{"id":3620,"date":"2020-11-04T14:20:35","date_gmt":"2020-11-04T06:20:35","guid":{"rendered":"http:\/\/www.sniper97.cn\/?p=3620"},"modified":"2020-11-04T14:20:35","modified_gmt":"2020-11-04T06:20:35","slug":"%e3%80%90%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e7%ac%94%e8%ae%b0%e3%80%91%e6%89%8b%e5%86%99transformer%e4%bb%a5%e5%8f%8a%e6%80%9d%e8%80%83","status":"publish","type":"post","link":"http:\/\/www.sniper97.cn\/index.php\/note\/deep-learning\/note-deep-learning\/3620\/","title":{"rendered":"\u3010\u6df1\u5ea6\u5b66\u4e60\u7b14\u8bb0\u3011\u624b\u5199Transformer\u4ee5\u53ca\u601d\u8003"},"content":{"rendered":"\n

\u6700\u8fd1\u5728\u624b\u5199Transformer<\/a>\uff0c\u6709\u4e86\u4e00\u4e9b\u7406\u89e3\uff0c\u4e4b\u524d\u867d\u7136\u611f\u89c9\u81ea\u5df1\u7406\u8bba\uff08Transformer\u7406\u8bba\u56fe\u89e3<\/a>\uff09\u90fd\u7406\u89e3\u4e86\uff0c\u4f46\u662f\u5199\u51fa\u6765\u4f9d\u7136\u53d1\u73b0\u6709\u5f88\u591a\u5730\u65b9\u4e0d\u592a\u660e\u767d\uff0c\u5f04\u660e\u767d\u4e4b\u540e\u5199\u51fa\u6765\u8bb0\u5f55\u4e00\u4e0b\u3002<\/p>\n\n\n

\u5148\u8bb0\u5f55\u4e00\u4e0b\u624b\u5199\u7684\u4e00\u4e9b\u7ec6\u8282\u5427\uff0c\u6700\u540e\u4f1a\u8d34\u4e0a\u5b8c\u6574\u7684github\u94fe\u63a5\u3002<\/p>\n\n\n

\u9996\u5148\u5c31\u662fTransformer\u7684\u7ed3\u6784\uff0c\u5de6\u4fa7\u7684encoder\u5c42\u548c\u53f3\u4fa7\u7684decoder\u5c42\uff0c\u5b9e\u9645\u7684Transformer\u7684encoder\u548cdecoder\u5c31\u662f\u7531N\u4e2aencoder\u5c42\u548cN\u4e2adocoder\u5c42\u7ec4\u6210\u3002<\/p>\n\n\n

\"\"<\/figure><\/div>\n\n\n

\u6211\u4eec\u7ee7\u7eed\u89c2\u5bdf\u6a21\u578b\uff0c\u53d1\u73b0\u591a\u5934\u6ce8\u610f\u529b\uff08\u6ce8\u610f\u529b\u57fa\u7840<\/a>\u3001\u6ce8\u610f\u529b\u7406\u8bba\u56fe\u89e3<\/a>\uff09\u6709\u4e24\u79cd\uff0c\u4e00\u79cd\u5c31\u662f\u5355\u7eaf\u7684Multi-Head Attention\uff0c\u53e6\u4e00\u79cd\u662fMasked Multi-Head Attention\u3002<\/p>\n\n\n

\u90a3\u4e48\u8fd9\u4e24\u79cd\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\u6709\u4ec0\u4e48\u533a\u522b\u5462\uff0c\u5f88\u663e\u7136\uff0c\u5c31\u662fMask\u673a\u5236\u3002<\/p>\n\n\n

\u8fd9\u4e2aMask\u673a\u5236\u5b9e\u9645\u4e0a\u662f\u4e00\u4e2a\u906e\u853d\u673a\u5236\uff0c\u7531\u4e8eTransformer\u7684\u5e76\u884c\u8981\u6c42\uff08\u4e8b\u5b9e\u4e0a\u8fd9\u4e5f\u662fTransformer\u76f8\u6bd4RNN\u7684\u4e00\u5927\u4f18\u52bf\uff09\uff0c\u9700\u8981\u5c06target\u8fdb\u884cembedding\u7136\u540e\u8f93\u5165\u5230decoder\u4e2d\uff0c\u56e0\u6b64\u5c31\u9700\u8981\u4f7f\u7528\u4e00\u4e2a\u673a\u5236\uff0c\u5c4f\u853d\u6389\u5f53\u524d\u9884\u6d4b\u4f4d\u53ca\u5176\u4e4b\u540e\u7684\u6587\u672c\uff0c\u6bd4\u5982\uff1a\u8f93\u51fa\u6587\u672c\u662f\u201c \u6211\u7231\u4f60 \u201d\uff0c\u90a3\u4e48\u5728\u9884\u6d4b\u201c \u7231 \u201d\u7684\u65f6\u5019\uff0c\u5f53\u524d\u7f51\u7edc\u5e94\u8be5\u53ea\u53ef\u89c1\u201c\u6211\u201d\u5b57\u3002<\/p>\n\n\n

\u90a3\u4e48\u662f\u600e\u4e48\u505a\u7684\u5462\uff1f\u5c31\u662f\u4f7f\u7528\u4e00\u4e2a\u4e0b\u4e09\u89d2\u77e9\u9635\uff0c\u7531\u4e8e\u6700\u7ec8\u7684embedding\u7684\u6587\u672c\u8fd8\u8981\u6709\u5f00\u59cb\u6807\u8bb0\uff08\u201c<BOS> \u6211 \u7231 \u4f60 <EOS>\u201d\uff09\uff0c\u56e0\u6b64\u5bf9\u4e8e\u8fd9\u53e5\u8bdd\u76845*5\u7684\u4e0b\u4e09\u89d2\u77e9\u9635\uff0c\u6b63\u597d\u53ef\u4ee5\u906e\u853d\u6389\u9884\u6d4b\u5b57\u540e\u9762\u7684\u5b57\uff0c\u6bd4\u5982\u77e9\u9635\u7b2c\u4e8c\u884c\u67092\u4e2a1\uff0c\u56e0\u6b64\u53ea\u6709\u201c <BOS> \u6211 \u201d\u5bf9\u7f51\u7edc\u53ef\u89c1\uff0c\u5e76\u8bd5\u56fe\u4f7f\u7f51\u7edc\u9884\u6d4b\u201c\u7231\u201d\u3002<\/p>\n\n\n

\u9664\u6b64\u4e4b\u5916\u8fd8\u6709\u4e00\u4e2amask\uff0c\u5c31\u662f\u8981mask\u6389padding\uff0c\u7531\u4e8e\u5bf9\u4e8e\u77ed\u6587\u672c\uff0c\u6211\u4eec\u9700\u8981\u5c06\u6587\u672cpadding\u5230\u5b9a\u957f\uff0c\u6211\u4eec\u5e76\u4e0d\u9700\u8981\u6ce8\u610f\u529b\u673a\u5236\u5173\u6ce8\u8fd9\u90e8\u5206padding\u4fe1\u606f\uff0c\u56e0\u6b64\u6211\u4eec\u540c\u6837\u53ef\u4ee5\u4f7f\u7528mask\u673a\u5236\u5f3a\u884c\u6e05\u7a7a\u6ce8\u610f\u529b\u7ed3\u679c\uff08\u4e580\u6216\u4e00\u4e2a\u6781\u5c0f\u7684\u6570\u5b57\uff09\u3002<\/p>\n\n\n

\u8be6\u89c1 Multi-Head Attention \u4ee3\u7801\u3002<\/p>\n\n\n

import tensorflow as tf


class MutiHeadAttention(tf.keras.layers.Layer):
    \"\"\"
<\/em>    <\/em>\u591a\u5934\u6ce8\u610f\u529b<\/em>
<\/em>    \"\"\"
<\/em>
<\/em>    <\/em>def __init__(self, d_model, num_heads):
        \"\"\"
<\/em>
<\/em>        <\/em>:param<\/em><\/strong> d_model: muti attn <\/em>\u8f93\u51fa\u7684\u7ef4\u5ea6
<\/em>        <\/em>:param<\/em><\/strong> num_heads: <\/em>\u591a\u5934\u4e2a\u6570<\/em>
<\/em>        \"\"\"
<\/em>        <\/em>super().__init__()

        assert d_model % num_heads == 0
        self.num_heads = num_heads
        self.d_model = d_model

        self.deep = self.d_model \/\/ self.num_heads

        self.WQ = tf.keras.layers.Dense(self.d_model)
        self.WK = tf.keras.layers.Dense(self.d_model)
        self.WV = tf.keras.layers.Dense(self.d_model)

        self.dense = tf.keras.layers.Dense(self.d_model)

    def split_heads(self, x, batch_size):
        \"\"\"
<\/em>        <\/em>\u5206\u5934\uff0c
<\/em>        <\/em>:param<\/em><\/strong> x:
<\/em>        <\/em>:param<\/em><\/strong> batch_size:
<\/em>        <\/em>:return<\/em><\/strong>:
<\/em>        \"\"\"
<\/em>        <\/em>x = tf.reshape(x, (batch_size, -1, self.num_heads, self.deep))
        # perm \u662f\u65b0\u7684\u7ef4\u5ea6\u7d22\u5f15\uff0c\u5047\u5982x\u7ef4\u5ea6\u662f[1,2,3,4]\uff0cperm=[0,2,1,3]\u8f6c\u7f6e\u540ex\u7ef4\u5ea6\u5c06\u53d8\u6210[1,3,2,4]
        return tf.transpose(x, perm=[0, 2, 1, 3])

    def call(self, q, k, v, mask):
        batch_size = tf.shape(q)[0]

        q = self.WQ(q)
        k = self.WK(k)
        v = self.WV(v)

        q = self.split_heads(q, batch_size)
        k = self.split_heads(k, batch_size)
        v = self.split_heads(v, batch_size)

        attention_output, attention_weight = dot_attention(q, k, v, mask)

        attention_output = tf.transpose(attention_output, perm=[0, 2, 1, 3])
        attention_output = tf.reshape(attention_output, (batch_size, -1, self.d_model))

        _output = self.dense(attention_output)
        return _output, attention_weight


#
#
def dot_attention(q, k, v, mask):
    \"\"\"
<\/em>    <\/em>\u70b9\u4e58
<\/em>    <\/em>:param<\/em><\/strong> q:
<\/em>    <\/em>:param<\/em><\/strong> k:
<\/em>    <\/em>:param<\/em><\/strong> v:
<\/em>    <\/em>:param<\/em><\/strong> mask:
<\/em>    <\/em>:return<\/em><\/strong>:
<\/em>    \"\"\"
<\/em>    <\/em>qk = tf.matmul(q, k, transpose_b=True)

    # \u8ba1\u7b97\u6839\u53f7\u4e0bdk\u5bf9qk\u8fdb\u884c\u7f29\u653e\uff0c\u7c7b\u4f3c\u4e8e emm \u5f52\u4e00\u5316
    dk = tf.cast(tf.shape(k)[-1], tf.float32)
    scaled_attention_logits = qk \/ tf.math.sqrt(dk)

    if mask is not None:
        scaled_attention_logits += (mask * -1e9)

    attention_weights = tf.nn.softmax(scaled_attention_logits, axis=-1)

    _output = tf.matmul(attention_weights, v)
    return _output, attention_weights<\/pre>\n\n\n
\u5176\u4ed6\u7684encoder\u5c42\u3001deocder\u5c42\u3001\u4ee5\u53ca\u6700\u7ec8\u7684encoder\u3001deocder\u8fd9\u91cc\u4e0d\u518d\u63cf\u8ff0\uff0c\u4e0d\u662f\u5f88\u590d\u6742\u3002<\/p>\n\n\n
\u5230\u8fd9\u91cc\u5176\u5b9e\u8fd8\u597d\uff0c\u4e0b\u9762\u7684\u95ee\u9898\u4e00\u5ea6\u8ba9\u6211\u6000\u7591\u6211\u7684\u7406\u89e3\u51fa\u4e86\u5f88\u5927\u7684\u95ee\u9898\u3002<\/p>\n\n\n
\u90fd\u5199\u5b8c\u4e4b\u540e\uff0c\u6211\u60f3\u6d4b\u8bd5\u4e00\u4e0b\u8fd9\u4e2a\u6a21\u578b\u7684\u6548\u679c\uff0c\u4e8e\u662f\u7ee7\u7eed\u624b\u5199\u4e86\u4e00\u4e0b\u6d4b\u8bd5\uff0c\u4f46\u662f\u53d1\u73b0\u4e86\u4e00\u4e2a\u95ee\u9898\uff0c\u8bad\u7ec3\u9636\u6bb5\u4f20\u7684target\uff0c\u5728\u6d4b\u8bd5\u9636\u6bb5\u4f20\u4ec0\u4e48\uff1f\u5982\u679c\u53ea\u4f20<BOS>\uff0c\u90a3\u4e48\u6700\u7ec8\u7684\u7ed3\u679c\u5c06\u5929\u82b1\u4e71\u5760\uff0c\u5e76\u4e14\u518d\u9884\u6d4b\u9636\u6bb5\u4f20\u5b8c\u6574\u7684target\u4e5f\u5c31\u6ca1\u6709\u610f\u4e49\u4e86\u3002<\/p>\n\n\n
\u7ecf\u8fc7\u4e00\u6bb5\u65f6\u95f4\u7684\u7814\u7a76\u53d1\u73b0\uff0cTransformer\u7684\u6d4b\u8bd5\u9636\u6bb5\u4f9d\u7136\u9700\u8981n\u6b21\u6a21\u578b\u8fed\u4ee3\uff0c\u5c31\u7c7b\u4f3c\u4e8eRNN\uff0c\u7528n-1\u6b21\u5f97\u8f93\u51fa\uff0c\u4f5c\u4e3aoutput target\uff0c\u4f20\u5165n\u6b21\uff0c\u7531\u6a21\u578b\u9884\u6d4b\u7b2cn\u4e2a\u8bcd\u3002<\/p>\n\n\n
\u8fd9\u662f\u7531\u4e8eTransformer\u7684teaching force\u51b3\u5b9a\u7684\uff0c\u56e0\u6b64\u5728Transformer\u7684\u8bad\u7ec3\u548c\u6d4b\u8bd5\u4e2d\u5c31\u5b58\u5728\u4e00\u4e2agap\u3002<\/p>\n\n\n
\u867d\u7136\u77e5\u9053\u4e4b\u540e\u611f\u89c9\u633a\u7b80\u5355\u7684\uff0c\u4f46\u662f\u5f53\u65f6\u786e\u5b9e\u56f0\u60d1\u4e86\u597d\u4e45\uff0c\u4e5f\u7b97\u638c\u63e1\u4e86\u4e00\u4e2a\u65b0\u77e5\u8bc6\uff0c\u5199\u4e0b\u6765 \u8bb0\u5f55\u4e00\u4e0b\u3002<\/p>\n\n\n
\u5b8c\u6574\u4ee3\u7801\u94fe\u63a5<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"
\u6700\u8fd1\u5728\u624b\u5199Transformer\uff0c\u6709\u4e86\u4e00\u4e9b\u7406\u89e3\uff0c\u4e4b\u524d\u867d\u7136\u611f\u89c9\u81ea\u5df1\u7406\u8bba\uff08Transformer\u7406\u8bba\u56fe\u89e3 […]<\/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":[9],"tags":[32,37,39],"views":9106,"_links":{"self":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/3620"}],"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=3620"}],"version-history":[{"count":0,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/3620\/revisions"}],"wp:attachment":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/media?parent=3620"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/categories?post=3620"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/tags?post=3620"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}