{"id":3732,"date":"2021-02-19T15:22:26","date_gmt":"2021-02-19T07:22:26","guid":{"rendered":"http:\/\/www.sniper97.cn\/?p=3732"},"modified":"2021-02-19T15:22:26","modified_gmt":"2021-02-19T07:22:26","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%e5%af%b9%e6%8a%97%e8%ae%ad%e7%bb%83","status":"publish","type":"post","link":"http:\/\/www.sniper97.cn\/index.php\/note\/deep-learning\/note-deep-learning\/3732\/","title":{"rendered":"\u3010\u6df1\u5ea6\u5b66\u4e60\u7b14\u8bb0\u3011\u5bf9\u6297\u8bad\u7ec3"},"content":{"rendered":"\n

\u5f53\u8bf4\u5230\u5bf9\u6297\u8bad\u7ec3\u65f6\uff0c\u6211\u4eec\u4e00\u822c\u60f3\u5230\u7684\u662f\u5bf9\u6297\u751f\u6210\u7f51\u7edc\uff08GAN\uff09\u548c\u5bf9\u6297\u653b\u51fb\u3001\u5bf9\u6297\u6837\u672c\u7b49\u3002<\/p>\n\n\n

\u8fd9\u91cc\u6211\u4eec\u4e3b\u8981\u8bb2\u5bf9\u6297\u653b\u51fb\u3001\u5bf9\u6297\u6837\u672c\u3002\u5b83\u4e3b\u8981\u5173\u5fc3\u6a21\u578b\u5728\u5c0f\u6270\u52a8\u4e0b\u7684\u7a33\u5065\u6027\u3002<\/p>\n\n\n

\u6bd4\u5982\uff0c\u5728cv\u9886\u57df\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u5728\u539f\u56fe\u50cf\u4e2d\u52a0\u5165\u566a\u70b9\uff0c\u4f46\u662f\u5e76\u4e0d\u5f71\u54cd\u539f\u56fe\u50cf\u7684\u6027\u8d28\u3002<\/p>\n\n\n

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

\u800c\u5728NLP\u9886\u57df\uff0c\u6211\u4eec\u5e76\u4e0d\u80fd\u76f4\u63a5\u7684\u901a\u8fc7\u5728\u8bcd\u7f16\u7801\u4e0a\u6dfb\u52a0\u566a\u70b9\uff0c\u56e0\u4e3a\u8bcd\u5d4c\u5165\u672c\u8d28\u4e0a\u5c31\u662fone-hot\uff0c\u5982\u679c\u5728one-hot\u4e0a\u589e\u52a0\u4e0a\u8ff0\u566a\u70b9\uff0c\u5c31\u4f1a\u5bf9\u539f\u53e5\u4ea7\u751f\u6b67\u4e49\u3002<\/p>\n\n\n

\u56e0\u6b64\uff0c\u4e00\u4e2a\u81ea\u7136\u7684\u60f3\u6cd5\u5c31\u662f\u5728embedding\u4e0a\u589e\u52a0\u6270\u52a8\uff0c\u4f46\u6709\u4e2a\u95ee\u9898\u662f\u6270\u52a8\u540e\u7684embedding\u4e0d\u4e00\u5b9a\u80fd\u5339\u914d\u5230\u539f\u6765\u7684emebdding\uff08\u4e5f\u5c31\u662f\u8bed\u4e49\u53ef\u80fd\u4f1a\u53d1\u751f\u53d8\u5316\uff09\uff0c\u8fd9\u6837\u6837\u672c\u5c31\u53d1\u751f\u4e86\u53d8\u5316\u3002<\/p>\n\n\n

\u4f46\u662f\u8fd9\u4e2a\u601d\u8def\u4f9d\u7136\u4f1a\u6709\u4e00\u5b9a\u7684\u63d0\u5347\u3002<\/p>\n\n\n

\u6211\u4eec\u53ef\u4ee5\u7b80\u5355\u5730\u76f4\u63a5\u5bf9embedding\u53c2\u6570\u77e9\u9635\u8fdb\u884c\u6270\u52a8\u6765\u83b7\u5f97\u6837\u672c\u7684\u591a\u6837\u6027\u3002<\/p>\n\n\n

keras\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n\n\n

def adversarial_training(model, embedding_name, epsilon=1):\n    \"\"\"\u7ed9\u6a21\u578b\u6dfb\u52a0\u5bf9\u6297\u8bad\u7ec3\n    \u5176\u4e2dmodel\u662f\u9700\u8981\u6dfb\u52a0\u5bf9\u6297\u8bad\u7ec3\u7684keras\u6a21\u578b\uff0cembedding_name\n    \u5219\u662fmodel\u91cc\u8fb9Embedding\u5c42\u7684\u540d\u5b57\u3002\u8981\u5728\u6a21\u578bcompile\u4e4b\u540e\u4f7f\u7528\u3002\n    \"\"\"\n    if model.train_function is None:  # \u5982\u679c\u8fd8\u6ca1\u6709\u8bad\u7ec3\u51fd\u6570\n        model._make_train_function()  # \u624b\u52a8make\n    old_train_function = model.train_function  # \u5907\u4efd\u65e7\u7684\u8bad\u7ec3\u51fd\u6570\n    # \u67e5\u627eEmbedding\u5c42\n    for output in model.outputs:\n        embedding_layer = search_layer(output, embedding_name)\n        if embedding_layer is not None:\n            break\n    if embedding_layer is None:\n        raise Exception('Embedding layer not found')\n    # \u6c42Embedding\u68af\u5ea6\n    embeddings = embedding_layer.embeddings  # Embedding\u77e9\u9635\n    gradients = K.gradients(model.total_loss, [embeddings])  # Embedding\u68af\u5ea6\n    gradients = K.zeros_like(embeddings) + gradients[0]  # \u8f6c\u4e3adense tensor\n    # \u5c01\u88c5\u4e3a\u51fd\u6570\n    inputs = (model._feed_inputs +\n              model._feed_targets +\n              model._feed_sample_weights)  # \u6240\u6709\u8f93\u5165\u5c42\n    embedding_gradients = K.function(\n        inputs=inputs,\n        outputs=[gradients],\n        name='embedding_gradients',\n    )  # \u5c01\u88c5\u4e3a\u51fd\u6570\n    def train_function(inputs):  # \u91cd\u65b0\u5b9a\u4e49\u8bad\u7ec3\u51fd\u6570\n        grads = embedding_gradients(inputs)[0]  # Embedding\u68af\u5ea6\n        delta = epsilon * grads \/ (np.sqrt((grads**2).sum()) + 1e-8)  # \u8ba1\u7b97\u6270\u52a8\n        K.set_value(embeddings, K.eval(embeddings) + delta)  # \u6ce8\u5165\u6270\u52a8\n        outputs = old_train_function(inputs)  # \u68af\u5ea6\u4e0b\u964d\n        K.set_value(embeddings, K.eval(embeddings) - delta)  # \u5220\u9664\u6270\u52a8\n        return outputs\n    model.train_function = train_function  # \u8986\u76d6\u539f\u8bad\u7ec3\u51fd\u6570<\/code><\/pre>\n\n\n
\u7136\u540e\u6211\u4eec\u53ea\u9700\u8981\uff1a<\/p>\n\n\n
# \u5199\u597d\u51fd\u6570\u540e\uff0c\u542f\u7528\u5bf9\u6297\u8bad\u7ec3\u53ea\u9700\u8981\u4e00\u884c\u4ee3\u7801\nadversarial_training(model, 'Embedding-Token', 0.5)<\/code><\/pre>\n\n\n
\u5373\u53ef\u6dfb\u52a0\u8fd9\u4e2a\u6270\u52a8\u3002<\/p>\n\n\n
\u4e0d\u8fc7\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u7531\u4e8e\u5728\u6dfb\u52a0\u6270\u52a8\u7684\u65f6\u5019\u8ba1\u7b97\u4e86\u4e00\u6b21\u68af\u5ea6\uff0c\u56e0\u6b64embedding\u5b9e\u9645\u4e0a\u7684\u65f6\u95f4\u589e\u52a0\u4e86\u4e00\u500d\u3002<\/p>\n\n\n
\u6d41\u7a0b\u56fe\u5982\u4e0b\uff1a<\/p>\n\n\n
\"\"<\/figure><\/div>\n\n\n
\u5b9e\u9645\u6548\u679c\u5728CLUE\u7684\u4e24\u4e2a\u5206\u7c7b\u4efb\u52a1<\/a>\u4e0a\u6709\u4e00\u5b9a\u7684\u6548\u679c<\/p>\n\n\n
\"\"<\/figure><\/div>\n\n\n
\u53e6\u4e00\u4e2a\u6709\u8da3\u7684\u5bf9\u6297\u65b9\u6cd5\u662f\u68af\u5ea6\u60e9\u7f5a\u3002<\/p>\n\n\n
\u7b80\u5355\u6765\u8bf4\u5c31\u662f\u5411loss\u4e2d\u6dfb\u52a0\u4e00\u4e2a\u201c\u68af\u5ea6\u60e9\u7f5a\u201d\u3002<\/p>\n\n\n
\u56fe\u5f62\u7406\u89e3\u5c31\u662f\uff1a<\/p>\n\n\n
\u5047\u8bbe\u6211\u4eec\u7684\u4efb\u52a1\u6709n\u4e2a\u5206\u7c7b\uff0c\u90a3\u4e48\u6211\u4eec\u7684\u6a21\u578b\u76f8\u5f53\u4e8e\u6316\u4e86n\u4e2a\u5751\u3002<\/p>\n\n\n
\"\"<\/figure><\/div>\n\n\n
\u800c\u68af\u5ea6\u60e9\u7f5a\uff0c\u5219\u8981\u6c42\u201c\u540c\u7c7b\u6837\u672c\u4e0d\u4ec5\u8981\u653e\u5728\u540c\u4e00\u4e2a\u5751\u5185\uff0c\u8fd8\u8981\u653e\u5728\u5751\u5e95\u201d\u3002\u8fd9\u5c31\u8981\u6c42\u6bcf\u4e2a\u5751\u5185\u957f\u8fd9\u6837\uff1a<\/p>\n\n\n
\"\"<\/figure><\/div>\n\n\n
\u7531\u4e8e\u4e0d\u5bb9\u6613\u53d1\u751f\u5e72\u6270\uff08\u5751\u5e95\u7a33\u5b9a\uff09\uff0c\u56e0\u6b64\u4e5f\u5c31\u8fbe\u5230\u4e86\u5bf9\u6297\u8bad\u7ec3\u7684\u76ee\u7684\u3002<\/p>\n\n\n
\u4ee3\u7801\uff1a<\/p>\n\n\n
def sparse_categorical_crossentropy(y_true, y_pred):\n    \"\"\"\u81ea\u5b9a\u4e49\u7a00\u758f\u4ea4\u53c9\u71b5\n    \u8fd9\u4e3b\u8981\u662f\u56e0\u4e3akeras\u81ea\u5e26\u7684sparse_categorical_crossentropy\u4e0d\u652f\u6301\u6c42\u4e8c\u9636\u68af\u5ea6\u3002\n    \"\"\"\n    y_true = K.reshape(y_true, K.shape(y_pred)[:-1])\n    y_true = K.cast(y_true, 'int32')\n    y_true = K.one_hot(y_true, K.shape(y_pred)[-1])\n    return K.categorical_crossentropy(y_true, y_pred)\ndef loss_with_gradient_penalty(y_true, y_pred, epsilon=1):\n    \"\"\"\u5e26\u68af\u5ea6\u60e9\u7f5a\u7684loss\n    \"\"\"\n    loss = K.mean(sparse_categorical_crossentropy(y_true, y_pred))\n    embeddings = search_layer(y_pred, 'Embedding-Token').embeddings\n    gp = K.sum(K.gradients(loss, [embeddings])[0].values**2)\n    return loss + 0.5 * epsilon * gp\nmodel.compile(\n    loss=loss_with_gradient_penalty,\n    optimizer=Adam(2e-5),\n    metrics=['sparse_categorical_accuracy'],\n)<\/code><\/pre>\n\n\n
\u6548\u679c\u4f9d\u7136\u4e0d\u9519\uff1a<\/p>\n\n\n
\"\"<\/figure><\/div>\n","protected":false},"excerpt":{"rendered":"
\u5f53\u8bf4\u5230\u5bf9\u6297\u8bad\u7ec3\u65f6\uff0c\u6211\u4eec\u4e00\u822c\u60f3\u5230\u7684\u662f\u5bf9\u6297\u751f\u6210\u7f51\u7edc\uff08GAN\uff09\u548c\u5bf9\u6297\u653b\u51fb\u3001\u5bf9\u6297\u6837\u672c\u7b49\u3002 \u8fd9\u91cc\u6211\u4eec\u4e3b\u8981\u8bb2\u5bf9\u6297 […]<\/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":[37,39],"views":7545,"_links":{"self":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/3732"}],"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=3732"}],"version-history":[{"count":0,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/3732\/revisions"}],"wp:attachment":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/media?parent=3732"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/categories?post=3732"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/tags?post=3732"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}