\u9996\u5148\u662f\u8bcd\u5411\u91cf\uff0c\u6559\u7a0b\u8bf4\u4e86\u4e00\u5806\uff0c\u610f\u601d\u5176\u5b9e\u5c31\u662f\u4e00\u4e2a\u8bcd\u7528\u4e00\u7ec4\u5411\u91cf\u8868\u793a\uff0c\u7136\u540e\u8bcd\u5178\u7684\u54c8\u5e0c\u8868\u7ed3\u6784\u67e5\u8be2\u4e0d\u591f\u5feb\uff0c\u6240\u4ee5\u7528\u72ec\u70ed\u7f16\u7801\uff0b\u77e9\u9635\u8ba1\u7b97\u5728GPU\u4e0a\u83b7\u5f97\u9ad8\u6548\u5904\u7406\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u53eb\u505aEmbedding Lookup<\/p>\n\n\n\n
CBOW\uff0c\u5148\u5728\u53e5\u5b50\u4e2d\u9009\u5b9a\u4e00\u4e2a\u4e2d\u5fc3\u8bcd\uff0c\u5e76\u628a\u5176\u5b83\u8bcd\u4f5c\u4e3a\u8fd9\u4e2a\u4e2d\u5fc3\u8bcd\u7684\u4e0a\u4e0b\u6587<\/p>\n\n\n\n skip-gram\uff0c\u540c\u6837\u5148\u9009\u5b9a\u4e00\u4e2a\u4e2d\u5fc3\u8bcd\uff0c\u5e76\u628a\u5176\u4ed6\u8bcd\u4f5c\u4e3a\u8fd9\u4e2a\u4e2d\u5fc3\u8bcd\u7684\u4e0a\u4e0b\u6587\u3002<\/p>\n\n\n\n Skip-gram\u7684\u7406\u60f3\u5b9e\u73b0<\/p>\n\n\n\n \u7406\u60f3\u5b9e\u73b0\u5c31\u662f\u6309\u7167\u4e0a\u9762\u7684\u793a\u610f\u56fe\u5b9e\u73b0\uff0c\u9700\u8981\u5b66\u4e60\u7684\u77e9\u9635\u7684\u662f\u8d1f\u8d23Embedding\u7684\u8fd9\u4e2a\u77e9\u9635W0\uff0c\u4ee5\u53ca\u518d\u8f6c\u6362\u6210onehot\u5411\u91cf\u7684\u77e9\u9635\u4e5f\u9700\u8981\u5b66\u4e60<\/p>\n\n\n\n Skip-gram\u7684\u5b9e\u9645\u5b9e\u73b0<\/p>\n\n\n\n \u7136\u800c\u5728\u5b9e\u9645\u60c5\u51b5\u4e2d\uff0cvocab_size\u901a\u5e38\u5f88\u5927\uff08\u51e0\u5341\u4e07\u751a\u81f3\u51e0\u767e\u4e07\uff09\uff0c\u5bfc\u81f4W0\u548cW1\u4e5f\u4f1a\u975e\u5e38\u5927\u3002\u5bf9\u4e8eW\u200b\u800c\u8a00\uff0c\u6240\u53c2\u4e0e\u7684\u77e9\u9635\u8fd0\u7b97\u5e76\u4e0d\u662f\u901a\u8fc7\u4e00\u4e2a\u77e9\u9635\u4e58\u6cd5\u5b9e\u73b0\uff0c\u800c\u662f\u901a\u8fc7\u6307\u5b9aID\uff0c\u5bf9\u53c2\u6570W0\u8fdb\u884c\u8bbf\u5b58\u7684\u65b9\u5f0f\u83b7\u53d6\u3002\u7136\u800c\u5bf9W1\u800c\u8a00\uff0c\u4ecd\u8981\u5904\u7406\u4e00\u4e2a\u975e\u5e38\u5927\u7684\u77e9\u9635\u8fd0\u7b97\uff08\u8ba1\u7b97\u8fc7\u7a0b\u975e\u5e38\u7f13\u6162\uff0c\u9700\u8981\u6d88\u8017\u5927\u91cf\u7684\u5185\u5b58\/\u663e\u5b58\uff09\u3002\u4e3a\u4e86\u7f13\u89e3\u8fd9\u4e2a\u95ee\u9898\uff0c\u901a\u5e38\u91c7\u53d6\u8d1f\u91c7\u6837\uff08negative_sampling\uff09\u7684\u65b9\u5f0f\u6765\u8fd1\u4f3c\u6a21\u62df\u591a\u5206\u7c7b\u4efb\u52a1\u3002\u6b64\u65f6\u65b0\u5b9a\u4e49\u7684W0\u548cW1\u200b\u5747\u4e3a\u5f62\u72b6\u4e3a[vocab_size, embedding_size]\u7684\u5f20\u91cf\u3002<\/p>\n\n\n\n \u4f7f\u7528\u98de\u6868\u5b9e\u73b0Skip-gram<\/p>\n\n\n\n 1\u5bfc\u5165\u76f8\u5173\u5e93<\/p>\n\n\n\n 2\u6570\u636e\u96c6<\/p>\n\n\n\n \u9996\u5148\uff0c\u627e\u5230\u4e00\u4e2a\u5408\u9002\u7684\u8bed\u6599\u7528\u4e8e\u8bad\u7ec3word2vec\u6a21\u578b\u3002\u4f7f\u7528text8\u6570\u636e\u96c6\uff0c\u8fd9\u4e2a\u6570\u636e\u96c6\u91cc\u5305\u542b\u4e86\u5927\u91cf\u4ece\u7ef4\u57fa\u767e\u79d1\u6536\u96c6\u5230\u7684\u82f1\u6587\u8bed\u6599\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u5982\u4e0b\u4ee3\u7801\u4e0b\u8f7d\u6570\u636e\u96c6\uff0c\u4e0b\u8f7d\u540e\u7684\u6587\u4ef6\u88ab\u4fdd\u5b58\u5728\u5f53\u524d\u76ee\u5f55\u7684\u201ctext8.txt\u201d\u6587\u4ef6\u5185\u3002<\/p>\n\n\n\n 3\u6570\u636e\u9884\u5904\u7406<\/p>\n\n\n\n \u8bdd\u8bf4\u6211\u4e0d\u662f\u5f88\u7406\u89e3<\/p>\n\n\n\n word2id_freq[word2id_dict[word]] = freq\u8fd9\u53e5\u4e3a\u4ec0\u4e48\u4e0d\u76f4\u63a5\u5199\u6210word2id_freq[curr_id] = freq<\/p>\n\n\n\n 4\u6784\u9020\u8bad\u7ec3\u6570\u636e<\/p>\n\n\n\n 5\u7f51\u7edc\u5b9a\u4e49<\/p>\n\n\n\n 6\u7f51\u7edc\u8bad\u7ec3<\/p>\n\n\n\n \u6682\u65f6\u4e0d\u4f1a\u505aNLP\u65b9\u5411\uff0ctorch\u7248\u672c\u4e4b\u540e\u66f4<\/p>\n","protected":false},"excerpt":{"rendered":" \u8df3\u8fc7\u76ee\u6807\u8bc6\u522b\uff0cyolo\u7ed3\u6784\u592a\u590d\u6742\u4e86\uff0c\u73b0\u5728\u7814\u7a76\u65b9\u5411\u4e5f\u7528\u4e0d\u5230 \u81ea\u7136\u8bed\u8a00\u5904\u7406\uff0cNLP\uff0c\u8fd9\u4e2a\u9886\u57df\u8ddf\u6211\u7684\u7814\u7a76\u65b9\u5411\u4e5f\u6ca1\u5565\u5173\u7cfb\uff0c\u4f46\u662f\u4e3a\u4e86\u5b66\u5faa\u73af\u795e …<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[21],"tags":[],"class_list":["post-474","post","type-post","status-publish","format-standard","hentry","category-21"],"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/posts\/474","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/comments?post=474"}],"version-history":[{"count":3,"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/posts\/474\/revisions"}],"predecessor-version":[{"id":483,"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/posts\/474\/revisions\/483"}],"wp:attachment":[{"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/media?parent=474"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/categories?post=474"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/tags?post=474"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}<\/figure>\n\n\n\n
<\/figure>\n\n\n\n
# encoding=utf8\nimport io\nimport os\nimport sys\nimport requests\nfrom collections import OrderedDict \nimport math\nimport random\nimport numpy as np\nimport paddle\nfrom paddle.nn import Embedding\nimport paddle.nn.functional as F<\/code><\/pre>\n\n\n\n# \u4e0b\u8f7d\u8bed\u6599\u7528\u6765\u8bad\u7ec3word2vec\ndef download():\n # \u53ef\u4ee5\u4ece\u767e\u5ea6\u4e91\u670d\u52a1\u5668\u4e0b\u8f7d\u4e00\u4e9b\u5f00\u6e90\u6570\u636e\u96c6\uff08dataset.bj.bcebos.com\uff09\n corpus_url = \"https:\/\/dataset.bj.bcebos.com\/word2vec\/text8.txt\"\n # \u4f7f\u7528python\u7684requests\u5305\u4e0b\u8f7d\u6570\u636e\u96c6\u5230\u672c\u5730\n web_request = requests.get(corpus_url)\n corpus = web_request.content\n # \u628a\u4e0b\u8f7d\u540e\u7684\u6587\u4ef6\u5b58\u50a8\u5728\u5f53\u524d\u76ee\u5f55\u7684text8.txt\u6587\u4ef6\u5185\n with open(\".\/text8.txt\", \"wb\") as f:\n f.write(corpus)\n f.close()\ndownload()\n# \u8bfb\u53d6text8\u6570\u636e\ndef load_text8():\n with open(\".\/text8.txt\", \"r\") as f:\n corpus = f.read().strip(\"\\n\")\n f.close()\n\n return corpus\n\ncorpus = load_text8()\n\n# \u6253\u5370\u524d500\u4e2a\u5b57\u7b26\uff0c\u7b80\u8981\u770b\u4e00\u4e0b\u8fd9\u4e2a\u8bed\u6599\u7684\u6837\u5b50\nprint(corpus[:500])<\/code><\/pre>\n\n\n\n# \u5bf9\u8bed\u6599\u8fdb\u884c\u9884\u5904\u7406\uff08\u5206\u8bcd\uff09\ndef data_preprocess(corpus):\n # \u7531\u4e8e\u82f1\u6587\u5355\u8bcd\u51fa\u73b0\u5728\u53e5\u9996\u7684\u65f6\u5019\u7ecf\u5e38\u8981\u5927\u5199\uff0c\u6240\u4ee5\u6211\u4eec\u628a\u6240\u6709\u82f1\u6587\u5b57\u7b26\u90fd\u8f6c\u6362\u4e3a\u5c0f\u5199\uff0c\n # \u4ee5\u4fbf\u5bf9\u8bed\u6599\u8fdb\u884c\u5f52\u4e00\u5316\u5904\u7406\uff08Apple vs apple\u7b49\uff09\n corpus = corpus.strip().lower()\n corpus = corpus.split(\" \")\n return corpus\n\ncorpus = data_preprocess(corpus)\nprint(corpus[:50])\n\n# \u6784\u9020\u8bcd\u5178\uff0c\u7edf\u8ba1\u6bcf\u4e2a\u8bcd\u7684\u9891\u7387\uff0c\u5e76\u6839\u636e\u9891\u7387\u5c06\u6bcf\u4e2a\u8bcd\u8f6c\u6362\u4e3a\u4e00\u4e2a\u6574\u6570id\ndef build_dict(corpus):\n # \u9996\u5148\u7edf\u8ba1\u6bcf\u4e2a\u4e0d\u540c\u8bcd\u7684\u9891\u7387\uff08\u51fa\u73b0\u7684\u6b21\u6570\uff09\uff0c\u4f7f\u7528\u4e00\u4e2a\u8bcd\u5178\u8bb0\u5f55\n word_freq_dict = dict()\n for word in corpus:\n if word not in word_freq_dict:\n word_freq_dict[word] = 0\n word_freq_dict[word] += 1\n\n # \u5c06\u8fd9\u4e2a\u8bcd\u5178\u4e2d\u7684\u8bcd\uff0c\u6309\u7167\u51fa\u73b0\u6b21\u6570\u6392\u5e8f\uff0c\u51fa\u73b0\u6b21\u6570\u8d8a\u9ad8\uff0c\u6392\u5e8f\u8d8a\u9760\u524d\n # \u4e00\u822c\u6765\u8bf4\uff0c\u51fa\u73b0\u9891\u7387\u9ad8\u7684\u9ad8\u9891\u8bcd\u5f80\u5f80\u662f\uff1aI\uff0cthe\uff0cyou\u8fd9\u79cd\u4ee3\u8bcd\uff0c\u800c\u51fa\u73b0\u9891\u7387\u4f4e\u7684\u8bcd\uff0c\u5f80\u5f80\u662f\u4e00\u4e9b\u540d\u8bcd\uff0c\u5982\uff1anlp\n word_freq_dict = sorted(word_freq_dict.items(), key = lambda x:x[1], reverse = True)\n \n # \u6784\u90203\u4e2a\u4e0d\u540c\u7684\u8bcd\u5178\uff0c\u5206\u522b\u5b58\u50a8\uff0c\n # \u6bcf\u4e2a\u8bcd\u5230id\u7684\u6620\u5c04\u5173\u7cfb\uff1aword2id_dict\n # \u6bcf\u4e2aid\u51fa\u73b0\u7684\u9891\u7387\uff1aword2id_freq\n # \u6bcf\u4e2aid\u5230\u8bcd\u7684\u6620\u5c04\u5173\u7cfb\uff1aid2word_dict\n word2id_dict = dict()\n word2id_freq = dict()\n id2word_dict = dict()\n\n # \u6309\u7167\u9891\u7387\uff0c\u4ece\u9ad8\u5230\u4f4e\uff0c\u5f00\u59cb\u904d\u5386\u6bcf\u4e2a\u5355\u8bcd\uff0c\u5e76\u4e3a\u8fd9\u4e2a\u5355\u8bcd\u6784\u9020\u4e00\u4e2a\u72ec\u4e00\u65e0\u4e8c\u7684id\n for word, freq in word_freq_dict:\n curr_id = len(word2id_dict)\n word2id_dict[word] = curr_id\n word2id_freq[word2id_dict[word]] = freq\n id2word_dict[curr_id] = word\n\n return word2id_freq, word2id_dict, id2word_dict\n\nword2id_freq, word2id_dict, id2word_dict = build_dict(corpus)\nvocab_size = len(word2id_freq)\nprint(\"there are totoally %d different words in the corpus\" % vocab_size)\nfor _, (word, word_id) in zip(range(50), word2id_dict.items()):\n print(\"word %s, its id %d, its word freq %d\" % (word, word_id, word2id_freq[word_id]))\n\n# \u628a\u8bed\u6599\u8f6c\u6362\u4e3aid\u5e8f\u5217\ndef convert_corpus_to_id(corpus, word2id_dict):\n # \u4f7f\u7528\u4e00\u4e2a\u5faa\u73af\uff0c\u5c06\u8bed\u6599\u4e2d\u7684\u6bcf\u4e2a\u8bcd\u66ff\u6362\u6210\u5bf9\u5e94\u7684id\uff0c\u4ee5\u4fbf\u4e8e\u795e\u7ecf\u7f51\u7edc\u8fdb\u884c\u5904\u7406\n corpus = [word2id_dict[word] for word in corpus]\n return corpus\n\n\n'''\u5f97\u5230word2id\u8bcd\u5178\u540e\uff0c\u8fd8\u9700\u8981\u8fdb\u4e00\u6b65\u5904\u7406\u539f\u59cb\u8bed\u6599\uff0c\u628a\u6bcf\u4e2a\u8bcd\u66ff\u6362\u6210\u5bf9\u5e94\u7684ID\uff0c\u4fbf\u4e8e\u795e\u7ecf\u7f51\u7edc\u8fdb\u884c\u5904\u7406\uff0c\u4ee3\u7801\u5982\u4e0b\uff1a'''\ncorpus = convert_corpus_to_id(corpus, word2id_dict)\nprint(\"%d tokens in the corpus\" % len(corpus))\nprint(corpus[:50])\n\n'''\u63a5\u4e0b\u6765\uff0c\u9700\u8981\u4f7f\u7528\u4e8c\u6b21\u91c7\u6837\u6cd5\u5904\u7406\u539f\u59cb\u6587\u672c\u3002\u4e8c\u6b21\u91c7\u6837\u6cd5\u7684\u4e3b\u8981\u601d\u60f3\u662f\u964d\u4f4e\u9ad8\u9891\u8bcd\u5728\u8bed\u6599\u4e2d\u51fa\u73b0\u7684\u9891\u6b21\u3002\u65b9\u6cd5\u662f\u968f\u673a\u5c06\u9ad8\u9891\u7684\u8bcd\u629b\u5f03\uff0c\u9891\u7387\u8d8a\u9ad8\uff0c\u88ab\u629b\u5f03\u7684\u6982\u7387\u5c31\u8d8a\u5927\uff1b\u9891\u7387\u8d8a\u4f4e\uff0c\u88ab\u629b\u5f03\u7684\u6982\u7387\u5c31\u8d8a\u5c0f\u3002\u6807\u70b9\u7b26\u53f7\u6216\u51a0\u8bcd\u8fd9\u6837\u7684\u9ad8\u9891\u8bcd\u5c31\u4f1a\u88ab\u629b\u5f03\uff0c\u4ece\u800c\u4f18\u5316\u6574\u4e2a\u8bcd\u8868\u7684\u8bcd\u5411\u91cf\u8bad\u7ec3\u6548\u679c\uff0c\u4ee3\u7801\u5982\u4e0b\uff1a'''\n# \u4f7f\u7528\u4e8c\u6b21\u91c7\u6837\u7b97\u6cd5\uff08subsampling\uff09\u5904\u7406\u8bed\u6599\uff0c\u5f3a\u5316\u8bad\u7ec3\u6548\u679c\ndef subsampling(corpus, word2id_freq):\n \n # \u8fd9\u4e2adiscard\u51fd\u6570\u51b3\u5b9a\u4e86\u4e00\u4e2a\u8bcd\u4f1a\u4e0d\u4f1a\u88ab\u66ff\u6362\uff0c\u8fd9\u4e2a\u51fd\u6570\u662f\u5177\u6709\u968f\u673a\u6027\u7684\uff0c\u6bcf\u6b21\u8c03\u7528\u7ed3\u679c\u4e0d\u540c\n # \u5982\u679c\u4e00\u4e2a\u8bcd\u7684\u9891\u7387\u5f88\u5927\uff0c\u90a3\u4e48\u5b83\u88ab\u9057\u5f03\u7684\u6982\u7387\u5c31\u5f88\u5927\n def discard(word_id):\n return random.uniform(0, 1) < 1 - math.sqrt(\n 1e-4 \/ word2id_freq[word_id] * len(corpus))\n\n corpus = [word for word in corpus if not discard(word)]\n return corpus\n\ncorpus = subsampling(corpus, word2id_freq)\nprint(\"%d tokens in the corpus\" % len(corpus))\nprint(corpus[:50])<\/code><\/pre>\n\n\n\n# \u6784\u9020\u6570\u636e\uff0c\u51c6\u5907\u6a21\u578b\u8bad\u7ec3\n# max_window_size\u4ee3\u8868\u4e86\u6700\u5927\u7684window_size\u7684\u5927\u5c0f\uff0c\u7a0b\u5e8f\u4f1a\u6839\u636emax_window_size\u4ece\u5de6\u5230\u53f3\u626b\u63cf\u6574\u4e2a\u8bed\u6599\n# negative_sample_num\u4ee3\u8868\u4e86\u5bf9\u4e8e\u6bcf\u4e2a\u6b63\u6837\u672c\uff0c\u6211\u4eec\u9700\u8981\u968f\u673a\u91c7\u6837\u591a\u5c11\u8d1f\u6837\u672c\u7528\u4e8e\u8bad\u7ec3\uff0c\n# \u4e00\u822c\u6765\u8bf4\uff0cnegative_sample_num\u7684\u503c\u8d8a\u5927\uff0c\u8bad\u7ec3\u6548\u679c\u8d8a\u7a33\u5b9a\uff0c\u4f46\u662f\u8bad\u7ec3\u901f\u5ea6\u8d8a\u6162\u3002 \ndef build_data(corpus, word2id_dict, word2id_freq, max_window_size = 3, negative_sample_num = 4):\n \n # \u4f7f\u7528\u4e00\u4e2alist\u5b58\u50a8\u5904\u7406\u597d\u7684\u6570\u636e\n dataset = []\n\n # \u4ece\u5de6\u5230\u53f3\uff0c\u5f00\u59cb\u679a\u4e3e\u6bcf\u4e2a\u4e2d\u5fc3\u70b9\u7684\u4f4d\u7f6e\n for center_word_idx in range(len(corpus)):\n # \u4ee5max_window_size\u4e3a\u4e0a\u9650\uff0c\u968f\u673a\u91c7\u6837\u4e00\u4e2awindow_size\uff0c\u8fd9\u6837\u4f1a\u4f7f\u5f97\u8bad\u7ec3\u66f4\u52a0\u7a33\u5b9a\n window_size = random.randint(1, max_window_size)\n # \u5f53\u524d\u7684\u4e2d\u5fc3\u8bcd\u5c31\u662fcenter_word_idx\u6240\u6307\u5411\u7684\u8bcd\n center_word = corpus[center_word_idx]\n\n # \u4ee5\u5f53\u524d\u4e2d\u5fc3\u8bcd\u4e3a\u4e2d\u5fc3\uff0c\u5de6\u53f3\u4e24\u4fa7\u5728window_size\u5185\u7684\u8bcd\u90fd\u53ef\u4ee5\u770b\u6210\u662f\u6b63\u6837\u672c\n positive_word_range = (max(0, center_word_idx - window_size), min(len(corpus) - 1, center_word_idx + window_size))\n positive_word_candidates = [corpus[idx] for idx in range(positive_word_range[0], positive_word_range[1]+1) if idx != center_word_idx]\n\n # \u5bf9\u4e8e\u6bcf\u4e2a\u6b63\u6837\u672c\u6765\u8bf4\uff0c\u968f\u673a\u91c7\u6837negative_sample_num\u4e2a\u8d1f\u6837\u672c\uff0c\u7528\u4e8e\u8bad\u7ec3\n for positive_word in positive_word_candidates:\n # \u9996\u5148\u628a\uff08\u4e2d\u5fc3\u8bcd\uff0c\u6b63\u6837\u672c\uff0clabel=1\uff09\u7684\u4e09\u5143\u7ec4\u6570\u636e\u653e\u5165dataset\u4e2d\uff0c\n # \u8fd9\u91cclabel=1\u8868\u793a\u8fd9\u4e2a\u6837\u672c\u662f\u4e2a\u6b63\u6837\u672c\n dataset.append((center_word, positive_word, 1))\n\n # \u5f00\u59cb\u8d1f\u91c7\u6837\n i = 0\n while i < negative_sample_num:\n negative_word_candidate = random.randint(0, vocab_size-1)\n\n if negative_word_candidate not in positive_word_candidates:\n # \u628a\uff08\u4e2d\u5fc3\u8bcd\uff0c\u6b63\u6837\u672c\uff0clabel=0\uff09\u7684\u4e09\u5143\u7ec4\u6570\u636e\u653e\u5165dataset\u4e2d\uff0c\n # \u8fd9\u91cclabel=0\u8868\u793a\u8fd9\u4e2a\u6837\u672c\u662f\u4e2a\u8d1f\u6837\u672c\n dataset.append((center_word, negative_word_candidate, 0))\n i += 1\n return dataset\ncorpus_light = corpus[:int(len(corpus)*0.2)]\ndataset = build_data(corpus_light, word2id_dict, word2id_freq)\nfor _, (center_word, target_word, label) in zip(range(50), dataset):\n print(\"center_word %s, target %s, label %d\" % (id2word_dict[center_word],\n id2word_dict[target_word], label))\n\n#\u8bad\u7ec3\u6570\u636e\u51c6\u5907\u597d\u540e\uff0c\u628a\u8bad\u7ec3\u6570\u636e\u90fd\u7ec4\u88c5\u6210mini-batch\uff0c\u5e76\u51c6\u5907\u8f93\u5165\u5230\u7f51\u7edc\u4e2d\u8fdb\u884c\u8bad\u7ec3\uff0c\u4ee3\u7801\u5982\u4e0b\uff1a\n# \u6784\u9020mini-batch\uff0c\u51c6\u5907\u5bf9\u6a21\u578b\u8fdb\u884c\u8bad\u7ec3\n# \u6211\u4eec\u5c06\u4e0d\u540c\u7c7b\u578b\u7684\u6570\u636e\u653e\u5230\u4e0d\u540c\u7684tensor\u91cc\uff0c\u4fbf\u4e8e\u795e\u7ecf\u7f51\u7edc\u8fdb\u884c\u5904\u7406\n# \u5e76\u901a\u8fc7numpy\u7684array\u51fd\u6570\uff0c\u6784\u9020\u51fa\u4e0d\u540c\u7684tensor\u6765\uff0c\u5e76\u628a\u8fd9\u4e9btensor\u9001\u5165\u795e\u7ecf\u7f51\u7edc\u4e2d\u8fdb\u884c\u8bad\u7ec3\ndef build_batch(dataset, batch_size, epoch_num):\n \n # center_word_batch\u7f13\u5b58batch_size\u4e2a\u4e2d\u5fc3\u8bcd\n center_word_batch = []\n # target_word_batch\u7f13\u5b58batch_size\u4e2a\u76ee\u6807\u8bcd\uff08\u53ef\u4ee5\u662f\u6b63\u6837\u672c\u6216\u8005\u8d1f\u6837\u672c\uff09\n target_word_batch = []\n # label_batch\u7f13\u5b58\u4e86batch_size\u4e2a0\u62161\u7684\u6807\u7b7e\uff0c\u7528\u4e8e\u6a21\u578b\u8bad\u7ec3\n label_batch = []\n\n for epoch in range(epoch_num):\n # \u6bcf\u6b21\u5f00\u542f\u4e00\u4e2a\u65b0epoch\u4e4b\u524d\uff0c\u90fd\u5bf9\u6570\u636e\u8fdb\u884c\u4e00\u6b21\u968f\u673a\u6253\u4e71\uff0c\u63d0\u9ad8\u8bad\u7ec3\u6548\u679c\n random.shuffle(dataset)\n \n for center_word, target_word, label in dataset:\n # \u904d\u5386dataset\u4e2d\u7684\u6bcf\u4e2a\u6837\u672c\uff0c\u5e76\u5c06\u8fd9\u4e9b\u6570\u636e\u9001\u5230\u4e0d\u540c\u7684tensor\u91cc\n center_word_batch.append([center_word])\n target_word_batch.append([target_word])\n label_batch.append(label)\n\n # \u5f53\u6837\u672c\u79ef\u6512\u5230\u4e00\u4e2abatch_size\u540e\uff0c\u6211\u4eec\u628a\u6570\u636e\u90fd\u8fd4\u56de\u56de\u6765\n # \u5728\u8fd9\u91cc\u6211\u4eec\u4f7f\u7528numpy\u7684array\u51fd\u6570\u628alist\u5c01\u88c5\u6210tensor\n # \u5e76\u4f7f\u7528python\u7684\u8fed\u4ee3\u5668\u673a\u5236\uff0c\u5c06\u6570\u636eyield\u51fa\u6765\n # \u4f7f\u7528\u8fed\u4ee3\u5668\u7684\u597d\u5904\u662f\u53ef\u4ee5\u8282\u7701\u5185\u5b58\n if len(center_word_batch) == batch_size:\n yield np.array(center_word_batch).astype(\"int64\"), \\\n np.array(target_word_batch).astype(\"int64\"), \\\n np.array(label_batch).astype(\"float32\")\n center_word_batch = []\n target_word_batch = []\n label_batch = []\n\n if len(center_word_batch) > 0:\n yield np.array(center_word_batch).astype(\"int64\"), \\\n np.array(target_word_batch).astype(\"int64\"), \\\n np.array(label_batch).astype(\"float32\")\n\nfor _, batch in zip(range(10), build_batch(dataset, 128, 3)):\n print(batch)<\/code><\/pre>\n\n\n\n#\u5b9a\u4e49skip-gram\u8bad\u7ec3\u7f51\u7edc\u7ed3\u6784\r\n#\u4f7f\u7528paddlepaddle\u76842.0.0\u7248\u672c\r\n#\u4e00\u822c\u6765\u8bf4\uff0c\u5728\u4f7f\u7528paddle\u8bad\u7ec3\u7684\u65f6\u5019\uff0c\u6211\u4eec\u9700\u8981\u901a\u8fc7\u4e00\u4e2a\u7c7b\u6765\u5b9a\u4e49\u7f51\u7edc\u7ed3\u6784\uff0c\u8fd9\u4e2a\u7c7b\u7ee7\u627f\u4e86paddle.nn.layer\r\nclass SkipGram(paddle.nn.Layer):\r\n def __init__(self, vocab_size, embedding_size, init_scale=0.1):\r\n # vocab_size\u5b9a\u4e49\u4e86\u8fd9\u4e2askipgram\u8fd9\u4e2a\u6a21\u578b\u7684\u8bcd\u8868\u5927\u5c0f\r\n # embedding_size\u5b9a\u4e49\u4e86\u8bcd\u5411\u91cf\u7684\u7ef4\u5ea6\u662f\u591a\u5c11\r\n # init_scale\u5b9a\u4e49\u4e86\u8bcd\u5411\u91cf\u521d\u59cb\u5316\u7684\u8303\u56f4\uff0c\u4e00\u822c\u6765\u8bf4\uff0c\u6bd4\u8f83\u5c0f\u7684\u521d\u59cb\u5316\u8303\u56f4\u6709\u52a9\u4e8e\u6a21\u578b\u8bad\u7ec3\r\n super(SkipGram, self).__init__()\r\n self.vocab_size = vocab_size\r\n self.embedding_size = embedding_size\r\n\r\n # \u4f7f\u7528Embedding\u51fd\u6570\u6784\u9020\u4e00\u4e2a\u8bcd\u5411\u91cf\u53c2\u6570\r\n # \u8fd9\u4e2a\u53c2\u6570\u7684\u5927\u5c0f\u4e3a\uff1a[self.vocab_size, self.embedding_size]\r\n # \u6570\u636e\u7c7b\u578b\u4e3a\uff1afloat32\r\n # \u8fd9\u4e2a\u53c2\u6570\u7684\u540d\u79f0\u4e3a\uff1aembedding_para\r\n # \u8fd9\u4e2a\u53c2\u6570\u7684\u521d\u59cb\u5316\u65b9\u5f0f\u4e3a\u5728[-init_scale, init_scale]\u533a\u95f4\u8fdb\u884c\u5747\u5300\u91c7\u6837\r\n self.embedding = Embedding( \r\n num_embeddings = self.vocab_size,\r\n embedding_dim = self.embedding_size,\r\n weight_attr=paddle.ParamAttr(\r\n initializer=paddle.nn.initializer.Uniform( \r\n low=-0.5\/embedding_size, high=0.5\/embedding_size)))\r\n\r\n # \u4f7f\u7528Embedding\u51fd\u6570\u6784\u9020\u53e6\u5916\u4e00\u4e2a\u8bcd\u5411\u91cf\u53c2\u6570\r\n # \u8fd9\u4e2a\u53c2\u6570\u7684\u5927\u5c0f\u4e3a\uff1a[self.vocab_size, self.embedding_size]\r\n # \u8fd9\u4e2a\u53c2\u6570\u7684\u521d\u59cb\u5316\u65b9\u5f0f\u4e3a\u5728[-init_scale, init_scale]\u533a\u95f4\u8fdb\u884c\u5747\u5300\u91c7\u6837\r\n self.embedding_out = Embedding(\r\n num_embeddings = self.vocab_size,\r\n embedding_dim = self.embedding_size,\r\n weight_attr=paddle.ParamAttr(\r\n initializer=paddle.nn.initializer.Uniform(\r\n low=-0.5\/embedding_size, high=0.5\/embedding_size)))\r\n\r\n # \u5b9a\u4e49\u7f51\u7edc\u7684\u524d\u5411\u8ba1\u7b97\u903b\u8f91\r\n # center_words\u662f\u4e00\u4e2atensor\uff08mini-batch\uff09\uff0c\u8868\u793a\u4e2d\u5fc3\u8bcd\r\n # target_words\u662f\u4e00\u4e2atensor\uff08mini-batch\uff09\uff0c\u8868\u793a\u76ee\u6807\u8bcd\r\n # label\u662f\u4e00\u4e2atensor\uff08mini-batch\uff09\uff0c\u8868\u793a\u8fd9\u4e2a\u8bcd\u662f\u6b63\u6837\u672c\u8fd8\u662f\u8d1f\u6837\u672c\uff08\u75280\u62161\u8868\u793a\uff09\r\n # \u7528\u4e8e\u5728\u8bad\u7ec3\u4e2d\u8ba1\u7b97\u8fd9\u4e2atensor\u4e2d\u5bf9\u5e94\u8bcd\u7684\u540c\u4e49\u8bcd\uff0c\u7528\u4e8e\u89c2\u5bdf\u6a21\u578b\u7684\u8bad\u7ec3\u6548\u679c\r\n def forward(self, center_words, target_words, label):\r\n # \u9996\u5148\uff0c\u901a\u8fc7embedding_para\uff08self.embedding\uff09\u53c2\u6570\uff0c\u5c06mini-batch\u4e2d\u7684\u8bcd\u8f6c\u6362\u4e3a\u8bcd\u5411\u91cf\r\n # \u8fd9\u91cccenter_words\u548ceval_words_emb\u67e5\u8be2\u7684\u662f\u4e00\u4e2a\u76f8\u540c\u7684\u53c2\u6570\r\n # \u800ctarget_words_emb\u67e5\u8be2\u7684\u662f\u53e6\u4e00\u4e2a\u53c2\u6570\r\n center_words_emb = self.embedding(center_words)\r\n target_words_emb = self.embedding_out(target_words)\r\n\r\n # \u6211\u4eec\u901a\u8fc7\u70b9\u4e58\u7684\u65b9\u5f0f\u8ba1\u7b97\u4e2d\u5fc3\u8bcd\u5230\u76ee\u6807\u8bcd\u7684\u8f93\u51fa\u6982\u7387\uff0c\u5e76\u901a\u8fc7sigmoid\u51fd\u6570\u4f30\u8ba1\u8fd9\u4e2a\u8bcd\u662f\u6b63\u6837\u672c\u8fd8\u662f\u8d1f\u6837\u672c\u7684\u6982\u7387\u3002\r\n word_sim = paddle.multiply(center_words_emb, target_words_emb)\r\n word_sim = paddle.sum(word_sim, axis=-1)\r\n word_sim = paddle.reshape(word_sim, shape=[-1])\r\n pred = F.sigmoid(word_sim)\r\n\r\n # \u901a\u8fc7\u4f30\u8ba1\u7684\u8f93\u51fa\u6982\u7387\u5b9a\u4e49\u635f\u5931\u51fd\u6570\uff0c\u6ce8\u610f\u6211\u4eec\u4f7f\u7528\u7684\u662fbinary_cross_entropy_with_logits\u51fd\u6570\r\n # \u5c06sigmoid\u8ba1\u7b97\u548ccross entropy\u5408\u5e76\u6210\u4e00\u6b65\u8ba1\u7b97\u53ef\u4ee5\u66f4\u597d\u7684\u4f18\u5316\uff0c\u6240\u4ee5\u8f93\u5165\u7684\u662fword_sim\uff0c\u800c\u4e0d\u662fpred\r\n \r\n loss = F.binary_cross_entropy_with_logits(word_sim, label)\r\n loss = paddle.mean(loss)\r\n\r\n # \u8fd4\u56de\u524d\u5411\u8ba1\u7b97\u7684\u7ed3\u679c\uff0c\u98de\u6868\u4f1a\u901a\u8fc7backward\u51fd\u6570\u81ea\u52a8\u8ba1\u7b97\u51fa\u53cd\u5411\u7ed3\u679c\u3002\r\n return pred, loss<\/code><\/pre>\n\n\n\n# \u5f00\u59cb\u8bad\u7ec3\uff0c\u5b9a\u4e49\u4e00\u4e9b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u9700\u8981\u4f7f\u7528\u7684\u8d85\u53c2\u6570\r\nbatch_size = 512\r\nepoch_num = 3\r\nembedding_size = 200\r\nstep = 0\r\nlearning_rate = 0.001\r\n\r\n#\u5b9a\u4e49\u4e00\u4e2a\u4f7f\u7528word-embedding\u67e5\u8be2\u540c\u4e49\u8bcd\u7684\u51fd\u6570\r\n#\u8fd9\u4e2a\u51fd\u6570query_token\u662f\u8981\u67e5\u8be2\u7684\u8bcd\uff0ck\u8868\u793a\u8981\u8fd4\u56de\u591a\u5c11\u4e2a\u6700\u76f8\u4f3c\u7684\u8bcd\uff0cembed\u662f\u6211\u4eec\u5b66\u4e60\u5230\u7684word-embedding\u53c2\u6570\r\n#\u6211\u4eec\u901a\u8fc7\u8ba1\u7b97\u4e0d\u540c\u8bcd\u4e4b\u95f4\u7684cosine\u8ddd\u79bb\uff0c\u6765\u8861\u91cf\u8bcd\u548c\u8bcd\u7684\u76f8\u4f3c\u5ea6\r\n#\u5177\u4f53\u5b9e\u73b0\u5982\u4e0b\uff0cx\u4ee3\u8868\u8981\u67e5\u8be2\u8bcd\u7684Embedding\uff0cEmbedding\u53c2\u6570\u77e9\u9635W\u4ee3\u8868\u6240\u6709\u8bcd\u7684Embedding\r\n#\u4e24\u8005\u8ba1\u7b97Cos\u5f97\u51fa\u6240\u6709\u8bcd\u5bf9\u67e5\u8be2\u8bcd\u7684\u76f8\u4f3c\u5ea6\u5f97\u5206\u5411\u91cf\uff0c\u6392\u5e8f\u53d6top_k\u653e\u5165indices\u5217\u8868\r\ndef get_similar_tokens(query_token, k, embed):\r\n W = embed.numpy()\r\n x = W[word2id_dict[query_token]]\r\n cos = np.dot(W, x) \/ np.sqrt(np.sum(W * W, axis=1) * np.sum(x * x) + 1e-9)\r\n flat = cos.flatten()\r\n indices = np.argpartition(flat, -k)[-k:]\r\n indices = indices[np.argsort(-flat[indices])]\r\n for i in indices:\r\n print('for word %s, the similar word is %s' % (query_token, str(id2word_dict[i])))\r\n\r\n# \u5c06\u6a21\u578b\u653e\u5230GPU\u4e0a\u8bad\u7ec3\r\npaddle.set_device('gpu:0')\r\n\r\n# \u901a\u8fc7\u6211\u4eec\u5b9a\u4e49\u7684SkipGram\u7c7b\uff0c\u6765\u6784\u9020\u4e00\u4e2aSkip-gram\u6a21\u578b\u7f51\u7edc\r\nskip_gram_model = SkipGram(vocab_size, embedding_size)\r\n\r\n# \u6784\u9020\u8bad\u7ec3\u8fd9\u4e2a\u7f51\u7edc\u7684\u4f18\u5316\u5668\r\nadam = paddle.optimizer.Adam(learning_rate=learning_rate, parameters = skip_gram_model.parameters())\r\n\r\n# \u4f7f\u7528build_batch\u51fd\u6570\uff0c\u4ee5mini-batch\u4e3a\u5355\u4f4d\uff0c\u904d\u5386\u8bad\u7ec3\u6570\u636e\uff0c\u5e76\u8bad\u7ec3\u7f51\u7edc\r\nfor center_words, target_words, label in build_batch(\r\n dataset, batch_size, epoch_num):\r\n # \u4f7f\u7528paddle.to_tensor\uff0c\u5c06\u4e00\u4e2anumpy\u7684tensor\uff0c\u8f6c\u6362\u4e3a\u98de\u6868\u53ef\u8ba1\u7b97\u7684tensor\r\n center_words_var = paddle.to_tensor(center_words)\r\n target_words_var = paddle.to_tensor(target_words)\r\n label_var = paddle.to_tensor(label)\r\n \r\n # \u5c06\u8f6c\u6362\u540e\u7684tensor\u9001\u5165\u98de\u6868\u4e2d\uff0c\u8fdb\u884c\u4e00\u6b21\u524d\u5411\u8ba1\u7b97\uff0c\u5e76\u5f97\u5230\u8ba1\u7b97\u7ed3\u679c\r\n pred, loss = skip_gram_model(\r\n center_words_var, target_words_var, label_var)\r\n\r\n # \u7a0b\u5e8f\u81ea\u52a8\u5b8c\u6210\u53cd\u5411\u8ba1\u7b97\r\n loss.backward()\r\n # \u7a0b\u5e8f\u6839\u636eloss\uff0c\u5b8c\u6210\u4e00\u6b65\u5bf9\u53c2\u6570\u7684\u4f18\u5316\u66f4\u65b0\r\n adam.step()\r\n # \u6e05\u7a7a\u6a21\u578b\u4e2d\u7684\u68af\u5ea6\uff0c\u4ee5\u4fbf\u4e8e\u4e0b\u4e00\u4e2amini-batch\u8fdb\u884c\u66f4\u65b0\r\n adam.clear_grad()\r\n\r\n # \u6bcf\u7ecf\u8fc7100\u4e2amini-batch\uff0c\u6253\u5370\u4e00\u6b21\u5f53\u524d\u7684loss\uff0c\u770b\u770bloss\u662f\u5426\u5728\u7a33\u5b9a\u4e0b\u964d\r\n step += 1\r\n if step % 1000 == 0:\r\n print(\"step %d, loss %.3f\" % (step, loss.numpy()[0]))\r\n\r\n # \u6bcf\u969410000\u6b65\uff0c\u6253\u5370\u4e00\u6b21\u6a21\u578b\u5bf9\u4ee5\u4e0b\u67e5\u8be2\u8bcd\u7684\u76f8\u4f3c\u8bcd\uff0c\u8fd9\u91cc\u6211\u4eec\u4f7f\u7528\u8bcd\u548c\u8bcd\u4e4b\u95f4\u7684\u5411\u91cf\u70b9\u79ef\u4f5c\u4e3a\u8861\u91cf\u76f8\u4f3c\u5ea6\u7684\u65b9\u6cd5\uff0c\u53ea\u6253\u5370\u4e865\u4e2a\u6700\u76f8\u4f3c\u7684\u8bcd\r\n if step % 10000 ==0:\r\n get_similar_tokens('movie', 5, skip_gram_model.embedding.weight)\r\n get_similar_tokens('one', 5, skip_gram_model.embedding.weight)\r\n get_similar_tokens('chip', 5, skip_gram_model.embedding.weight)<\/code><\/pre>\n\n\n\n