EMBEDDING_FILE)\n",
115 |     "        for line in f:\n",
116 |     "            #split up line into an indexed array\n",
117 |     "            values = line.split()\n",
118 |     "            #first index is word\n",
119 |     "            word = values[0]\n",
120 |     "            #store the rest of the values in the array as a new array\n",
121 |     "            coefs = np.asarray(values[1:], dtype='float32')\n",
122 |     "            embeddings_index[word] = coefs #50 dimensions\n",
123 |     "        f.close()\n",
124 |     "        print('Loaded %s word vectors.' % len(embeddings_index))\n",
125 |     "    else:\n",
126 |     "        embeddings_index = dict()\n",
127 |     "        for word in word2vecDict.wv.vocab:\n",
128 |     "            embeddings_index[word] = word2vecDict.word_vec(word)\n",
129 |     "        print('Loaded %s word vectors.' % len(embeddings_index))\n",
130 |     "    \n",
131 |     "    gc.collect()\n",
132 |     "    #我们得到嵌入权值的均值和标准差，这样我们就可以保持\n",
133 |     "    #对于我们自己随机产生的权值的其余部分，同样的统计数据。\n",
134 |     "    all_embs = np.stack(list(embeddings_index.values()))\n",
135 |     "    emb_mean, emb_std = all_embs.mean(), all_embs.std()\n",
136 |     "    \n",
137 |     "    nb_words = len(tokenizer.word_index)\n",
138 |     "    #我们讲设置Embedding的尺寸为我们复制的预训练的维度\n",
139 |     "    #Embedding矩阵大小将为 词汇中词数 X Embedding Size\n",
140 |     "    embedding_matrix = np.random.normal(emb_mean, emb_std, (nb_words, embed_size))\n",
141 |     "    gc.collect()\n",
142 |     "    \n",
143 |     "    #使用新创建的嵌入矩阵，我们将用两个矩阵中的单词填充它\n",
144 |     "    #我们自己的字典和加载预训练嵌入。\n",
145 |     "    embeddedCount = 0\n",
146 |     "    for word, i in tokenizer.word_index.items():\n",
147 |     "        i -= 1\n",
148 |     "        #然后我们看看这个词是否在GLove的字典里，如果是，得到相应的权重\n",
149 |     "        embedding_vector = embeddings_index.get(word)\n",
150 |     "        #并存储在Embedding矩阵中，我们稍后将用其进行训练。\n",
151 |     "        if embedding_vector is not None:\n",
152 |     "            embedding_matrix[i] = embedding_vector\n",
153 |     "            embeddedCount += 1\n",
154 |     "    print('total embedded:', embeddedCount,'common words')\n",
155 |     "    \n",
156 |     "    del(embeddings_index)\n",
157 |     "    gc.collect()\n",
158 |     "    \n",
159 |     "    #最后，返回Embedding矩阵\n",
160 |     "    return embedding_matrix\n",
161 |     "    "
162 |    ]
163 |   },
164 |   {
165 |    "cell_type": "code",
166 |    "execution_count": 28,
167 |    "metadata": {},
168 |    "outputs": [],
169 |    "source": [
170 |     "#该函数将返回一个新的嵌入矩阵，\n",
171 |     "#该矩阵中包含了我们常用单词的预训练嵌入中加载的权重，\n",
172 |     "#以及随机初始化的数字，该矩阵中其余权重的均值和标准差相同。"
173 |    ]
174 |   },
175 |   {
176 |    "cell_type": "code",
177 |    "execution_count": 29,
178 |    "metadata": {},
179 |    "outputs": [],
180 |    "source": [
181 |     "#让我们继续，从Word2Vec加载第一个嵌入"
182 |    ]
183 |   },
184 |   {
185 |    "cell_type": "code",
186 |    "execution_count": 30,
187 |    "metadata": {},
188 |    "outputs": [
189 |     {
190 |      "name": "stdout",
191 |      "output_type": "stream",
192 |      "text": [
193 |       "Loaded 3000000 word vectors.\n",
194 |       "total embedded: 66078 common words\n"
195 |      ]
196 |     }
197 |    ],
198 |    "source": [
199 |     "embedding_matrix = loadEmbeddingMatrix('word2vec')"
200 |    ]
201 |   },
202 |   {
203 |    "cell_type": "code",
204 |    "execution_count": 31,
205 |    "metadata": {},
206 |    "outputs": [
207 |     {
208 |      "data": {
209 |       "text/plain": [
210 |        "(210337, 300)"
211 |       ]
212 |      },
213 |      "execution_count": 31,
214 |      "metadata": {},
215 |      "output_type": "execute_result"
216 |     }
217 |    ],
218 |    "source": [
219 |     "embedding_matrix.shape"
220 |    ]
221 |   },
222 |   {
223 |    "cell_type": "markdown",
224 |    "metadata": {},
225 |    "source": [
226 |     "With the embedding weights, we can proceed to build a LSTM layer. The whole architecture is pretty much the same as the previous one I have done in the earlier kernel here, except that I have turned the LSTM into a bidirectional one, and added a dropout factor to it. \n",
227 |     "\n",
228 |     "We start off with defining our input layer. By indicating an empty space after comma, we are telling Keras to infer the number automatically."
229 |    ]
230 |   },
231 |   {
232 |    "cell_type": "code",
233 |    "execution_count": 32,
234 |    "metadata": {},
235 |    "outputs": [],
236 |    "source": [
237 |     "inp = Input(shape=(maxlen, )) #maxlen=200 as defined earlier"
238 |    ]
239 |   },
240 |   {
241 |    "cell_type": "markdown",
242 |    "metadata": {},
243 |    "source": [
244 |     "Next, we pass it to our Embedding layer, where we use the \"weights\" parameter to indicate the use of the pretrained embedding weights we have loaded and the \"trainable\" parameter to tell Keras not to retrain the embedding layer."
245 |    ]
246 |   },
247 |   {
248 |    "cell_type": "code",
249 |    "execution_count": 33,
250 |    "metadata": {},
251 |    "outputs": [],
252 |    "source": [
253 |     "x = Embedding(len(tokenizer.word_index), embedding_matrix.shape[1],weights=[embedding_matrix],trainable=False)(inp)"
254 |    ]
255 |   },
256 |   {
257 |    "cell_type": "markdown",
258 |    "metadata": {},
259 |    "source": [
260 |     "Next, we pass it to a LSTM unit. But this time round, we will be using a Bidirectional LSTM instead because there are several kernels which shows a decent gain in accuracy by using Bidirectional LSTM.\n",
261 |     "\n",
262 |     "How does Bidirectional LSTM work? \n",
263 |     "\n",
264 |     "![](https://i.imgur.com/jaKiP0S.png)\n",
265 |     "\n",
266 |     "Imagine that the LSTM is split between 2 hidden states for each time step. As the sequence of words is being feed into the LSTM in a forward fashion, there's another reverse sequence that is feeding to the different hidden state at the same time. You might noticed later at the model summary that the output dimension of LSTM layer has doubled to 120 because 60 dimensions are used for forward, and another 60 are used for reverse.\n",
267 |     "\n",
268 |     "The greatest advantage in using Bidirectional LSTM is that when it runs backwards you preserve information from the future and using the two hidden states combined, you are able in any point in time to preserve information from both past and future.\n"
269 |    ]
270 |   },
271 |   {
272 |    "cell_type": "code",
273 |    "execution_count": 34,
274 |    "metadata": {},
275 |    "outputs": [
276 |     {
277 |      "name": "stdout",
278 |      "output_type": "stream",
279 |      "text": [
280 |       "WARNING:tensorflow:From F:\\Anaconda\\envs\\tensorflow-gpu\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py:1242: calling reduce_sum (from tensorflow.python.ops.math_ops) with keep_dims is deprecated and will be removed in a future version.\n",
281 |       "Instructions for updating:\n",
282 |       "keep_dims is deprecated, use keepdims instead\n"
283 |      ]
284 |     }
285 |    ],
286 |    "source": [
287 |     "x = Bidirectional(LSTM(60, return_sequences=True,name='lstm_layer',dropout=0.1,recurrent_dropout=0.1))(x)"
288 |    ]
289 |   },
290 |   {
291 |    "cell_type": "markdown",
292 |    "metadata": {},
293 |    "source": [
294 |     "Okay! With the LSTM behind us, we'll feed the output into the rest of the layers which we have done so in the previous kernel. "
295 |    ]
296 |   },
297 |   {
298 |    "cell_type": "code",
299 |    "execution_count": 35,
300 |    "metadata": {},
301 |    "outputs": [
302 |     {
303 |      "name": "stdout",
304 |      "output_type": "stream",
305 |      "text": [
306 |       "WARNING:tensorflow:From F:\\Anaconda\\envs\\tensorflow-gpu\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py:1208: calling reduce_max (from tensorflow.python.ops.math_ops) with keep_dims is deprecated and will be removed in a future version.\n",
307 |       "Instructions for updating:\n",
308 |       "keep_dims is deprecated, use keepdims instead\n"
309 |      ]
310 |     }
311 |    ],
312 |    "source": [
313 |     "x = GlobalMaxPool1D()(x)"
314 |    ]
315 |   },
316 |   {
317 |    "cell_type": "code",
318 |    "execution_count": 36,
319 |    "metadata": {},
320 |    "outputs": [],
321 |    "source": [
322 |     "x = Dropout(0.1)(x)"
323 |    ]
324 |   },
325 |   {
326 |    "cell_type": "code",
327 |    "execution_count": 37,
328 |    "metadata": {},
329 |    "outputs": [],
330 |    "source": [
331 |     "x = Dense(50, activation='relu')(x)"
332 |    ]
333 |   },
334 |   {
335 |    "cell_type": "code",
336 |    "execution_count": 38,
337 |    "metadata": {},
338 |    "outputs": [],
339 |    "source": [
340 |     "x = Dropout(0.1)(x)"
341 |    ]
342 |   },
343 |   {
344 |    "cell_type": "code",
345 |    "execution_count": 39,
346 |    "metadata": {},
347 |    "outputs": [],
348 |    "source": [
349 |     "x = Dense(6, activation='sigmoid')(x)"
350 |    ]
351 |   },
352 |   {
353 |    "cell_type": "code",
354 |    "execution_count": 40,
355 |    "metadata": {},
356 |    "outputs": [
357 |     {
358 |      "name": "stdout",
359 |      "output_type": "stream",
360 |      "text": [
361 |       "WARNING:tensorflow:From F:\\Anaconda\\envs\\tensorflow-gpu\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py:1344: calling reduce_mean (from tensorflow.python.ops.math_ops) with keep_dims is deprecated and will be removed in a future version.\n",
362 |       "Instructions for updating:\n",
363 |       "keep_dims is deprecated, use keepdims instead\n"
364 |      ]
365 |     }
366 |    ],
367 |    "source": [
368 |     "model = Model(inputs=inp, outputs=x)\n",
369 |     "model.compile(loss='binary_crossentropy',\n",
370 |     "             optimizer='adam',\n",
371 |     "             metrics=['accuracy'])"
372 |    ]
373 |   },
374 |   {
375 |    "cell_type": "code",
376 |    "execution_count": 41,
377 |    "metadata": {},
378 |    "outputs": [],
379 |    "source": [
380 |     "#在训练之前查看整个网络的架构是一个好主意，因为您不想把宝贵的时间浪费在错误的设置上。"
381 |    ]
382 |   },
383 |   {
384 |    "cell_type": "code",
385 |    "execution_count": 42,
386 |    "metadata": {},
387 |    "outputs": [
388 |     {
389 |      "name": "stdout",
390 |      "output_type": "stream",
391 |      "text": [
392 |       "_________________________________________________________________\n",
393 |       "Layer (type)                 Output Shape              Param #   \n",
394 |       "=================================================================\n",
395 |       "input_1 (InputLayer)         (None, 200)               0         \n",
396 |       "_________________________________________________________________\n",
397 |       "embedding_1 (Embedding)      (None, 200, 300)          63101100  \n",
398 |       "_________________________________________________________________\n",
399 |       "bidirectional_1 (Bidirection (None, 200, 120)          173280    \n",
400 |       "_________________________________________________________________\n",
401 |       "global_max_pooling1d_1 (Glob (None, 120)               0         \n",
402 |       "_________________________________________________________________\n",
403 |       "dropout_1 (Dropout)          (None, 120)               0         \n",
404 |       "_________________________________________________________________\n",
405 |       "dense_1 (Dense)              (None, 50)                6050      \n",
406 |       "_________________________________________________________________\n",
407 |       "dropout_2 (Dropout)          (None, 50)                0         \n",
408 |       "_________________________________________________________________\n",
409 |       "dense_2 (Dense)              (None, 6)                 306       \n",
410 |       "=================================================================\n",
411 |       "Total params: 63,280,736\n",
412 |       "Trainable params: 179,636\n",
413 |       "Non-trainable params: 63,101,100\n",
414 |       "_________________________________________________________________\n"
415 |      ]
416 |     }
417 |    ],
418 |    "source": [
419 |     "model.summary()"
420 |    ]
421 |   },
422 |   {
423 |    "cell_type": "markdown",
424 |    "metadata": {},
425 |    "source": [
426 |     "最后，我们启动了训练过程，目标是运行4个epoch，批大小为32。我们将训练和验证损失保存在一个变量中，这样我们可以查看是否存在过拟合。"
427 |    ]
428 |   },
429 |   {
430 |    "cell_type": "code",
431 |    "execution_count": 43,
432 |    "metadata": {},
433 |    "outputs": [],
434 |    "source": [
435 |     "#batch_size = 32\n",
436 |     "#epochs = 4\n",
437 |     "#hist = model.fit(X_t,y, batch_size=batch_size, epochs=epochs, validation_split=0.1)"
438 |    ]
439 |   },
440 |   {
441 |    "cell_type": "markdown",
442 |    "metadata": {},
443 |    "source": [
444 |     "模型的训练将比Kaggle kenel所允许的要长。我已经预先运行过了，这就是你们应该大致看到的结果"
445 |    ]
446 |   },
447 |   {
448 |    "cell_type": "markdown",
449 |    "metadata": {},
450 |    "source": [
451 |     "Train on 143613 samples, validate on 15958 samples\n",
452 |     "\n",
453 |     "Epoch 1/4\n",
454 |     "143613/143613 [==============================] - 2938s 20ms/step - loss: 0.0843 - acc: 0.9739 - val_loss: 0.0630 - val_acc: 0.9786\n",
455 |     "\n",
456 |     "Epoch 2/4\n",
457 |     "143613/143613 [==============================] - 3332s 23ms/step - loss: 0.0573 - acc: 0.9805 - val_loss: 0.0573 - val_acc: 0.9803\n",
458 |     "\n",
459 |     "Epoch 3/4\n",
460 |     "143613/143613 [==============================] - 3119s 22ms/step - loss: 0.0513 - acc: 0.9819 - val_loss: 0.0511 - val_acc: 0.9817\n",
461 |     "\n",
462 |     "Epoch 4/4\n",
463 |     "143613/143613 [==============================] - 3137s 22ms/step - loss: 0.0477 - acc: 0.9827 - val_loss: 0.0498 - val_acc: 0.9820\n"
464 |    ]
465 |   },
466 |   {
467 |    "cell_type": "code",
468 |    "execution_count": 44,
469 |    "metadata": {},
470 |    "outputs": [],
471 |    "source": [
472 |     "#loadEmbeddingMatrix('word2vec')"
473 |    ]
474 |   },
475 |   {
476 |    "cell_type": "markdown",
477 |    "metadata": {},
478 |    "source": [
479 |     "with"
480 |    ]
481 |   },
482 |   {
483 |    "cell_type": "code",
484 |    "execution_count": 45,
485 |    "metadata": {},
486 |    "outputs": [],
487 |    "source": [
488 |     "#loadEmbeddingMatrix('glove') #for GLOVE or\n",
489 |     "#loadEmbeddingMatrix('fasttext') #for fasttext"
490 |    ]
491 |   },
492 |   {
493 |    "cell_type": "markdown",
494 |    "metadata": {},
495 |    "source": [
496 |     "加载来自不同来源的预训练嵌入。\n",
497 |     "为了我们的基准。我已经预运行了它并收集了所有的结果。"
498 |    ]
499 |   },
500 |   {
501 |    "cell_type": "markdown",
502 |    "metadata": {},
503 |    "source": [
504 |     "**GLOVE:**\n",
505 |     "\n",
506 |     "Train on 143613 samples, validate on 15958 samples\n",
507 |     "\n",
508 |     "Epoch 1/4\n",
509 |     "143613/143613 [==============================] - 2470s 17ms/step - loss: 0.1160 - acc: 0.9656 - val_loss: 0.0935 - val_acc: 0.9703\n",
510 |     "\n",
511 |     "Epoch 2/4\n",
512 |     "143613/143613 [==============================] - 2448s 17ms/step - loss: 0.0887 - acc: 0.9721 - val_loss: 0.0800 - val_acc: 0.9737\n",
513 |     "\n",
514 |     "Epoch 3/4\n",
515 |     "143613/143613 [==============================] - 2410s 17ms/step - loss: 0.0799 - acc: 0.9745 - val_loss: 0.0753 - val_acc: 0.9757\n",
516 |     "\n",
517 |     "Epoch 4/4\n",
518 |     "143613/143613 [==============================] - 2398s 17ms/step - loss: 0.0753 - acc: 0.9760 - val_loss: 0.0724 - val_acc: 0.9768\n"
519 |    ]
520 |   },
521 |   {
522 |    "cell_type": "markdown",
523 |    "metadata": {},
524 |    "source": [
525 |     "**Fasttext:**\n",
526 |     "\n",
527 |     "Train on 143613 samples, validate on 15958 samples\n",
528 |     "\n",
529 |     "Epoch 1/4\n",
530 |     "143613/143613 [==============================] - 2800s 19ms/step - loss: 0.0797 - acc: 0.9757 - val_loss: 0.0589 - val_acc: 0.9795\n",
531 |     "\n",
532 |     "Epoch 2/4\n",
533 |     "143613/143613 [==============================] - 2756s 19ms/step - loss: 0.0561 - acc: 0.9808 - val_loss: 0.0549 - val_acc: 0.9804\n",
534 |     "\n",
535 |     "Epoch 3/4\n",
536 |     "143613/143613 [==============================] - 2772s 19ms/step - loss: 0.0507 - acc: 0.9819 - val_loss: 0.0548 - val_acc: 0.9811\n",
537 |     "\n",
538 |     "Epoch 4/4\n",
539 |     "143613/143613 [==============================] - 2819s 20ms/step - loss: 0.0474 - acc: 0.9828 - val_loss: 0.0507 - val_acc: 0.9817\n"
540 |    ]
541 |   },
542 |   {
543 |    "cell_type": "markdown",
544 |    "metadata": {},
545 |    "source": [
546 |     "当然，同样的baseline模型不使用任何预先训练的嵌入，直接从之前的kenel中获得，除了我们运行了4个epochs:"
547 |    ]
548 |   },
549 |   {
550 |    "cell_type": "markdown",
551 |    "metadata": {},
552 |    "source": [
553 |     "Train on 143613 samples, validate on 15958 samples\n",
554 |     "\n",
555 |     "Epoch 1/4\n",
556 |     "143613/143613 [==============================] - 5597s 39ms/step - loss: 0.0633 - acc: 0.9788 - val_loss: 0.0480 - val_acc: 0.9825\n",
557 |     "\n",
558 |     "Epoch 2/4\n",
559 |     "143613/143613 [==============================] - 5360s 37ms/step - loss: 0.0448 - acc: 0.9832 - val_loss: 0.0464 - val_acc: 0.9828\n",
560 |     "\n",
561 |     "Epoch 3/4\n",
562 |     "143613/143613 [==============================] - 5352s 37ms/step - loss: 0.0390 - acc: 0.9848 - val_loss: 0.0470 - val_acc: 0.9829\n",
563 |     "\n",
564 |     "Epoch 4/4\n",
565 |     "129984/143613 [==============================] - 5050s 37ms/step - loss: 0.0386 - acc: 0.9858 - val_loss: 0.0478 - val_acc: 0.9830"
566 |    ]
567 |   },
568 |   {
569 |    "cell_type": "markdown",
570 |    "metadata": {},
571 |    "source": [
572 |     "如果我们把损失画成图表就容易多了"
573 |    ]
574 |   },
575 |   {
576 |    "cell_type": "code",
577 |    "execution_count": 46,
578 |    "metadata": {},
579 |    "outputs": [],
580 |    "source": [
581 |     "all_losses = {\n",
582 |     "'word2vec_loss': [0.084318213647104789,\n",
583 |     "  0.057314205012433353,\n",
584 |     "  0.051338302593577821,\n",
585 |     "  0.047672802178572039],\n",
586 |     " 'word2vec_val_loss': [0.063002561892695971,\n",
587 |     "  0.057253835496480658,\n",
588 |     "  0.051085027624451551,\n",
589 |     "  0.049801279793734249],\n",
590 |     "'glove_loss': [0.11598931579683543,\n",
591 |     "  0.088738223480436862,\n",
592 |     "  0.079895263566000005,\n",
593 |     "  0.075343037429358703],\n",
594 |     " 'glove_val_loss': [0.093467933030432285,\n",
595 |     "  0.080007083813922117,\n",
596 |     "  0.075349041991106688,\n",
597 |     "  0.072366507668134517],\n",
598 |     " 'fasttext_loss': [0.079714499498945865,\n",
599 |     "  0.056074704045674786,\n",
600 |     "  0.050703874653286324,\n",
601 |     "  0.047420131195761134],\n",
602 |     " 'fasttext_val_loss': [0.058888281775148932,\n",
603 |     "  0.054906051694414926,\n",
604 |     "  0.054768857866843601,\n",
605 |     "  0.050697043558286421],\n",
606 |     " 'baseline_loss': [0.063304489498915865,\n",
607 |     "  0.044864004045674786,\n",
608 |     "  0.039013874651286124,\n",
609 |     "  0.038630130175761134],\n",
610 |     " 'baseline_val_loss': [0.048044281075148932,\n",
611 |     "  0.046414051594414926,\n",
612 |     "  0.047058757860843601,\n",
613 |     "  0.047886043558285421]\n",
614 |     "}"
615 |    ]
616 |   },
617 |   {
618 |    "cell_type": "code",
619 |    "execution_count": 47,
620 |    "metadata": {},
621 |    "outputs": [
622 |     {
623 |      "data": {
624 |       "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYwAAAEWCAYAAAB1xKBvAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nUcAQEZFYFDBERCSW/wccbdUbvdT71QAAAABJRU5ErkJggg==\n",
625 |       "text/plain": [
626 |        "<matplotlib.figure.Figure at 0x207933d4518>"
627 |       ]
628 |      },
629 |      "metadata": {},
630 |      "output_type": "display_data"
631 |     }
632 |    ],
633 |    "source": [
634 |     "#f, ax = plt.subplots(1)\n",
635 |     "epochRange = np.arange(1,5,1)\n",
636 |     "plt.plot(epochRange,all_losses['word2vec_loss'])\n",
637 |     "plt.plot(epochRange,all_losses['glove_loss'])\n",
638 |     "plt.plot(epochRange,all_losses['fasttext_loss'])\n",
639 |     "plt.plot(epochRange,all_losses['baseline_loss'])\n",
640 |     "plt.title('Training loss for different embeddings')\n",
641 |     "plt.ylabel('loss')\n",
642 |     "plt.xlabel('epoch')\n",
643 |     "plt.legend(['Word2Vec', 'GLOVE','FastText','Baseline'], loc='upper left')\n",
644 |     "plt.show()"
645 |    ]
646 |   },
647 |   {
648 |    "cell_type": "markdown",
649 |    "metadata": {},
650 |    "source": [
651 |     "看起来baseline的训练损失是最小的。但是在我们结束这个案例并选择baseline模型作为赢家之前，这个情节并不能说明全部，因为在baseline模型中似乎存在一些过拟合。<br>从第二阶段开始，随着验证损失大于训练损失，过拟合开始出现"
652 |    ]
653 |   },
654 |   {
655 |    "cell_type": "code",
656 |    "execution_count": 48,
657 |    "metadata": {},
658 |    "outputs": [
659 |     {
660 |      "data": {
661 |       "image/png": "iVBORw0KGgoAAAANSUhEUgAAAZIAAAEWCAYAAABMoxE0AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAElFTkSuQmCC\n",
662 |       "text/plain": [
663 |        "<matplotlib.figure.Figure at 0x20793e589b0>"
664 |       ]
665 |      },
666 |      "metadata": {},
667 |      "output_type": "display_data"
668 |     }
669 |    ],
670 |    "source": [
671 |     "epochRange = np.arange(1,5,1)\n",
672 |     "plt.plot(epochRange,all_losses['baseline_loss'])\n",
673 |     "plt.plot(epochRange,all_losses['baseline_val_loss'])\n",
674 |     "plt.title('Training Vs Validation loss for baseline model')\n",
675 |     "plt.ylabel('loss')\n",
676 |     "plt.xlabel('epoch')\n",
677 |     "plt.legend(['Training', 'Validation'], loc='upper left')\n",
678 |     "plt.show()"
679 |    ]
680 |   },
681 |   {
682 |    "cell_type": "markdown",
683 |    "metadata": {},
684 |    "source": [
685 |     "剩下的呢?让我们画出所有的训练/验证损失图来进行比较。"
686 |    ]
687 |   },
688 |   {
689 |    "cell_type": "code",
690 |    "execution_count": 49,
691 |    "metadata": {},
692 |    "outputs": [
693 |     {
694 |      "data": {
695 |       "image/png": "iVBORw0KGgoAAAANSUhEUgAABIsAAARuCAYAAABTBrdfAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
696 |       "text/plain": [
697 |        "<matplotlib.figure.Figure at 0x207940979b0>"
698 |       ]
699 |      },
700 |      "metadata": {},
701 |      "output_type": "display_data"
702 |     }
703 |    ],
704 |    "source": [
705 |     "f, ((ax1, ax2), (ax3, ax4)) = plt.subplots(2, 2, sharex='col', sharey='row',figsize=(20, 20))\n",
706 |     "\n",
707 |     "plt.title('Training Vs Validation loss for all embeddings')\n",
708 |     "ax1.plot(epochRange,all_losses['baseline_loss'])\n",
709 |     "ax1.plot(epochRange,all_losses['baseline_val_loss'])\n",
710 |     "ax1.set_title('Baseline')\n",
711 |     "ax1.set_ylim(0.03, 0.12)\n",
712 |     "\n",
713 |     "ax2.plot(epochRange,all_losses['word2vec_loss'])\n",
714 |     "ax2.plot(epochRange,all_losses['word2vec_val_loss'])\n",
715 |     "ax2.set_title('Word2Vec')\n",
716 |     "ax2.set_ylim(0.03, 0.12)\n",
717 |     "\n",
718 |     "ax3.plot(epochRange,all_losses['glove_loss'])\n",
719 |     "ax3.plot(epochRange,all_losses['glove_val_loss'])\n",
720 |     "ax3.set_title('GLOVE')\n",
721 |     "ax3.set_ylim(0.03, 0.12)\n",
722 |     "\n",
723 |     "\n",
724 |     "ax4.plot(epochRange,all_losses['fasttext_loss'])\n",
725 |     "ax4.plot(epochRange,all_losses['fasttext_val_loss'])\n",
726 |     "ax4.set_title('FastText')\n",
727 |     "ax4.set_ylim(0.03, 0.12)\n",
728 |     "\n",
729 |     "plt.show()"
730 |    ]
731 |   },
732 |   {
733 |    "cell_type": "markdown",
734 |    "metadata": {},
735 |    "source": [
736 |     "考虑到所有的损失，很容易看出哪个是最好的选择。虽然这只手套看起来仍有一些空间，但与其他手套相比，它的损耗是很大的。另一方面，Word2Vec和FastText分别在第4和第3个时代开始过度匹配。那么你会选择哪一个作为赢家呢?在我看来，仍然是基线模型。\n",
737 |     "那么问题出在哪里呢?难道预先训练的嵌入不应该改进吗?因为它是用来自大量特征丰富语料库的数十亿单词进行训练的。\n",
738 |     "一种可能性是，这些预先训练的嵌入没有针对文本i进行训练"
739 |    ]
740 |   },
741 |   {
742 |    "cell_type": "code",
743 |    "execution_count": 50,
744 |    "metadata": {},
745 |    "outputs": [],
746 |    "source": [
747 |     "wordCount = {'word2vec':66078,'glove':81610,'fasttext':59613,'baseline':210337}"
748 |    ]
749 |   },
750 |   {
751 |    "cell_type": "code",
752 |    "execution_count": 51,
753 |    "metadata": {},
754 |    "outputs": [
755 |     {
756 |      "data": {
757 |       "image/png": "iVBORw0KGgoAAAANSUhEUgAAAZgAAAEICAYAAABiXeIWAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\ninDAmJlZEQ4YMzMrwgFjZmZF/H9S3otlqtogAgAAAABJRU5ErkJggg==\n",
758 |       "text/plain": [
759 |        "<matplotlib.figure.Figure at 0x20794074d30>"
760 |       ]
761 |      },
762 |      "metadata": {},
763 |      "output_type": "display_data"
764 |     }
765 |    ],
766 |    "source": [
767 |     "ind = np.arange(0,4,1)  # the x locations for the groups\n",
768 |     "width = 0.35       # the width of the bars\n",
769 |     "\n",
770 |     "plt.title('Number of common words used in different embeddings')\n",
771 |     "embNames = list(wordCount.keys())\n",
772 |     "embVals = list(wordCount.values())\n",
773 |     "plt.barh(ind,embVals,align='center', height=0.5, color='m',tick_label=embNames)\n",
774 |     "plt.show()"
775 |    ]
776 |   },
777 |   {
778 |    "cell_type": "markdown",
779 |    "metadata": {},
780 |    "source": [
781 |     "从上面的条形图中可以明显看出，baseline包含的单词最多<br>\n",
782 |     "因为嵌入层是使用数据集中的单词进行训练的。<br>重要的结论是，预先训练的嵌入只包含大约60,000个共同的单词(少于基线的一半)，\n",
783 |     "<br>而由这些预先训练的权重构建的嵌入层不能很好地表示训练数据。\n",
784 |     "<br>虽然构建您自己的嵌入需要更长的时间，但它可能是值得的，因为它是专门为您的上下文构建的。"
785 |    ]
786 |   },
787 |   {
788 |    "cell_type": "markdown",
789 |    "metadata": {},
790 |    "source": [
791 |     "这就完成了这个内核!我希望有人在这个内核中学到了一些东西。\n",
792 |     "<br>如果你发现了一个错误，请在下面的评论中告诉我。\n",
793 |     "<br>感谢阅读，祝比赛顺利!"
794 |    ]
795 |   },
796 |   {
797 |    "cell_type": "markdown",
798 |    "metadata": {},
799 |    "source": [
800 |     "**TODO:**\n",
801 |     "1. There are many pretrained embeddings in Kaggle, and they are trained in different contexts of text corpus. You could try out other pretrained embeddings that is more suitable to the dataset in our competition.\n",
802 |     "2. Introduce LSTM drop out and recurrent drop out in baseline model, and tune the dropout rate to decrease overfitting."
803 |    ]
804 |   },
805 |   {
806 |    "cell_type": "code",
807 |    "execution_count": null,
808 |    "metadata": {},
809 |    "outputs": [],
810 |    "source": []
811 |   }
812 |  ],
813 |  "metadata": {
814 |   "kernelspec": {
815 |    "display_name": "tensorflow-gpu",
816 |    "language": "python",
817 |    "name": "tensorflow-gpu"
818 |   },
819 |   "language_info": {
820 |    "codemirror_mode": {
821 |     "name": "ipython",
822 |     "version": 3
823 |    },
824 |    "file_extension": ".py",
825 |    "mimetype": "text/x-python",
826 |    "name": "python",
827 |    "nbconvert_exporter": "python",
828 |    "pygments_lexer": "ipython3",
829 |    "version": "3.6.2"
830 |   }
831 |  },
832 |  "nbformat": 4,
833 |  "nbformat_minor": 2
834 | }
835 | 


--------------------------------------------------------------------------------
/use_tutorial.md:
--------------------------------------------------------------------------------
  1 | ## 简介
  2 | 我是NLP小白一枚。从1.16开始接触Quora比赛，以下是我对**预训练词向量**的一些归纳~(大部分是照搬kernel的 ...)
  3 | - 预训练词向量会带来优势吗？(通过kernel里的例子)
  4 | - 使用预训练词向量之前的清洗
  5 | - 如何使用预训练词向量
  6 | 
  7 | ## 1. 预训练词向量真会给你带来额外的优势吗？
  8 | 
  9 | ![word2vec](https://qph.fs.quoracdn.net/main-qimg-3e812fd164a08f5e4f195000fecf988f)
 10 | 
 11 | - 参考的[kernel](https://www.kaggle.com/sbongo/do-pretrained-embeddings-give-you-the-extra-edge)
 12 | - 嵌入通常表示基于在文本语料库中一起出现的频率的单词的几何编码。下面描述的word嵌入的各种实现在构造方式上有所不同。
 13 | 
 14 | ### 1.1 Word2Vec
 15 | -  主要思想：在每个单词的**上下文中**训练一个模型，这样类似的单词就会有类似的数字表示。
 16 | -  就像一个正常的前馈网络，其中你有一组自变量和一个你试图预测的目标因变量，你首先把你的句子分解成单词(tokenize)，并根据窗口大小创建一些单词对。其中一个组合可以是一对词，如(cat, purr)其中cat是自变量(X) purr是我们要预测的目标因变量(Y)
 17 | -  我们将‘cat’通过一个初始化为随机权值的嵌入层送入NN，并将其通过softmax层传递，最终目的是预测‘purr’。优化方法如SGD最小化损失函数“(target word | context words)”，该函数的目的是在给定上下文单词的情况下最小化预测目标单词的损失。如果我们用足够的时间来做这件事，嵌入层中的权值最终将代表单词向量的词汇，也就是单词在这个几何向量空间中的“坐标”。
 18 | ![Word2Vec示意](https://i.imgur.com/R8VLFs2.png)
 19 | - 注：上面的例子采用了skip模型。对于连续单词包(CBOW)，我们基本上是在给定上下文的情况下预测一个单词  
 20 | 
 21 | ### 1.2 GLOVE
 22 | 
 23 | ​        GLOVE的工作原理类似于Word2Vec。上面可以看到Word2Vec是一个“预测”模型，它预测给定单词的上下文，GLOVE通过构造一个共现矩阵(words X context)来学习，该矩阵主要计算单词在上下文中出现的频率。因为它是一个巨大的矩阵，我们分解这个矩阵来得到一个低维的表示。有很多细节是相互配合的，但这只是粗略的想法。
 24 | 
 25 | ### 1.3 FastText
 26 | ​        FastText与上面的两个嵌入有很大的不同。Word2Vec和GLOVE将每个单词作为最小的训练单元，而FastText使用n-gram字符作为最小的单元。例如，单词vector，“apple”，可以分解为单词vector的不同单位，如“ap”，“app”，“ple”。使用FastText的最大好处是，它可以为罕见的单词，甚至是在训练过程中没有看到的单词生成更好的嵌入，因为n-gram字符向量与其他单词共享。这是Word2Vec和GLOVE无法做到的。
 27 | 
 28 | ### 1.4 See
 29 | 实验基于Kaggle的比赛Toxic Comment Classification Challenge，请参见kernel。
 30 | 我也实验了该kernel，也可以看我的[中文版该kernel](https://github.com/Cooper111/pretrained_embedding_tutorial/blob/master/Toxic%20Comment%20Classification%20Challenge/Do_Pretrained_Embeddings_Give_You_The_Extra_Edge%EF%BC%9F.ipynb)。
 31 | 注：其中值得一提的技巧
 32 | - 该kernel开头给出了另一个keras教程，其中一步，通过查看句子长度图分布来设置maxlen
 33 | - 不同的预训练词向量基于不同的语料训练而成，画出训练和验证损失来评估性能
 34 | - 使用函数封装制作Embedding_matrix
 35 | - 查看预训练词向量包含的喂入文本词汇占总喂入文本词汇的比率或柱状统计
 36 | 
 37 | ## 2.使用预训练词向量之前的清洗
 38 | - 参考的[kernel](https://www.kaggle.com/christofhenkel/how-to-preprocessing-when-using-embeddings)
 39 | - 实验基于Kaggle的比赛Quora Insincere Questions Classification，请参见kernel。
 40 | 我也实验了该kernel，也可以看我的[中文版该kernel](https://github.com/Cooper111/pretrained_embedding_tutorial/blob/master/Quora/how-to-preprocessing-when-using-embeddings.ipynb)。
 41 | 
 42 | ### 2.1 清洗的两条法则
 43 | - 看情况使用stopwords操作，不一定得按照标准的清洗流程
 44 | - 使用于训练的语料尽可能接近预训练的语料，就是未登录词尽量少
 45 | 
 46 | ### 2.2 几个好用的函数
 47 | - 计算包含的单词的出现次数，返回w2id的字典
 48 | ```
 49 | 函数：
 50 | #我将使用下面的函数来跟踪我们的训练词汇，它将遍历我们的所有文本并计算包含的单词的出现次数。
 51 | def build_vocab(sentences, verbose = True):
 52 |     # 参数  sentences list of list of words，就是二维的
 53 |     # 返回值 对应  词和词的次数 的字典
 54 |     vocab = {}
 55 |     for sentence in tqdm(sentences, disable = (not verbose)):
 56 |         for word in sentence:
 57 |             try:
 58 |                 vocab[word] += 1
 59 |             except KeyError:
 60 |                 vocab[word] = 1
 61 |     return vocab
 62 | 使用方法：
 63 | #因此，让我们填充词汇表并显示前5个元素及其计数。注意，现在我们可以使用progess_apply查看进度条
 64 | sentences = train['question_text'].progress_apply(lambda x: x.split()).values
 65 | vocab = build_vocab(sentences)
 66 | print({k: vocab[k] for k in list(vocab)[:5]})
 67 | ```
 68 | 
 69 | - 检查词汇表和嵌入之间的交集
 70 | 
 71 | ```
 72 | #接下来，我定义一个函数来检查词汇表和嵌入之间的交集。它将输出一个out of vocabulary (oov)单词列表，我们可以使用它来改进我们的预处理
 73 | import operator
 74 | def check_coverage(vocab, embeddings_index):
 75 |     a = {}
 76 |     oov = {}
 77 |     k = 0
 78 |     i = 0
 79 |     for word in tqdm(vocab):
 80 |         try:
 81 |             a[word] = embeddings_index[word]
 82 |             k += vocab[word]
 83 |         except:
 84 |             oov[word] = vocab[word]
 85 |             i += vocab[word]
 86 |             pass
 87 |     
 88 |     print('Found embeddings for {:.2%} of vocab'.format(len(a) /  len(vocab)))
 89 |     print('Found embeddings for {:.2%} of all text'.format(k / (k + i)))
 90 |     sorted_x = sorted(oov.items(), key=operator.itemgetter(1))[::-1]#取axis=1维度进行排序，并换为逆序
 91 |     
 92 |     return sorted_x
 93 | 
 94 | 使用方法：
 95 | oov = check_coverage(vocab, embedding_index)
 96 | 
 97 | 样例输出：
 98 | 100%|█████| 253623/253623 [00:00<00:00, 329145.72it/s]
 99 | Found embeddings for 57.38% of vocab
100 | Found embeddings for 89.99% of all text
101 | 
102 | 然后可以 oov[:10]这样查看次数靠前的未登录词
103 | ```
104 | - 清洗操作
105 | 
106 | ```
107 | train['question_text'] = train['question_text'].progress_apply(lambda x: clean_numbers(x))#progress_apply可以显示进度，tqdm也是
108 | sentences = train['question_text'].progress_apply(lambda x: x.split())#分割英文句子成词汇列表
109 | vocab = build_vocab(sentences)#构建词汇表
110 | ```
111 | 
112 | ### 2.3 清洗流程
113 | - ①加载数据，导入预训练词向量
114 | - ②对于Pandas.Series实例应用apply清洗函数
115 | >e.g.   train['question_text'] =  train['question_text'].progress_apply(lambda x: clean_text(x))
116 | 
117 | - ③构造词汇表
118 | - ④检查词汇表和嵌入之间的交集
119 | - ⑤查看未登录词靠前的几个，构造新的清洗函数
120 | - ②③④⑤循环，例子见2.2清洗操作
121 | 
122 | ### 2.4 清洗思路
123 | - ‘a’,'to'这种停用词在训练GoogleNews预训练词向量训练时被删除了,所以会出现在oov靠前，在最后对用于训练的文本剔除这些词(Kernel里这么做的)
124 | - 特殊符号的去除视情况而定，一些空格替换，一些去除（效果： 未登录词类占总词汇类比76%to43%）
125 | - 数字再处理，预训练词向量里所有大于9的数字都被hashs替换了。即成为# #,123变成# # #或15.80€变成# #,# #€。因此，让我们模拟这个预处理步骤来进一步改进我们的嵌入式覆盖率（效果：未登录词词类占总词汇类 43%to40%）
126 | - 口语用词替换成标准词汇（和思路第一步一起的效果：未登录词量占总词量 11% to 1%）
127 | 
128 | ## 3. 如何使用预训练词向量
129 | - 参考的[kernel](https://www.kaggle.com/sudalairajkumar/a-look-at-different-embeddings)
130 | - 实验基于Kaggle的比赛Quora Insincere Questions Classification，请参见kernel。
131 | 我也实验了该kernel，也可以看我的[中文版该kernel](https://github.com/Cooper111/pretrained_embedding_tutorial/blob/master/Quora/A%20look%20at%20different%20embeddings.!.ipynb)。
132 | 
133 | ### 3.1 使用流程
134 | - 将文本划分训练集和验证集
135 | 
136 |   ```
137 |   ## split to train and val
138 |   train_df, val_df = train_test_split(train_df, test_size=0.1, random_state=2018)
139 |   ```
140 | 
141 | 
142 | - 填充文本缺失值
143 | 
144 |   ```
145 |   ## fill up the missing values
146 |   train_X = train_df['question_text'].fillna('_na_').values
147 |   val_X = val_df['question_text'].fillna('_na_').values
148 |   test_X = test_df['question_text'].fillna('_na_').values
149 |   ```
150 | 
151 |   
152 | 
153 | - 使用Tokenizer讲文本转换为向量序列
154 | 
155 |   ```
156 |   ## Tokenize the sentences
157 |   tokenizer = Tokenizer(num_words=max_features)
158 |   tokenizer.fit_on_texts(list(train_X))
159 |   train_X = tokenizer.texts_to_sequences(train_X)
160 |   val_X = tokenizer.texts_to_sequences(val_X)
161 |   test_X = tokenizer.texts_to_sequences(test_X)
162 |   ```
163 | 
164 |   
165 | 
166 | - 根据max_len来Pad_sequences,固定句子长度
167 | 
168 |   ```
169 |   ## Pad the sentences
170 |   train_X = pad_sequences(train_X, maxlen=maxlen)
171 |   val_X = pad_sequences(val_X, maxlen=maxlen)
172 |   test_X = pad_sequences(test_X, maxlen=maxlen)
173 |   ```
174 | 
175 |   
176 | 
177 | - 构造词向量嵌入矩阵
178 |   法①
179 | 
180 |   ```
181 |   EMBEDDING_FILE = '../input/embeddings/glove.840B.300d/glove.840B.300d.txt'
182 |   def get_coefs(word, *arr):
183 |       return word, np.asarray(arr, dtype='float32')
184 |   embeddings_index = dict(get_coefs(*o.split(" ")) for o in open(EMBEDDING_FILE, encoding='utf-8') if len(o)>100 )
185 |   
186 |   all_embs = np.stack(embeddings_index.values())
187 |   emb_mean, emb_std = all_embs.mean(), all_embs.std()
188 |   embed_size = all_embs.shape[1]
189 |   
190 |   word_index = tokenizer.word_index
191 |   nb_words = min(max_features, len(word_index))
192 |   embedding_matrix = np.random.normal(emb_mean, emb_std, (nb_words, embed_size))
193 |   for word, i in word_index.items():
194 |       if i >= max_features:
195 |           continue
196 |       embedding_vector = embeddings_index.get(word)
197 |       if embedding_vector is not None:
198 |           embedding_matrix[i] = embedding_vector
199 |   ```
200 |   法②
201 | 
202 |   ```
203 |   def loadEmbeddingMatrix(typeToLoad):
204 |       #根据Embedding的不同，从Kaggle加载不同的嵌入文件
205 |       #我们要实验的矩阵
206 |       if(typeToLoad=='glove'):
207 |           EMBEDDING_FILE = '../input/glove.twitter.27B.25d.txt'
208 |           embed_size = 25
209 |       elif(typeToLoad=='word2vec'):
210 |           word2vecDict = word2vec.KeyedVectors.load_word2vec_format('../../Quora/input/embeddings/GoogleNews-vectors-negative300/GoogleNews-vectors-negative300.bin', binary=True)
211 |           embed_size = 300
212 |       elif(typeToLoad=='fasttext'):
213 |           EMBEDDING_FILE='../input/wiki.simple.vec'
214 |           embed_size = 300
215 |       
216 |       if(typeToLoad=='glove' or typeToLoad=='fasttext'):
217 |           embeddings_index = dict()
218 |           #通过遍历文件的每一行，将Embedding权重转移到字典中。
219 |           f = open(EMBEDDING_FILE)
220 |           for line in f:
221 |               #split up line into an indexed array
222 |               values = line.split()
223 |               #first index is word
224 |               word = values[0]
225 |               #store the rest of the values in the array as a new array
226 |               coefs = np.asarray(values[1:], dtype='float32')
227 |               embeddings_index[word] = coefs #50 dimensions
228 |           f.close()
229 |           print('Loaded %s word vectors.' % len(embeddings_index))
230 |       else:
231 |           embeddings_index = dict()
232 |           for word in word2vecDict.wv.vocab:
233 |               embeddings_index[word] = word2vecDict.word_vec(word)
234 |           print('Loaded %s word vectors.' % len(embeddings_index))
235 |       
236 |       gc.collect()
237 |       #我们得到嵌入权值的均值和标准差，这样我们就可以保持
238 |       #对于我们自己随机产生的权值的其余部分，同样的统计数据。
239 |       all_embs = np.stack(list(embeddings_index.values()))
240 |       emb_mean, emb_std = all_embs.mean(), all_embs.std()
241 |       
242 |       nb_words = len(tokenizer.word_index)
243 |       #我们讲设置Embedding的尺寸为我们复制的预训练的维度
244 |       #Embedding矩阵大小将为 词汇中词数 X Embedding Size
245 |       embedding_matrix = np.random.normal(emb_mean, emb_std, (nb_words, embed_size))
246 |       gc.collect()
247 |       
248 |       #使用新创建的嵌入矩阵，我们将用两个矩阵中的单词填充它
249 |       #我们自己的字典和加载预训练嵌入。
250 |       embeddedCount = 0
251 |       for word, i in tokenizer.word_index.items():
252 |           i -= 1
253 |           #然后我们看看这个词是否在GLove的字典里，如果是，得到相应的权重
254 |           embedding_vector = embeddings_index.get(word)
255 |           #并存储在Embedding矩阵中，我们稍后将用其进行训练。
256 |           if embedding_vector is not None:
257 |               embedding_matrix[i] = embedding_vector
258 |               embeddedCount += 1
259 |       print('total embedded:', embeddedCount,'common words')
260 |       
261 |       del(embeddings_index)
262 |       gc.collect()
263 |       
264 |       #最后，返回Embedding矩阵
265 |       return embedding_matrix
266 |       
267 |   ```
268 | 
269 |   
270 | 
271 | - 训练模型
272 | 
273 |   ```
274 |   model.fit(train_X, train_y, batch_size=512, epochs=2, validation_data=(val_X, val_y))
275 |   ```
276 | 
277 |   
278 | 
279 | - 得到验证样本预测和F1得分的最佳阈值,得到测试集预测
280 | 
281 |   ```
282 |   pred_noemb_val_y = model.predict([val_X], batch_size=1024, verbose=1)
283 |   for thresh in np.arange(0.1, 0.501, 0.01):
284 |       thresh = np.round(thresh, 2)
285 |       print("F1 score at threshold {0} is {1}".format(thresh, metrics.f1_score(val_y, (pred_noemb_val_y>thresh).astype(int))))
286 |    #得到测试集预测
287 |    pred_noemb_test_y = model.predict([test_X], batch_size=1024, verbose=1)
288 |   ```
289 | 
290 |   
291 | 
292 | - 清理一些内存
293 | 
294 |   ```
295 |   del model, inp, x
296 |   import gc
297 |   gc.collect()
298 |   time.sleep(10)
299 |   ```
300 | 
301 | 
302 | 
303 | ## 3.2 词向量优化方法
304 | - 通过统一模型对不同预训练词向量和无预训练词向量，绘制损失曲线来选取
305 | - 尝试给定各个结果以权重进行ensemble
306 | 
307 | 
308 | ## TODO
309 | - 学习大佬的总结~
310 | - 复现大佬的代码~
311 | - 吸取建议，然后补全词向量优化方法~
312 | - 祝各位比赛取的好成绩，我先去旅游到28号再继续···
313 | 
314 | 
315 | <hr>
316 | 学号： 071-沈凯文


--------------------------------------------------------------------------------