{"id":2593,"date":"2018-10-10T13:57:48","date_gmt":"2018-10-10T05:57:48","guid":{"rendered":"http:\/\/www.yewen.us\/blog\/?p=2593"},"modified":"2018-10-10T13:57:48","modified_gmt":"2018-10-10T05:57:48","slug":"append-machine-learning-3-linear-regression","status":"publish","type":"post","link":"https:\/\/www.yewen.us\/blog\/2018\/10\/append-machine-learning-3-linear-regression\/","title":{"rendered":"\u8865\u9057: \u673a\u5668\u5b66\u4e60\u624b\u8bb0\u7cfb\u5217 3 \u7ebf\u6027\u56de\u5f52\u6837\u4f8b\u7a0b\u5e8f"},"content":{"rendered":"

\n \u8ddd\u79bb 2012 \u7684\u4e24\u4e09\u5e74\u540e\uff08\u8fd9\u7bc7\u7684\u8349\u7a3f\u65f6\u95f4\uff09\u53c8\u8fc7\u4e86\u4e24\u4e09\u5e74\uff0c\u8fd9\u4e2a\u8865\u9057\u770b\u8d77\u6765\u4e5f\u70c2\u5c3e\u4e86 -.-\n<\/p><\/blockquote>\n

\u4e4b\u524d\u5728\u673a\u5668\u5b66\u4e60\u624b\u8bb0\u7cfb\u5217 3: \u7ebf\u6027\u56de\u5f52\u548c\u6700\u5c0f\u4e8c\u4e58\u6cd5<\/a>\u540e\u9762\u7559\u4e86\u4e2a\u95ee\u9898, \u4e5f\u7ed9\u4e86\u7ed3\u679c, \u4f46\u662f\u5f53\u65f6\u8bf4\u597d\u7684\u7a0b\u5e8f\u4ee3\u7801\u5e76\u6ca1\u7ed9\u51fa\u6765, \u90a3\u4e2a\u624b\u8bb0\u7cfb\u5217\u7684\u5751\u611f\u89c9\u586b\u4e0d\u4e0a\u4e86, \u4f46\u662f\u5df2\u7ecf\u5228\u597d\u7684\u5c0f\u5751\u8fd8\u662f\u586b\u4e0a\u5427<\/p>\n

\u73b0\u5728\u5df2\u7ecf\u6709\u5f88\u591a\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u548c\u6559\u7a0b\u6765\u6559\u8fd9\u4e2a\uff0c\u81ea\u5df1\u4e5f\u5fd8\u5f97\u5dee\u4e0d\u591a\u4e86\uff0c\u5c31\u4e0d\u73ed\u95e8\u5f04\u65a7\u88f8\u5199\u3002\u63a8\u8350\u770b\u4e00\u4e0b \u52a8\u624b\u5b66\u6df1\u5ea6\u5b66\u4e60 http:\/\/zh.gluon.ai\/index.html<\/a>\uff0cDeep Learning \u9886\u57df\u5927\u795e \u674e\u6c90 \u7b49\u4eba\u5728\u7ef4\u62a4\uff08\u6211\u80fd\u51d1\u4e0d\u8981\u8138\u7684\u8e6d\u70ed\u5ea6\u8bf4\u4e0b\u8fd9\u662f\u524d\u767e\u5ea6\u540c\u4e8b\u6211\u4eec\u8fd8\u4e00\u8d77\u5403\u996d\u6253\u724c\u6765\u7740\u4e48\uff09\u3002\u5228\u7684\u5c0f\u5751\u5c31\u6309 \u7ebf\u6027\u56de\u5f52\u7684\u4ece\u96f6\u5f00\u59cb\u5b9e\u73b0 http:\/\/zh.gluon.ai\/chapter_deep-learning-basics\/linear-regression-scratch.html<\/a> \u91cc\u7684\u505a\u6cd5\u6765\u5b9e\u73b0<\/p>\n

\u5148\u91cd\u590d\u4e0b\u95ee\u9898<\/p>\n

\u5982\u4e0b\u5f0f\u5b50\u91cc\u4e0d\u540c\u7684\u963f\u62c9\u4f2f\u6570\u5b57\u53ea\u662f\u4e00\u4e2a\u7b26\u53f7, \u5b9e\u9645\u8868\u793a\u7684\u53ef\u80fd\u662f\u5176\u4ed6\u6570\u5b57
\n967621 = 3
\n797321 = 1
\n378581 = 4
\n422151 = 0
\n535951 = 1
\n335771 = 0<\/p>\n

\u6839\u636e\u4e0a\u8ff0\u5f0f\u5b50, \u5224\u65ad\u4e0b\u5f0f\u7b49\u4e8e?
\n565441 = ?<\/p><\/blockquote>\n

\u8fd9\u9898\u7684\u8111\u7b4b\u6025\u8f6c\u5f2f\u7248\u672c\u7b54\u6848\u662f\u770b\u6bcf\u4e2a\u6570\u5b57\u6709\u51e0\u4e2a\u5708\uff0c\u5c31\u4ee3\u8868\u51e0\uff0c\u8fd9\u6837 1\/2\/3\/4\/5\/7 \u90fd\u662f 0 \u4e2a\u5708\uff0c6\/9 \u662f 1 \u4e2a\u5708\uff0c8 \u662f 2 \u4e2a\u5708\uff0c\u6240\u4ee5\u6700\u540e 565441 \u91cc\u9762\u53ea\u6709 6 \u6709 1 \u4e2a\u5708\uff0c\u7b54\u6848\u4e3a 1<\/p>\n

\u6309 gluon \u4e0a\u7684\u6559\u7a0b\u6211\u4eec\u4e5f\u6765\u8d70\u4e00\u904d\uff0c\u88c5\u73af\u5883\u4ec0\u4e48\u7684\u5c31\u770b gluon \u4e86\uff0c\u5148\u5f15\u5165\u8981\u7528\u7684\u5305<\/p>\n

\r\nfrom mxnet import autograd, nd\r\n<\/pre>\n
\u771f\u6b63\u505a\u7ebf\u6027\u56de\u5f52\u662f\u6ca1\u6cd5\u53ea\u7528\u8fd9\u4e48\u4e00\u70b9\u6570\u636e\u6765\u6a21\u62df\u7684\uff0c\u6240\u4ee5\u6211\u4eec\u8981\u5148\u6839\u636e\u771f\u5b9e\u503c\u6765\u6784\u9020\u4e00\u4e9b\u6570\u636e\uff08\u8fd9\u91cc\u8ddf gluon \u4e0d\u4e00\u6837\u7684\u662f\u6211\u6ca1\u6709 bias \u56e0\u5b50 b\uff0c\u540e\u9762\u4e5f\u8bf7\u4e00\u5e76\u6ce8\u610f\uff09<\/p>\n
\nnum_inputs = 9          # \u7279\u5f81\u6570\uff0c\u5f53\u524d\u95ee\u9898\u91cc\u7684\u53d8\u91cf\u6570 1-9\nnum_examples = 1000     # \u6837\u4f8b\u6570\uff0c\u6211\u4eec\u4f1a\u968f\u673a\u751f\u6210\u591a\u5c11\u4efd\u6837\u4f8b\u6765\u5b66\u4e60\ntrue_w = nd.array([0, 0, 0, 0, 0, 1, 0, 2, 1])  # \u771f\u5b9e\u503c\nfeatures = nd.random.normal(scale=1, shape=(num_examples, num_inputs))  # \u968f\u673a\u751f\u6210\u6570\u636e\u96c6\nlabels = nd.dot(features, true_w)                                   # \u6570\u636e\u96c6\u5bf9\u5e94\u7684\u7ed3\u679c\n<\/pre>\n
\u521d\u59cb\u5316\u6a21\u578b\u53c2\u6570\u5e76\u521b\u5efa\u68af\u5ea6<\/p>\n
\nw = nd.random.normal(scale=0.01, shape=(9, 1))\nw.attach_grad()\n<\/pre>\n
\u5b9a\u4e49\u6a21\u578b\uff0c\u6211\u4eec\u5c31\u662f\u505a\u7684\u77e9\u9635\u4e58\u6cd5<\/p>\n
\ndef linreg(X, w):\n    return nd.dot(X, w)\n<\/pre>\n
\u5b9a\u4e49\u635f\u5931\u51fd\u6570\uff0c\u7528\u5e73\u65b9\u635f\u5931<\/p>\n
\ndef squared_loss(y_hat, y):\n    return (y_hat - y.reshape(y_hat.shape)) ** 2 \/ 2\n<\/pre>\n
\u5b9a\u4e49\u4f18\u5316\u7b97\u6cd5\uff0c\u7528\u5c0f\u6279\u91cf\u968f\u673a\u68af\u5ea6\u4e0b\u964d\uff08\u56e0\u4e3a\u6211\u4eec\u53ea\u7528\u4e86\u4e00\u4e2a\u5927\u53c2\u6570 w\uff0c\u6240\u4ee5\u8fd8\u662f\u6bd4 gluon \u7684\u6837\u4f8b\u7b80\u5355\uff09<\/p>\n
\ndef sgd(param, lr, batch_size):\n    param[:] = param - lr * param.grad \/ batch_size\n<\/pre>\n
\u8bad\u7ec3\uff0c\u53d6\u6b65\u957f lr \u4e3a 0.01\uff0c\u8f6e\u6b21\u4e3a 1000 \u8f6e<\/p>\n
\ndef train():\n    lr = 0.01\n    num_epochs = 1000\n    net = linreg\n    loss = squared_loss\n\n    for epoch in range(num_epochs):\n        with autograd.record():\n            l = loss(net(features, w), labels)\n        l.backward()\n        sgd(w, lr, labels.size)\n        train_l = loss(net(features, w), labels)\n        if epoch % 100 == 99:\n            print("epoch {}, loss {}, w {}".format(epoch + 1, train_l.mean().asnumpy(), w))\n<\/pre>\n
\u9a8c\u8bc1\u4e0b\u7ed3\u679c\u770b\u770b<\/p>\n
\nif __name__ == "__main__":\n    train()\n    test = nd.array([1, 0, 0, 2, 2, 1, 0, 0, 0])    # \u6d4b\u8bd5\u96c6\uff0c565441\n    print(nd.dot(test, w))\n<\/pre>\n
\u968f\u4fbf\u8dd1\u4e86\u4e00\u6b21\u8f93\u51fa\u5982\u4e0b\uff0c\u6ce8\u610f\u6a21\u578b\u91cc\u6bcf\u4e2a\u503c\u7684\u79d1\u5b66\u8ba1\u6570\u6cd5\u7684\u6307\u6570<\/p>\n
\nepoch 1000, loss [  5.72006487e-09], w\n[[ -6.20802666e-06]\n [  1.62000088e-05]\n [ -1.03610901e-05]\n [  7.82768348e-06]\n [  2.59973749e-05]\n [  9.99964714e-01]\n [  1.86312645e-05]\n [  1.99990368e+00]\n [  1.00001490e+00]]\n<NDArray 9x1 @cpu(0)>\n\n[ 1.00002611]\n<NDArray 1 @cpu(0)>\n<\/pre>\n
\u5ffd\u7565\u7cbe\u5ea6\u95ee\u9898\uff0c\u53ef\u4ee5\u8ba4\u4e3a\u7b26\u5408\u771f\u5b9e\u7ed3\u679c<\/p>\n
\u5168\u90e8\u4ee3\u7801\u8be6\u89c1 https:\/\/gist.github.com\/whusnoopy\/af0aa6fd276ace8a7c4d483e586e936d<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"
\u8ddd\u79bb 2012 \u7684\u4e24\u4e09\u5e74\u540e\uff08\u8fd9\u7bc7\u7684\u8349\u7a3f\u65f6\u95f4\uff09\u53c8\u8fc7\u4e86\u4e24\u4e09\u5e74\uff0c\u8fd9\u4e2a\u8865\u9057\u770b\u8d77\u6765\u4e5f\u70c2\u5c3e\u4e86 -.- \u4e4b\u524d\u5728\u673a\u5668\u5b66\u4e60\u624b\u8bb0\u7cfb […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"jetpack_post_was_ever_published":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[9],"tags":[677,310,386],"class_list":["post-2593","post","type-post","status-publish","format-standard","hentry","category-machine-learning","tag-gluon","tag-310","tag-386"],"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"jetpack_shortlink":"https:\/\/wp.me\/p4aR5e-FP","_links":{"self":[{"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/posts\/2593","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/comments?post=2593"}],"version-history":[{"count":8,"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/posts\/2593\/revisions"}],"predecessor-version":[{"id":3424,"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/posts\/2593\/revisions\/3424"}],"wp:attachment":[{"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/media?parent=2593"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/categories?post=2593"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/tags?post=2593"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}