{"id":3810,"date":"2021-04-28T16:18:37","date_gmt":"2021-04-28T08:18:37","guid":{"rendered":"http:\/\/www.sniper97.cn\/?p=3810"},"modified":"2021-04-28T16:18:37","modified_gmt":"2021-04-28T08:18:37","slug":"%e3%80%90%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e7%ac%94%e8%ae%b0%e3%80%91bert%e7%9a%84mlm%e4%bb%bb%e5%8a%a1%e5%ae%9e%e7%8e%b0%e6%96%b9%e5%bc%8f","status":"publish","type":"post","link":"http:\/\/www.sniper97.cn\/index.php\/note\/deep-learning\/note-deep-learning\/3810\/","title":{"rendered":"\u3010\u6df1\u5ea6\u5b66\u4e60\u7b14\u8bb0\u3011bert\u7684mlm\u4efb\u52a1\u5b9e\u73b0\u65b9\u5f0f"},"content":{"rendered":"\n

\u4e4b\u524d\u5bf9\u4e8ebert\u7684MLM\u4efb\u52a1\u4e00\u76f4\u90fd\u662f\u6a21\u68f1\u4e24\u53ef\uff0c\u4eca\u5929\u5bf9\u4e8e\u5b9e\u73b0\u7ec6\u8282\u8fdb\u884c\u4e86\u8865\u5168\u3002<\/p>\n\n\n

\u60f3\u770b\u7ed3\u8bba\u7684\u76f4\u63a5\u62c9\u5230\u6700\u540e\u3002<\/strong><\/p>\n\n\n

mlm\u7684\u6d41\u7a0b\u5b9e\u9645\u4e0a\u662f\u4f7f\u7528token\uff0c\u901a\u8fc7\u53d8\u6362\u83b7\u5f97\u4e00\u4e2avocab size\u5927\u5c0f\u7684\u8f93\u51fa\uff0c\u4ece\u800c\u5224\u65ad\u9884\u6d4b\u4f4d\u7f6e\u7684\u5355\u8bcd\u3002<\/p>\n\n\n

\u5176\u5b9e\u4e00\u5f00\u59cb\u662f\u5728\u7814\u7a76bert4keras\u65f6\uff0c\u53d1\u73b0\u8f93\u51fa\u7684\u53c2\u6570\u548c\u6211\u60f3\u8c61\u4e2d\u7684\u5bf9\u4e0d\u4e0a\u554a\uff0c\u7136\u540e\u5411\u4e0b\u6df1\u7a76\u4ee3\u7801\u3002\u6700\u540e\u5728transorfmers\u4e2d\u627e\u5230\u4e86\u7ed3\u679c\u3002<\/p>\n\n\n

\u672c\u6587\u5c31\u6253\u7b97\u6839\u636e\u6211\u7684\u601d\u8def\u5386\u7a0b\u6765\u7f16\u5199\u3002<\/p>\n\n\n

\u6700\u8fd1\u5728\u505abert4keras\u7684\u6587\u6863<\/a>\uff0c\u505a\u5230AutoRegressiveDecoder\u7684\u65f6\u5019\u9047\u5230\u4e86\u4e4b\u524d\u7684\u4e00\u4e2a\u95ee\u9898\uff0c\u5f53\u4efb\u52a1\u4e3alm\u6216\u8005unilm\u65f6\uff0cbert\u4f1a\u88ab\u5207\u6362\u5230mlm\u6a21\u5f0f\uff0c\u4e5f\u5c31\u9700\u8981\u5bf9token\u8fdb\u884c\u9884\u6d4b\u3002<\/p>\n\n\n

\u589e\u52a0\u7684\u7ed3\u6784\u4e3a\uff1a<\/p>\n\n\n

__________________________________________________________________________________________________\nTransformer-11-FeedForward-Norm (None, None, 768)    1536        Transformer-11-FeedForward-Add[0]\n__________________________________________________________________________________________________\nMLM-Dense (Dense)               (None, None, 768)    590592      Transformer-11-FeedForward-Norm[0\n__________________________________________________________________________________________________\nMLM-Norm (LayerNormalization)   (None, None, 768)    1536        MLM-Dense[0][0]\n__________________________________________________________________________________________________\nMLM-Bias (BiasAdd)              (None, None, 21128)  21128       Embedding-Token[1][0]\n__________________________________________________________________________________________________\nMLM-Activation (Activation)     (None, None, 21128)  0           MLM-Bias[0][0]\n__________________________________________________________________________________________________\ncross_entropy (CrossEntropy)    (None, None, 21128)  0           Input-Token[0][0]\n                                                                 MLM-Activation[0][0]\n<\/code><\/pre>\n\n\n
\u6211\u4eec\u53ef\u4ee5\u770b\u5230\uff0c\u591a\u51fa\u6765\u4e86 MLM-Dense\u3001MLM-Norm \u3001MLM-Bias\u51e0\u5904\u3002 <\/p>\n\n\n
\u90a3\u6211\u5c31\u5f88\u5947\u602a\uff0c\u8fd9\u4e2a\u7ed3\u6784\u662f\u600e\u4e48\u9884\u6d4btoken\u7684\uff1fMLM-Dense\u53ea\u662f\u4e00\u4e2a\u5f88\u57fa\u7840\u7684dense\uff0c768*768+768\uff0cbert4keras\u4e2d\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n\n\n
        if self.with_mlm:\n            # Masked Language Model\u90e8\u5206\n            x = outputs[0]\n            x = self.apply(\n                inputs=x,\n                layer=Dense,\n                units=self.embedding_size,\n                activation=self.hidden_act,\n                kernel_initializer=self.initializer,\n                name='MLM-Dense'\n            )<\/code><\/pre>\n\n\n
\u56e0\u6b64\uff0c\u8f93\u51fa\u8fd8\u662f768\u7ef4\u5ea6\u7684\uff0c\u4e3a\u4ec0\u4e48\u53ef\u4ee5\u548c21128\u7684MLM-Bias\u8fdb\u884c\u76f8\u52a0\uff1f<\/p>\n\n\n
\uff08\u50b2\uff0c\u6211\u7a81\u7136\u53d1\u73b0\u53ef\u4ee5\u4eceMLM-Bias\u540e\u9762\u7684Embedding-Token[1][0] \u53d1\u73b0\u732b\u817b\uff0c\u7b97\u4e86\uff0c\u5148\u5047\u88c5\u4e0d\u77e5\u9053\uff09<\/p>\n\n\n
\u767e\u601d\u4e0d\u5f97\u5176\u89e3\uff0c\u4e8e\u662f\u53bb\u67e5\u770btransformers\u7684\u6e90\u7801\u3002<\/p>\n\n\n
transorfmers\u7684torch\u5b9e\u73b0\uff1a<\/p>\n\n\n
class BertLMPredictionHead(nn.Module):\n    def __init__(self, config):\n        super().__init__()\n        self.transform = BertPredictionHeadTransform(config)\n        # The output weights are the same as the input embeddings, but there is\n        # an output-only bias for each token.\n        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)\n        self.bias = nn.Parameter(torch.zeros(config.vocab_size))\n        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`\n        self.decoder.bias = self.bias\n    def forward(self, hidden_states):\n        hidden_states = self.transform(hidden_states)\n        hidden_states = self.decoder(hidden_states)\n        return hidden_states<\/code><\/pre>\n\n\n
\u6211\u4eec\u53ef\u4ee5\u770b\u5230\uff0c\u521d\u59cb\u5316\u4e86\u4e00\u4e2aBertPredictionHeadTransform\u548c\u4e00\u4e2aLinear\uff0c Linear \u4e3ahidden size * vocab size\u3002<\/p>\n\n\n
\u8fd9\u91cc\u662f\u4e0d\u662f\u79d1\u5b66\u4e86\u4e00\u70b9\uff1f\u6700\u8d77\u7801\u8f93\u51fa\u662f\u4e00\u4e2avocab size\u5927\u5c0f\u7684\u53d8\u91cf\u4e86\uff0c\u7136\u540e\u518d\u52a0\u4e00\u4e2abias\u3002<\/p>\n\n\n
\u90a3\u6211\u4eec\u73b0\u5728\u770b\u770b BertPredictionHeadTransform<\/p>\n\n\n
class BertPredictionHeadTransform(nn.Module):\n    def __init__(self, config):\n        super().__init__()\n        self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n        if isinstance(config.hidden_act, str):\n            self.transform_act_fn = ACT2FN[config.hidden_act]\n        else:\n            self.transform_act_fn = config.hidden_act\n        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n    def forward(self, hidden_states):\n        hidden_states = self.dense(hidden_states)\n        hidden_states = self.transform_act_fn(hidden_states)\n        hidden_states = self.LayerNorm(hidden_states)\n        return hidden_states<\/code><\/pre>\n\n\n
\u6211\u4eec\u53d1\u73b0\u8fd9\u91cc\u521d\u59cb\u5316\u4e86\u4e00\u4e2a768*768+768\u7684Linear\uff0c\u662f\u4e0d\u662f\u773c\u524d\u4e00\u4eae\uff1f\u8fd9\u597d\u50cf\u6570\u76ee\u5bf9\u4e0a\u4e86\u3002<\/p>\n\n\n
\u7136\u540e\u5c31\u662fLN\u4e00\u4e0b\uff0c\u7136\u540e\u5c31\u662f\u4e00\u4e2a hidden size * vocab size \u7684Linear\uff08\u4e0a\u9762\u90a3\u6bb5\u4ee3\u7801\uff09\u3002<\/p>\n\n\n
\u81f3\u6b64\uff0c\u4e5f\u5c31\u8f93\u51fa\u5230\u4e86vocab size\u5927\u5c0f\u7684\u5411\u91cf\uff0c\u5c31\u53ef\u4ee5\u8f93\u51fa\u5bf9\u5e94\u7684token\u4e86\u3002<\/p>\n\n\n
\u4f46\u662f\u8fd9\u91cc\u5462\uff0c\u6709\u4e00\u4e2a\u7ec6\u8282\u4e0d\u592a\u660e\u663e\uff0c\u6211\u4eec\u7ee7\u7eed\u67e5\u770btf\u7684\u4ee3\u7801\uff1a<\/p>\n\n\n
\u4e0d\u4e0a\u4ee3\u7801\u4e86\uff0c\u592a\u957f\u4e86\uff0c\u540e\u9762\u4f1a\u6709\u622a\u56fe<\/strong>\u3002\u60f3\u770b\u8df3\u8f6cgithub<\/a><\/p>\n\n\n
\u957f\u4e86\u597d\u591a\uff0c\u6211\u4eec\u7ee7\u7eed\u628aTFBertPredictionHeadTransform\u7684\u4ee3\u7801\u4e00\u5e76\u5728\u4e0b\u9762\u5c55\u793a\u51fa\u6765\u3002<\/p>\n\n\n
\"\"<\/figure>\n\n\n
\u53ef\u4ee5\u770b\u5230\u8fd9\u91cc\u521d\u59cb\u5316\u4e86\u4e00\u4e2a768*768+768\u7684Dense\uff08\u7ea2\uff09\u3002\u9690\u85cf\u5c42\u901a\u8fc7\u8fd9\u4e2adense\uff0c\u6fc0\u6d3b\u4e4b\u540e\uff0c\u8fdb\u884cLN\u64cd\u4f5c\uff08\u9ec4\uff09\u3002<\/p>\n\n\n
\u90a3\u4e48\u6574\u4e2a\u6d41\u7a0b\u5c31\u53ef\u4ee5\u53d8\u4e3a\uff1a<\/p>\n\n\n
\"\"<\/figure><\/div>\n\n\n
\u867d\u7136\u6700\u540e\u4e5f\u662f\u8f93\u51fa\u7684vocab size\uff0c\u4f46\u662f\u4f9d\u7136\u4e0d\u80fd\u89e3\u51b3\u6211\u4eec\u3002<\/p>\n\n\n
\u4f46\u662f\u6211\u4eec\u67e5\u770b\u7ec6\u8282\uff1a<\/p>\n\n\n
\u5728call\u4e2d\u6211\u4eec\u53ef\u4ee5\u53d1\u73b0\uff1a\uff08\u91cd\u70b9\u6765\u4e86\u50b2\uff01\uff09<\/p>\n\n\n
\u9996\u5148\uff0c\u5c31\u662f\u5c06bert\u7684\u8f93\u51fa( seq length *768)\u901a\u8fc7\u4e00\u4e2a768*768+768\u53d8\u6362( seq length *768)\uff0c\u7136\u540e\u901a\u8fc7\u77e9\u9635\u76f8\u4e58\uff0c\u53d8\u6362\u6210seq length*vocab size\uff08reshape\uff09\uff0c\u968f\u540e\u5bf9\u6bcf\u4e00\u4e2a\u5355\u8bcd\u6dfb\u52a0\u4e00\u4e2a\u504f\u7f6ebias\u968f\u540e\u8f93\u51fa\u3002<\/p>\n\n\n
\u800c\u77e9\u9635\u53d8\u6362\u662f\uff08 seq length * 768\uff09*\uff08768* vocab size\uff09\u6700\u540e\u5f97\u5230\uff08 seq length *vocab size\uff09\u3002\u800c\u6700\u5173\u952e\u7684\u8fd9\u4e2a \uff08768* vocab size\uff09 \uff0c\u6211\u4eec\u901a\u8fc7\u89c2\u5bdf\u53d8\u91cf\u540d\uff1ainput_embeddings.weight\uff0c\u5c31\u53ef\u4ee5\u53d1\u73b0\uff0c\u5b9e\u9645\u4e0a\u8fd9\u4e2a\u5c31\u662f\u8f93\u5165\u7684embedding\uff0c\u53ea\u4e0d\u8fc7\u8f6c\u7f6e\u4e00\u4e0b\uff0c\u76f8\u5f53\u4e8e\u8f93\u5165\u53d8\u8f93\u51fa\uff08\u76f8\u6bd4embedding\u8fc7\u7a0b\uff09\uff0c\u56e0\u6b64\u5728\u8fd9\u91cc\u5b9e\u9645\u4e0a\u662f\u53c2\u6570\u590d\u7528\u3002<\/p>\n\n\n
\u4e5f\u5c31\u662f\u8bf4\uff0c\u6211\u4eec\u5047\u8bbe\u6709\u4e00\u4e2a\u5b57\u201c\u4f60\u201d\uff0c\u5c06 [101,770,102]\uff08  [CLS,\u4f60,SEP] \uff09\u8f93\u5165\u5230embedding\u4e2d\uff0c\u6700\u540e\u8f93\u51fa3*768\u7ef4\u5411\u91cf\u8fdb\u5165bert\u4e2d\u8ba1\u7b97\uff0c\u6700\u540e\u5c06bert\u8f93\u51fa\u53cd\u5411\u9001\u5165embedding\u4e2d\uff0c\u8bd5\u56fe\u8ba9 embedding \u7ee7\u7eed\u8f93\u51fa [101,770,102] \u7684\u8fc7\u7a0b\u3002<\/p>\n\n\n
\u6240\u4ee5\uff0c\u5230\u73b0\u5728\u6211\u4eec\u5c31\u7406\u6e05\u4e86mlm\u7684\u601d\u8def\uff1a<\/strong><\/h4>\n\n\n
\u5c06bert\u7684\u8f93\u51fa\uff08512*768\uff09\u901a\u8fc7\u4e00\u4e2a768*768+768\u7684\u53d8\u6362\uff0c\u53d8\u6210\uff08512*768\uff09\uff0c\u7136\u540e\u901a\u8fc7\u4e0eembedding\uff0821128*768\uff09\u7684\u8f6c\u7f6e\uff08768*21128\uff09\u76f8\u4e58\uff0c\u83b7\u5f97\uff08512*21128\uff09\u7684\u5411\u91cf\uff0c\u518d\u52a0\u4e0a\u5bf9\u6bcf\u4e00\u4e2a\u5b57\u7684\u4e00\u4e2a\u504f\u7f6e\uff081\uff0c21128\uff09\uff0c\u4ece\u800c\u53ef\u4ee5\u83b7\u5f97\u6bcf\u4e00\u4e2a\u4f4d\u7f6e\u7684\u8f93\u51fa\u3002<\/p>\n\n\n
PS<\/strong>:\u6709\u6ca1\u6709\u4ea7\u751f\u4e00\u4e2a\u7591\u95ee\uff08\u53cd\u6b63\u6211\u6709\uff09\uff1f \u76f4\u89c9\u4e0a\uff0cembedding->token\u548ctoken->embedding\u5e94\u8be5\u662f\u4e00\u4e2a\u53ef\u9006\u8fc7\u7a0b\uff0c\u4e3a\u4ec0\u4e48\u8fd9\u91cc\u8981\u4f7f\u7528 768*768+768 \u7684Dense\u8fd8\u8981\u6dfb\u52a0\u4e00\u4e2a \uff081\uff0c21128\uff09 \u7684\u504f\u7f6e\u5462\uff1f<\/p>\n\n\n
\u56e0\u4e3a\u3002\u3002\u6211\u628a\u95ee\u9898\u60f3\u7b80\u5355\u4e86\uff0c\u8fd9\u4e2a\u76f4\u89c9\u4e0d\u51c6\uff0c embedding->token\u548ctoken->embedding \u5e76\u4e0d\u662f\u4e00\u4e2a\u53ef\u9006\u8fc7\u7a0b\uff0c\u5bf9\u4e8eEmbedding\u5c42\u5f53\u7136\u53ef\u9006\uff0c\u53ef\u662f\u6211\u4eec\u73b0\u5728\u8981\u5904\u7406\u7684\u662f\u7ecf\u8fc7\u4e8612\u5c42transformer\u5904\u7406\u540e\u7684\u201cembedding\u201d\uff0c\u56e0\u6b64\u6211\u4eec\u9700\u8981\u8fd9\u4e2aDense\u548cbias\u3002<\/p>\n\n\n
\n\n\n
2021.09.16\u66f4\uff1a\u4eca\u5929\u770b\u5230\u4e86\u4e00\u4e2a\u5173\u4e8ebert\u7684\u540d\u8bcd\uff1aweight tying\uff0c\u770b\u89e3\u91ca\u611f\u89c9\u5c31\u662f\u8fd9\u91cc\u7684\u4e1c\u897f\u3002\u67e5\u4e86\u4e00\u4e0b\u5176\u5b9e\u5c31\u662f\u8fd9\u91cc\u7684\u53c2\u6570\u5171\u4eab\u3002<\/p>\n\n\n
\u4f46\u662f\uff0c\u867d\u7136\u662f\u5171\u4eab\uff0c\u4f46\u662f\u4ed6\u4eec\u5e76\u4e0d\u662f\u540c\u4e00\u4e2a\u5c42\uff0c\u56e0\u4e3a\u4ed6\u4eec\u90fd\u6709\u7740\u5404\u81ea\u7684Dense\u5c42\u3002\u4f7f\u7528\u8fd9\u79cd\u65b9\u6cd5\u65e2\u53ef\u4ee5\u51cf\u5c11\u53c2\u6570\u91cf\uff0c\u4e5f\u53ef\u4ee5\u52a0\u5feb\u8bad\u7ec3\u901f\u5ea6\u3002<\/p>\n\n\n
2021.11.13\u66f4\uff1a\u82cf\u5efa\u6797\u5e08\u5144\u5728\u6700\u8fd1\u7684\u535a\u5ba2\u91cc\u4e5f\u8bf4\u660e\u4e86\u4e3a\u4ec0\u4e48MLM\u8981\u591a\u52a0Dense<\/a>\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u4e4b\u524d\u5bf9\u4e8ebert\u7684MLM\u4efb\u52a1\u4e00\u76f4\u90fd\u662f\u6a21\u68f1\u4e24\u53ef\uff0c\u4eca\u5929\u5bf9\u4e8e\u5b9e\u73b0\u7ec6\u8282\u8fdb\u884c\u4e86\u8865\u5168\u3002 \u60f3\u770b\u7ed3\u8bba\u7684\u76f4\u63a5\u62c9\u5230\u6700\u540e\u3002 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_mi_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"footnotes":""},"categories":[18,9],"tags":[31,37,39],"views":19602,"_links":{"self":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/3810"}],"collection":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/comments?post=3810"}],"version-history":[{"count":0,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/3810\/revisions"}],"wp:attachment":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/media?parent=3810"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/categories?post=3810"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/tags?post=3810"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}