![\"\"](\"http:\/\/www.gisdeveloper.co.kr\/wp-content\/uploads\/2019\/08\/model2.png\")
<\/p>\n\uc774 \uae00\uc740 \ud150\uc11c\ud50c\ub85c\ub098 \ud30c\uc774\ud1a0\uce58\uc640 \uac19\uc740 \uac19\uc740 \ub525\ub7ec\ub2dd \ub77c\uc774\ube0c\ub7ec\ub9ac\uc758 \uae30\ubc18\uc744 \uc774\ud574\ud558\uac70\ub098, \uc9c1\uc811 \uac1c\ubc1c\ud558\uace0\uc790 \ud560\ub54c \ucc38\uace0\ud560 \ub9cc\ud55c \uae00\ub85c \uc791\uc131\ud55c \ucf54\ub4dc\ub294 \uc774\ud574\ub97c \uc704\ud574 \ub098\uc774\ube0c\ud558\uac8c \uc791\uc131\ud588\ub2e4 . \uc544\ub798\ub294 \uc774 \uae00\uc5d0\uc11c \uc0ac\uc6a9\ud560 \ubaa8\ub378\ub85c \ub2e8\uc21c\ud654\ub97c \uc704\ud574 \uc740\ub2c9\uce35\uc740 \uc5c6\uace0 \uc785\ub825\uce35\uacfc \ucd9c\ub825\uce35\ub9cc\uc774 \uc874\uc7ac\ud55c\ub2e4.<\/p>\n
<\/p>\n
\uc785\ub825\uac12 x\uc640 \ud574\ub2f9 \uc785\ub825\uac12\uc5d0 \ub300\ud55c \ub77c\ubca8\uac12 t\ub294 \ub2e4\uc74c\uacfc \uac19\ub2e4.<\/p>\n
\r\nx = np.array([20.1, 32.2])\r\nt = np.array([0, 0, 1])\r\n<\/pre>\n\ub370\uc774\ud130\uac00 \ub2e8\uc77c \ud56d\ubaa9\uc778\ub370, \uc2e4\uc81c\ub294 \uadf8 \uac1c\uc218\uac00 \uc0c1\ub2f9\uc774 \ub9ce\uc744 \uac83\uc774\ub2e4. \uac00\uc911\uce58 W\ub294 \uc785\ub825\uac12\uc758 \ud2b9\uc131\uac1c\uc218\uac00 2\uac1c\uc774\uace0 \uce28\ub825\uac12\uc758 \ubd84\ub958\uc218\uac00 3\uac1c\uc774\ubbc0\ub85c 2×3 \ud589\ub82c\uc774\uba70, \ucd08\uae30\uac12\uc740 \uc544\ub798\ucc98\ub7fc \ub09c\uc218\ub85c \uc7a1\ub294\ub2e4.<\/p>\n
\r\nW = np.random.randn(2, 3)\r\n<\/pre>\n\ubaa8\ub378\uc5d0 \ub300\ud55c \uc785\ub825 \ub370\uc774\ud130\uac00 \uc815\ud574\uc84c\uc73c\ubbc0\ub85c, \ucd08\uae30\uc5d0 \ub09c\uc218\ub85c \uc815\ud55c \uac00\uc911\uce58\uac12\uc744 \ubcf4\uc815\ud558\uae30 \uc704\ud55c \ubc29\ubc95\uc778 \uacbd\uc0ac\ud558\uac15\ubc95\uc5d0 \ub300\ud55c \ud568\uc218\uc778 gradient\ub97c \ub2e4\uc74c\ucc98\ub7fc \uc0ac\uc6a9\ud560 \uc218 \uc788\ub2e4.<\/p>\n
\r\ndW = gradient(x, t, W)\r\nprint(dW)\r\n<\/pre>\ngradient \ud568\uc218\ub294 \uc778\uc790\ub85c \uc785\ub825\uac12\uacfc \ub77c\ubca8\uac12 \uadf8\ub9ac\uace0 \ubcf4\uc815\ud560 \uac00\uc911\uce58\uac12\uc774 \uc800\uc7a5\ub41c \ud589\ub82c\uc744 \ubc1b\uc73c\uba70, \uac00\uc911\uce58\uac12\ub4e4\uc744 \ubcf4\uc815\ud560 \uacbd\uc0ac\ub3c4(\ubbf8\ubd84\uac12)\ub97c \uac00\uc911\uce58 \ud589\ub82c\uacfc \ub3d9\uc77c\ud55c \ud06c\uae30\ub85c \ubc18\ud658\ud55c\ub2e4. \uba3c\uc800 \uc190\uc2e4\ud568\uc218 loss\ub294 \ub2e4\uc74c\uacfc \uac19\ub2e4.<\/p>\n
\r\ndef predict(input, weight):\r\n return np.dot(input, weight)\r\n\r\ndef softmax(input):\r\n input = input - np.max(input)\r\n return np.exp(input) \/ np.sum(np.exp(input))\r\n\r\ndef cee(activated, label):\r\n label = label.argmax(axis=0)\r\n result = -np.sum(np.log(activated[label] + 1e-7))\r\n\r\n return result \r\n\r\ndef loss(input, label, weight):\r\n output = predict(input, weight)\r\n activated = softmax(output)\r\n loss = cee(activated, label)\r\n\r\n return loss\r\n<\/pre>\n\uc190\uc2e4\ud568\uc218\ub294 \uad50\ucc28 \uc5d4\ud2b8\ub85c\ud53c \uc624\ucc28(Cross Entropy Error, CEE) \ud568\uc218\ub97c \uc0ac\uc6a9\ud588\uc73c\uba70, \ucd9c\ub825\uce35\uc5d0 \ub300\ud55c \ud65c\uc131\ud654 \ud568\uc218\ub294 Softmax\ub97c \uc0ac\uc6a9\ud55c \uac83\uc744 \uc54c \uc218 \uc788\ub2e4. \uc774\uc81c gradient \ud568\uc218\ub294 \uc544\ub798\uc640 \uac19\ub2e4.<\/p>\n
\r\ndef gradient(input, label, weight):\r\n h = 1e-4\r\n grad = np.zeros_like(weight)\r\n \r\n it = np.nditer(weight, flags=['multi_index'])\r\n while not it.finished:\r\n idx = it.multi_index\r\n v = weight[idx]\r\n\r\n weight[idx] = v + h\r\n fxh1 = loss(input, label, weight)\r\n \r\n weight[idx] = v - h \r\n fxh2 = loss(input, label, weight)\r\n \r\n grad[idx] = (fxh1 - fxh2) \/ (2*h)\r\n \r\n weight[idx] = v\r\n\r\n it.iternext() \r\n \r\n return grad\r\n<\/pre>\n\uac00\uc911\uce58\uc5d0 \ub300\ud55c W\uc758 \uac01 \uc694\uc18c\ubcc4\ub85c \ud3b8\ubbf8\ubd84\uc744 \uae30\uc6b8\uae30 \ubc29\ud5a5\uc744 \uad6c\ud558\uace0 \uc788\ub2e4. \uc774 \uae30\uc6b8\uae30 \ubc29\ud5a5\uc73c\ub85c \uc77c\uc815\ud55c \uae38\uc774\ub9cc\ud07c \uac00\uc911\uce58\uac12\uc744 \uc774\ub3d9\ud574 \uc8fc\ub294 \uac83\uc744 \ubc18\ubcf5\ud558\uba74 \ucd5c\uc18c\uc758 \uc190\uc2e4\uac12\uc744 \uac16\ub294 \uac00\uc911\uce58\ub4e4\uc758 \ubaa8\uc74c\uc744 \uc5bb\uc744 \uc218 \uc788\uac8c \ub41c\ub2e4. \uae30\uc6b8\uae30 \ubc29\ud5a5\uc744 \uad6c\ud558\uae30 \uc704\ud55c \ubc29\ubc95\uc73c\ub85c \ud3b8\ubbf8\ubd84\uc5d0 \ub300\ud55c \uc218\uce58\ud574\uc11d \uae30\ubc95\uc744 \ud65c\uc6a9\ud588\uc73c\ub098, \uc2e4\uc81c \ud150\uc11c\ud50c\ub85c\uc6b0\ub098 \ud30c\uc774\ud1a0\uce58 \ub4f1\uacfc \uac19\uc740 \uba38\uc2e0\ub7ec\ub2dd \ub77c\uc774\ube0c\ub7ec\ub9ac\uc5d0\uc11c\ub294 \uc5ed\uc804\ud30c\uae30\ubc95\uc744 \ud65c\uc6a9\ud55c\ub2e4.<\/p>\n","protected":false},"excerpt":{"rendered":"
\uc774 \uae00\uc740 \ud150\uc11c\ud50c\ub85c\ub098 \ud30c\uc774\ud1a0\uce58\uc640 \uac19\uc740 \uac19\uc740 \ub525\ub7ec\ub2dd \ub77c\uc774\ube0c\ub7ec\ub9ac\uc758 \uae30\ubc18\uc744 \uc774\ud574\ud558\uac70\ub098, \uc9c1\uc811 \uac1c\ubc1c\ud558\uace0\uc790 \ud560\ub54c \ucc38\uace0\ud560 \ub9cc\ud55c \uae00\ub85c \uc791\uc131\ud55c \ucf54\ub4dc\ub294 \uc774\ud574\ub97c \uc704\ud574 \ub098\uc774\ube0c\ud558\uac8c \uc791\uc131\ud588\ub2e4 . \uc544\ub798\ub294 \uc774 \uae00\uc5d0\uc11c \uc0ac\uc6a9\ud560 \ubaa8\ub378\ub85c \ub2e8\uc21c\ud654\ub97c \uc704\ud574 \uc740\ub2c9\uce35\uc740 \uc5c6\uace0 \uc785\ub825\uce35\uacfc \ucd9c\ub825\uce35\ub9cc\uc774 \uc874\uc7ac\ud55c\ub2e4. \uc785\ub825\uac12 x\uc640 \ud574\ub2f9 \uc785\ub825\uac12\uc5d0 \ub300\ud55c \ub77c\ubca8\uac12 t\ub294 \ub2e4\uc74c\uacfc \uac19\ub2e4. x = np.array([20.1, 32.2]) t = np.array([0, 0, 1]) \ub370\uc774\ud130\uac00 \ub2e8\uc77c … <\/p>\n