{"id":3999,"date":"2024-07-05T01:03:35","date_gmt":"2024-07-04T17:03:35","guid":{"rendered":"https:\/\/17aitech.com\/?p=3999"},"modified":"2024-10-08T15:16:52","modified_gmt":"2024-10-08T07:16:52","slug":"%e3%80%90%e8%af%be%e7%a8%8b%e6%80%bb%e7%bb%93%e3%80%91day14%ef%bc%9amtcnn%e8%bf%87%e7%a8%8b%e7%9a%84%e6%b7%b1%e5%85%a5%e7%90%86%e8%a7%a3","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=3999","title":{"rendered":"\u3010\u8bfe\u7a0b\u603b\u7ed3\u3011Day14\uff1aMTCNN\u8fc7\u7a0b\u7684\u6df1\u5165\u7406\u89e3"},"content":{"rendered":"<div id=\"ez-toc-container\" class=\"ez-toc-v2_0_85 ez-toc-wrap-left-text counter-hierarchy ez-toc-counter ez-toc-light-blue ez-toc-container-direction\">\n<div class=\"ez-toc-title-container\">\n<p class=\"ez-toc-title\" style=\"cursor:inherit\">\u6587\u7ae0\u76ee\u5f55<\/p>\n<span class=\"ez-toc-title-toggle\"><a href=\"#\" class=\"ez-toc-pull-right ez-toc-btn ez-toc-btn-xs ez-toc-btn-default ez-toc-toggle\" aria-label=\"Toggle Table of Content\"><span class=\"ez-toc-js-icon-con\"><span class=\"\"><span class=\"eztoc-hide\" style=\"display:none;\">Toggle<\/span><span class=\"ez-toc-icon-toggle-span\"><svg style=\"fill: #999;color:#999\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" class=\"list-377408\" width=\"20px\" height=\"20px\" viewBox=\"0 0 24 24\" fill=\"none\"><path d=\"M6 6H4v2h2V6zm14 0H8v2h12V6zM4 11h2v2H4v-2zm16 0H8v2h12v-2zM4 16h2v2H4v-2zm16 0H8v2h12v-2z\" fill=\"currentColor\"><\/path><\/svg><svg style=\"fill: #999;color:#999\" class=\"arrow-unsorted-368013\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"10px\" height=\"10px\" viewBox=\"0 0 24 24\" version=\"1.2\" baseProfile=\"tiny\"><path d=\"M18.2 9.3l-6.2-6.3-6.2 6.3c-.2.2-.3.4-.3.7s.1.5.3.7c.2.2.4.3.7.3h11c.3 0 .5-.1.7-.3.2-.2.3-.5.3-.7s-.1-.5-.3-.7zM5.8 14.7l6.2 6.3 6.2-6.3c.2-.2.3-.5.3-.7s-.1-.5-.3-.7c-.2-.2-.4-.3-.7-.3h-11c-.3 0-.5.1-.7.3-.2.2-.3.5-.3.7s.1.5.3.7z\"\/><\/svg><\/span><\/span><\/span><\/a><\/span><\/div>\n<nav><ul class='ez-toc-list ez-toc-list-level-1 ' ><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-1\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E5%89%8D%E8%A8%80\" >\u524d\u8a00<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-2\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E9%A2%84%E5%A4%84%E7%90%86%E8%BF%87%E7%A8%8B%E5%88%86%E6%9E%90%E7%90%86%E8%A7%A3\" >\u9884\u5904\u7406\u8fc7\u7a0b\u5206\u6790\u7406\u89e3<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-3\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E6%A0%87%E6%B3%A8%E6%95%B0%E6%8D%AE%E6%96%87%E4%BB%B6\" >\u6807\u6ce8\u6570\u636e\u6587\u4ef6<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-4\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E9%A2%84%E5%A4%84%E7%90%86%E8%BF%87%E7%A8%8B\" >\u9884\u5904\u7406\u8fc7\u7a0b<\/a><ul class='ez-toc-list-level-4' ><li class='ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-5\" href=\"https:\/\/17aitech.com\/?p=3999\/#process_annotation%E5%87%BD%E6%95%B0%E8%A7%A3%E6%9E%90\" >process_annotation\u51fd\u6570\u89e3\u6790<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-6\" href=\"https:\/\/17aitech.com\/?p=3999\/#generate_crop_boxes%E5%87%BD%E6%95%B0%E8%A7%A3%E6%9E%90\" >generate_crop_boxes\u51fd\u6570\u89e3\u6790<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-7\" href=\"https:\/\/17aitech.com\/?p=3999\/#process_crop_box%E5%87%BD%E6%95%B0%E8%A7%A3%E6%9E%90\" >process_crop_box\u51fd\u6570\u89e3\u6790<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-8\" href=\"https:\/\/17aitech.com\/?p=3999\/#save_samples%E5%87%BD%E6%95%B0%E8%A7%A3%E6%9E%90\" >save_samples\u51fd\u6570\u89e3\u6790<\/a><\/li><\/ul><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-9\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E8%AE%AD%E7%BB%83%E8%BF%87%E7%A8%8B%E5%88%86%E6%9E%90%E7%90%86%E8%A7%A3\" >\u8bad\u7ec3\u8fc7\u7a0b\u5206\u6790\u7406\u89e3<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-10\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E6%89%B9%E9%87%8F%E5%8C%96%E6%89%93%E5%8C%85%E6%95%B0%E6%8D%AE\" >\u6279\u91cf\u5316\u6253\u5305\u6570\u636e<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-11\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E6%9E%84%E5%BB%BA%E6%A8%A1%E5%9E%8B\" >\u6784\u5efa\u6a21\u578b<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-12\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E7%AD%B9%E5%A4%87%E8%AE%AD%E7%BB%83\" >\u7b79\u5907\u8bad\u7ec3<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-13\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E5%AE%9A%E4%B9%89%E8%AE%AD%E7%BB%83%E8%BF%87%E7%A8%8B\" >\u5b9a\u4e49\u8bad\u7ec3\u8fc7\u7a0b<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-14\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E5%BC%80%E5%A7%8B%E8%AE%AD%E7%BB%83\" >\u5f00\u59cb\u8bad\u7ec3<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-15\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E9%A2%84%E6%B5%8B%E8%BF%87%E7%A8%8B%E5%88%86%E6%9E%90%E7%90%86%E8%A7%A3\" >\u9884\u6d4b\u8fc7\u7a0b\u5206\u6790\u7406\u89e3<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-16\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E5%88%9D%E5%A7%8B%E5%8C%96%E5%92%8C%E9%A2%84%E5%A4%84%E7%90%86\" >\u521d\u59cb\u5316\u548c\u9884\u5904\u7406<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-17\" href=\"https:\/\/17aitech.com\/?p=3999\/#P-Net%E7%BD%91%E7%BB%9C%E9%A2%84%E6%B5%8B\" >P-Net\u7f51\u7edc\u9884\u6d4b<\/a><ul class='ez-toc-list-level-4' ><li class='ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-18\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E6%9E%84%E5%BB%BA%E5%9B%BE%E5%83%8F%E9%87%91%E5%AD%97%E5%A1%94\" >\u6784\u5efa\u56fe\u50cf\u91d1\u5b57\u5854:<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-19\" href=\"https:\/\/17aitech.com\/?p=3999\/#P-Net%E5%89%8D%E5%90%91%E9%A2%84%E6%B5%8B\" >P-Net\u524d\u5411\u9884\u6d4b<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-20\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E5%8F%8D%E5%90%91%E6%B1%82%E8%A7%A3box%E6%A1%86\" >\u53cd\u5411\u6c42\u89e3box\u6846<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-21\" href=\"https:\/\/17aitech.com\/?p=3999\/#NMS%E6%B1%82%E6%9C%80%E7%BB%88%E6%A1%86\" >NMS\u6c42\u6700\u7ec8\u6846<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-22\" href=\"https:\/\/17aitech.com\/?p=3999\/#R-Net%E7%BD%91%E7%BB%9C%E9%A2%84%E6%B5%8B\" >R-Net\u7f51\u7edc\u9884\u6d4b<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-23\" href=\"https:\/\/17aitech.com\/?p=3999\/#O-Net%E7%BD%91%E7%BB%9C%E9%A2%84%E6%B5%8B\" >O-Net\u7f51\u7edc\u9884\u6d4b<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-24\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E9%81%97%E7%95%99%E5%BE%85%E6%8E%A2%E7%B4%A2%E9%97%AE%E9%A2%98\" >\u9057\u7559\u5f85\u63a2\u7d22\u95ee\u9898<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-25\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E5%86%85%E5%AE%B9%E5%B0%8F%E7%BB%93\" >\u5185\u5bb9\u5c0f\u7ed3<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-26\" href=\"https:\/\/17aitech.com\/?p=3999\/#%E5%8F%82%E8%80%83%E8%B5%84%E6%96%99\" >\u53c2\u8003\u8d44\u6599<\/a><\/li><\/ul><\/nav><\/div>\n<h2><span class=\"ez-toc-section\" id=\"%E5%89%8D%E8%A8%80\"><\/span>\u524d\u8a00<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>\u5728\u4e0a\u4e00\u7ae0<a href=\"https:\/\/17aitech.com\/?p=2421\">\u3010\u8bfe\u7a0b\u603b\u7ed3\u3011Day13\uff08\u4e0b\uff09\uff1a\u4eba\u8138\u8bc6\u522b\u548cMTCNN\u6a21\u578b<\/a>\u4e2d\uff0c\u6211\u4eec\u521d\u6b65\u4e86\u89e3\u4e86\u4eba\u8138\u8bc6\u522b\u7684\u6982\u5ff5\u4ee5\u53caMTCNN\u7684\u7f51\u7edc\u7ed3\u6784\uff0c\u501f\u52a9\u5f00\u6e90\u9879\u76ee\u7684\u4ee3\u7801\uff0c\u521d\u6b65\u5728\u672c\u5730\u5b9e\u73b0\u4e86MTCNN\u7684\u6570\u636e\u9884\u5904\u7406\u3001\u8bad\u7ec3\u548c\u9884\u6d4b\u8fc7\u7a0b\u3002\u672c\u7ae0\u5185\u5bb9\uff0c\u6211\u4eec\u5c06\u6df1\u5165MTCNN\u7684\u4ee3\u7801\uff0c\u7406\u89e3\u6570\u636e\u9884\u5904\u7406\u3001\u8bad\u7ec3\u548c\u9884\u6d4b\u8fc7\u7a0b\u3002<\/p>\n<h2><span class=\"ez-toc-section\" id=\"%E9%A2%84%E5%A4%84%E7%90%86%E8%BF%87%E7%A8%8B%E5%88%86%E6%9E%90%E7%90%86%E8%A7%A3\"><\/span>\u9884\u5904\u7406\u8fc7\u7a0b\u5206\u6790\u7406\u89e3<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<h3><span class=\"ez-toc-section\" id=\"%E6%A0%87%E6%B3%A8%E6%95%B0%E6%8D%AE%E6%96%87%E4%BB%B6\"><\/span>\u6807\u6ce8\u6570\u636e\u6587\u4ef6<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>\u9996\u5148\uff0c\u6211\u4eec\u5148\u4e86\u89e3\u4e00\u4e0bCelebA\u6570\u636e\u96c6\u7684\u6807\u6ce8\u6587\u4ef6\u5185\u5bb9\uff1a<\/p>\n<p>list_landmarks_celeba\u6587\u4ef6\u5185\u5bb9<\/p>\n<pre><code class=\"language-python\">202599\nlefteye_x lefteye_y righteye_x righteye_y nose_x nose_y leftmouth_x leftmouth_y rightmouth_x rightmouth_y\n000001.jpg 165  184  244  176  196  249  194  271  266  260\n000002.jpg 140  204  220  204  168  254  146  289  226  289\n000003.jpg 244  104  264  105  263  121  235  134  251  140<\/code><\/pre>\n<ul>\n<li>\u7b2c\u4e00\u884c\uff1a\u4ee3\u8868\u6570\u636e\u7684\u6570\u91cf\uff0c\u4f8b\u5982\uff1a202599\u6761\u6570\u636e<\/li>\n<li>\u7b2c\u4e8c\u884c\uff1a\u4ee3\u8868\u6570\u636e\u7684\u8868\u5934\u4fe1\u606f\uff0c\u4f8b\u5982\uff1alefteye_x\u662f\u5de6\u773c\u7684x\u5750\u6807\uff0clefteye_y\u662f\u53f3\u773c\u7684y\u5750\u6807\u3002<\/li>\n<li>\u7b2c\u4e09\u884c\uff1a\u4ee3\u8868\u4e00\u6761\u6807\u6ce8\u6570\u636e\uff0c\u5bf9\u5e94\u56fe\u7247\u4e2d5\u4e2a\u5173\u952e\u70b9\u7684\u5750\u6807\u4f4d\u7f6e(\u8fd9\u91cc\u7684\u5750\u6807\u4f4d\u7f6e\u5bf9\u5e94\u662f\u5728\u539f\u56fe\u4e2d\u7684\u5750\u6807\u4f4d\u7f6e)<\/li>\n<\/ul>\n<p>list_bbox_celeba\u6587\u4ef6\u5185\u5bb9<\/p>\n<pre><code class=\"language-shell\">202599\nimage_id x_1 y_1 width height\n000001.jpg    95  71 226 313\n000002.jpg    72  94 221 306\n000003.jpg   216  59  91 126\n000004.jpg   622 257 564 781<\/code><\/pre>\n<ul>\n<li>\u7b2c\u4e00\u884c\uff1a\u540c\u6837\u4ee3\u8868\u6570\u636e\u7684\u6570\u91cf<\/li>\n<li>\u7b2c\u4e8c\u884c\uff1a\u540c\u6837\u4ee3\u8868\u6570\u636e\u7684\u8868\u5934<\/li>\n<li>\u7b2c\u4e09\u884c\uff1a\u4ee3\u8868\u4e00\u6761\u4eba\u8138\u6846\u7684\u6570\u636e\uff0cx_1\u4ee3\u8868\u5de6\u4e0a\u89d2\u70b9\u7684x\u5750\u6807\uff0cy_1\u4ee3\u8868\u53f3\u4e0a\u89d2y\u5750\u6807\uff0cwidth\u662f\u6846\u7684\u5bbd\u5ea6\uff0cheight\u662f\u6846\u7684\u9ad8\u5ea6<\/li>\n<\/ul>\n<h3><span class=\"ez-toc-section\" id=\"%E9%A2%84%E5%A4%84%E7%90%86%E8%BF%87%E7%A8%8B\"><\/span>\u9884\u5904\u7406\u8fc7\u7a0b<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>generate_samples\u7684\u4e3b\u8981\u8fc7\u7a0b\u662f\uff1a<\/p>\n<ol>\n<li>\u521b\u5efa\u6837\u672c\u6570\u636e\u7684\u4fdd\u5b58\u76ee\u5f55<\/li>\n<li>\u8bfb\u53d6\u6807\u6ce8\u4fe1\u606f<\/li>\n<li>\u901a\u8fc7for\u5faa\u73af\uff0c\u4f9d\u6b21\u904d\u5386\u6bcf\u4e00\u884c\u7684\u6807\u6ce8\u4fe1\u606f<\/li>\n<li>(<strong>\u6838\u5fc3\u90e8\u5206<\/strong>)\u5904\u7406\u5355\u884c\u6807\u6ce8\u4fe1\u606f\uff0c\u751f\u6210\u6b63\u8d1f\u6837\u672c\uff0c\u5373<code>process_annotation<\/code>\u51fd\u6570<\/li>\n<li>\u4fdd\u5b58\u6b63\u8d1f\u6837\u672c\u5230\u5bf9\u5e94\u76ee\u5f55\u3002<\/li>\n<\/ol>\n<blockquote>\n<ul>\n<li>\u7531\u4e8eMTCNN\u539f\u59cb\u4ee3\u7801\u4e2d\u7684\u9884\u5904\u7406\u8fc7\u7a0b\u53ef\u8bfb\u6027\u4e0d\u9ad8\uff0c\u6240\u4ee5\u6211\u5c06generate_samples\u8fdb\u884c\u4e86\u91cd\u6784\uff0c\u91cd\u6784\u540e\u4ee3\u7801\u53ef\u8bfb\u6027\u4f1a\u66f4\u9ad8\u4e00\u4e9b\u3002<\/li>\n<li>\u91cd\u6784\u540e\u7684\u5b8c\u6574\u4ee3\u7801\u8bf7\u89c1<a href=\"https:\/\/github.com\/domonic18\/detect_face_mtcnn\">Github\u4ed3\u5e93<\/a><\/li>\n<\/ul>\n<\/blockquote>\n<pre><code class=\"language-python\">def generate_samples(face_size, max_samples=-1):\n    &quot;&quot;&quot;\n    \u751f\u6210\u6307\u5b9a\u5927\u5c0f\u7684\u4eba\u8138\u6837\u672c,\u5e76\u4fdd\u5b58\u5230\u6587\u4ef6\u4e2d\u3002\n\n    \u53c2\u6570:\n    face_size (int): \u751f\u6210\u7684\u4eba\u8138\u56fe\u50cf\u5c3a\u5bf8\n    max_samples (int): \u6700\u5927\u751f\u6210\u6837\u672c\u6570\u91cf,\u8bbe\u7f6e\u4e3a -1 \u8868\u793a\u4e0d\u9650\u5236\n    &quot;&quot;&quot;\n    if not os.path.exists(DST_PATH):\n        os.makedirs(DST_PATH)\n\n    paths, base_path = create_directories(DST_PATH, face_size)\n    # \u65b0\u5efa\u6807\u6ce8\u6587\u4ef6\n    files = open_label_files(base_path)\n\n    # \u6837\u672c\u8ba1\u6570\n    counters = {&#039;positive&#039;: 0, &#039;negative&#039;: 0, &#039;part&#039;: 0}\n\n    # \u8bfb\u53d6\u6807\u6ce8\u4fe1\u606f\n    with open(LANMARKS_PATH) as f:\n        landmarks_list = f.readlines()\n    with open(LABEL_PATH) as f:\n        anno_list = f.readlines()\n\n    for i, (anno_line, landmarks) in enumerate(zip(anno_list, landmarks_list)):\n        print(f&quot;positive:{counters[&#039;positive&#039;]}, \\\n                negative:{counters[&#039;negative&#039;]}, \\\n                part:{counters[&#039;part&#039;]}&quot;)\n\n        # \u8df3\u8fc7\u524d\u4e24\u884c\n        if i &lt; 2:\n            continue\n\n        # \u5982\u679c\u5904\u7406\u4e86\u6307\u5b9a\u6570\u91cf\u7684\u6837\u672c,\u5219\u9000\u51fa\u5faa\u73af\n        if max_samples &gt; 0 and i &gt; max_samples:\n            break\n\n        # \u5904\u7406\u5355\u884c\u6807\u6ce8\u4fe1\u606f,\u751f\u6210\u6b63\u8d1f\u6837\u672c\n        samples = process_annotation(\n            face_size, anno_line, landmarks\n        )\n\n        # \u4fdd\u5b58\u6b63\u8d1f\u6837\u672c\u5230\u6587\u4ef6\n        save_samples(\n            samples,\n            files, base_path, counters\n        )\n\n    for file in files.values():\n        file.close()<\/code><\/pre>\n<p>\u4ee3\u7801\u89e3\u6790\uff1a<\/p>\n<ul>\n<li>\u8be5\u51fd\u6570\u7684\u4e3b\u8981\u529f\u80fd\u662f\uff1a\n<ul>\n<li>\u8bfb\u53d6\u6807\u6ce8\u6587\u4ef6list_landmarks_celeba\u548clist_bbox_celeba<\/li>\n<li>\u8c03\u7528<code>process_annotation()<\/code>\u51fd\u6570\u5904\u7406\u6807\u6ce8\u4fe1\u606f\uff0c\u751f\u6210\u6b63\u8d1f\u6837\u672c<\/li>\n<li>\u8c03\u7528<code>save_samples()<\/code>\u51fd\u6570\u4fdd\u5b58\u6b63\u8d1f\u6837\u672c\u5230\u6587\u4ef6<\/li>\n<\/ul>\n<\/li>\n<li>\u5176\u4e2d\u6838\u5fc3\u903b\u8f91\u662f<code>process_annotation()<\/code>\u51fd\u6570\uff0c\u63a5\u4e0b\u6765\u6211\u4eec\u4eceprocess_annotation\u51fd\u6570\u5165\u624b\uff0c\u68b3\u7406\u5176\u4e3b\u8981\u903b\u8f91\u3002<\/li>\n<\/ul>\n<h4><span class=\"ez-toc-section\" id=\"process_annotation%E5%87%BD%E6%95%B0%E8%A7%A3%E6%9E%90\"><\/span>process_annotation\u51fd\u6570\u89e3\u6790<span class=\"ez-toc-section-end\"><\/span><\/h4>\n<pre><code class=\"language-python\">\ndef process_annotation(face_size, anno_line, landmarks):\n    &quot;&quot;&quot;\n    \u5904\u7406\u5355\u884c\u6ce8\u91ca\u4fe1\u606f,\u751f\u6210\u6b63\u8d1f\u6837\u672c\u3002\n\n    \u53c2\u6570:\n    anno_line (str): \u4e00\u884c\u6ce8\u91ca\u4fe1\u606f,\u683c\u5f0f\u4e3a &quot;image_filename x1 y1 w h&quot;\n    face_size (int): \u751f\u6210\u7684\u4eba\u8138\u56fe\u50cf\u5c3a\u5bf8\n    landmarks (str): \u5173\u952e\u70b9\u6807\u6ce8\u5b57\u7b26\u4e32\n\n    \u8fd4\u56de:\n    samples (list): \u751f\u6210\u7684\u6837\u672c\u5217\u8868\n    &quot;&quot;&quot;\n    # 5\u4e2a\u5173\u952e\u70b9\n    _landmarks = landmarks.split()\n\n    # \u4f7f\u7528\u5217\u8868\u89e3\u6790\u548c\u89e3\u5305\u4e00\u6b21\u6027\u83b7\u53d6\u6240\u6709\u5173\u952e\u70b9\u7684\u5750\u6807\n    landmarks = [float(x) for x in _landmarks[1:11]]\n\n    # \u89e3\u6790\u6ce8\u91ca\u884c,\u83b7\u53d6\u56fe\u50cf\u6587\u4ef6\u540d\u548c\u4eba\u8138\u4f4d\u7f6e\u4fe1\u606f\n    strs = parse_annotation_line(anno_line)\n    image_filename = strs[0].strip()\n    x1, y1, w, h = map(int, strs[1:])\n\n    # \u6807\u7b7e\u77eb\u6b63\n    x1, y1, x2, y2, w, h = adjust_bbox(x1, y1, w, h)\n    # \u5224\u65ad\u5750\u6807\u662f\u5426\u7b26\u5408\u8981\u6c42\n    if max(w, h) &lt; 40 or x1 &lt; 0 or x2 &lt; 0 or y1 &lt; 0 or y2 &lt; 0:\n        # \u4e0d\u7b26\u5408\u8981\u6c42\u7684\u56fe\u7247\uff0c\u8fd4\u56de[]\u4e0d\u505a\u5904\u7406\n        return []\n\n    boxes = [[x1, y1, x2, y2]]\n\n    # \u8ba1\u7b97\u4eba\u8138\u4e2d\u5fc3\u70b9\u5750\u6807\n    cx = w \/ 2 + x1\n    cy = h \/ 2 + y1\n\n    # \u6700\u5927\u8fb9\u957f\n    max_side = max(w, h)\n\n    # \u6253\u5f00\u56fe\u50cf\u6587\u4ef6\n    image_filepath = os.path.join(IMG_PATH, image_filename)\n    with Image.open(image_filepath) as img:\n        # \u89e3\u6790\u51fa\u5bbd\u5ea6\u548c\u9ad8\u5ea6\n        img_w, img_h = img.size\n        # \u751f\u6210\u5019\u9009\u7684\u88c1\u526a\u6846\n        samples = []\n        for crop_box in generate_crop_boxes(cx, cy, max_side, img_w, img_h):\n            # \u5904\u7406\u6bcf\u4e2a\u5019\u9009\u88c1\u526a\u6846,\u751f\u6210\u6b63\u8d1f\u6837\u672c\n            sample = process_crop_box(img, face_size, max_side, crop_box, boxes, landmarks )\n            if sample:\n                samples.append(sample)\n\n    return samples<\/code><\/pre>\n<p>\u4ee3\u7801\u89e3\u6790\uff1a<\/p>\n<ul>\n<li>process_annotation\u7684\u4e3b\u8981\u529f\u80fd\u662f\uff1a\n<ul>\n<li>\u8ba1\u7b97\u4eba\u8138\u4e2d\u5fc3\u70b9\u5750\u6807\u548c\u6700\u5927\u8fb9\u957f<\/li>\n<li>\u8c03\u7528<code>generate_crop_boxes()<\/code>\u751f\u6210\u5019\u9009\u88c1\u526a\u6846<\/li>\n<\/ul>\n<\/li>\n<li>\u6838\u5fc3\u903b\u8f91\u5728<code>generate_crop_boxes()<\/code>\u751f\u6210\u5019\u9009\u88c1\u526a\u6846\u548c<code>process_crop_box()<\/code>\u5904\u7406\u5019\u9009\u88c1\u526a\u6846\u4e24\u4e2a\u51fd\u6570\uff0c\u6211\u4eec\u4f9d\u6b21\u5206\u6790\u8fd9\u4e24\u4e2a\u51fd\u6570\u3002<\/li>\n<\/ul>\n<h4><span class=\"ez-toc-section\" id=\"generate_crop_boxes%E5%87%BD%E6%95%B0%E8%A7%A3%E6%9E%90\"><\/span>generate_crop_boxes\u51fd\u6570\u89e3\u6790<span class=\"ez-toc-section-end\"><\/span><\/h4>\n<pre><code class=\"language-python\">def generate_crop_boxes(cx, cy, max_side, img_w, img_h):\n    &quot;&quot;&quot;\n    \u6839\u636e\u7ed9\u5b9a\u7684\u4eba\u8138\u4e2d\u5fc3\u70b9\u5750\u6807\u548c\u5c3a\u5bf8,\u751f\u62105\u4e2a\u5019\u9009\u7684\u88c1\u526a\u6846\u3002\n\n    \u53c2\u6570:\n    cx (float): \u4eba\u8138\u4e2d\u5fc3\u70b9\u7684 x \u5750\u6807\n    cy (float): \u4eba\u8138\u4e2d\u5fc3\u70b9\u7684 y \u5750\u6807\n    max_side (int): \u4eba\u8138\u6846\u7684\u6700\u5927\u8fb9\u957f\n    img_w (int): \u56fe\u50cf\u5bbd\u5ea6\n    img_h (int): \u56fe\u50cf\u9ad8\u5ea6\n\n    \u8fd4\u56de:\n    crop_boxes (list): \u4e00\u4e2a\u5305\u542b5\u4e2a\u88c1\u526a\u6846\u5750\u6807\u7684\u5217\u8868,\u6bcf\u4e2a\u88c1\u526a\u6846\u7684\u683c\u5f0f\u4e3a [x1, y1, x2, y2]\n    max_sides (list): \u4e00\u4e2a\u5305\u542b5\u4e2a\u88c1\u526a\u6846\u6700\u5927\u8fb9\u957f\u7684\u5217\u8868\n    &quot;&quot;&quot;\n\n    crop_boxes = []\n    max_sides = []\n    for _ in range(5):\n        # \u968f\u673a\u504f\u79fb\u4e2d\u5fc3\u70b9\u5750\u6807\u4ee5\u53ca\u8fb9\u957f\n        seed = float_num[np.random.randint(0, len(float_num))]\n\n        # \u6700\u5927\u8fb9\u957f\u968f\u673a\u504f\u79fb\n        _max_side = max_side + np.random.randint(int(-max_side * seed), int(max_side * seed))\n\n        # \u4e2d\u5fc3\u70b9x\u5750\u6807\u968f\u673a\u504f\u79fb\n        _cx = cx + np.random.randint(int(-cx * seed), int(cx * seed))\n\n        # \u4e2d\u5fc3\u70b9y\u5750\u6807\u968f\u673a\u504f\u79fb\n        _cy = cy + np.random.randint(int(-cy * seed), int(cy * seed))\n\n        # \u5f97\u5230\u504f\u79fb\u540e\u7684\u5750\u6807\u503c\uff08\u65b9\u6846\uff09\n        _x1 = _cx - _max_side \/ 2\n        _y1 = _cy - _max_side \/ 2\n        _x2 = _x1 + _max_side\n        _y2 = _y1 + _max_side\n\n        # \u504f\u79fb\u8fc7\u5927\uff0c\u504f\u51fa\u56fe\u50cf\u4e86\uff0c\u6b64\u65f6\uff0c\u4e0d\u80fd\u7528\uff0c\u5e94\u8be5\u518d\u6b21\u5c1d\u8bd5\u504f\u79fb\n        if _x1 &lt; 0 or _y1 &lt; 0 or _x2 &gt; img_w or _y2 &gt; img_h:\n            continue\n\n        # \u6dfb\u52a0\u88c1\u526a\u6846\u5750\u6807\u548c\u6700\u5927\u8fb9\u957f\u5230\u5217\u8868\u4e2d\n        crop_boxes.append(np.array([_x1, _y1, _x2, _y2]))\n        max_sides.append(_max_side)\n\n    return crop_boxes, max_sides<\/code><\/pre>\n<p>\u4e3a\u4e86\u4fbf\u4e8e\u7406\u89e3\uff0c\u6211\u4eec\u901a\u8fc7\u4e00\u6bb5\u6d4b\u8bd5\u4ee3\u7801\uff0c\u5c06\u4e0a\u8ff0\u968f\u673a\u751f\u6210\u7684\u88c1\u526a\u6846\u753b\u51fa\u6765\uff0c\u4ee5\u4fbf\u66f4\u52a0\u5f62\u8c61\u5730\u770b\u5230\u88c1\u526a\u6846\u3002<br \/>\n<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u968f\u673a\u751f\u6210\u7684\u88c1\u526a\u6846.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u968f\u673a\u751f\u6210\u7684\u88c1\u526a\u6846.png\" alt=\"\" \/><\/a><\/p>\n<ul>\n<li>\u7ea2\u8272\u6846\uff1a\u4ee3\u8868\u6807\u6ce8\u6570\u636e\u5bf9\u5e94\u7684\u4eba\u8138\u6846<\/li>\n<li>\u84dd\u8272\u6846\uff1a\u4ee3\u8868\u968f\u673a\u751f\u6210\u7684\u88c1\u526a\u6846<\/li>\n<\/ul>\n<blockquote>\n<p>\u7bc7\u5e45\u539f\u56e0\uff0c\u6d4b\u8bd5\u4ee3\u7801\u4e0d\u518d\u8d58\u8ff0\uff0c\u76f8\u5173\u4ee3\u7801\u53ef\u4ee5\u5728github\u4ee3\u7801\u4ed3\u5e93\u4e0b\u7684\\doc\\CelebA.ipynb\u627e\u5230\u3002<\/p>\n<\/blockquote>\n<p>\u4ee3\u7801\u89e3\u6790\uff1a<\/p>\n<ul>\n<li>\n<p><code>seed<\/code>\u662f\u5728\u5df2\u5b9a\u4e49\u7684<code>float_num<\/code>\u4e2d\u8fdb\u884c\u968f\u673a\u53d6\u6570\uff0c\u540e\u7eed\u7528\u4e8e\u968f\u673a\u504f\u79fb\u7684\u7cfb\u6570\u3002\u67e5\u770bfloat_num\u4e2d\u6570\u503c\u7684\u5206\u5e03\uff0c\u6b63\uff1a\u504f\uff1a\u8d1f=2\uff1a2\uff1a6=1\uff1a1\uff1a3\uff0c\u8fd9\u53ef\u4ee5\u4f7f\u5f97\u8d1f\u6837\u672c\u7684\u6bd4\u4f8b\u66f4\u591a\u4e00\u4e9b\u3002<\/p>\n<pre><code class=\"language-python\">float_num = [0.1, 0.1, 0.3, 0.5, 0.95, 0.95, 0.99, 0.99, 0.99, 0.99]<\/code><\/pre>\n<\/li>\n<li>\n<p><code>_max_side<\/code>\u662f\u4eba\u8138\u6846(\u7ea2\u6846)\u6700\u5927\u8fb9\u957f\u7684\u968f\u673a\u504f\u79fb\uff0c\u8ba1\u7b97\u65b9\u6cd5\u662f<code>np.random.randint(int(-max_side * seed), int(max_side * seed)) <\/code>\u751f\u6210\u4e00\u4e2a\u968f\u673a\u6574\u6570,\u8303\u56f4\u5728 [-max_side <em> seed, max_side <\/em> seed] \u4e4b\u95f4<\/p>\n<blockquote>\n<p>\u4f8b\u5982,\u5982\u679c\u539f\u59cb\u4eba\u8138\u6846\u7684\u6700\u5927\u8fb9\u957f\u662f 100 \u50cf\u7d20,\u800c seed \u53d6\u503c\u4e3a 0.2,\u90a3\u4e48 _max_side \u7684\u53d6\u503c\u8303\u56f4\u5c31\u4f1a\u5728 100 + (-100 <em> 0.2) ~ 100 + (100 <\/em> 0.2) \u4e4b\u95f4,\u4e5f\u5c31\u662f 80 ~ 120 \u50cf\u7d20\u4e4b\u95f4\u3002<\/p>\n<\/blockquote>\n<\/li>\n<li>\n<p><code>_cx<\/code>\u548c<code>_cy<\/code>\u662f\u6839\u636e\u4eba\u8138\u6846\u4e2d\u5fc3\u70b9<code>cx<\/code>\u548c<code>cy<\/code>\u968f\u673a\u504f\u79fb\u5f97\u5230\u65b0\u4e2d\u5fc3\u70b9<\/p>\n<\/li>\n<\/ul>\n<p>\u5c06\u4e0a\u8ff0\u7684\u53d8\u91cf\u753b\u56fe\u7406\u89e3\u5982\u4e0b\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u8ba1\u7b97\u504f\u5dee.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u8ba1\u7b97\u504f\u5dee.png\" alt=\"\" \/><\/a><\/p>\n<h4><span class=\"ez-toc-section\" id=\"process_crop_box%E5%87%BD%E6%95%B0%E8%A7%A3%E6%9E%90\"><\/span>process_crop_box\u51fd\u6570\u89e3\u6790<span class=\"ez-toc-section-end\"><\/span><\/h4>\n<pre><code class=\"language-python\">def process_crop_box(img, face_size, _max_side, crop_box, boxes, landmarks):\n    &quot;&quot;&quot;\n    \u5904\u7406\u5355\u4e2a\u88c1\u526a\u6846,\u751f\u6210\u6b63\u8d1f\u6837\u672c\u3002\n\n    \u53c2\u6570:\n    img (Image): \u539f\u59cb\u56fe\u50cf\n    crop_box (list): \u88c1\u526a\u6846\u5750\u6807 [x1, y1, x2, y2]\n    boxes (list): \u4eba\u8138\u6846\u5750\u6807\u5217\u8868\n    face_size (int): \u751f\u6210\u7684\u4eba\u8138\u56fe\u50cf\u5c3a\u5bf8\n\n    \u8fd4\u56de:\n    sample (dict): \u6837\u672c\u4fe1\u606f {&#039;image&#039;: image, &#039;label&#039;: label, &#039;bbox_offsets&#039;: offsets, &#039;landmark_offsets&#039;: landmark_offsets}\n    &quot;&quot;&quot;\n    x1, y1, x2, y2 = boxes[0][:4]\n    _x1, _y1, _x2, _y2 = crop_box[:4]\n    px1, py1, px2, py2, px3, py3, px4, py4, px5, py5 = landmarks\n\n    offset_x1 = (x1 - _x1) \/ _max_side\n    offset_y1 = (y1 - _y1) \/ _max_side\n    offset_x2 = (x2 - _x2) \/ _max_side\n    offset_y2 = (y2 - _y2) \/ _max_side\n\n    offset_px1 = (px1 - _x1) \/ _max_side\n    offset_py1 = (py1 - _y1) \/ _max_side\n    offset_px2 = (px2 - _x1) \/ _max_side\n    offset_py2 = (py2 - _y1) \/ _max_side\n    offset_px3 = (px3 - _x1) \/ _max_side\n    offset_py3 = (py3 - _y1) \/ _max_side\n    offset_px4 = (px4 - _x1) \/ _max_side\n    offset_py4 = (py4 - _y1) \/ _max_side\n    offset_px5 = (px5 - _x1) \/ _max_side\n    offset_py5 = (py5 - _y1) \/ _max_side\n\n    face_crop = img.crop(crop_box)\n    face_resize = face_crop.resize((face_size, face_size), Image.Resampling.LANCZOS)\n\n    iou = IOU(torch.tensor([x1, y1, x2, y2]), torch.tensor([crop_box[:4]]))\n\n    if iou &gt; 0.7:  # \u6b63\u6837\u672c\n        label = 1\n    elif 0.4 &lt; iou &lt; 0.6:  # \u90e8\u5206\u6837\u672c\n        label = 2\n    elif iou &lt; 0.2:  # \u8d1f\u6837\u672c\n        label = 0\n    else:\n        return None  # \u4e0d\u7b26\u5408\u4efb\u4f55\u6761\u4ef6\u7684\u6837\u672c\u4e0d\u5904\u7406\n\n    return {\n        &#039;image&#039;: face_resize,\n        &#039;label&#039;: label,\n        &#039;bbox_offsets&#039;: (offset_x1, offset_y1, offset_x2, offset_y2),\n        &#039;landmark_offsets&#039;: (offset_px1, offset_py1, offset_px2, offset_py2, offset_px3, offset_py3, offset_px4, offset_py4, offset_px5, offset_py5)\n    }<\/code><\/pre>\n<p>\u4ee3\u7801\u89e3\u6790\uff1a<\/p>\n<ul>\n<li>\u7b2c\u4e00\u6b65\uff0c\u63a5\u6536\u4f20\u5165\u7684\u53c2\u6570boxes\u3001crop_box\u4ee5\u53calandmarks\uff0c\u5206\u522b\u89e3\u6790\u539f\u56fe\u4eba\u6846\u3001\u4e00\u4e2a\u968f\u673a\u88c1\u526a\u6846\u7684\u5750\u6807\u548c\u5173\u952e\u70b9\u5750\u6807<\/li>\n<li>(<strong>\u7cbe\u9ad3<\/strong>)\u7b2c\u4e8c\u6b65\uff0c\u6839\u636e\u539f\u56fe\u4eba\u6846(\u7ea2\u8272) &#8211; \u968f\u673a\u88c1\u526a\u6846(\u84dd\u56fe) \/ \u968f\u673a\u88c1\u526a\u6846\u7684\u6700\u5927\u8fb9\u957f\uff0c\u6c42\u5f97\u504f\u79fb\u7387<br \/>\n<blockquote>\n<p>\u6653\u534e\u8001\u5e08\u63d0\u5230\u8fd9\u6bb5\u4ee3\u7801\u975e\u5e38\u7cbe\u5999\uff0c\u4ed4\u7ec6\u7406\u89e3\uff0c\u5176\u7cbe\u5999\u4e4b\u5904\u5728\u4e8e\uff1a<br \/>\n<strong>\u7b2c\u4e00\u70b9<\/strong>\uff1a\u56e0\u4e3aP\u7f51\u7edc\u3001R\u7f51\u7edc\u3001O\u7f51\u7edc\u8981\u5904\u7406\u7684\u662f\u6b63\u65b9\u5f62\u65b9\u56fe\uff0c\u6240\u4ee5\u5728<code>generate_crop_boxes()<\/code>\u4e2d\u4f7f\u7528max_side\u7684\u505a\u6cd5\uff0c\u53ef\u4ee5\u5c06\u56fe\u7247\u8f6c\u4e3a\u6b63\u65b9\u5f62\uff0c\u65b9\u4fbf\u540e\u7eed\u7684\u8bad\u7ec3\u800c\u4e0d\u7528\u518d\u8fdb\u884c\u88c1\u526a\u64cd\u4f5c\uff1b<br \/>\n<strong>\u7b2c\u4e8c\u70b9<\/strong>\uff1a\u56e0\u4e3a\u673a\u5668\u5b66\u4e60\u8981\u8fdb\u884c\u6570\u636e<a href=\"https:\/\/17aitech.com\/?p=2006#toc-18\">\u5f52\u4e00\u5316<\/a>\u5904\u7406\uff0c\u800c<code>offset_x1 = (x1 - _x1) \/ _max_side<\/code>\u8ba1\u7b97\u7684\u504f\u79fb\u7387\u521a\u597d\u662f\u4ee50\u4e3a\u4e2d\u5fc3\uff0c\u6ee1\u8db3\u4e86\u5f52\u4e00\u5316\u7684\u8981\u6c42\uff1b<br \/>\n<strong>\u7b2c\u4e09\u70b9<\/strong>\uff1a\u56e0\u4e3a<a href=\"https:\/\/17aitech.com\/?p=2006#toc-21\">\u4fe1\u606f\u8574\u542b\u5728\u6570\u636e\u7684\u76f8\u5bf9\u5927\u5c0f<\/a>\u7684\u539f\u5219\uff0c\u6240\u4ee5\u5728\u6570\u636e\u5b58\u50a8\u8bb0\u5f55\u65f6\uff0c\u53ea\u8981\u76f8\u5bf9\u5927\u5c0f\u4e0d\u4e22\uff0c\u4fe1\u606f\u662f\u4e0d\u4f1a\u7f3a\u5931\u7684\u3002\u90a3\u4e48\u901a\u8fc7\u4ee5\u4e0a\u505a\u6cd5\uff0c\u6211\u4eec\u539f\u672c\u8981\u8bb0\u5f55\u7684\u4fe1\u606f\u4e2a\u6570\u505a\u4e86\u4f18\u5316\u51cf\u5c11\uff0c\u4ece<code>4\u4e2a(\u539f\u56fe\u5750\u6807)<\/code>+<code>4\u4e2a(\u88c1\u526a\u56fe\u5750\u6807)<\/code>+<code>10\u4e2a(5\u4e2a\u5173\u952e\u70b9\u5750\u6807)<\/code>\u53d8\u4e3a<code>4\u4e2a(\u88c1\u526a\u56fe\u504f\u79fb\u91cf)<\/code>+<code>1\u4e2a(\u6700\u5927\u8fb9\u957f)<\/code>+<code>10\u4e2a(\u5173\u952e\u70b9\u504f\u79fb\u91cf)<\/code>,\u8fd9\u4f1a\u964d\u4f4e\u6570\u636e\u7684\u5b58\u50a8\u4ee3\u4ef7\u548c\u8ba1\u7b97\u4ee3\u4ef7\uff0c\u540c\u65f6\u53c8\u53ef\u4ee5\u901a\u8fc7\u53cd\u89e3\u968f\u65f6\u6c42\u5f97\u524d\u9762\u7684\u5750\u6807\u503c\u3002<br \/>\n<strong>\u7b2c\u56db\u70b9<\/strong>\uff1a\u56e0\u4e3a<code>generate_crop_boxes()<\/code>\u662f\u901a\u8fc7\u968f\u673a\u751f\u6210\u4e0d\u540c\u7684\u88c1\u526a\u6846\uff0c\u968f\u673a\u751f\u6210\u8fc7\u7a0b\u4e2d\u53ef\u80fd\u751f\u6210\u66f4\u597d\u5957\u4f4f\u8138\u7684\u6846\uff0c\u8fd9\u4e00\u8fc7\u7a0b\u672c\u8eab\u5c31\u662fanchor-free\u968f\u673a\u751f\u957f\u7684\u601d\u60f3\u4f53\u73b0(\u5982\u4e0b\u56fe\u5de6\u4fa7);\u76f8\u6bd4\u8f83\u5982\u679c\u8bad\u7ec3\u6570\u636e\u4e0d\u505a\u968f\u673a\u751f\u6210\u88c1\u526a\u6846(\u5982\u4e0b\u56fe\u53f3\u4fa7)\uff0c\u76f4\u63a5\u8bad\u7ec3\u7684\u6807\u6ce8\u6846\u7684\u5750\u6807\u4f4d\u7f6e\uff0c\u90a3\u4e48\u673a\u5668\u5728\u9884\u6d4b\u65f6\u5c31\u53ea\u6709\u5728\u8138\u6846\u65f6\u4e0e\u8bad\u7ec3\u7684\u76f8\u5951\u5408\u65f6\u624d\u80fd\u8bc6\u522b\u662f\u8138\u90e8\uff0c\u8fd9\u5e94\u8be5\u662fanchor-base\u7684\u601d\u60f3\u3002<br \/>\n<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u968f\u673a\u751f\u6210\u4e0e\u53ea\u6709\u6807\u6ce8\u6846.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u968f\u673a\u751f\u6210\u4e0e\u53ea\u6709\u6807\u6ce8\u6846.png\" alt=\"\" \/><\/a><\/p>\n<\/blockquote>\n<\/li>\n<li>\u7b2c\u4e09\u6b65\uff0c\u8ba1\u7b97\u539f\u56fe\u4e0e\u88c1\u526a\u56fe\u7684iou\uff0c\u5982\u679c&gt;0.7\u5219\u4e3a\u6b63\u6837\u672c\uff1b\u4ecb\u4e8e0.4~0.6\u4e4b\u95f4\u662f\u504f\u6837\u672c\uff1b\u5982\u679c&lt;0.2\u5219\u4e3a\u8d1f\u6837\u672c<\/li>\n<\/ul>\n<h4><span class=\"ez-toc-section\" id=\"save_samples%E5%87%BD%E6%95%B0%E8%A7%A3%E6%9E%90\"><\/span>save_samples\u51fd\u6570\u89e3\u6790<span class=\"ez-toc-section-end\"><\/span><\/h4>\n<p>\u6700\u540e\u5c06\u751f\u6210\u7684\u6837\u672c\uff0c\u4fdd\u5b58\u5230\u5bf9\u5e94\u7684\u6587\u4ef6\u4e2d\u3002<\/p>\n<pre><code class=\"language-python\">def save_samples(samples, files, base_path, counters):\n    &quot;&quot;&quot;\n    \u4fdd\u5b58\u6b63\u8d1f\u6837\u672c\u5230\u6587\u4ef6\u4e2d\u3002\n\n    \u53c2\u6570:\n    samples (list): \u6837\u672c\u5217\u8868, \u6bcf\u4e2a\u5143\u7d20\u4e3a\u4e00\u4e2a\u5b57\u5178, \u5305\u542b &#039;image&#039;, &#039;label&#039;, &#039;bbox_offsets&#039;, &#039;landmark_offsets&#039;\n    files (dict): \u5305\u542b\u6b63\u8d1f\u6837\u672c\u8f93\u51fa\u6587\u4ef6\u7684\u5b57\u5178\n    base_path (str): \u8f93\u51fa\u6587\u4ef6\u7684\u57fa\u7840\u8def\u5f84\n    counters (dict): \u6837\u672c\u8ba1\u6570\u5668\u5b57\u5178\n    &quot;&quot;&quot;\n    for sample in samples:\n        image = sample[&#039;image&#039;]\n        label = sample[&#039;label&#039;]\n        bbox_offsets = sample[&#039;bbox_offsets&#039;]\n        landmark_offsets = sample[&#039;landmark_offsets&#039;]\n\n        if label == 1:\n            category = &#039;positive&#039;\n            counters[&#039;positive&#039;] += 1\n        elif label == 2:\n            category = &#039;part&#039;\n            counters[&#039;part&#039;] += 1\n        else:\n            category = &#039;negative&#039;\n            counters[&#039;negative&#039;] += 1\n\n        filename = f&quot;{category}\/{counters[category]}.jpg&quot;\n        image.save(os.path.join(base_path, filename))\n\n        try:\n            bbox_str = &#039; &#039;.join(map(str, bbox_offsets))\n            landmark_str = &#039; &#039;.join(map(str, landmark_offsets))\n            files[category].write(f&quot;{filename} {label} {bbox_str} {landmark_str}\\n&quot;)\n        except IOError as e:\n            print(f&quot;Error writing to file: {e}&quot;)\n<\/code><\/pre>\n<p>\u4fdd\u5b58\u540e\u7684\u6570\u636e\u5185\u5bb9\u683c\u5f0f\u4e3a\uff1a<\/p>\n<pre><code class=\"language-shell\">negative\/1.jpg 0 -2.1306818181818183 -2.909090909090909 -0.8238636363636364 -0.8863636363636364 -1.6420454545454546 -1.9772727272727273 -0.7443181818181818 -2.0681818181818183 -1.2897727272727273 -1.2386363636363635 -1.3125 -0.9886363636363636 -0.4943181818181818 -1.1136363636363635\nnegative\/2.jpg 0 -0.14335664335664336 -1.9125874125874125 0.24125874125874125 -1.0944055944055944 0.15034965034965034 -1.3531468531468531 0.7097902097902098 -1.3531468531468531 0.34615384615384615 -1.0034965034965035 0.19230769230769232 -0.7587412587412588 0.7517482517482518 -0.7587412587412588<\/code><\/pre>\n<h2><span class=\"ez-toc-section\" id=\"%E8%AE%AD%E7%BB%83%E8%BF%87%E7%A8%8B%E5%88%86%E6%9E%90%E7%90%86%E8%A7%A3\"><\/span>\u8bad\u7ec3\u8fc7\u7a0b\u5206\u6790\u7406\u89e3<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>\u8bad\u7ec3\u8fc7\u7a0b\u53c2\u7167<a href=\"https:\/\/17aitech.com\/?p=2178\">\u6df1\u5ea6\u5b66\u4e60\u7684\u57fa\u672c\u6d41\u7a0b<\/a>\uff0c\u5206\u8981\u8fdb\u884c\u6279\u91cf\u5316\u6253\u5305\u6570\u636e\u3001\u6784\u5efa\u6a21\u578b\u3001\u5b9a\u4e49\u635f\u5931\u51fd\u6570\u3001\u5b9a\u4e49\u8bad\u7ec3\u8fc7\u7a0b\u3002<\/p>\n<h3><span class=\"ez-toc-section\" id=\"%E6%89%B9%E9%87%8F%E5%8C%96%E6%89%93%E5%8C%85%E6%95%B0%E6%8D%AE\"><\/span>\u6279\u91cf\u5316\u6253\u5305\u6570\u636e<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<pre><code class=\"language-python\"># \u6784\u5efa\u81ea\u5b9a\u4e49\u7684\u8138\u90e8\u8bc6\u522b\u6570\u636e\u96c6\n# \u6570\u636e\u96c6\u4f7f\u7528CelebA\u6570\u636e\u96c6\u7684\u56fe\u7247\u548c\u6807\u7b7e\nimport torch\nimport os\nimport numpy as np\nfrom torch.utils.data import DataLoader, Dataset \nfrom PIL import Image\n\nclass FaceDataset(Dataset):\n    def __init__(self, path):\n        super().__init__()\n        self.path = path\n        self.datasets = []\n        self._read_annos()\n\n    def _read_annos(self):\n        with open(os.path.join(self.path, &quot;positive.txt&quot;)) as f:\n            self.datasets.extend(f.readlines())\n\n        with open(os.path.join(self.path, &quot;negative.txt&quot;)) as f:\n            self.datasets.extend(f.readlines())\n\n        with open(os.path.join(self.path, &quot;part.txt&quot;)) as f:\n            self.datasets.extend(f.readlines())\n\n    def __len__(self):\n        return len(self.datasets)\n\n    def __getitem__(self, idx):\n        strs = self.datasets[idx].strip().split()\n        # \u6587\u4ef6\u540d\u5b57\n        img_name = strs[0]\n\n        # \u53d6\u51fa\u7c7b\u522b\n        cls = torch.tensor([int(strs[1])], dtype=torch.float32)\n\n        # \u5c06\u6240\u6709\u504f\u7f6e\u8f6c\u4e3afloat\u7c7b\u578b\n        strs[2:] = [float(x) for x in strs[2:]]\n\n        # bbox\u7684\u504f\u7f6e\n        offset = torch.tensor(strs[2:6], dtype=torch.float32)\n\n        # landmark\u7684\u504f\u7f6e\n        point = torch.tensor(strs[6:16], dtype=torch.float32)\n\n        # \u6253\u5f00\u56fe\u50cf\n        img = Image.open(os.path.join(self.path, img_name))\n\n        # \u6570\u636e\u8c03\u6574\u5230 [-1, 1]\u4e4b\u95f4\n        img_data = torch.tensor((np.array(img) \/ 255. - 0.5) \/ 0.5, dtype=torch.float32)\n        # [H, W, C] --&gt; [C, H ,W]\n        img_data = img_data.permute(2, 0, 1)\n\n        return img_data, cls, offset, point<\/code><\/pre>\n<p>\u4ee3\u7801\u89e3\u6790\uff1a<\/p>\n<ul>\n<li>\u4ee5\u4e0a\u4ee3\u7801\u4f7f\u7528\u7684\u662f\u6807\u51c6\u7684\u81ea\u5b9a\u4e49\u6570\u636e\u96c6\u65b9\u5f0f\uff0c\u5373\uff1a\u58f0\u660e\u4e00\u4e2a\u7c7b\u7ee7\u627fDataset\u7c7b\uff0c\u540c\u65f6\u5b9e\u73b0\u56de\u8c03\u51fd\u6570<strong>len<\/strong>()\u548c<strong>getitem<\/strong>()\u5373\u53ef\u3002<\/li>\n<\/ul>\n<h3><span class=\"ez-toc-section\" id=\"%E6%9E%84%E5%BB%BA%E6%A8%A1%E5%9E%8B\"><\/span>\u6784\u5efa\u6a21\u578b<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>\u8fd9\u90e8\u5206\u5185\u5bb9\u5df2\u5728<a href=\"https:\/\/17aitech.com\/?p=2421\">\u3010\u8bfe\u7a0b\u603b\u7ed3\u3011Day13\uff08\u4e0b\uff09\uff1a\u4eba\u8138\u8bc6\u522b\u548cMTCNN\u6a21\u578b<\/a>\u5b9a\u4e49\uff0c\u672c\u6587\u4e0d\u518d\u8d58\u8ff0\u3002<\/p>\n<blockquote>\n<p>\u9057\u7559\u95ee\u9898\uff1a\u6709\u4e9b\u6587\u7ae0\u4e2d\u90e8\u5206MTCNN\u4f7f\u7528\u7684\u6fc0\u6d3b\u51fd\u6570\u662fnn.PReLU\uff0c\u5f85\u4e86\u89e3\u4e0eReLU\u7684\u533a\u522b\u3002<\/p>\n<\/blockquote>\n<h3><span class=\"ez-toc-section\" id=\"%E7%AD%B9%E5%A4%87%E8%AE%AD%E7%BB%83\"><\/span>\u7b79\u5907\u8bad\u7ec3<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<pre><code class=\"language-python\">class Trainer:\n    def __init__(self, net, param_path, data_path):\n        # \u68c0\u6d4b\u662f\u5426\u6709GPU\n        self.device = &#039;cuda:0&#039; if torch.cuda.is_available() else &quot;cpu&quot;\n        # \u628a\u6a21\u578b\u642c\u5230device\n        self.net = net.to(self.device)\n\n        self.param_path = param_path\n\n        # \u6253\u5305\u6570\u636e\n        self.datasets = FaceDataset(data_path)\n\n        # \u5b9a\u4e49\u635f\u5931\u51fd\u6570\uff1a\u7c7b\u522b\u5224\u65ad\uff08\u5206\u7c7b\u4efb\u52a1\uff09\n        self.cls_loss_func = torch.nn.BCELoss()\n\n        # \u5b9a\u4e49\u635f\u5931\u51fd\u6570\uff1a\u6846\u7684\u504f\u7f6e\u56de\u5f52\n        self.offset_loss_func = torch.nn.MSELoss()\n\n        # \u5b9a\u4e49\u635f\u5931\u51fd\u6570\uff1a\u5173\u952e\u70b9\u7684\u504f\u7f6e\u56de\u5f52\n        self.point_loss_func = torch.nn.MSELoss()\n\n        # \u5b9a\u4e49\u4f18\u5316\u5668\n        self.optimizer = torch.optim.Adam(params=self.net.parameters(), lr=1e-3)\n\n    def compute_loss(self, out_cls, out_offset, out_point, cls, offset, point, landmark):\n        # \u9009\u53d6\u7f6e\u4fe1\u5ea6\u4e3a0\uff0c1\u7684\u6b63\u8d1f\u6837\u672c\u6c42\u7f6e\u4fe1\u5ea6\u635f\u5931\n        cls_mask = torch.lt(cls, 2)\n        cls_loss = self.cls_loss_func(torch.masked_select(out_cls, cls_mask), \n                                      torch.masked_select(cls, cls_mask))\n\n        # \u9009\u53d6\u6b63\u6837\u672c\u548c\u90e8\u5206\u6837\u672c\u6c42\u504f\u79fb\u7387\u7684\u635f\u5931\n        offset_mask = torch.gt(cls, 0)\n        offset_loss = self.offset_loss_func(torch.masked_select(out_offset, offset_mask),\n                                            torch.masked_select(offset, offset_mask))\n\n        if landmark:\n            point_loss = self.point_loss_func(torch.masked_select(out_point, offset_mask),\n                                              torch.masked_select(point, offset_mask))\n            return cls_loss, offset_loss, point_loss\n        else:\n            return cls_loss, offset_loss, None<\/code><\/pre>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u5b9a\u4e49\u635f\u5931\u51fd\u6570.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u5b9a\u4e49\u635f\u5931\u51fd\u6570.png\" alt=\"\" \/><\/a><\/p>\n<p>\u4ee3\u7801\u89e3\u6790\uff1a<br \/>\n\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u56e0\u4e3a\u6211\u4eec\u8981\u9884\u6d4b\u7684\u5185\u5bb9\u65e2\u6709\u5206\u7c7b\u95ee\u9898\uff0c\u4e5f\u6709\u56de\u5f52\u95ee\u9898\uff0c\u6240\u4ee5\u9700\u8981\u6839\u636e\u5bf9\u5e94\u60c5\u51b5\u9009\u62e9\u635f\u5931\u51fd\u6570\u3002<\/p>\n<ul>\n<li>\u5728\u7c7b\u522b\u65b9\u9762\uff0c\u6211\u4eec\u7684\u6837\u672c\u867d\u7136\u6709\u6b63\u3001\u8d1f\u3001\u504f\u6837\u672c\uff0c\u4f46\u662f\u6211\u4eec\u8981\u8bad\u7ec3\u6a21\u578b\u7684\u4e3b\u8981\u662f\uff1a\u662f\u8138\u6216\u4e0d\u662f\u8138\uff0c\u6240\u4ee5\u5176\u672c\u8d28\u662f\u4e2a\u4e8c\u5206\u7c7b\u95ee\u9898\uff0c\u635f\u5931\u51fd\u6570\u9009\u62e9\u4e86BCELoss\uff1b<\/li>\n<li>\u5728\u8138\u6846\u504f\u79fb\u91cf\u548c\u5173\u952e\u70b9\u504f\u79fb\u91cf\uff0c\u90fd\u5c5e\u4e8e\u56de\u5f52\u95ee\u9898\uff0c\u6240\u4ee5\u635f\u5931\u51fd\u6570\u9009\u62e9\u4e86MSELoss\u3002<\/li>\n<\/ul>\n<blockquote>\n<ul>\n<li><strong>Mean Squared Error (MSE) Loss:<\/strong><br \/>\n\u9002\u7528\u573a\u666f: \u56de\u5f52\u95ee\u9898,\u5e0c\u671b\u9884\u6d4b\u503c\u548c\u771f\u5b9e\u503c\u4e4b\u95f4\u7684\u5dee\u5f02\u6700\u5c0f\u5316\u3002\u4f8b\u5982\u623f\u4ef7\u9884\u6d4b\u3001\u80a1\u7968\u4ef7\u683c\u9884\u6d4b\u7b49\u3002<\/li>\n<li><strong>Binary Cross-Entropy (BCE) Loss:<\/strong><br \/>\n\u9002\u7528\u573a\u666f: \u4e8c\u5206\u7c7b\u95ee\u9898,\u9884\u6d4b\u7ed3\u679c\u4e3a0\u62161\u3002\u4f8b\u5982\u5783\u573e\u90ae\u4ef6\u5206\u7c7b\u3001\u56fe\u50cf\u4e8c\u5206\u7c7b\u7b49\u3002<\/li>\n<li><strong>Categorical Cross-Entropy (CCE) Loss:<\/strong><br \/>\n\u9002\u7528\u573a\u666f: \u591a\u5206\u7c7b\u95ee\u9898,\u9884\u6d4b\u7ed3\u679c\u4e3a\u591a\u4e2a\u7c7b\u522b\u4e2d\u7684\u4e00\u4e2a\u3002\u4f8b\u5982\u56fe\u50cf\u5206\u7c7b\u3001\u6587\u672c\u5206\u7c7b\u7b49\u3002<\/li>\n<li><strong>Focal Loss:<\/strong><br \/>\n\u9002\u7528\u573a\u666f: \u7c7b\u522b\u4e0d\u5e73\u8861\u7684\u5206\u7c7b\u95ee\u9898,\u53ef\u4ee5\u63d0\u9ad8\u6a21\u578b\u5bf9\u4e8e\u96be\u5206\u7c7b\u6837\u672c\u7684\u5173\u6ce8\u5ea6\u3002<\/li>\n<li><strong>Dice Loss:<\/strong><br \/>\n\u9002\u7528\u573a\u666f: \u56fe\u50cf\u5206\u5272\u4efb\u52a1,\u53ef\u4ee5\u63d0\u9ad8\u6a21\u578b\u5bf9\u4e8e\u8fb9\u754c\u533a\u57df\u7684\u5173\u6ce8\u5ea6\u3002<\/li>\n<\/ul>\n<\/blockquote>\n<h3><span class=\"ez-toc-section\" id=\"%E5%AE%9A%E4%B9%89%E8%AE%AD%E7%BB%83%E8%BF%87%E7%A8%8B\"><\/span>\u5b9a\u4e49\u8bad\u7ec3\u8fc7\u7a0b<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<pre><code class=\"language-python\">def train(self, epochs, landmark=False):\n        &quot;&quot;&quot;\n            - \u65ad\u70b9\u7eed\u4f20 --&gt; \u77ed\u70b9\u7eed\u8bad\n            - transfer learning \u8fc1\u79fb\u5b66\u4e60\n            - pretrained model \u9884\u8bad\u7ec3\n\n        :param epochs: \u8bad\u7ec3\u7684\u8f6e\u6570\n        :param landmark: \u662f\u5426\u4e3alandmark\u4efb\u52a1\n        :return:\n        &quot;&quot;&quot;\n\n        # \u52a0\u8f7d\u4e0a\u6b21\u8bad\u7ec3\u7684\u53c2\u6570\n        if os.path.exists(self.param_path):\n            self.net.load_state_dict(torch.load(self.param_path))\n            print(&quot;\u52a0\u8f7d\u53c2\u6570\u6587\u4ef6,\u7ee7\u7eed\u8bad\u7ec3 ...&quot;)\n        else:\n            print(&quot;\u6ca1\u6709\u53c2\u6570\u6587\u4ef6,\u5168\u65b0\u8bad\u7ec3 ...&quot;)\n\n        # \u5c01\u88c5\u6570\u636e\u52a0\u8f7d\u5668\n        dataloader = DataLoader(self.datasets, batch_size=32, shuffle=True)\n\n        # \u5b9a\u4e49\u5217\u8868\u5b58\u50a8\u635f\u5931\u503c\n        cls_losses = []\n        offset_losses = []\n        point_losses = []\n        total_losses = []\n\n        for epoch in range(epochs):\n            # \u8bad\u7ec3\u4e00\u8f6e\n            for i, (img_data, _cls, _offset, _point) in enumerate(dataloader):\n                # \u6570\u636e\u642c\u5bb6 [32, 3, 12, 12]\n                img_data = img_data.to(self.device)\n                _cls = _cls.to(self.device)\n                _offset = _offset.to(self.device)\n                _point = _point.to(self.device)\n\n                if landmark:\n                    # O-Net\u8f93\u51fa\u4e09\u4e2a\n                    out_cls, out_offset, out_point = self.net(img_data)\n                    out_point = out_point.view(-1, 10)\n                else:\n                    # O-Net\u8f93\u51fa\u4e24\u4e2a\n                    out_cls, out_offset = self.net(img_data)\n                    out_point = None\n\n                # [B, C, H, W] \u8f6c\u6362\u4e3a [B, C]\n                out_cls = out_cls.view(-1, 1)\n                out_offset = out_offset.view(-1, 4)\n\n                if landmark:\n                    out_point = out_point.view(-1, 10)\n\n                # \u8ba1\u7b97\u635f\u5931\n                cls_loss, offset_loss, point_loss = self.compute_loss(out_cls, out_offset, out_point,\n                                                                    _cls, _offset, _point, landmark)\n\n                if landmark:\n                    loss = cls_loss + offset_loss + point_loss\n                else:\n                    loss = cls_loss + offset_loss\n\n                # \u6253\u5370\u635f\u5931\n                if landmark:\n                    print(f&quot;Epoch [{epoch+1}\/{epochs}], loss:{loss.item():.4f}, cls_loss:{cls_loss.item():.4f}, &quot;\n                        f&quot;offset_loss:{offset_loss.item():.4f}, point_loss:{point_loss.item():.4f}&quot;)\n                else:\n                    print(f&quot;Epoch [{epoch+1}\/{epochs}], loss:{loss.item():.4f}, cls_loss:{cls_loss.item():.4f}, &quot;\n                        f&quot;offset_loss:{offset_loss.item():.4f}&quot;)\n\n                # \u5b58\u50a8\u635f\u5931\u503c\n                cls_losses.append(cls_loss.item())\n                offset_losses.append(offset_loss.item())\n                if landmark:\n                    point_losses.append(point_loss.item())\n                total_losses.append(loss.item())\n\n                # \u6e05\u7a7a\u68af\u5ea6\n                self.optimizer.zero_grad()\n\n                # \u68af\u5ea6\u56de\u4f20\n                loss.backward()\n\n                # \u4f18\u5316\n                self.optimizer.step()\n\n            # \u4fdd\u5b58\u6a21\u578b\uff08\u53c2\u6570\uff09\n            # torch.save(self.net.state_dict(), self.param_path)\n\n            # \u4fdd\u5b58\u6574\u4e2a\u6a21\u578b\n            torch.save(self.net, self.param_path)\n\n        # \u7ed8\u5236\u635f\u5931\u66f2\u7ebf\n        self.plot_losses(cls_losses, offset_losses, point_losses, total_losses, landmark)\n\n        print(&quot;\u8bad\u7ec3\u5b8c\u6210!&quot;)<\/code><\/pre>\n<p>\u4ee3\u7801\u89e3\u6790\uff1a<br \/>\n\u5728\u8bad\u7ec3\u4fdd\u5b58\u6a21\u578b\u6587\u4ef6\u65f6\uff0c\u4e00\u822c\u6709\u4e24\u79cd\u65b9\u5f0f\uff1a<\/p>\n<ul>\n<li>\u65b9\u5f0f\u4e00\uff1a\u4fdd\u5b58\u6574\u4e2a\u795e\u7ecf\u7f51\u7edc\u7684\u7684\u7ed3\u6784\u4fe1\u606f\u548c\u6a21\u578b\u53c2\u6570\u4fe1\u606f<\/li>\n<\/ul>\n<pre><code class=\"language-python\"># \u4fdd\u5b58\u8bad\u7ec3\u5b8c\u7684\u7f51\u7edc\u7684\u5404\u5c42\u53c2\u6570\uff08\u5373weights\u548cbias)\ntorch.save(net.state_dict(),path):\n\n# \u52a0\u8f7d\u4fdd\u5b58\u5230path\u4e2d\u7684\u5404\u5c42\u53c2\u6570\u5230\u795e\u7ecf\u7f51\u7edc\nnet.load_state_dict(torch.load(path)):<\/code><\/pre>\n<ul>\n<li>\u65b9\u5f0f\u4e8c\uff1a\u53ea\u4fdd\u5b58\u795e\u7ecf\u7f51\u7edc\u7684\u8bad\u7ec3\u6a21\u578b\u53c2\u6570<\/li>\n<\/ul>\n<pre><code class=\"language-python\"># \u4fdd\u5b58\u6574\u4e2a\u6a21\u578b\u65b9\u6cd5\ntorch.save(net,path):\n\n# \u52a0\u8f7d\u6574\u4e2a\u6a21\u578b\u65b9\u6cd5\nnet = torch.load(path)<\/code><\/pre>\n<p>\u4e00\u822c\u60c5\u51b5\u4e0b\uff0c\u5b98\u65b9\u63a8\u8350\u4f7f\u7528\u7b2c\u4e00\u79cd\u65b9\u5f0f\uff0c\u539f\u56e0\u662f\u4fdd\u5b58\u5185\u5bb9\u7684\u5c11\uff0c\u901f\u5ea6\u4e5f\u5feb\uff0c\u6240\u4ee5\u6211\u5728\u539f\u6709\u7684\u793a\u4f8b\u4ee3\u7801\u6539\u4e3a\u4e86\u65b9\u5f0f\u4e00\u3002<\/p>\n<blockquote>\n<p>\u6211\u5728\u4fdd\u5b58\u548c\u52a0\u8f7d\u6a21\u578b\u65f6\uff0c\u8fd8\u9047\u5230\u4e24\u4e2a\u95ee\u9898\uff0c\u4f5c\u4e3a\u7ecf\u9a8c\u5206\u4eab\u8bb0\u5f55\u4e0b\u6765\uff1a<\/p>\n<ol>\n<li>\u5982\u679c\u5728GPU\u673a\u5668\u4e0a\u8bad\u7ec3\u597d\u6a21\u578b\uff0c\u7136\u540e\u5728\u4f7f\u7528CPU\u673a\u5668\u4e0a\u52a0\u8f7d\u6a21\u578b\u9884\u6d4b\u662f\u4e0d\u53ef\u4ee5\u7684\uff0ctorch\u4f1a\u63d0\u793a\u9519\u8bef\u3002<\/li>\n<li>\u5982\u679cCPU\u673a\u5668\u662fapple M3\u82af\u7247\uff0c\u4e0a\u8ff0\u4fdd\u5b58\u548c\u52a0\u8f7d\u6a21\u578b\u4f1a\u9047\u5230(Segmentation Fault)\u7684\u9519\u8bef,\u67e5\u770bgithub\u4e0a\u7684issue\uff0c\u8be5\u95ee\u9898\u6682\u672a\u89e3\u51b3\u3002<\/li>\n<\/ol>\n<\/blockquote>\n<h3><span class=\"ez-toc-section\" id=\"%E5%BC%80%E5%A7%8B%E8%AE%AD%E7%BB%83\"><\/span>\u5f00\u59cb\u8bad\u7ec3<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<pre><code class=\"language-python\">from train import model_mtcnn as nets\nimport os\nimport train.train as train\n\nif __name__ == &#039;__main__&#039;:\n\n    current_path = os.path.dirname(os.path.abspath(__file__))\n    # \u6743\u91cd\u5b58\u653e\u5730\u5740\n    base_path = os.path.join(current_path, &quot;model&quot;)\n    model_path = os.path.join(base_path, &quot;p_net.pth&quot;)\n\n    # \u6570\u636e\u5b58\u653e\u5730\u5740\n    data_path = os.path.join(current_path, &quot;datasets\/train\/12&quot;)\n\n    # \u5982\u679c\u6ca1\u6709\u8fd9\u4e2a\u53c2\u6570\u5b58\u653e\u76ee\u5f55\uff0c\u5219\u521b\u5efa\u4e00\u4e2a\u76ee\u5f55\n    if not os.path.exists(base_path):\n        os.makedirs(base_path)\n\n    # \u6784\u5efa\u6a21\u578b\n    pnet = nets.PNet()\n\n    # \u5f00\u59cb\u8bad\u7ec3\n    t = train.Trainer(pnet, model_path, data_path)\n\n    t.train(100)\n<\/code><\/pre>\n<p>\u4ee3\u7801\u89e3\u6790\uff1a<\/p>\n<ul>\n<li>\u56e0\u4e3a\u6211\u4eec\u9700\u8981\u5206\u522b\u8bad\u7ec3P-Net\u3001R-Net\u3001O-Net\uff0c\u6240\u4ee5\u5206\u522b\u5b9e\u73b0\u4e86train_pnet.py\u3001train_rnet.py\u3001train_onet.py<\/li>\n<\/ul>\n<h2><span class=\"ez-toc-section\" id=\"%E9%A2%84%E6%B5%8B%E8%BF%87%E7%A8%8B%E5%88%86%E6%9E%90%E7%90%86%E8%A7%A3\"><\/span>\u9884\u6d4b\u8fc7\u7a0b\u5206\u6790\u7406\u89e3<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>\u9884\u6d4b\u8fc7\u7a0b\u5927\u4f53\u4e0a\u53ef\u4ee5\u5206\u4e3a\uff1a\u521d\u59cb\u5316\u3001\u9884\u5904\u7406\u3001P-Net\u9884\u6d4b\u3001R-Net\u9884\u6d4b\u3001O-Net\u9884\u6d4b\u4e94\u4e2a\u90e8\u5206\u3002<\/p>\n<h3><span class=\"ez-toc-section\" id=\"%E5%88%9D%E5%A7%8B%E5%8C%96%E5%92%8C%E9%A2%84%E5%A4%84%E7%90%86\"><\/span>\u521d\u59cb\u5316\u548c\u9884\u5904\u7406<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<pre><code class=\"language-python\">class Detector(object):\n    def __init__(self,\n                 pnet_path,\n                 rnet_path,\n                 onet_path,\n                 softnms=False,\n                 thresholds=(0.6, 0.6, 0.95),\n                 factor=0.709):\n        &quot;&quot;&quot;\n            \u521d\u59cb\u5316\n        &quot;&quot;&quot;\n        # \u4e09\u4e2a\u7f51\u7edc\u7684\u7f6e\u4fe1\u5ea6\u9608\u503c\n        self.thresholds = thresholds\n\n        # \u7f29\u653e\u56e0\u5b50\n        self.factor = factor\n\n        # \u662f\u5426\u542f\u7528softnms\n        self.softnms = softnms\n\n        # \u6784\u5efa\u6a21\u578b\n        self.pnet = nets.PNet().to(device)\n        self.rnet = nets.RNet().to(device)\n        self.onet = nets.ONet().to(device)\n\n        # \u52a0\u8f7d\u53c2\u6570\n        # self.pnet.load_state_dict(torch.load(pnet_path))\n        # self.rnet.load_state_dict(torch.load(rnet_path))\n        # self.onet.load_state_dict(torch.load(onet_path))\n\n        # \u52a0\u8f7d\u6574\u4e2a\u6a21\u578b\n        self.pnet = torch.load(pnet_path).to(device)\n        self.rnet = torch.load(rnet_path).to(device)\n        self.onet = torch.load(onet_path).to(device)\n\n        # \u8bbe\u4e3a\u8bc4\u4f30\u6a21\u5f0f\n        self.pnet.eval()\n        self.rnet.eval()\n        self.onet.eval()\n\n        # \u8ddf\u8bad\u7ec3\u65f6\u505a\u76f8\u540c\u7684\u9884\u5904\u7406\n        self.img_transfrom = transforms.Compose([\n            # \u8f6c\u5f20\u91cf [0, 1]\n            transforms.ToTensor(),\n            # [0,1] --&gt; [-1, 1]\n            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])\n        ])\n<\/code><\/pre>\n<p>\u4ee3\u7801\u89e3\u6790\uff1a<br \/>\n\u8fd9\u90e8\u5206\u4ee3\u7801\u7684\u4f5c\u7528\u4e3b\u8981\u662f\uff1a<\/p>\n<ol>\n<li>\u58f0\u660e\u4e86\u4e00\u4e2aDetector\u7c7b\uff0c\u7528\u4e8e\u5b9e\u73b0\u4eba\u8138\u68c0\u6d4b\u5668<\/li>\n<li>\u521d\u59cb\u5316<strong>init<\/strong>\u51fd\u6570\u4e2d\uff0c\u52a0\u8f7d\u8bad\u7ec3\u597d\u7684\u6a21\u578b<\/li>\n<li>\u5c06\u56fe\u50cf\u8f6c\u6362\u4e3a\u5f20\u91cf\uff1a\u5c06\u56fe\u50cf\u4ecePIL\u683c\u5f0f\u6216NumPy\u6570\u7ec4\u683c\u5f0f\u8f6c\u6362\u4e3aPyTorch\u5f20\u91cf\uff0c\u5e76\u5c06\u50cf\u7d20\u503c\u4ece0-255\u7684\u8303\u56f4\u7f29\u653e\u52300-1\u7684\u8303\u56f4\u3002<\/li>\n<li>\u6807\u51c6\u5316\uff1a\u5bf9\u5f20\u91cf\u8fdb\u884c\u6807\u51c6\u5316\u5904\u7406\uff0c\u5c06\u50cf\u7d20\u503c\u4ece0-1\u7684\u8303\u56f4\u8f6c\u6362\u5230-1\u52301\u7684\u8303\u56f4\u3002<\/li>\n<\/ol>\n<blockquote>\n<ul>\n<li><code>transforms.Compose<\/code><br \/>\ntransforms.Compose\u662f\u4e00\u4e2a\u5de5\u5177\uff0c\u5b83\u63a5\u53d7\u4e00\u4e2a\u7531\u591a\u4e2a\u56fe\u50cf\u53d8\u6362\u64cd\u4f5c\u7ec4\u6210\u7684\u5217\u8868\uff0c\u5e76\u5c06\u8fd9\u4e9b\u64cd\u4f5c\u6309\u987a\u5e8f\u5e94\u7528\u5230\u8f93\u5165\u56fe\u50cf\u4e0a\u3002\u8fd9\u6837\u53ef\u4ee5\u5c06\u591a\u4e2a\u56fe\u50cf\u5904\u7406\u6b65\u9aa4\u7ec4\u5408\u5728\u4e00\u8d77\uff0c\u5f62\u6210\u4e00\u4e2a\u9884\u5904\u7406\u7ba1\u9053\u3002<\/li>\n<li><code>transforms.ToTensor()<\/code><br \/>\ntransforms.ToTensor()\u5c06\u4e00\u4e2aPIL\u56fe\u50cf\u6216NumPy\u6570\u7ec4\u8f6c\u6362\u4e3a\u4e00\u4e2aPyTorch\u5f20\u91cf\uff08Tensor\uff09\u3002<br \/>\n<strong>\u8f93\u5165<\/strong>\uff1a\u4e00\u4e2aPIL\u56fe\u50cf\u6216NumPy\u6570\u7ec4\uff0c\u50cf\u7d20\u503c\u8303\u56f4\u4e3a0-255\u3002<br \/>\n<strong>\u8f93\u51fa<\/strong>\uff1a\u4e00\u4e2aPyTorch\u5f20\u91cf\uff0c\u50cf\u7d20\u503c\u8303\u56f4\u4e3a0-1\u3002<\/li>\n<li><code>transforms.Normalize(mean, std)<\/code><br \/>\ntransforms.Normalize(mean, std)\u5bf9\u5f20\u91cf\u8fdb\u884c\u6807\u51c6\u5316\u5904\u7406\u3002\u5b83\u5c06\u6bcf\u4e2a\u901a\u9053\u7684\u50cf\u7d20\u503c\u51cf\u53bb\u7ed9\u5b9a\u7684\u5747\u503c\uff08mean\uff09\uff0c\u7136\u540e\u9664\u4ee5\u7ed9\u5b9a\u7684\u6807\u51c6\u5dee\uff08std\uff09\u3002\u8fd9\u4e00\u6b65\u901a\u5e38\u7528\u4e8e\u4f7f\u8f93\u5165\u6570\u636e\u5177\u6709\u96f6\u5747\u503c\u548c\u5355\u4f4d\u65b9\u5dee\uff0c\u4ece\u800c\u6709\u52a9\u4e8e\u52a0\u901f\u795e\u7ecf\u7f51\u7edc\u7684\u8bad\u7ec3\u548c\u63d0\u9ad8\u6a21\u578b\u7684\u6536\u655b\u6027\u3002<br \/>\n\u5728\u4e0a\u8ff0\u4ee3\u7801\u4e2d\uff0cmean=[0.5, 0.5, 0.5]\u548cstd=[0.5, 0.5, 0.5]\u8868\u793a\u5bf9\u6bcf\u4e2a\u901a\u9053\uff08\u7ea2\u3001\u7eff\u3001\u84dd\uff09\u8fdb\u884c\u76f8\u540c\u7684\u6807\u51c6\u5316\u5904\u7406\u3002<\/li>\n<\/ul>\n<\/blockquote>\n<h3><span class=\"ez-toc-section\" id=\"P-Net%E7%BD%91%E7%BB%9C%E9%A2%84%E6%B5%8B\"><\/span>P-Net\u7f51\u7edc\u9884\u6d4b<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<pre><code class=\"language-python\">def pnet_detect(self, image):\n        &quot;&quot;&quot;\n            P-Net \u68c0\u6d4b\n\n            - \u4f20\u5165\u7684\u662f \u539f\u59cb\u56fe\u50cf \u7684 Image\u5bf9\u8c61\n\n        &quot;&quot;&quot;\n\n        boxes = []\n\n        print(&quot;P-Net \u68c0\u6d4b&quot;)\n        # \u9075\u4ece\u5750\u6807\u4e60\u60ef H, W --&gt; W, H\n\n        w, h = image.size\n        print(&quot;\u539f\u59cb\u56fe\u50cf\uff1a&quot;, image.size)\n\n        min_side = min(w, h)\n\n        scale = 1\n        scale_time = 1\n        # \u7edf\u8ba1PNet\u68c0\u6d4b\u7684\u6570\u91cf\n        nums = 0\n        # \u56fe\u50cf\u91d1\u5b57\u5854\n        while min_side &gt; 12:\n            # \u56fe\u50cf\u9884\u5904\u7406\n            img_data = self.img_transfrom(image).to(device)\n            # \u6dfb\u52a0\u6279\u91cf\u7ef4\u5ea6 [3, h, w] --&gt; [1, 3, h, w ]\n            img_data.unsqueeze_(0)\n            # print(f&quot;\u7b2c {scale_time} \u6b21\u68c0\u6d4b &quot;)\n            # print(&quot;\u8f93\u5165\u56fe\u50cf\uff1a&quot;, img_data.shape)\n            # \u901a\u8fc7pnet\u6a21\u578b\uff0c\u4e5f\u5c31\u662f\u6b63\u5411\u4f20\u64ad\n            with torch.no_grad():\n                _cls, _offset = self.pnet(img_data)\n\n            # [1, 1, H, W]\n            print(&quot;\u8f93\u51fa\u7f6e\u4fe1\u5ea6\uff1a&quot;, _cls.shape)\n            print(&quot;\u8f93\u51fabbox\u504f\u7f6e\uff1a&quot;, _offset.shape)\n\n            nums += _cls.shape[-1] * _cls.shape[-2]\n\n            # \u5c06\u6570\u636e\u642c\u5230CPU\u4e0a\u8ba1\u7b97 [1, 1, 295, 445]\n            _cls = _cls[0][0].data.cpu() # [295, 445]\n            _offset = _offset[0].data.cpu() # [4, 295, 445]\n\n            indexes = torch.nonzero(_cls &gt; self.thresholds[0])\n\n            # \u5904\u7406\n            boxes.extend(self.box(indexes, _cls, _offset, scale))\n\n            # \u8ba1\u7b97\u65b0\u7684\u5c3a\u5bf8\n            scale *= self.factor\n            _w = int(w * scale)\n            _h = int(h * scale)\n\n            # \u7f29\u653e\u56fe\u50cf\uff08\u6784\u5efa\u56fe\u50cf\u91d1\u5b57\u5854\uff09\n            image = image.resize((_w, _h))\n            min_side = min(_w, _h)\n            scale_time += 1\n        # \u6253\u5370PNet\u7684\u68c0\u6d4b\u6b21\u6570\n        print(nums)\n        # \u6ca1\u6709\u505a\u53bb\u91cd\u590d\n        if self.softnms:\n            return tool.soft_nms(torch.stack(boxes).numpy(), 0.3)\n\n        # return tool.nms(torch.stack(boxes).numpy(), 0.3)\n        boxes = torch.stack(boxes)\n        return boxes[nms(boxes[:, :4], boxes[:, 4], 0.3)].numpy()<\/code><\/pre>\n<p>\u4ee3\u7801\u89e3\u6790\uff1a<br \/>\n\u8fd9\u90e8\u5206\u4ee3\u7801\u5305\u542b\u5185\u5bb9\u8f83\u591a\uff0c\u6211\u4eec\u5206\u89e3\u4e3a\u4ee5\u4e0b\u51e0\u90e8\u5206\u5206\u522b\u7406\u89e3\uff1a<\/p>\n<h4><span class=\"ez-toc-section\" id=\"%E6%9E%84%E5%BB%BA%E5%9B%BE%E5%83%8F%E9%87%91%E5%AD%97%E5%A1%94\"><\/span>\u6784\u5efa\u56fe\u50cf\u91d1\u5b57\u5854:<span class=\"ez-toc-section-end\"><\/span><\/h4>\n<pre><code class=\"language-python\">        while min_side &gt; 12:\n            # \u56fe\u50cf\u9884\u5904\u7406\n            img_data = self.img_transfrom(image).to(device)\n            # \u6dfb\u52a0\u6279\u91cf\u7ef4\u5ea6 [3, h, w] --&gt; [1, 3, h, w ]\n            img_data.unsqueeze_(0)\n\n            #\uff08\u4e2d\u95f4\u4ee3\u7801\u7701\u7565....\uff09\n\n            # \u8ba1\u7b97\u65b0\u7684\u5c3a\u5bf8\n            scale *= self.factor\n            _w = int(w * scale)\n            _h = int(h * scale)\n\n            # \u7f29\u653e\u56fe\u50cf\uff08\u6784\u5efa\u56fe\u50cf\u91d1\u5b57\u5854\uff09\n            image = image.resize((_w, _h))\n            min_side = min(_w, _h)\n            scale_time += 1<\/code><\/pre>\n<ul>\n<li>\u901a\u8fc7<code>while<\/code>\u5faa\u73af\uff0c\u8ba1\u7b97\u8f93\u5165\u56fe\u50cf\u7684\u6700\u5c0f\u8fb9\u957fmin_side\uff0c\u76f4\u5230\u6700\u5c0f\u8fb9\u957f\u5c0f\u4e8e12\u9000\u51fa\u5faa\u73af\u3002<\/li>\n<li>\u5728\u6bcf\u6b21\u5faa\u73af\u4e2d,\u901a\u8fc7<code>image.resize<\/code>\u5c06\u5f53\u524d\u56fe\u50cf\u7f29\u653e\u81f3\u65b0\u7684\u5c3a\u5bf8(_w, _h),\u5e76\u5c06\u5176\u8d4b\u503c\u7ed9image\u53d8\u91cf\u3002<br \/>\n<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u56fe\u50cf\u91d1\u5b57\u5854\u793a\u4f8b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u56fe\u50cf\u91d1\u5b57\u5854\u793a\u4f8b.png\" alt=\"\" \/><\/a><\/li>\n<\/ul>\n<h4><span class=\"ez-toc-section\" id=\"P-Net%E5%89%8D%E5%90%91%E9%A2%84%E6%B5%8B\"><\/span>P-Net\u524d\u5411\u9884\u6d4b<span class=\"ez-toc-section-end\"><\/span><\/h4>\n<pre><code class=\"language-python\"># (\u7701\u7565...)\n            # \u56fe\u50cf\u9884\u5904\u7406\n            img_data = self.img_transfrom(image).to(device)\n            # \u6dfb\u52a0\u6279\u91cf\u7ef4\u5ea6 [3, h, w] --&gt; [1, 3, h, w ]\n            img_data.unsqueeze_(0)\n            # print(f&quot;\u7b2c {scale_time} \u6b21\u68c0\u6d4b &quot;)\n            # print(&quot;\u8f93\u5165\u56fe\u50cf\uff1a&quot;, img_data.shape)\n            # \u901a\u8fc7pnet\u6a21\u578b\uff0c\u4e5f\u5c31\u662f\u6b63\u5411\u4f20\u64ad\n            with torch.no_grad():\n                _cls, _offset = self.pnet(img_data)\n\n            # [1, 1, H, W]\n            print(&quot;\u8f93\u51fa\u7f6e\u4fe1\u5ea6\uff1a&quot;, _cls.shape)\n            print(&quot;\u8f93\u51fabbox\u504f\u7f6e\uff1a&quot;, _offset.shape)\n\n            # \u5c06\u6570\u636e\u642c\u5230CPU\u4e0a\u8ba1\u7b97 [1, 1, 295, 445]\n            _cls = _cls[0][0].data.cpu() # [295, 445]\n            _offset = _offset[0].data.cpu() # [4, 295, 445]\n\n            indexes = torch.nonzero(_cls &gt; self.thresholds[0])\n# (\u7701\u7565...)<\/code><\/pre>\n<ul>\n<li>\u5c06\u51c6\u5907\u597d\u7684img_data\u8f93\u5165\u5230self.pnet\u6a21\u578b\u4e2d\u8fdb\u884c\u524d\u5411\u4f20\u64ad,\u5f97\u5230\u8f93\u51fa\u7684\u7f6e\u4fe1\u5ea6_cls\u548c\u8fb9\u754c\u6846\u504f\u79fb\u91cf_offset\u3002<\/li>\n<li><code>_cls<\/code> \u662f P-Net \u8f93\u51fa\u7684\u5206\u7c7b\u7ed3\u679c\u5f20\u91cf\uff0c\u8868\u793a\u6bcf\u4e2a\u50cf\u7d20\u70b9\u7684\u7f6e\u4fe1\u5ea6\uff08confidence score\uff09\uff0c\u5373\u8be5\u4f4d\u7f6e\u662f\u5426\u5305\u542b\u4eba\u8138\u7684\u6982\u7387\u3002<\/li>\n<li><code>self.thresholds[0]<\/code> \u662f\u4e00\u4e2a\u9608\u503c\uff0c\u5bf9\u5e94<code>thresholds=(0.6, 0.6, 0.95)<\/code>\u4e2d\u76840.6\u3002<\/li>\n<li><code>_cls &gt; self.thresholds[0]<\/code> \u4f1a\u751f\u6210\u4e00\u4e2a\u5e03\u5c14\u5f20\u91cf\uff0c\u5f62\u72b6\u4e0e _cls \u76f8\u540c\uff0c\u5143\u7d20\u503c\u4e3a True \u8868\u793a\u8be5\u4f4d\u7f6e\u7684\u7f6e\u4fe1\u5ea6\u5927\u4e8e\u9608\u503c\uff0cFalse \u8868\u793a\u7f6e\u4fe1\u5ea6\u5c0f\u4e8e\u6216\u7b49\u4e8e\u9608\u503c\u3002<\/li>\n<li><code>torch.nonzero()<\/code> \u662f PyTorch \u7684\u4e00\u4e2a\u51fd\u6570\uff0c\u7528\u4e8e\u83b7\u53d6\u8f93\u5165\u5f20\u91cf\u4e2d\u6240\u6709\u975e\u96f6\u5143\u7d20\u7684\u7d22\u5f15,\u5b83\u4f1a\u8fd4\u56de\u4e00\u4e2a\u4e8c\u7ef4\u5f20\u91cf\uff0c\u5176\u4e2d\u6bcf\u4e00\u884c\u8868\u793a _cls \u4e2d\u4e00\u4e2a\u5927\u4e8e\u9608\u503c\u7684\u4f4d\u7f6e\u7684\u7d22\u5f15\u3002<\/li>\n<li><code>indexes<\/code> \u4fdd\u5b58\u7684\u662f _cls \u5f20\u91cf\u4e2d\u6240\u6709\u5927\u4e8e self.thresholds[0] \u7684\u5143\u7d20\u7684\u7d22\u5f15\u4f4d\u7f6e\u3002<\/li>\n<\/ul>\n<pre><code class=\"language-python\">import torch\n# \u5047\u8bbe_cls \u7684\u503c\u4e3a\uff1a\n_cls = torch.tensor([[[[0.1, 0.7, 0.4],\n                      [0.8, 0.2, 0.9],\n                      [0.5, 0.3, 0.6]]]])\nself_thresholds = [0.5]\n\n_cls = _cls[0][0].data.cpu()  # [3, 3]\n_cls &gt; self.thresholds[0]\n# tensor([[False,  True, False],\n#         [ True, False,  True],\n#         [False, False,  True]])\n\ntorch.nonzero(_cls &gt; self.thresholds[0])\n# tensor([[0, 1],\n#         [1, 0],\n#         [1, 2],\n#         [2, 2]])\n# [0, 1]\u5bf9\u5e94\u7684\u503c\u662f 0.7\uff0c\u5b83\u5728 _cls\u4e2d\u7684\u4f4d\u7f6e\u662f(0, 0)\uff0c\u6240\u4ee5\u8fd4\u56de\u7684\u662f [0, 1]\u3002\n# [0, 0]\u5bf9\u5e94\u7684\u503c\u662f 0.8\uff0c\u5b83\u5728 _cls\u4e2d\u7684\u4f4d\u7f6e\u662f(0, 1)\uff0c\u6240\u4ee5\u8fd4\u56de\u7684\u662f [0, 0]\u3002<\/code><\/pre>\n<p>\u5728\u6c42\u5f97indexes\u540e\uff0c\u901a\u8fc7<code>boxes.extend(self.box(indexes, _cls, _offset, scale))<\/code>\u53cd\u5411\u6c42\u89e3\u4eba\u8138\u7684\u6846\u3002<\/p>\n<h4><span class=\"ez-toc-section\" id=\"%E5%8F%8D%E5%90%91%E6%B1%82%E8%A7%A3box%E6%A1%86\"><\/span>\u53cd\u5411\u6c42\u89e3box\u6846<span class=\"ez-toc-section-end\"><\/span><\/h4>\n<pre><code class=\"language-python\">def box(self, indexes, cls, offset, scale, stride=2, side_len=12):\n\n        # \u53cd\u5411\u89e3\u7801\uff0c\u6620\u5c04\u5230\u539f\u59cb\u56fe\u50cf\u4e0a\n        # P-Net \u53cd\u89e3\u5750\u6807\n        # \u5de6\u4e0a\u89d2\u5750\u6807\n        # anchor \u7684\u5750\u6807\u662f\u6b7b\u7684\n\n        # \u6c42\u6620\u5c04\u5230\u539f\u56fe\u4e2d\u7684\n        # \u5de6\u4e0a\u89d2\u7684\u5750\u6807\n        _x1 = (indexes[:, 1] * stride) \/ scale\n        _y1 = (indexes[:, 0] * stride) \/ scale\n\n        # \u53f3\u4e0b\u89d2\u5750\u6807\n        _x2 = (indexes[:, 1] * stride + side_len) \/ scale\n        _y2 = (indexes[:, 0] * stride + side_len) \/ scale\n\n        # \u8fb9\u957f\n        side = _x2 - _x1\n\n        # \u53d6\u51fa\u6709\u6548\u6846\n        offset = offset[:, indexes[:, 0], indexes[:, 1]]\n\n        # \u901a\u8fc7\u8bef\u5dee\u548c\u6807\u51c6\u6846\uff0c\u8ba1\u7b97\u51fa\u771f\u5b9e\u9884\u6d4b\u6846\n        x1 = (_x1 + side * offset[0])\n        y1 = (_y1 + side * offset[1])\n        x2 = (_x2 + side * offset[2])\n        y2 = (_y2 + side * offset[3])\n\n        # (n,)\n        cls = cls[indexes[:, 0], indexes[:, 1]]\n\n        # (n, 5)\n        result = torch.stack([x1, y1, x2, y2, cls], dim=1)\n        return result<\/code><\/pre>\n<p>\u5982\u56fe\u6240\u793a<br \/>\n<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u53cd\u5411\u6c42\u89e3\u793a\u610f\u56fe.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u53cd\u5411\u6c42\u89e3\u793a\u610f\u56fe.png\" alt=\"\" \/><\/a><\/p>\n<ul>\n<li><code>indexes[:, 1]<\/code>: \u8fd9\u662f\u4ece indexes \u4e2d\u63d0\u53d6\u7684\u7b2c\u4e8c\u5217\uff0c\u8868\u793a\u8fd9\u4e9b\u5143\u7d20\u5728\u7279\u5f81\u56fe\u5bbd\u5ea6\u65b9\u5411\u4e0a\u7684\u7d22\u5f15\u3002<\/li>\n<li><code>stride<\/code>: \u8fd9\u662f P-Net \u7684\u6b65\u957f\uff0c\u8868\u793a\u5728\u7279\u5f81\u56fe\u4e0a\u6bcf\u79fb\u52a8\u4e00\u6b65\u5728\u539f\u56fe\u4e0a\u5bf9\u5e94\u7684\u50cf\u7d20\u8ddd\u79bb\u3002<\/li>\n<li><code>scale<\/code>: \u8fd9\u662f\u5f53\u524d\u56fe\u50cf\u91d1\u5b57\u5854\u7684\u7f29\u653e\u6bd4\u4f8b\uff0c\u8868\u793a\u7279\u5f81\u56fe\u76f8\u5bf9\u4e8e\u539f\u59cb\u56fe\u50cf\u7684\u7f29\u653e\u56e0\u5b50\u3002<\/li>\n<li>\u901a\u8fc7<code>(indexes[:, 1] * stride) \/ scale<\/code>\u5c06\u7279\u5f81\u56fe\u4e0a\u7684\u5750\u6807\u6620\u5c04\u56de\u539f\u56fe\u5750\u6807\u7cfb<\/li>\n<li>\u901a\u8fc7\u8ba1\u7b97side\u8fb9\u957f\u548coffset\uff0c\u53cd\u5411\u6c42\u89e3\u51fa\u771f\u5b9e\u9884\u6d4b\u6846\u7684\u5750\u6807\u3002<\/li>\n<li>\u6700\u540e\u8fd4\u56de\u771f\u5b9e\u9884\u6d4b\u6846\u7684\u5750\u6807\u4ee5\u53ca\u7f6e\u4fe1\u5ea6\u3002<\/li>\n<\/ul>\n<h4><span class=\"ez-toc-section\" id=\"NMS%E6%B1%82%E6%9C%80%E7%BB%88%E6%A1%86\"><\/span>NMS\u6c42\u6700\u7ec8\u6846<span class=\"ez-toc-section-end\"><\/span><\/h4>\n<pre><code class=\"language-python\">return boxes[nms(boxes[:, :4], boxes[:, 4], 0.3)].numpy()<\/code><\/pre>\n<p>\u5bf9\u7279\u5f81\u6846\u8fdb\u884c\u6392\u5e8f\uff0c\u901a\u8fc7NMS\u6c42\u5f97\u6700\u7ec8\u7684\u6846\u3002<\/p>\n<h3><span class=\"ez-toc-section\" id=\"R-Net%E7%BD%91%E7%BB%9C%E9%A2%84%E6%B5%8B\"><\/span>R-Net\u7f51\u7edc\u9884\u6d4b<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<pre><code class=\"language-python\">def rnet_detect(self, image, pnet_boxes):\n        &quot;&quot;&quot;\n            R-Net \u68c0\u6d4b\n        &quot;&quot;&quot;\n        boxes = []\n        img_dataset = []\n        # \u53d6\u51faPNet\u7684\u6846\uff0c\u8f6c\u4e3a\u6b63\u65b9\u5f62\uff0c\u8f6c\u6210tensor\uff0c\u65b9\u4fbf\u540e\u9762\u7528tensor\u53bb\u7d22\u5f15\n        square_boxes = torch.from_numpy(tool.convert_to_square(pnet_boxes))\n        for box in square_boxes:\n            _x1 = int(box[0])\n            _y1 = int(box[1])\n            _x2 = int(box[2])\n            _y2 = int(box[3])\n            # crop\u88c1\u526a\u7684\u65f6\u5019\u8d85\u51fa\u539f\u56fe\u5927\u5c0f\u7684\u5750\u6807\u4f1a\u81ea\u52a8\u586b\u5145\u4e3a\u9ed1\u8272\n            img_crop = image.crop([_x1, _y1, _x2, _y2])\n            # \u8f6c\u4e3a24*24\uff0c\u4e5f\u5c31\u662fRNet\u8f93\u5165\n            img_crop = img_crop.resize((24, 24))\n\n            img_data = self.img_transfrom(img_crop).to(device)\n            img_dataset.append(img_data)\n\n        # (n,1) (n,4)\n        _cls, _offset = self.rnet(torch.stack(img_dataset))\n\n        _cls = _cls.data.cpu()\n        _offset = _offset.data.cpu()\n\n        # (14,)\n        indexes = torch.nonzero(_cls &gt; self.thresholds[1])[:, 0]\n\n        # (n,5)\n        box = square_boxes[indexes]\n\n        # (n,)\n        _x1 = box[:, 0]\n        _y1 = box[:, 1]\n        _x2 = box[:, 2]\n        _y2 = box[:, 3]\n\n        side = _x2 - _x1\n        # (n,4)\n        offset = _offset[indexes]\n        # (n,)\n        x1 = _x1 + side * offset[:, 0]\n        y1 = _y1 + side * offset[:, 1]\n        x2 = _x2 + side * offset[:, 2]\n        y2 = _y2 + side * offset[:, 3]\n        # (n,)\n        cls = _cls[indexes][:, 0]\n\n        # np.array([x1, y1, x2, y2, cls]) (5,n)\n        boxes.extend(torch.stack([x1, y1, x2, y2, cls], dim=1))\n        if len(boxes) == 0:\n            return np.array([])\n\n        boxes = torch.stack(boxes)\n        return boxes[nms(boxes[:, :4], boxes[:, 4], 0.3)].numpy()<\/code><\/pre>\n<p>\u4ee3\u7801\u89e3\u6790\uff1a<br \/>\nR-Net\u9884\u6d4b\u8fc7\u7a0b\u4e0eP-Net\u7c7b\u4f3c\uff0c\u53ea\u662fR-Net\u7684\u8f93\u5165\u9664\u4e86\u56fe\u50cf\u4e4b\u5916\u8fd8\u6709P-Net\u7684\u9884\u6d4b\u7ed3\u679c\u3002<\/p>\n<h3><span class=\"ez-toc-section\" id=\"O-Net%E7%BD%91%E7%BB%9C%E9%A2%84%E6%B5%8B\"><\/span>O-Net\u7f51\u7edc\u9884\u6d4b<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<pre><code class=\"language-python\">def onet_detect(self, image, rnet_boxes):\n        &quot;&quot;&quot;\n            O-Net \u68c0\u6d4b\n\n        &quot;&quot;&quot;\n        boxes = []\n        img_dataset = []\n        square_boxes = tool.convert_to_square(rnet_boxes)\n        for box in square_boxes:\n            _x1 = int(box[0])\n            _y1 = int(box[1])\n            _x2 = int(box[2])\n            _y2 = int(box[3])\n            img_crop = image.crop([_x1, _y1, _x2, _y2])\n            img_crop = img_crop.resize((48, 48))\n            img_data = self.img_transfrom(img_crop).to(device)\n            img_dataset.append(img_data)\n\n        _cls, _offset, _point = self.onet(torch.stack(img_dataset))\n        _cls = _cls.data.cpu().numpy()\n        _offset = _offset.data.cpu().numpy()\n        _point = _point.data.cpu().numpy()\n\n        # 0.95\n        indexes, _ = np.where(_cls &gt; self.thresholds[2])\n\n        # (n,5)\n        box = square_boxes[indexes]\n\n        # (n,)\n        _x1 = box[:, 0]\n        _y1 = box[:, 1]\n        _x2 = box[:, 2]\n        _y2 = box[:, 3]\n\n        side = _x2 - _x1\n\n        # (n,4)\n        offset = _offset[indexes]\n\n        # (n,)\n        x1 = _x1 + side * offset[:, 0]\n        y1 = _y1 + side * offset[:, 1]\n        x2 = _x2 + side * offset[:, 2]\n        y2 = _y2 + side * offset[:, 3]\n\n        # (n,)\n        cls = _cls[indexes][:, 0]\n        # (n,10)\n        point = _point[indexes]\n        px1 = _x1 + side * point[:, 0]\n        py1 = _y1 + side * point[:, 1]\n        px2 = _x1 + side * point[:, 2]\n        py2 = _y1 + side * point[:, 3]\n        px3 = _x1 + side * point[:, 4]\n        py3 = _y1 + side * point[:, 5]\n        px4 = _x1 + side * point[:, 6]\n        py4 = _y1 + side * point[:, 7]\n        px5 = _x1 + side * point[:, 8]\n        py5 = _y1 + side * point[:, 9]\n\n        # np.array([x1, y1, x2, y2, cls, px1, py1, px2, py2, px3, py3, px4, py4, px5, py5]) (15,n)\n        boxes.extend(np.stack([x1, y1, x2, y2, cls, px1, py1, px2, py2, px3, py3, px4, py4, px5, py5], axis=1))\n\n        if len(boxes) == 0:\n            return np.array([])\n\n        # return tool.nms(np.stack(boxes), 0.3, isMin=True)\n        return NMS(np.stack(boxes), 0.3, isMin=True)<\/code><\/pre>\n<p>\u4ee3\u7801\u89e3\u6790\uff1a<br \/>\nO-Net\u9884\u6d4b\u8fc7\u7a0b\u4e0eR-Net\u7c7b\u4f3c\uff0c\u53ea\u662fO-Net\u7f51\u7edc\u9664\u4e86\u9884\u6d4b\u4eba\u7269\u6846\u4e4b\u5916\uff0c\u8fd8\u9700\u8981\u9884\u6d4b\u4e94\u4e2a\u5173\u952e\u70b9\uff0c\u4e94\u4e2a\u5173\u952e\u70b9\u7684\u53cd\u5411\u6c42\u89e3\u4e0eBox\u7684\u53cd\u5411\u6c42\u89e3\u8fc7\u7a0b\u7c7b\u4f3c\u3002<\/p>\n<h2><span class=\"ez-toc-section\" id=\"%E9%81%97%E7%95%99%E5%BE%85%E6%8E%A2%E7%B4%A2%E9%97%AE%E9%A2%98\"><\/span>\u9057\u7559\u5f85\u63a2\u7d22\u95ee\u9898<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>\u95ee\u98981\uff1a\u4e3a\u4ec0\u4e48\u8981\u6709\u504f\u6837\u672c<\/p>\n<p>\u95ee\u98982\uff1a\u4e3a\u4ec0\u4e48\u8bad\u7ec3\u8fc7\u7a0b\u635f\u5931\u7387\u662f\u4e0d\u5b9a\u7684<\/p>\n<p>\u95ee\u98983\uff1a\u5982\u4f55\u589e\u52a0\u65b0\u7684\u6807\u7b7e\u8fdb\u884c\u8bad\u7ec3<\/p>\n<p>\u95ee\u98984\uff1a\u5982\u4f55\u5c06\u4e09\u4e2a\u6a21\u578b\u5408\u6210\u4e00\u4e2a\u8fc7\u7a0b<\/p>\n<h2><span class=\"ez-toc-section\" id=\"%E5%86%85%E5%AE%B9%E5%B0%8F%E7%BB%93\"><\/span>\u5185\u5bb9\u5c0f\u7ed3<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<ul>\n<li>\u6570\u636e\u9884\u5904\u7406\n<ul>\n<li>\u6807\u6ce8\u7684\u539f\u59cb\u6570\u636e\u662f\u4e0d\u80fd\u76f4\u63a5\u704c\u5165\u6a21\u578b\u8bad\u7ec3\u7684\uff0c\u9700\u8981\u8fdb\u884c\u9884\u5904\u7406\uff0c\u5305\u62ec\u5f52\u4e00\u5316\u3001\u88c1\u526a\u3001\u65cb\u8f6c\u3001\u7ffb\u8f6c\u3001\u7f29\u653e\u7b49\u3002<\/li>\n<li>\u6570\u636e\u7684\u5b58\u50a8\u3001\u8ba1\u7b97\u9700\u8981\u8003\u7a76\u8ba1\u7b97\u4ee3\u4ef7\uff0c\u8fd9\u4e5f\u662f\u6700\u8003\u9a8c\u95ee\u9898\u601d\u8003\u548c\u89e3\u51b3\u80fd\u529b\u7684\u5730\u65b9\u3002<\/li>\n<li>\u8c28\u8bb0&quot;\u4fe1\u606f\u8574\u542b\u5728\u6570\u636e\u7684\u76f8\u5bf9\u5927\u5c0f&quot;\u4e4b\u4e2d\uff0c\u6240\u4ee5\u65e0\u8bba\u662f\u6570\u636e\u5f52\u4e00\u5316\u8fd8\u662f\u6570\u636e\u5b58\u50a8\uff0c\u53ef\u4ee5\u8003\u8651\u901a\u8fc7\u8ba1\u7b97\u6570\u636e\u6bd4\u4f8b(\u76f8\u5bf9\u4e8e\u539f\u6765\u7684\u6bd4\u4f8b)\u6765\u8fdb\u884c\u5904\u7406\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\u6a21\u578b\u8bad\u7ec3\n<ul>\n<li>MTCNN\u7684\u8bad\u7ec3\u8fc7\u7a0b\u9700\u8981\u4e09\u4e2a\u7f51\u7edc\uff0c\u5206\u522b\u662fP-Net\u3001R-Net\u3001O-Net\uff0c\u8bad\u7ec3\u8fc7\u7a0b\u9700\u8981\u5206\u522b\u8bad\u7ec3\u4e09\u4e2a\u7f51\u7edc\u3002<\/li>\n<li>MTCNN\u7684\u8bad\u7ec3\u8fc7\u7a0b\u4e0e\u5176\u4ed6\u6df1\u5ea6\u5b66\u4e60\u7684\u8bad\u7ec3\u8fc7\u7a0b\u7c7b\u4f3c\uff0c\u5305\u62ec\u81ea\u5b9a\u4e49datasets\u3001\u6a21\u578b\u642d\u5efa\u3001\u7b79\u5907\u8bad\u7ec3\u3001\u8bad\u7ec3\u53ca\u4fdd\u5b58\u6a21\u578b\u7b49\u3002<\/li>\n<li>\u6a21\u578b\u4fdd\u5b58\u65f6\u6709\u4e24\u79cd\u65b9\u5f0f\uff0c\u4e00\u79cd\u662f\u4fdd\u5b58\u6a21\u578b\u53c2\u6570\uff0c\u53e6\u4e00\u79cd\u662f\u4fdd\u5b58\u6a21\u578b\u7ed3\u6784+\u6a21\u578b\u53c2\u6570\uff0c\u5b98\u65b9\u63a8\u8350\u4f7f\u7528\u540e\u8005\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\u6a21\u578b\u9884\u6d4b\n<ul>\n<li>\u9884\u6d4b\u8fc7\u7a0b\u5927\u4f53\u4e0a\u53ef\u4ee5\u5206\u4e3a\uff1a\u521d\u59cb\u5316\u3001\u9884\u5904\u7406\u3001P-Net\u9884\u6d4b\u3001R-Net\u9884\u6d4b\u3001O-Net\u9884\u6d4b\u4e94\u4e2a\u90e8\u5206\u3002<\/li>\n<li>P-Net\u9884\u6d4b\u65f6\u4f1a\u751f\u6210\u4e0d\u540c\u7684\u56fe\u50cf\u91d1\u5b57\u5854\uff0c\u7136\u540e\u8fdb\u884c\u9884\u6d4b\uff1b\u9884\u6d4b\u540e\u4f1a\u8fd4\u56de\u7f6e\u4fe1\u5ea6 <code>_cls<\/code>\u548c\u4eba\u8138\u6846\u504f\u79fb\u91cf<code>_offset<\/code><\/li>\n<li>\u53cd\u89e3\u6c42\u4eba\u8138\u6846box\u65f6\uff0c\u4f1a\u5148\u901a\u8fc7\u7f6e\u4fe1\u5ea6\u6c42\u5f97\u7279\u5f81\u56fe\u7684\u5bbd\u9ad8\uff0c\u7136\u540e\u518d\u901a\u8fc7\u6b65\u957f\u3001\u7f29\u653e\u7cfb\u6570\u3001\u504f\u79fb\u91cf\u7b49\uff0c\u6c42\u5f97\u539f\u56fe\u4e2d\u7684\u771f\u5b9e\u6846\u3002<\/li>\n<li>R-Net\u4e0eO-Net\u7684\u9884\u6d4b\u8fc7\u7a0b\u7c7b\u4f3c\uff0cO-Net\u7684\u9884\u6d4b\u8fc7\u7a0b\u4f1a\u589e\u52a0\u5173\u952e\u70b9\u7684\u9884\u6d4b\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\u5176\u4ed6\n<ul>\n<li>\u4ee3\u7801\u5b9e\u73b0\u8fd8\u662f\u8981\u9075\u7167\u7f16\u7a0b\u89c4\u8303(\u4f8b\u5982\uff1a\u5355\u4e2a\u51fd\u6570\u7684\u4ee3\u7801\u884c\u63a7\u5236\u5728200\u884c\u5de6\u53f3)\uff0c\u63d0\u5347\u4ee3\u7801\u7684\u53ef\u8bfb\u6027\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h2><span class=\"ez-toc-section\" id=\"%E5%8F%82%E8%80%83%E8%B5%84%E6%96%99\"><\/span>\u53c2\u8003\u8d44\u6599<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p><a href=\"https:\/\/zhuanlan.zhihu.com\/p\/446737300\">\u77e5\u4e4e\uff1aBCE\u3001CE\u3001MSE\u635f\u5931\u51fd\u6570<\/a><\/p>\n<p><a href=\"https:\/\/blog.csdn.net\/qq_37803694\/article\/details\/127340818\">CSDN\uff1aPytorch\u4fdd\u5b58torch.save()\u548c\u52a0\u8f7dtorch.load()\u6a21\u578b<\/a><\/p>\n<p><a href=\"https:\/\/blog.csdn.net\/dongjinkun\/article\/details\/136535782\">CSDN\uff1aPyTorch\u57fa\u7840\uff0820\uff09<\/a><\/p>\n<p align=\"center\">\u6b22\u8fce\u5173\u6ce8\u516c\u4f17\u53f7\u4ee5\u83b7\u5f97\u6700\u65b0\u7684\u6587\u7ae0\u548c\u65b0\u95fb<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/\u626b\u7801_\u641c\u7d22\u8054\u5408\u4f20\u64ad\u6837\u5f0f-\u767d\u8272\u7248.bmp\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/\u626b\u7801_\u641c\u7d22\u8054\u5408\u4f20\u64ad\u6837\u5f0f-\u767d\u8272\u7248.bmp\" alt=\"\" \/><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u524d\u8a00 \u5728\u4e0a\u4e00\u7ae0\u3010\u8bfe\u7a0b\u603b\u7ed3\u3011Day13\uff08\u4e0b [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":4058,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"aside","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"default","adv-header-id-meta":"","stick-header-meta":"default","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[28],"tags":[65,66],"class_list":["post-3999","post","type-post","status-publish","format-aside","has-post-thumbnail","hentry","category-blog","tag-65","tag-66","post_format-post-format-aside"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>\u3010\u8bfe\u7a0b\u603b\u7ed3\u3011Day14\uff1aMTCNN\u8fc7\u7a0b\u7684\u6df1\u5165\u7406\u89e3 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=3999\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=3999#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=3999\"},\"author\":{\"name\":\"Dongming\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/3d23bb6f7f115fcefc9ae7803a691739\"},\"headline\":\"\u3010\u8bfe\u7a0b\u603b\u7ed3\u3011Day14\uff1aMTCNN\u8fc7\u7a0b\u7684\u6df1\u5165\u7406\u89e3\",\"datePublished\":\"2024-07-04T17:03:35+00:00\",\"dateModified\":\"2024-10-08T07:16:52+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=3999\"},\"wordCount\":236,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=3999#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/07\\\/\u5c01\u9762-1.jpg\",\"keywords\":[\"\u6df1\u5ea6\u5b66\u4e60\",\"\u8ba1\u7b97\u673a\u89c6\u89c9\"],\"articleSection\":[\"\u535a\u5ba2\u6587\u7ae0\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=3999#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=3999\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=3999\",\"name\":\"\u3010\u8bfe\u7a0b\u603b\u7ed3\u3011Day14\uff1aMTCNN\u8fc7\u7a0b\u7684\u6df1\u5165\u7406\u89e3 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=3999#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=3999#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/07\\\/\u5c01\u9762-1.jpg\",\"datePublished\":\"2024-07-04T17:03:35+00:00\",\"dateModified\":\"2024-10-08T07:16:52+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/3d23bb6f7f115fcefc9ae7803a691739\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=3999#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=3999\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=3999#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/07\\\/\u5c01\u9762-1.jpg\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/07\\\/\u5c01\u9762-1.jpg\",\"width\":1024,\"height\":576},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=3999#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u3010\u8bfe\u7a0b\u603b\u7ed3\u3011Day14\uff1aMTCNN\u8fc7\u7a0b\u7684\u6df1\u5165\u7406\u89e3\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\",\"url\":\"https:\\\/\\\/17aitech.com\\\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/17aitech.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/3d23bb6f7f115fcefc9ae7803a691739\",\"name\":\"Dongming\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/member\\\/avatars\\\/238a0b923820dcc5.1732798681.jpg\",\"url\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/member\\\/avatars\\\/238a0b923820dcc5.1732798681.jpg\",\"contentUrl\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/member\\\/avatars\\\/238a0b923820dcc5.1732798681.jpg\",\"caption\":\"Dongming\"},\"description\":\"\u89c1\u5929\u5730\uff0c\u89c1\u4f17\u751f\uff0c\u89c1\u81ea\u5df1\u3002\",\"sameAs\":[\"http:\\\/\\\/17aitech.com\"],\"url\":\"https:\\\/\\\/17aitech.com\\\/?page_id=33738&user=1\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u3010\u8bfe\u7a0b\u603b\u7ed3\u3011Day14\uff1aMTCNN\u8fc7\u7a0b\u7684\u6df1\u5165\u7406\u89e3 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=3999","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=3999#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=3999"},"author":{"name":"Dongming","@id":"https:\/\/17aitech.com\/#\/schema\/person\/3d23bb6f7f115fcefc9ae7803a691739"},"headline":"\u3010\u8bfe\u7a0b\u603b\u7ed3\u3011Day14\uff1aMTCNN\u8fc7\u7a0b\u7684\u6df1\u5165\u7406\u89e3","datePublished":"2024-07-04T17:03:35+00:00","dateModified":"2024-10-08T07:16:52+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=3999"},"wordCount":236,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=3999#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u5c01\u9762-1.jpg","keywords":["\u6df1\u5ea6\u5b66\u4e60","\u8ba1\u7b97\u673a\u89c6\u89c9"],"articleSection":["\u535a\u5ba2\u6587\u7ae0"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=3999#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=3999","url":"https:\/\/17aitech.com\/?p=3999","name":"\u3010\u8bfe\u7a0b\u603b\u7ed3\u3011Day14\uff1aMTCNN\u8fc7\u7a0b\u7684\u6df1\u5165\u7406\u89e3 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=3999#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=3999#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u5c01\u9762-1.jpg","datePublished":"2024-07-04T17:03:35+00:00","dateModified":"2024-10-08T07:16:52+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/3d23bb6f7f115fcefc9ae7803a691739"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=3999#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=3999"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=3999#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u5c01\u9762-1.jpg","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/\u5c01\u9762-1.jpg","width":1024,"height":576},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=3999#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u3010\u8bfe\u7a0b\u603b\u7ed3\u3011Day14\uff1aMTCNN\u8fc7\u7a0b\u7684\u6df1\u5165\u7406\u89e3"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/3d23bb6f7f115fcefc9ae7803a691739","name":"Dongming","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"\/\/17aitech.com\/wp-content\/uploads\/member\/avatars\/238a0b923820dcc5.1732798681.jpg","url":"\/\/17aitech.com\/wp-content\/uploads\/member\/avatars\/238a0b923820dcc5.1732798681.jpg","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/member\/avatars\/238a0b923820dcc5.1732798681.jpg","caption":"Dongming"},"description":"\u89c1\u5929\u5730\uff0c\u89c1\u4f17\u751f\uff0c\u89c1\u81ea\u5df1\u3002","sameAs":["http:\/\/17aitech.com"],"url":"https:\/\/17aitech.com\/?page_id=33738&user=1"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/3999","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=3999"}],"version-history":[{"count":12,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/3999\/revisions"}],"predecessor-version":[{"id":33238,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/3999\/revisions\/33238"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/media\/4058"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3999"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3999"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3999"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}