{"id":665,"date":"2021-04-28T03:23:00","date_gmt":"2021-04-28T03:23:00","guid":{"rendered":"https:\/\/tensorzen.online\/?p=665"},"modified":"2024-05-21T07:53:35","modified_gmt":"2024-05-21T07:53:35","slug":"pytorch%e5%ae%9e%e7%8e%b0%e6%9c%80%e5%a4%a7%e4%bc%bc%e7%84%b6%e4%bc%b0%e8%ae%a1","status":"publish","type":"post","link":"https:\/\/tensorzen.blog\/?p=665","title":{"rendered":"PyTorch\u5b9e\u73b0\u6700\u5927\u4f3c\u7136\u4f30\u8ba1"},"content":{"rendered":"\n

\u6211\u5077\u5077\u751f\u6210\u4e00\u6ce2\u6570\u636e$x$\uff0c\u751f\u62101000\u4e2a\u5427\uff0c\u8fd9\u4e9b\u6570\u636e\u662f\u4ece\u4e00\u4e2a\u5206\u5e03\u53d6\u6837\u7684\uff0c\u65e2\u7136\u662f\u5077\u5077\u751f\u6210\u7684\uff0c\u90a3\u80af\u5b9a\u4e0d\u80fd\u544a\u8bc9\u8fd9\u4e2a\u5206\u5e03\u662f\u4ec0\u4e48~~~<\/p>\n\n\n\n

\u806a\u660e\u7684\u4f60\u80af\u5b9a\u5148\u753b\u4e2a\u6563\u70b9\u56fe\u770b\u770b\u5927\u6982\u662f\u4ec0\u4e48\u5206\u5e03\uff1a<\/p>\n\n\n\n

<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
import<\/span> matplotlib<\/span>.<\/span>pyplot <\/span>as<\/span> plt<\/span><\/span>\nplt<\/span>.<\/span>hist<\/span>(<\/span>x<\/span>,<\/span> <\/span>bins<\/span>=<\/span>20<\/span>)<\/span><\/span><\/code><\/pre><\/div>\n\n\n
\n
\"\"<\/figure>\n<\/div>\n\n\n
\u6253\u773c\u4e00\u770b\u8fd9\u4e0d\u5c31\u662f\u6b63\u6001\u5206\u5e03\u5417\uff0c\u6b63\u6001\u5206\u5e03\u7684\u8bdd\u9700\u8981\u77e5\u9053\u4e24\u4e2a\u53c2\u6570$\\mu, \\sigma$\uff0c\u4e5f\u5c31\u662f\u5747\u503c\u548c\u6807\u51c6\u5dee\uff0c\u4e8e\u662f\u806a\u660e\u7684\u4f60\u4f1a\u5b9a\u4e49\u4e24\u4e2a\u53d8\u91cf\uff0c\u8fd9\u4e24\u4e2a\u53d8\u91cf\u5c31\u662f\u4f60\u8981learning\u7684\u53c2\u6570\u3002<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
mu <\/span>=<\/span> torch<\/span>.<\/span>tensor<\/span>(<\/span>1.0<\/span>,<\/span> <\/span>requires_grad<\/span> <\/span>=<\/span> <\/span>True<\/span>)<\/span><\/span>\nsigma <\/span>=<\/span> torch<\/span>.<\/span>tensor<\/span>(<\/span>1.0<\/span>,<\/span> <\/span>requires_grad<\/span> <\/span>=<\/span> <\/span>True<\/span>)<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
\u4e3a\u4e86\u65b9\u4fbf\uff0c\u4f60\u53ef\u80fd\u4e0d\u60f3\u81ea\u5df1\u5b9e\u73b0SGD\u6765\u4f18\u5316\u8fd9\u4fe9\u53c2\u6570\u4e86\uff0c\u4f60\u4f1a\u9009\u62e9\u4f7f\u7528torch\u5185\u7f6e\u7684\u4f18\u5316\u5668\u6765\u4f18\u5316\u8fd9\u4fe9\u53c2\u6570\uff1a<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
optimizer <\/span>=<\/span> torch<\/span>.<\/span>optim<\/span>.<\/span>SGD<\/span>([<\/span>mu<\/span>,<\/span> sigma<\/span>],<\/span> <\/span>lr<\/span>=<\/span>2e-2<\/span>)<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
lr\u662f\u5b66\u4e60\u7387\uff0c\u8bbe\u5b9a\u7684\u5c0f\u4e00\u70b9\u4f1a\u597d\u70b9\u3002pytoch\u5185\u7f6e\u7684\u5927\u591a\u6570\u5e38\u7528\u7684\u5206\u5e03\u51fd\u6570\uff0c\u4efb\u541b\u6311\u9009\uff1a<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
q <\/span>=<\/span> torch<\/span>.<\/span>distributions<\/span>.<\/span>Normal<\/span>(<\/span>loc<\/span>=<\/span>mu<\/span>,<\/span> <\/span>scale<\/span>=<\/span>sigma<\/span>)<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
\u8fd9\u6837\u5c31\u76f8\u5f53\u4e8e\u5b9a\u4e49\u4e86\u4e00\u4e2a\u5747\u503c\u4e3a1\uff0c\u6807\u51c6\u5dee\u4e3a1\u7684\u6b63\u6001\u5206\u5e03$q(x)$\uff0c\u4f60\u7684\u76ee\u6807\u662f\u901a\u8fc7\u4f7f\u7528\u6700\u5927\u4f3c\u7136\u4f30\u8ba1\u6765\u5f97\u5230$\\mu, \\sigma$\u5148\u5199\u4e2a\u4f3c\u7136\u51fd\u6570\uff1a<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
negative_log_likelihood <\/span>=<\/span> <\/span>-<\/span>1<\/span> <\/span>*<\/span> torch<\/span>.<\/span>sum<\/span>(<\/span>q<\/span>.<\/span>log_prob<\/span>(<\/span>x_batch<\/span>))<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
pytorch\u5e76\u6ca1\u6709\u63d0\u4f9bprob\u65b9\u6cd5\u8fd4\u56de\u6982\u7387\uff0c\u53cd\u6b63\u4f60\u4e5f\u4e0d\u4f1a\u7528\u6982\u7387\u76f8\u4e58\uff0c\u90fd\u662f\u6982\u7387\u7684\u5bf9\u6570\u76f8\u52a0\uff0c\u4e8e\u662f\u5bf9log_prob\u6c42\u548c\u5c31\u53ef\u4ee5\u4e86\u3002\u63a5\u4e0b\u6765\u8ba1\u7b97$\\mu, \\sigma$\u7684\u68af\u5ea6<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
negative_log_likelihood<\/span>.<\/span>backward<\/span>()<\/span><\/span>\noptimizer<\/span>.<\/span>step<\/span>()<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
\u987a\u4fbf\u8c03\u7528\u4e00\u4e0boptimizer.step()\u65b9\u6cd5\u6765\u66f4\u65b0\u4e00\u4e0b$\\mu, \\sigma$\uff0c\u6267\u884c\u591a\u6b21\u4e4b\u540e\uff0c\u5c31\u80fd\u5f97\u5230\u76ee\u6807\u4e86\u3002<\/p>\n\n\n\n
\u81f3\u4e8e\uff0c\u5077\u5077\u751f\u6210\u6570\u636e\u7684\u8fc7\u7a0b\uff0c\u5176\u5b9e\u662f\u7528\u5747\u503c\u4e3a-4\uff0c\u65b9\u5dee\u4e3a2\u7684\u6b63\u6001\u5206\u5e03\u751f\u6210\u4e861000\u4e2a\u6837\u672c\u3002<\/p>\n\n\n\n
\u4e8e\u662f\u7ecf\u8fc7\u4e0a\u8ff0\u7684\u5b66\u4e60\u8fc7\u7a0b\uff0c$\\mu, \\sigma$\u4f1a\u5728-4,2\u524d\u540e\u5f98\u5f8a\uff0c\u56e0\u4e3a\u6bd5\u7adf\u662f\u53d6\u6837\u7684\u6570\u636e\uff0c\u4e0e\u771f\u5b9e\u5206\u5e03\u8fd8\u662f\u6709\u5dee\u8ddd\u7684~<\/p>\n\n\n\n
\u5b8c\u6574\u4ee3\u7801\uff1a<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
import<\/span> numpy <\/span>as<\/span> np<\/span><\/span>\nfrom<\/span> scipy<\/span>.<\/span>stats <\/span>import<\/span> norm<\/span><\/span>\nimport<\/span> torch<\/span><\/span>\nimport<\/span> matplotlib<\/span>.<\/span>pyplot <\/span>as<\/span> plt<\/span><\/span>\n#<\/span> \u751f\u6210\u6570\u636e<\/span><\/span>\nx <\/span>=<\/span> np<\/span>.<\/span>random<\/span>.<\/span>normal<\/span>(<\/span>loc<\/span> <\/span>=<\/span> <\/span>-<\/span>4<\/span>,<\/span> <\/span>scale<\/span> <\/span>=<\/span> <\/span>2<\/span>,<\/span> <\/span>size<\/span> <\/span>=<\/span> <\/span>1000<\/span>)<\/span><\/span>\nplt<\/span>.<\/span>hist<\/span>(<\/span>x<\/span>,<\/span> <\/span>bins<\/span>=<\/span>20<\/span>)<\/span><\/span>\nx <\/span>=<\/span> torch<\/span>.<\/span>tensor<\/span>(<\/span>x<\/span>)<\/span><\/span>\n#<\/span> MLE<\/span><\/span>\nmu <\/span>=<\/span> torch<\/span>.<\/span>tensor<\/span>(<\/span>1.0<\/span>,<\/span> <\/span>requires_grad<\/span> <\/span>=<\/span> <\/span>True<\/span>)<\/span><\/span>\nsigma <\/span>=<\/span> torch<\/span>.<\/span>tensor<\/span>(<\/span>1.0<\/span>,<\/span> <\/span>requires_grad<\/span> <\/span>=<\/span> <\/span>True<\/span>)<\/span><\/span>\noptimizer <\/span>=<\/span> torch<\/span>.<\/span>optim<\/span>.<\/span>SGD<\/span>([<\/span>mu<\/span>,<\/span> sigma<\/span>],<\/span> <\/span>lr<\/span>=<\/span>2e-2<\/span>)<\/span><\/span>\n<\/span>\n#<\/span> SGD<\/span><\/span>\nidx <\/span>=<\/span> <\/span>list<\/span>(<\/span>range<\/span>(<\/span>len<\/span>(<\/span>x<\/span>)))<\/span><\/span>\nfor<\/span> epoch <\/span>in<\/span> <\/span>range<\/span>(<\/span>2<\/span>):<\/span><\/span>\n    np<\/span>.<\/span>random<\/span>.<\/span>shuffle<\/span>(<\/span>idx<\/span>)<\/span><\/span>\n    <\/span>for<\/span> i <\/span>in<\/span> <\/span>range<\/span>(<\/span>0<\/span>,<\/span>len<\/span>(<\/span>idx<\/span>),<\/span>10<\/span>):<\/span><\/span>\n        x_batch <\/span>=<\/span> x<\/span>[<\/span>idx<\/span>[<\/span>i<\/span>:<\/span>i<\/span>+<\/span>10<\/span>]]<\/span><\/span>\n        optimizer<\/span>.<\/span>zero_grad<\/span>()<\/span><\/span>\n        q <\/span>=<\/span> torch<\/span>.<\/span>distributions<\/span>.<\/span>Normal<\/span>(<\/span>loc<\/span>=<\/span>mu<\/span>,<\/span> <\/span>scale<\/span>=<\/span>sigma<\/span>)<\/span><\/span>\n        negative_log_likelihood <\/span>=<\/span> <\/span>-<\/span>1<\/span> <\/span>*<\/span> torch<\/span>.<\/span>sum<\/span>(<\/span>q<\/span>.<\/span>log_prob<\/span>(<\/span>x_batch<\/span>))<\/span><\/span>\n        negative_log_likelihood<\/span>.<\/span>backward<\/span>()<\/span><\/span>\n        optimizer<\/span>.<\/span>step<\/span>()<\/span><\/span>\n<\/span>\nprint<\/span>(<\/span>"<\/span>{}<\/span>,<\/span>{}<\/span>"<\/span>.<\/span>format<\/span>(<\/span>mu<\/span>.<\/span>detach<\/span>(),<\/span> sigma<\/span>.<\/span>detach<\/span>()))<\/span><\/span><\/code><\/pre><\/div>\n","protected":false},"excerpt":{"rendered":"
\u83b7\u5f97\u4e00\u6279\u6570\u636e\u540e\u600e\u4e48\u4f7f\u7528\u6700\u5927\u4f3c\u7136\u4f30\u8ba1\u6765\u83b7\u5f97\u9884\u4f30\u7684\u53c2\u6570\u5462\uff1f\u7528PyTorch\u6765\u5b9e\u73b0\u4e0b\uff5e<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[],"class_list":["post-665","post","type-post","status-publish","format-standard","hentry","category-machine-learning"],"_links":{"self":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/665","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=665"}],"version-history":[{"count":8,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/665\/revisions"}],"predecessor-version":[{"id":675,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/665\/revisions\/675"}],"wp:attachment":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=665"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=665"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=665"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}