{"id":1885,"date":"2020-07-24T21:51:31","date_gmt":"2020-07-24T12:51:31","guid":{"rendered":"https:\/\/developers.agirobots.com\/jp\/?p=1885"},"modified":"2025-07-30T17:27:15","modified_gmt":"2025-07-30T08:27:15","slug":"depthwise-pointwise-convolution","status":"publish","type":"post","link":"https:\/\/developers.agirobots.com\/jp\/depthwise-pointwise-convolution\/","title":{"rendered":"Depthwise Separable Convolution\u306b\u3064\u3044\u3066\u5206\u304b\u308a\u3084\u3059\u304f\u89e3\u8aac\uff01"},"content":{"rendered":"\n

\u7573\u307f\u8fbc\u307f\u5c64\u306e\u8a08\u7b97\u30b3\u30b9\u30c8\u3092\u5927\u5e45\u306b\u8efd\u6e1b\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u308b\u7406\u60f3\u7684\u306a\u7573\u307f\u8fbc\u307f\u624b\u6cd5\u304c\u3042\u308b\u3068\u8a00\u308f\u308c\u305f\u3089\u4f55\u304c\u601d\u3044\u6d6e\u304b\u3073\u307e\u3059\u304b\uff1f
\u8272\u3005\u306a\u89e3\u6c7a\u624b\u6cd5\u304c\u63d0\u6848\u3055\u308c\u3066\u3044\u307e\u3059\u304c\u3001\u4eca\u56de\u306f\u3001\u305d\u306e\u4e2d\u3067\u3082MobileNet\u3068\u547c\u3070\u308c\u308b\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3067\u4f7f\u7528\u3055\u308c\u3066\u3044\u308bDepthwise Separable Convolution\u306b\u3064\u3044\u3066\u8aac\u660e\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

MobileNet\u3068\u306f\u30b9\u30de\u30fc\u30c8\u30d5\u30a9\u30f3\u306a\u3069\u306e\u8a08\u7b97\u8cc7\u6e90\u306e\u5c0f\u3055\u3044\u30de\u30b7\u30f3\u3067\u4f7f\u3046\u3053\u3068\u3092\u60f3\u5b9a\u3057\u3001\u8a08\u7b97\u30b3\u30b9\u30c8\u304c\u5c0f\u3055\u304f\u306a\u308b\u3088\u3046\u306b\u8a2d\u8a08\u3055\u308c\u305f\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3067\u3059\u3002\u5f93\u6765\u306e\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3067\u4f7f\u7528\u3055\u308c\u308b\u7573\u307f\u8fbc\u307f\u5c64\u306f\u8a08\u7b97\u30b3\u30b9\u30c8\u304c\u9ad8\u3044\u305f\u3081\u9ad8\u6027\u80fd\u306aGPU\u3092\u7a4d\u3093\u3060\u30de\u30b7\u30f3\u3067\u3057\u304b\u73fe\u5b9f\u7684\u306a\u6642\u9593\u5185\u3067\u5229\u7528\u3059\u308b\u3053\u3068\u306f\u3067\u304d\u307e\u305b\u3093\u3067\u3057\u305f\u3002\u305d\u3053\u3067\u3001\u7573\u307f\u8fbc\u307f\u5c64\u306e\u6027\u80fd\u3092\u6975\u529b\u843d\u3068\u3055\u305a\u306b\u8a08\u7b97\u30b3\u30b9\u30c8\u3092\u6539\u5584\u3057\u3088\u3046\u3068\u3057\u3066\u63d0\u6848\u3055\u308c\u305f\u7573\u307f\u8fbc\u307f\u624b\u6cd5\u306e\uff11\u3064\u304c\u4eca\u56de\u8aac\u660e\u3059\u308bDepthwise Separable Convolution\u3067\u3059\u3002Depthwise Separable Convolution\u306e\u8981\u7d20\u6280\u8853\u306f\u3001Depthwise Convolution\u3068Pointwise Convolution\u306e2\u3064\u3060\u3051\u3068\u306a\u3063\u3066\u304a\u308a\u3001\u672c\u8a18\u4e8b\u3092\u304a\u8aad\u307f\u3044\u305f\u3060\u3051\u308c\u3070\u5341\u5206\u306b\u7406\u89e3\u304c\u3067\u304d\u308b\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u305d\u308c\u3067\u306f\u3001Depthwise Separable Convolution\u306b\u3064\u3044\u3066\u5206\u304b\u308a\u3084\u3059\u304f\u89e3\u8aac\u3057\u3066\u3044\u304d\u307e\u3059\uff01<\/p>\n\n\n\n

\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af<\/h2>\n\n\n\n

\u307e\u305a\u6700\u521d\u306b\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\uff08Convolutional Neural Network: CNN\uff09\u306b\u3064\u3044\u3066\u7c21\u5358\u306b\u8aac\u660e\u3057\u307e\u3059\u3002\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306b\u3064\u3044\u3066\u8a73\u3057\u304f\u77e5\u308a\u305f\u3044\u65b9\u306f\u4ee5\u4e0b\u306e\u8a18\u4e8b\u3092\u53c2\u8003\u306b\u3057\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n\n\n\n

\n
\"\"<\/dt>

\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u7406\u8ad6\u3068PyTorch\u306b\u3088\u308bLeNet\u306e\u5b9f\u88c5<\/p>

\u5f53\u8a18\u4e8b\u3067\u306f\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\uff08Convolutional Neural Network;CNN\uff09\u306e\u57fa\u672c\u7684\u306a\u7406\u8ad6\u3068PyToch\u3092\u4f7f\u3063\u305f\u5b9f\u88c5\u306b\u3064\u3044\u3066\u89e3\u8aac\u3057\u3066\u3044\u304d\u307e\u3059\u3002 \u304a\u304a\u307e\u304b\u306a\u69cb\u6210\u306f\u3001\u7573\u307f\u8fbc ... <\/p><\/div><\/dd><\/dl><\/div><\/a><\/div>\n<\/div><\/figure>\n\n\n\n

\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3068\u306f\u3001\u4e3b\u306b\u753b\u50cf\u5206\u91ce\u3067\u4f7f\u7528\u3055\u308c\u308b\u6df1\u5c64\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3067\u3059\u3002\u7573\u307f\u8fbc\u307f\u5c64\u3068\u30d7\u30fc\u30ea\u30f3\u30b0\u5c64\u3092\u7528\u3044\u3066\u753b\u50cf\u7279\u5fb4\u3092\u62bd\u51fa\u3057\u3001\u56de\u5e30\u3084\u5206\u985e\u306a\u3069\u3092\u5b9f\u73fe\u3057\u307e\u3059\u3002\u4e0b\u56f3\u306b\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u6982\u7565\u3092\u793a\u3057\u307e\u3059\u3002<\/p>\n\n\n

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

\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3067\u306f\u3001\u7573\u307f\u8fbc\u307f\u5c64\u3068\u30d7\u30fc\u30ea\u30f3\u30b0\u5c64\u3092\u7e70\u308a\u8fd4\u3057\u9069\u7528\u3057\u307e\u3059\u304c\u3001\u4f55\u5ea6\u3082\u9069\u7528\u3059\u308b\u904e\u7a0b\u3067\u3001\u753b\u50cf\u7279\u5fb4\u304c\u62bd\u51fa\u3055\u308c\u3001\u6700\u7d42\u7684\u306b\u306f1\u6b21\u5143\u306e\u30d9\u30af\u30c8\u30eb\u306b\u3057\u3066\u5168\u7d50\u5408\u5c64\uff08fully connected layer\uff09\u306b\u5165\u529b\u3059\u308b\u306e\u304c\u5b9a\u77f3\u3068\u306a\u3063\u3066\u3044\u307e\u3059\u3002\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u8d77\u6e90\u306f\u30cd\u30aa\u30b3\u30b0\u30cb\u30c8\u30ed\u30f3\uff08Neocognitron\uff09\u3068\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u3053\u308c\u306f\u5b66\u7fd2\u306b\u81ea\u5df1\u7d44\u7e54\u5316\u3092\u7528\u3044\u3066\u304a\u308a\u3001\u73fe\u5728\u306e\u5b66\u7fd2\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306e\u67a0\u7d44\u307f\u3067\u5b66\u7fd2\u3067\u304d\u306a\u3044\u305f\u3081\u3001\u305d\u308c\u3092\u73fe\u5728\u306e\u6a5f\u68b0\u5b66\u7fd2\u624b\u6cd5\u306e\u67a0\u7d44\u307f\u3067\u5b66\u7fd2\u53ef\u80fd\u306b\u3057\u305f\u306e\u304cLeNet\u3068\u306a\u3063\u3066\u3044\u307e\u3059\u3002\u3053\u306e\u3088\u3046\u306a\u7406\u7531\u304b\u3089LeNet\u304c\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u8d77\u6e90\u3067\u3042\u308b\u3068\u4e3b\u5f35\u3059\u308b\u4eba\u3082\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306f\u3001\u6df1\u5c64\u5b66\u7fd2\u30d6\u30fc\u30e0\u306e\u706b\u4ed8\u3051\u5f79\u3067\u3082\u3042\u308a\u307e\u3059\u30022012\u5e74\u306b\u884c\u308f\u308c\u305fILSVRC\u3068\u547c\u3070\u308c\u308b\u753b\u50cf\u8a8d\u8b58\u30b3\u30f3\u30da\u3067AlexNet\u3068\u3044\u3046\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u304c\u5927\u304d\u306a\u6210\u679c\u3092\u4e0a\u3052\u305f\u3053\u3068\u3067\u3001\u6df1\u5c64\u5b66\u7fd2\u30d6\u30fc\u30e0\u306b\u706b\u304c\u4ed8\u304d\u307e\u3057\u305f\u3002AlexNet\u306e\u767b\u5834\u4ee5\u964d\u3001VGG\u3084GeegleNet\u3001ResNet\u306a\u3069\u69d8\u3005\u306a\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u304c\u767b\u5834\u3057\u307e\u3057\u305f\u3002\u4e0b\u56f3\u306f\u3001ILSVRC\u306b\u304a\u3051\u308b\u512a\u52dd\u30e2\u30c7\u30eb\u306e\u5909\u9077\u3067\u3059\u3002\u7cbe\u5ea6\u306e\u5411\u4e0a\u306f\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u7dcf\u6570\u306e\u5897\u52a0\u3068\u5927\u304d\u304f\u95a2\u4fc2\u304c\u3042\u308a\u307e\u3059\u3002\u57fa\u672c\u7684\u306b\u7dcf\u6570\u304c\u5897\u52a0\u3059\u308c\u3070\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u6570\u3082\u5897\u52a0\u3059\u308b\u305f\u3081\u3001\u5927\u304d\u306a\u8a08\u7b97\u8cc7\u6e90\u304c\u5fc5\u8981\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n

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

\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u5909\u9077\u306b\u3064\u3044\u3066\u306f\u3001\u4ee5\u4e0b\u306e\u52d5\u753b\u3067\u8a73\u3057\u304f\u8aac\u660e\u3057\u3066\u3044\u307e\u3059\u306e\u3067\u305c\u3072\u3054\u8996\u8074\u304f\u3060\u3055\u3044\uff01<\/p>\n\n\n\n

\u7573\u307f\u8fbc\u307f\u5c64\u306e\u51e6\u7406<\/h2>\n\n\n\n

\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3092\u69cb\u6210\u3059\u308b\u91cd\u8981\u306a\u69cb\u6210\u8981\u7d20\u306b\u7573\u307f\u8fbc\u307f\u5c64\u3068\u30d7\u30fc\u30ea\u30f3\u30b0\u5c64\u304c\u3042\u308a\u307e\u3059\u3002\u7573\u307f\u8fbc\u307f\u5c64\u3067\u306f\u753b\u50cf\u306e\u7279\u5fb4\u3092\u691c\u51fa\u3057\u3001\u30d7\u30fc\u30ea\u30f3\u30b0\u5c64\u3067\u306f\u691c\u51fa\u3055\u308c\u305f\u7279\u5fb4\u306e\u4f4d\u7f6e\u305a\u308c\u3092\u8a31\u5bb9\u3059\u308b\u50cd\u304d\u304c\u3042\u308a\u307e\u3059\u3002\u3053\u3053\u3067\u306f\u3001\u7573\u307f\u8fbc\u307f\u5c64\u306b\u7126\u70b9\u3092\u5f53\u3066\u3001\u305d\u306e\u7573\u307f\u8fbc\u307f\u51e6\u7406\u306b\u3064\u3044\u3066\u89e3\u8aac\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

\u7573\u307f\u8fbc\u307f\u5c64\u306f\u8907\u6570\u306e\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u304b\u3089\u69cb\u6210\u3055\u308c\u307e\u3059\u3002\u307e\u305f\u3001\u5404\u3005\u306e\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306f\u30c1\u30e3\u30cd\u30eb\u65b9\u5411\u306e\u6210\u5206\u3092\u6301\u3063\u3066\u304a\u308a\u3001\u3053\u308c\u306f\u8907\u6570\u679a\u306e\u30d5\u30a3\u30eb\u30bf\u304b\u3089\u69cb\u6210\u3055\u308c\u3066\u3044\u308b\u3053\u3068\u3092\u610f\u5473\u3057\u307e\u3059\u3002\u4e0b\u56f3\u306e\u3088\u3046\u306b\u3001\u5165\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u3068\u3057\u3066\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u304c3\u306eRGB\u753b\u50cf\u3092\u3001\u7573\u307f\u8fbc\u307f\u51e6\u7406\u3092\u9069\u7528\u3057\u3066\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u304c\uff14\u306e\u7279\u5fb4\u30de\u30c3\u30d7\u306b\u5909\u63db\u3055\u308c\u308b\u5834\u5408\u3092\u4eee\u5b9a\u3057\u307e\u3059\u3002\u4e0b\u56f3\u306f\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u5c64\u306b\u3088\u308b\u7573\u307f\u8fbc\u307f\u51e6\u7406\u3092\u56f3\u793a\u3057\u305f\u3082\u306e\u3067\u3001\u80cc\u666f\u304c\u9752\u8272\u3067\u5857\u3089\u308c\u3066\u3044\u308b\u9818\u57df\u304c\u7573\u307f\u8fbc\u307f\u5c64\u306b\u3088\u308b\u51e6\u7406\u3092\u8868\u3057\u3066\u3044\u307e\u3059\u3002\u3053\u306e\u5834\u5408\u306b\u4f7f\u7528\u3055\u308c\u308b\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306e\u6570\u306f\u3001\u5909\u63db\u5f8c\u306e\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u3068\u540c\u3058\u306b\u306a\u308b\u306e\u3067\uff14\u3064\u3067\u3059\u3002\u5404\u3005\u306e\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306e\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u306f\u5165\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u3068\u540c\u3058\u306a\u306e\u3067\uff13\u3064\u3067\u3059\u3002\u3059\u306a\u308f\u3061\u3001\u5165\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u304b\u3089\u3001\u5404\u3005\u306e\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306e\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u304c\u6c7a\u307e\u308a\u3001\u51fa\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u304b\u3089\u3001\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306e\u6570\u304c\u6c7a\u307e\u308a\u307e\u3059\u3002\u4e00\u822c\u7684\u306b\u3001\u3053\u308c\u3089\u306f\u4e00\u610f\u306b\u6c7a\u307e\u308a\u307e\u3059\u3002<\/p>\n\n\n

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

\u4e0a\u306e\u56f3\u3092\u4e00\u822c\u5316\u3057\u305f\u3082\u306e\u3092\u4ee5\u4e0b\u306b\u793a\u3057\u307e\u3059\u3002\u4e0b\u56f3\u3067\u306f\u3001\u5165\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u30b5\u30a4\u30ba\u3092\\((H^{l-1}, W^{l-1}, C^{l-1})\\)\u3001\u51fa\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u30b5\u30a4\u30ba\u3092\\((H^l, W^l, C^l)\\)\u3068\u3057\u3066\u3044\u307e\u3059\uff08Channel Last\u8868\u8a18\uff09\u3002\u3053\u306e\u3068\u304d\u3001\u5404\u3005\u306e\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306e\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u306f\\(C^{l-1}\\)\u3001\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306e\u6570\u306f\\(C^l\\)\u306b\u306a\u308a\u307e\u3059\u3002\u3053\u308c\u306f\u4ee5\u524d\u306e\u8a18\u4e8b<\/a>\u3067\u3082\u89e3\u8aac\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n

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

\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3067\u306f\u3001\u5c64\u304c\u6df1\u304f\u306a\u308b\u306b\u3064\u308c\u3066\u7279\u5fb4\u62bd\u51fa\u304c\u9032\u307f\u3001\u51fa\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u304c\u5927\u304d\u304f\u306a\u308a\u307e\u3059\u3002\u3059\u308b\u3068\u3001\u8a08\u7b97\u30b3\u30b9\u30c8\u304c\u81a8\u5927\u306b\u306a\u308b\u3053\u3068\u304c\u76f4\u611f\u3067\u3082\u5206\u304b\u308b\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u6700\u8fd1\u306f\u30b9\u30de\u30fc\u30c8\u30d5\u30a9\u30f3\u306a\u3069\u306e\u30e2\u30d0\u30a4\u30eb\u6a5f\u5668\u306eCPU\u3084GPU\u3001NPU\u306a\u3069\u306e\u30d7\u30ed\u30bb\u30c3\u30b5\u306e\u6027\u80fd\u5411\u4e0a\u304c\u76ee\u899a\u307e\u3057\u3044\u3067\u3059\u304c\u3001\u5f53\u6642\u306f\u305d\u3053\u307e\u3067\u3067\u3042\u308a\u307e\u305b\u3093\u3067\u3057\u305f\u3002\u305d\u306e\u70ba\u3001\u8a08\u7b97\u30b3\u30b9\u30c8\u3092\u5c0f\u3055\u304f\u3059\u308b\u5de5\u592b\u304c\u5fc5\u8981\u306b\u306a\u308a\u307e\u3057\u305f\u3002\u69d8\u3005\u306a\u5de5\u592b\u304c\u8003\u3048\u3089\u308c\u307e\u3059\u304c\u3001\u4eca\u56de\u8aac\u660e\u3059\u308bDepthwise Separable Convolution\u3067\u306f\u3001\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u51e6\u7406\u3092Depthwise Convolution\u3068Pointwise Convolution\u306b\u5206\u3051\u308b\u3053\u3068\u3067\u3001\u5165\u51fa\u529b\u306e\u30b5\u30a4\u30ba\u306f\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u3068\u540c\u69d8\u3067\u3042\u308a\u306a\u304c\u3089\u3001\u5927\u5e45\u306a\u8a08\u7b97\u30b3\u30b9\u30c8\u306e\u6539\u5584\u3092\u5b9f\u73fe\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

Depthwise Separable Convolution\u306e\u51e6\u7406<\/h2>\n\n\n\n

Depthwise Separable Convolution\u3068\u306f\u3001\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u5c64\u306e\u51e6\u7406\u3092Depthwise Convolution\u3068Pointwise Convolution\u306b\u5206\u3051\u305f\u3082\u306e\u3067\u3059\u3002\u305d\u308c\u3067\u306f\u3001\u305d\u308c\u305e\u308c\u306e\u7573\u307f\u8fbc\u307f\u51e6\u7406\u304c\u3069\u306e\u3088\u3046\u306b\u884c\u308f\u308c\u308b\u306e\u304b\u8aac\u660e\u3057\u3066\u3044\u304d\u307e\u3059\u3002<\/p>\n\n\n\n

\u4e0b\u56f3\u306bDepthwise Separable Convolution\u306b\u3088\u308b\u7573\u307f\u8fbc\u307f\u3092\u793a\u3057\u307e\u3059\u3002\u5148\u307b\u3069\u793a\u3057\u305f\u56f3\u306b\u304a\u3044\u3066\u80cc\u666f\u304c\u9752\u8272\u306e\u9818\u57df\u304c\uff12\u3064\u306b\u5206\u304b\u308c\u3066\u3044\u307e\u3059\u3002\u3082\u3063\u3068\u3082\u5206\u304b\u308a\u3084\u3059\u3044\u3068\u3053\u308d\u3092\u8a00\u3046\u3068\u3001\u524d\u534a\u306e\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306e\u6570\u304c1\u3064\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002Depthwise Separable Convolution\u3067\u306f\u524d\u534a\u90e8\u5206\u306f\u51fa\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u306b\u4f9d\u5b58\u305b\u305a\u5e38\u306b\uff11\u3064\u3067\u3059\u3002\u3088\u304f\u898b\u308b\u3068\u5f8c\u534a\u306e\u51e6\u7406\u3082\u82e5\u5e72\u5909\u308f\u3063\u3066\u3044\u307e\u3059\u3002\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u5c64\u3067\u306f\u3001\u524d\u534a\u306e\u5404\u3005\u306e\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u3067\u8a08\u7b97\u3055\u308c\u305f\u7d50\u679c\u3092\u30c1\u30e3\u30f3\u30cd\u30eb\u65b9\u5411\u306b\u8db3\u3057\u5408\u308f\u305b\u308b\uff08\u5168\u91cd\u307f\uff11\u306b\u3088\u308b\u91cd\u307f\u4ed8\u548c\u3068\u3044\u3046\u89e3\u91c8\u306f\u53ef\u80fd\uff09\u51e6\u7406\u3092\u3057\u307e\u3059\u304c\u3001Depthwise Separable Convolution\u3067\u306f\u5f8c\u534a\u3067\u5b66\u7fd2\u306b\u3088\u308a\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u53ef\u80fd\u306a\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u6301\u3064\u91cd\u307f\u4ed8\u548c\u3068\u3057\u3066\u8a08\u7b97\u3092\u3057\u307e\u3059\u3002\u524d\u534a\u3067\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306e\u6570\u3092\u6e1b\u3089\u3057\u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u3092\u5927\u5e45\u306b\u30ab\u30c3\u30c8\u3057\u3001\u5f8c\u534a\u306b\u5b66\u7fd2\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u8ffd\u52a0\u3057\u305f\u306e\u3067\u3059\u3002\u3053\u308c\u306b\u3088\u308a\u524d\u534a\u306e\u8a08\u7b97\u30b3\u30b9\u30c8\u306e\u6539\u5584\u306f\u5927\u304d\u304f\u3001\u4f8b\u3048\u3070\u3001\u51fa\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u304c256\u306a\u3089\u3001256\u500b\u306e\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u304c\u5fc5\u8981\u3060\u3063\u305f\u306e\u304c\u3001Depthwise Separable Convolution\u3067\u306f1\u3064\u3067\u3088\u304f\u306a\u308a\u307e\u3057\u305f\u3002Depthwise Separable Convolution\u3067\u306f\u3001\u524d\u534a\u306e\u7573\u307f\u8fbc\u307f\u51e6\u7406\u3092Depthwise Convolution\u3001\u5f8c\u534a\u306e\u7573\u307f\u8fbc\u307f\u51e6\u7406\u3092Pointwise Convolution\u3068\u547c\u3073\u307e\u3059\u3002Depthwise\u3068\u306f\u6df1\u3055\u5358\u4f4d\u3001Pointwise\u3068\u306f\u70b9\u5358\u4f4d\u3068\u8a33\u3059\u3053\u3068\u304c\u3067\u304d\u308b\u3088\u3046\u306b\u3001Depthwise Convolution\u306f\u6df1\u3055\u5358\u4f4d\u3067\u7573\u307f\u8fbc\u307f\u3092\u3001Pointwise Convolution\u306f\u70b9\u5358\u4f4d\uff08\uff1d\u30d4\u30af\u30bb\u30eb\u5358\u4f4d\uff09\u3067\u7573\u307f\u8fbc\u307f\u3092\u3057\u307e\u3059\u3002<\/p>\n\n\n

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

\u3053\u308c\u3092\u5148\u307b\u3069\u3068\u540c\u69d8\u306b\u4e00\u822c\u5316\u3057\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002\u4e0b\u56f3\u306e\u3088\u3046\u306b\u305d\u308c\u305e\u308c\u306e\u30b5\u30a4\u30ba\u3092\u6587\u5b57\u3067\u7f6e\u304d\u307e\u3059\u3002\u307e\u305a\u3001\u5165\u529b\u7279\u5fb4\u30de\u30c3\u30d7\\((H^{l-1}, W^{l-1}, C^{l-1})\\)\u306f\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u304c\\(C^{l-1}\\)\u306eDepthwise Convolution\u306b\u3088\u308a\u3001\\((H^l, W^l, C^{l-1})\\)\u306b\u5909\u63db\u3055\u308c\u307e\u3059\u3002\u6b21\u306b\u3001\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\\(C^{l-1}\\)\u306e\\(C^l\\)\u500b\u306ePointwise\u3000Convolution\u306b\u3088\u308a\u3001\u7279\u5fb4\u30de\u30c3\u30d7\\((H^l, W^l, C^l)\\)\u304c\u51fa\u529b\u3055\u308c\u307e\u3059\u3002<\/p>\n\n\n

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

\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u3068Depthwise Separable Convolution\u306e\u6bd4\u8f03<\/h2>\n\n\n\n

\u57fa\u672c\u4e8b\u9805<\/h3>\n\n\n\n

\u4e0b\u56f3\u306f\u3001\u5165\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u304b\u3089\u51fa\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u3078\u5909\u63db\u3059\u308b\u3068\u304d\u306e\u3001\u7573\u307f\u8fbc\u307f\u5c64\u306e\u5165\u529b\u9818\u57df\u3068\u51fa\u529b\u9818\u57df\u306e\u95a2\u4fc2\u3092\u793a\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u56f3\u4e2d\u3001Convolution Layer\u3068\u8a18\u8f09\u304c\u3042\u308b\u90e8\u5206\u306f\u3001\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u5c64\u306e\u51e6\u7406\u3092\u8868\u3057\u3066\u304a\u308a\u3001\u5165\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u4e00\u7b87\u6240\u306e\u30d1\u30c3\u30c1\\((K,K,C^{l-1})\\)\u306b\u30d5\u30a3\u30eb\u30bf\u3092\u9069\u7528\u3059\u308b\u3068\u3001\u51fa\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e1\u7b87\u6240\\((1,1,1)\\)\u306e\u5024\u306b\u5199\u50cf\u3055\u308c\u308b\u3053\u3068\u304c\u5206\u304b\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

Depthwise Convolution Layer\u3068\u8a18\u8f09\u304c\u3042\u308b\u90e8\u5206\u306f\u3001Depthwise Convolution\u306e\u51e6\u7406\u3092\u8868\u3057\u3066\u304a\u308a\u3001\u5165\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u4e00\u7b87\u6240\u306e\u30d1\u30c3\u30c1\\((K,K,C^{l-1})\\)\u306b\u30d5\u30a3\u30eb\u30bf\u3092\u9069\u7528\u3059\u308b\u3068\u3001\u51fa\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e1\u7b87\u6240\u306e\u30d1\u30c3\u30c1\\((1,1,C^l)\\)\u306b\u5199\u50cf\u3055\u308c\u308b\u3053\u3068\u304c\u5206\u304b\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

Pointwise Convolution Layer\u3068\u8a18\u8f09\u304c\u3042\u308b\u90e8\u5206\u306f\u3001Pointwise Convolution\u306e\u51e6\u7406\u3092\u8868\u3057\u3066\u304a\u308a\u3001\u5165\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u4e00\u7b87\u6240\u306e\u30d1\u30c3\u30c1\\((1,1,C^{l-1})\\)\u306b\u30d5\u30a3\u30eb\u30bf\u3092\u9069\u7528\u3059\u308b\u3068\u3001\u51fa\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e1\u7b87\u6240\\((1,1,1)\\)\u306e\u5024\u306b\u5199\u50cf\u3055\u308c\u308b\u3053\u3068\u304c\u5206\u304b\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

Depthwise Convolution\u306e\u5f8c\u306bPointwise Convolution\u3092\u884c\u3046\u3068\u3001\\((K,K,C^{l-1})\\)\u304b\u3089\\((1,1,1)\\)\u3078\u306e\u5909\u63db\u3068\u3001\u5165\u51fa\u529b\u306e\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u95a2\u4fc2\u306f\u3001\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u3068\u540c\u3058\u306b\u306a\u308b\u306e\u3067\u3001\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u5c64\u3092\u3053\u308c\u3067\u7f6e\u304d\u63db\u3048\u308b\u3053\u3068\u304c\u53ef\u80fd\u306b\u306a\u308a\u307e\u3059\u3002\u3053\u308c\u304c\u3001Depthwise Separable Convolution\u3067\u3059\u3002<\/p>\n\n\n

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

\u305d\u308c\u3067\u306f\u3001\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u3068Depthwise Separable Convolution\u306b\u3064\u3044\u3066\u3001\u8a08\u7b97\u5f0f\u3084\u8a08\u7b97\u30b3\u30b9\u30c8\u306b\u3064\u3044\u3066\u78ba\u8a8d\u3057\u6bd4\u8f03\u3057\u3066\u307f\u307e\u3059\u3002<\/p>\n\n\n\n

\u8a08\u7b97\u5f0f<\/h3>\n\n\n\n

\u305d\u308c\u3067\u306f\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u3068Depthwise Separable Convolution\u306e\u8a08\u7b97\u5f0f\u3092\u5c0e\u51fa\u3057\u3066\u3044\u304d\u307e\u3059\u3002\u307e\u305a\u306f\u3001\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u306b\u304a\u3044\u3066\u5165\u529b\u304a\u3088\u3073\u51fa\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u3001\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u3092\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u4eee\u5b9a\u3057\u307e\u3059\uff08Channel Last\u5f62\u5f0f\u3067\u8a18\u8f09\u3057\u3066\u3044\u307e\u3059\uff09\u3002<\/p>\n\n\n\n

    \n
  • \u5165\u529b\u7279\u5fb4\u30de\u30c3\u30d7\\(I^{l-1}\\): \\((H^{l-1}, W^{l-1}, C^{l-1})\\)<\/li>\n\n\n\n
  • \u51fa\u529b\u7279\u5fb4\u30de\u30c3\u30d7\\(I^l\\): \\((H^l, W^l, C^l)\\)<\/li>\n\n\n\n
  • \u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\\(F\\): \\((K, K, C^{l-1})\\)\uff08\u305f\u3060\u3057\\(K=2N^{l-1} + 1\\)\uff09<\/li>\n<\/ul>\n\n\n\n

    \u3053\u306e\u3068\u304d\u3001\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u5c64\u306f\u4ee5\u4e0b\u306b\u793a\u3059\u8a08\u7b97\u5f0f\u3067\u8868\u3055\u308c\u307e\u3059\u3002<\/p>\n\n\n\n

    $$
    I_{xym}^l=\\sum_{k=1}^{C^{l-1}}\\sum_{i=-N^{l-1}}^{N^{l-1}}\\sum_{j=-N^{l-1}}^{N^{l-1}}I^{l-1}_{sx+i, sy+j, k}F_{i+N^{l-1}, j+N^{l-1},k,m}
    $$<\/p>\n\n\n\n

    \\(s\\)\u306f\u30b9\u30c8\u30e9\u30a4\u30c9\u3001\\(k\\)\u306f\u5165\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306b\u304a\u3044\u3066\u6ce8\u76ee\u3057\u3066\u3044\u308b\u30c1\u30e3\u30f3\u30cd\u30eb\u3001\\(m\\)\u306f\u51fa\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306b\u304a\u3044\u3066\u6ce8\u76ee\u3057\u3066\u3044\u308b\u30c1\u30e3\u30f3\u30cd\u30eb\u3067\u3059\u3002\\(\\sum_{k=1}^{C^{l-1}}\\)\u306f\u5404\u3005\u306e\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306b\u3088\u3063\u3066\u8a08\u7b97\u3055\u308c\u305f\u5024\u3092\u30c1\u30e3\u30f3\u30cd\u30eb\u65b9\u5411\u306b\u91cd\u307f\uff11\u3067\u8db3\u3057\u5408\u308f\u305b\u308b\u64cd\u4f5c\u3067\u3059\uff08\u5168\u3066\u306e\u91cd\u307f\u304c\uff11\u306ePointwise Convolution\u306b\u5bfe\u5fdc\uff09\u3002\\(\\sum_{i=-N^{l-1}}^{N^{l-1}}\\)\u3068\\(\\sum_{j=-N^{l-1}}^{N^{l-1}}\\)\u306f\u3001\u7a7a\u9593\u65b9\u5411\u306e\u7573\u307f\u8fbc\u307f\u3092\u3057\u3066\u3044\u307e\u3059\uff08Depthwise Convolution\u306b\u985e\u4f3c\uff09\u3002<\/p>\n\n\n\n

    \u305d\u308c\u3067\u306f\u3001\u3053\u306e\u8a08\u7b97\u5f0f\u3092\u3082\u3068\u306b\u3057\u3066Depthwise Separable Convolution\u306e\u8a08\u7b97\u5f0f\u3092\u5c0e\u51fa\u3057\u307e\u3059\u3002\u307e\u305a\u3001Depthwise Convolution\u3067\u306f\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306e\u6570\u306f\uff11\u3064\u306a\u306e\u3067\u3001\\(F_{i+N^{l-1}, j+N^{l-1},k,m}\\)\u304b\u3089\\(m\\)\u3092\u53d6\u308a\u9664\u304d\u3001\\(F_{i+N^{l-1}, j+N^{l-1},k}\\)\u306b\u3057\u307e\u3059\u3002Pointwise Convolution\u306f\u3001\u91cd\u307f\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u6301\u3061\u3001\u307e\u305f\u3001\u51fa\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u6df1\u3055\\(m\\)\u306b\u4f4d\u7f6e\u3059\u308b\u5024\u3092\u8a08\u7b97\u3059\u308b\u305f\u3081\u3001\\(m\\)\u756a\u76ee\u306ePointwise Convolution\u306e\\(k\\)\u756a\u76ee\u306e\u91cd\u307f\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\\(w_{k,m}\\)\u3068\u7f6e\u304d\u307e\u3059\u3002\u3053\u306e\u3068\u304d\u3001Depthwise Separable Convolution\u306e\u8a08\u7b97\u5f0f\u306f\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    $$
    I_{xym}^l=\\sum_{k=1}^{C^{l-1}}w_{k,m}\\left\\{\\sum_{i=-N^{l-1}}^{N^{l-1}}\\sum_{j=-N^{l-1}}^{N^{l-1}}I^{l-1}_{sx+i, sy+j, k}F_{i+N^{l-1}, j+N^{l-1},k}\\right\\}
    $$<\/p>\n\n\n\n

    \\(\\{\\}\\)\u306e\u4e2d\u304cDepthwise Convolution\u3067\u5916\u5074\u304cPointwise Convolution\uff08\u6b63\u78ba\u306b\u306fDepthwise Convolution\u3067\u8a08\u7b97\u3055\u308c\u305f\u3082\u306e\u306e\u3046\u3061\u3001\\(k\\)\u756a\u76ee\u306e\u30c1\u30e3\u30f3\u30cd\u30eb\u306e\\((x, y)\\)\u6210\u5206\uff09\u3067\u3059\u3002\\(m\\)\u304c\u5916\u5074\u3078\u79fb\u52d5\u3057\u305f\u306e\u3067\u3001\u7e70\u308a\u8fd4\u3057\u306e\u8a08\u7b97\u56de\u6570\u304c\u5927\u5e45\u306b\u6e1b\u5c11\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

    \u8a08\u7b97\u30b3\u30b9\u30c8<\/h3>\n\n\n\n

    \u305d\u308c\u3067\u306f\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u3068Depthwise Separable Convolution\u306e\u8a08\u7b97\u30b3\u30b9\u30c8\u3092\u6bd4\u8f03\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

    \u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u5c64\u306e\u4e57\u7b97\u56de\u6570\u304a\u3088\u3073\u30d1\u30e9\u30e1\u30fc\u30bf\u6570<\/h4>\n\n\n\n

    1\u3064\u306e\u30d1\u30c3\u30c1\u306b1\u3064\u306e\u30d5\u30a3\u30eb\u30bf\u3092\u9069\u7528\u3059\u308b\u3068\u304d\u306e\u4e57\u7b97\u56de\u6570\u306f\\(K^2C^{l-1}\\)\u3067\u3059\u3002\u5165\u529b\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u30b5\u30a4\u30ba\u3092\\((H^{l-1}, W^{l-1}, C^{l-1})\\)\u3001\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306e\u500b\u6570\u3092\\(C^l\\)\u3001\u30b9\u30c8\u30e9\u30a4\u30c9\u3092\uff11\u3068\u3059\u308b\u3068\u3001\u5168\u30d1\u30c3\u30c1\u6570\u306f\\(H^{l-1}W^{l-1}\\)\u3067\u306a\u306e\u3067\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306e\u4e57\u7b97\u56de\u6570\u306f\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    $$
    K^2C^{l-1}H^{l-1}W^{l-1}\u00d7C^l
    $$<\/p>\n\n\n\n

    \u305d\u3057\u3066\u3001\u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u306f1\u3064\u306e\u30d5\u30a3\u30eb\u30bf\u304c\u3082\u3064\u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u00d7\u30d5\u30a3\u30eb\u30bf\u306e\u500b\u6570\u306a\u306e\u3067<\/p>\n\n\n\n

    $$
    K^2C^{l-1}\u00d7C^l
    $$<\/p>\n\n\n\n

    \u3068\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    Depthwise Convolution\u306e\u4e57\u7b97\u56de\u6570\u304a\u3088\u3073\u30d1\u30e9\u30e1\u30fc\u30bf\u6570<\/h4>\n\n\n\n

    Depthwise Convolution\u306f\u3001\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u306e\u500b\u6570\u306f1\u3064\u306a\u306e\u3067\u3001\u4ee5\u4e0b\u306e\u5f0f\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    $$
    K^2C^{l-1}H^{l-1}W^{l-1}
    $$<\/p>\n\n\n\n

    \u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u306f<\/p>\n\n\n\n

    $$
    K^2C^{l-1}
    $$<\/p>\n\n\n\n

    \u3068\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    Pointwise Convolution\u306e\u4e57\u7b97\u56de\u6570\u304a\u3088\u3073\u30d1\u30e9\u30e1\u30fc\u30bf\u6570<\/h4>\n\n\n\n

    Pointwise Convolution\u306f\u30011x1\u306e\u7573\u307f\u8fbc\u307f\u306a\u306e\u3067\u3001\u4ee5\u4e0b\u306e\u3088\u3046\u306a\u5f0f\u3067\u8868\u3055\u308c\u307e\u3059\u3002<\/p>\n\n\n\n

    $$
    C^{l-1}H^lW^l\u00d7C^l
    $$<\/p>\n\n\n\n

    \u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u306f<\/p>\n\n\n\n

    $$
    C^{l-1}C^l
    $$<\/p>\n\n\n\n

    \u3068\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    Depthwise Separable Convolution\u306e\u9b45\u529b<\/h4>\n\n\n\n

    \u5148\u307b\u3069\u6c42\u3081\u305f\u4e57\u7b97\u56de\u6570\u3068\u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u3092\u8868\u306b\u307e\u3068\u3081\u308b\u3068\u4ee5\u4e0b\u306e\u8868\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n

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

    \u3053\u308c\u3092\u7528\u3044\u3066\u3001\u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u3068Depthwise Separable Convolution\u306e\u4e57\u7b97\u56de\u6570\u3068\u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u3092\u6bd4\u8f03\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

    \u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\uff08Standard-Conv\u3068\u8868\u8a18\uff09\u3001Depthwise Separable Convolution\uff08Depth-Sep-Conv\u3068\u8868\u8a18\uff09\u306e\u4e57\u7b97\u91cf\u53ca\u3073\u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u306e\u6bd4\u3092\u8a08\u7b97\u3057\u3066\u307f\u307e\u3059\u3002\u4ee5\u4e0b\u3067\u306f\u8a08\u7b97\u5f0f\u3092\u7c21\u5358\u306b\u3059\u308b\u305f\u3081\u306b\u3001\\(W^{l-1} = H^{l-1} = W^l = H^l = D\\)\u3068\u7f6e\u3044\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

    $$
    \\begin{eqnarray}
    \\begin{array}{rcl}
    &&\\frac{Depth-Sep-Conv\\mbox{\u306e\u4e57\u7b97\u56de\u6570}}{Standard-Conv\\mbox{\u306e\u4e57\u7b97\u56de\u6570}}\\\\
    &=&\\frac{K^2C^{l-1}D^2 + C^{l-1}D^2\u00d7C^l}{K^2C^{l-1}D^2\u00d7C^l}\\\\
    &=&\\frac{1}{C^l} + \\frac{1}{K^2}
    \\end{array}
    \\end{eqnarray}
    $$<\/p>\n\n\n\n

    \u307e\u305f\u3001\u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u306e\u6bd4\u8f03\u3092\u884c\u3046\u3068<\/p>\n\n\n\n

    $$
    \\begin{eqnarray}
    \\begin{array}{rcl}
    &&\\frac{Depth-Sep-Conv\\mbox{\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u6570}}{Standard-Conv\\mbox{\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u6570}}\\\\
    &=&\\frac{K^2C^{l-1} + C^{l-1}C^l}{K^2C^{l-1}\u00d7C^l}\\\\
    &=&\\frac{1}{C^l} + \\frac{1}{K^2}
    \\end{array}
    \\end{eqnarray}
    $$<\/p>\n\n\n\n

    \u3068\u306a\u308b\u3053\u3068\u304c\u5206\u304b\u308a\u307e\u3059\u3002\u4e21\u65b9\u3068\u3082\u6bd4\u7387\u306f\u540c\u3058\u306b\u306a\u308a\u307e\u3059\u3002Pointwise Convolution\u306e\u500b\u6570\\(C^l\\)\u306e\u9006\u6570\u3068Depthwise Convolution\u306e\u30ab\u30fc\u30cd\u30eb\u30b5\u30a4\u30ba\u306e2\u4e57\u306e\u9006\u6570\u3092\u7dda\u5f62\u7d50\u5408\u3057\u305f\u3082\u306e\u3068\u3057\u3066\u8868\u3055\u308c\u307e\u3059\u3002<\/p>\n\n\n\n

    \u5177\u4f53\u7684\u306b\u8a08\u7b97\u3092\u3057\u3066\u307f\u307e\u3057\u3087\u3046\u3002ILSVRC\u306a\u3069\u3067\u4f7f\u7528\u3055\u308c\u308b\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306f\u7279\u5fb4\u30de\u30c3\u30d7\u306e\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\u304c100\u3092\u8d85\u3059\u3053\u3068\u304c\u591a\u3044\u306e\u3067\u3001\u7b2c1\u9805\u76ee\\(\\frac{1}{C^l}\\simeq 0\\)\u3068\u3067\u304d\u307e\u3059\u3002\u30d5\u30a3\u30eb\u30bf\u30b5\u30a4\u30ba\u30925x5\u306b\u3059\u308b\u3068\\(\\frac{1}{K^2}=\\frac{1}{25}\\)\u306b\u306a\u308a\u307e\u3059\u3002\u3059\u306a\u308f\u3061\u3001\u53e4\u5178\u7684\u306a\u7573\u307f\u8fbc\u307f\u30d5\u30a3\u30eb\u30bf\u3092\u4f7f\u7528\u3059\u308b\u3088\u308a\u3082\u3001\u5358\u7d14\u8a08\u7b97\u306725\u500d\u306e\u8efd\u91cf\u5316\u306b\u6210\u529f\u3067\u304d\u308b\u3053\u3068\u304c\u5206\u304b\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    \u307e\u3068\u3081<\/h2>\n\n\n\n

    \u4e00\u822c\u7684\u306a\u7573\u307f\u8fbc\u307f\u3092Depthwise Convolution\u3068Pointwise Convolution\u306b\u5206\u5272\u3057\u3001Depthwise Convolution\u3092\uff11\u7a2e\u985e\u306b\u3059\u308b\u3068\u3067\u8a08\u7b97\u30b3\u30b9\u30c8\u3092\u5927\u5e45\u306b\u5c0f\u3055\u304f\u3059\u308b\u3053\u3068\u306b\u6210\u529f\u305fDepthwise Separable Convolution\u5c64\u306f\u3001MobileNet\u3092\u4e2d\u5fc3\u306b\u4f7f\u7528\u3055\u308c\u3066\u304a\u308a\u3001\u6027\u80fd\u9762\u3067\u3082\u5927\u5dee\u306a\u3044\u3001\u7d44\u8fbc\u307f\u30b7\u30b9\u30c6\u30e0\u3067\u753b\u50cf\u8a8d\u8b58\u3092\u3059\u308b\u5fc5\u8981\u6027\u304c\u3042\u308b\u5834\u5408\u306f\u3001\u7a4d\u6975\u7684\u306b\u4f7f\u7528\u3057\u3066\u3044\u304d\u305f\u3044\u3067\u3059\u306d\u3002<\/p>\n\n\n\n

    \u53c2\u8003\u6587\u732e<\/h2>\n\n\n\n

    Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam, \"MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,\" arXiv, 2017.<\/p>\n","protected":false},"excerpt":{"rendered":"

    \u5f53\u8a18\u4e8b\u3067\u306f\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\uff08Convolutional Neural Network;CNN\uff09\u306e\u57fa\u672c\u7684\u306a\u7406\u8ad6\u3068PyToch\u3092\u4f7f\u3063\u305f\u5b9f\u88c5\u306b\u3064\u3044\u3066\u89e3\u8aac\u3057\u3066\u3044\u304d\u307e\u3059\u3002 \u304a\u304a\u307e\u304b\u306a\u69cb\u6210\u306f\u3001\u7573\u307f\u8fbc … <\/p>\n","protected":false},"author":1,"featured_media":1895,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_vk_print_noindex":"","sitemap_hide":"","_veu_custom_css":"","veu_display_promotion_alert":"common","vkexunit_cta_each_option":"","_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[170,171,106,174],"tags":[147,145,149],"class_list":["post-1885","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-blog","category-learn","category-ml","category-dl","tag-147","tag-145","tag-149"],"veu_head_title_object":{"title":"","add_site_title":""},"jetpack_featured_media_url":"https:\/\/developers.agirobots.com\/jp\/wp-content\/uploads\/2020\/07\/image-46.png","jetpack-related-posts":[],"jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/posts\/1885","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/comments?post=1885"}],"version-history":[{"count":79,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/posts\/1885\/revisions"}],"predecessor-version":[{"id":10553,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/posts\/1885\/revisions\/10553"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/media\/1895"}],"wp:attachment":[{"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/media?parent=1885"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/categories?post=1885"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/tags?post=1885"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}