\ub525\ub7ec\ub2dd\uc5d0\uc11c \uc2dc\uacc4\uc5f4 \ub370\uc774\ud130\ub098 \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc640 \uac19\uc740 \uc21c\ucc28\uc801 \ub370\uc774\ud130\ub97c \ubaa8\ub378\ub9c1\ud558\uae30 \uc704\ud574 \uc21c\ud658 \uc2e0\uacbd\ub9dd(RNN, Recurrent Neural Network)\uc744 \ub110\ub9ac \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \uc774 \uc911\uc5d0\uc11c Gated Recurrent Unit (GRU)\uc740 RNN\uc758 \ubcc0\ud615\uc73c\ub85c, \uc7a5\uae30 \uc758\uc874\uc131 \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574 \uac1c\ubc1c\ub41c \uac83\uc774\uba70, LSTM(Long Short-Term Memory)\uc640 \uc720\uc0ac\ud55c \uad6c\uc870\ub97c \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \ud3ec\uc2a4\ud305\uc5d0\uc11c\ub294 GRU\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \uc774\ub97c \ud30c\uc774\ud1a0\uce58\ub85c \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc124\uba85\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
GRU\ub294 2014\ub144 Kyunghyun Cho\uc5d0 \uc758\ud574 \uc81c\uc548\ub41c \uad6c\uc870\ub85c, \uc785\ub825 \uc815\ubcf4\uc640 \uc774\uc804 \uc0c1\ud0dc\uc758 \uc815\ubcf4\ub97c \uacb0\ud569\ud558\uc5ec \ud604\uc7ac \uc0c1\ud0dc\ub97c \uacb0\uc815\ud558\ub294 \ub370 \uc788\uc5b4, \ub354 \uac04\ub2e8\ud558\uace0 \uc5f0\uc0b0\ub7c9\uc774 \uc801\uc740 \ubc29\uc2dd\uc73c\ub85c \uc791\ub3d9\ud569\ub2c8\ub2e4. GRU\ub294 \ub450 \uac1c\uc758 \uc8fc\uc694 \uac8c\uc774\ud2b8\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4:<\/p>\n
GRU\uc758 \uc8fc\uc694 \uc218\uc2dd\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
1. \uc785\ub825 \ubca1\ud130 \uc5ec\uae30\uc11c 2. \uc0c8\ub85c\uc6b4 \uc740\ub2c9 \uc0c1\ud0dc \uc5ec\uae30\uc11c \uc774\uc81c \ud30c\uc774\ud1a0\uce58\ub97c \uc0ac\uc6a9\ud558\uc5ec GRU \uc140\uc744 \uc9c1\uc811 \uad6c\ud604\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc544\ub798\uc758 \uc608\uc81c \ucf54\ub4dc\uc5d0\uc11c\ub294 GRU\uc758 \uae30\ubcf8 \uc791\ub3d9 \ubc29\uc2dd\uc744 \uc774\ud574\ud558\uae30 \uc27d\uac8c \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<\/p>\n \uc704 \ucf54\ub4dc\ub294 \uac04\ub2e8\ud55c GRU \uc140\uc758 \uad6c\uc870\ub97c \uad6c\ud604\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774\uc81c GRU \uc140\uc744 \ud14c\uc2a4\ud2b8\ud558\uae30 \uc704\ud55c \uc608\uc81c \ucf54\ub4dc\ub97c \uc791\uc131\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n \uc704\uc758 \ucf54\ub4dc\ub97c \ud1b5\ud574 GRU \uc140\uc758 \ub3d9\uc791\uc744 \ud655\uc778\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc784\uc758\uc758 \uc785\ub825\uc744 \uc0dd\uc131\ud558\uace0, \ucd08\uae30 \uc740\ub2c9 \uc0c1\ud0dc\ub97c 0\uc73c\ub85c \uc124\uc815\ud55c \ud6c4, GRU \uc140\uc744 \ud1b5\ud574 \ud604\uc7ac \uc740\ub2c9 \uc0c1\ud0dc \uc774\uc81c GRU \uc140\uc744 \uc774\uc6a9\ud558\uc5ec RNN \ubaa8\ub378\uc744 \uc804\uccb4\uc801\uc73c\ub85c \uad6c\uc131\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n \uc704\uc758 \uc774\uc81c GRU \ubaa8\ub378\uc744 \ud14c\uc2a4\ud2b8\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n \uc704 \ucf54\ub4dc\ub97c \ud1b5\ud574 GRU \ubaa8\ub378\uc774 \uc8fc\uc5b4\uc9c4 \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\uc5d0 \ub300\ud574 \ucd9c\ub825\uc744 \uc0dd\uc131\ud558\ub294 \uacfc\uc815\uc744 \ud655\uc778\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n GRU\ub294 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \ud65c\uc6a9\ub429\ub2c8\ub2e4. \ud2b9\ud788, \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP) \uc791\uc5c5\uc5d0\uc11c \ud6a8\uacfc\uc801\uc73c\ub85c \uc0ac\uc6a9\ub418\uba70, \uae30\uacc4 \ubc88\uc5ed, \uac10\uc815 \ubd84\uc11d, \ud14d\uc2a4\ud2b8 \uc0dd\uc131 \ub4f1 \uc5ec\ub7ec \uc751\uc6a9 \ubd84\uc57c\uc5d0\uc11c \uc774\uc6a9\ub429\ub2c8\ub2e4. GRU\uc640 \uac19\uc740 \uc21c\ud658 \uad6c\uc870\ub294 \uc5f0\uc18d\ub41c \uc2dc\uac04\uc801 \uc758\uc874\uc131\uc744 \ubaa8\ub378\ub9c1\ud558\ub294 \ub370 \uac15\ub825\ud55c \uc7a5\uc810\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4.<\/p>\n GRU\ub294 LSTM\ubcf4\ub2e4 \uac04\ub2e8\ud558\uba74\uc11c\ub3c4 \uc88b\uc740 \uc131\ub2a5\uc744 \ubc1c\ud718\ud558\ub294 \uacbd\uc6b0\uac00 \ub9ce\uae30 \ub54c\ubb38\uc5d0, \ub370\uc774\ud130\uc758 \ud2b9\uc131\uacfc \ubb38\uc81c\uc758 \uc131\uaca9\uc5d0 \ub530\ub77c \uc801\uc808\ud55c \uc120\ud0dd\uc744 \ud558\ub294 \uac83\uc774 \uc911\uc694\ud569\ub2c8\ub2e4.<\/p>\n \ubcf8 \ud3ec\uc2a4\ud305\uc5d0\uc11c\ub294 GRU\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \ud30c\uc774\ud1a0\uce58\ub97c \uc774\uc6a9\ud55c GRU \uc140 \ubc0f RNN \ubaa8\ub378\uc758 \uad6c\ud604 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. GRU\ub294 \ubcf5\uc7a1\ud55c \uc21c\ucc28\uc801 \ub370\uc774\ud130 \ucc98\ub9ac\uc5d0 \uc720\uc6a9\ud55c \uad6c\uc870\ub85c, \ub2e4\uc591\ud55c \ub525\ub7ec\ub2dd \ubaa8\ub378\uc5d0 \ud1b5\ud569\ub418\uc5b4 \uc560\ud50c\ub9ac\ucf00\uc774\uc158\uc744 \ubc1c\uc804\uc2dc\ud0ac \uc218 \uc788\uc2b5\ub2c8\ub2e4. GRU\uc5d0 \ub300\ud55c \uc774\ud574\ub294 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubc0f \uc2dc\uacc4\uc5f4 \ubd84\uc11d\uc5d0 \ub300\ud55c \ud1b5\ucc30\ub825\uc744 \uc81c\uacf5\ud558\uba70, \uc2e4\ubb34\uc5d0\uc11c \ubc1c\uc0dd\ud560 \uc218 \uc788\ub294 \ubb38\uc81c\uc758 \ud574\uacb0\uc5d0 \ub3c4\uc6c0\uc774 \ub429\ub2c8\ub2e4.<\/p>\n \uc774\uc81c \uc5ec\ub7ec\ubd84\ub3c4 GRU\ub97c \ud65c\uc6a9\ud558\uc5ec \uc790\uc2e0\uc758 \ud504\ub85c\uc81d\ud2b8\uc5d0 \uc801\uc6a9\ud574 \ubcf4\uae30\ub97c \ubc14\ub78d\ub2c8\ub2e4!<\/p>\nx_t<\/code>\uc640 \uc774\uc804 \uc740\ub2c9 \uc0c1\ud0dc h_{t-1}<\/code>\uc5d0 \ub300\ud574, \ub9ac\uc14b \uac8c\uc774\ud2b8 r_t<\/code>\uc640 \uc5c5\ub370\uc774\ud2b8 \uac8c\uc774\ud2b8 z_t<\/code>\ub97c \uc815\uc758\ud569\ub2c8\ub2e4.<\/p>\nr_t = \u03c3(W_r * x_t + U_r * h_{t-1})\nz_t = \u03c3(W_z * x_t + U_z * h_{t-1})<\/code><\/pre>\nW_r<\/code>, W_z<\/code>\ub294 \uac00\uc911\uce58 \ub9e4\uac1c\ubcc0\uc218, U_r<\/code>, U_z<\/code>\ub294 \uc774\uc804 \uc0c1\ud0dc\uc5d0 \ub300\ud55c \uac00\uc911\uce58\uc785\ub2c8\ub2e4. \u03c3<\/code>\ub294 \uc2dc\uadf8\ubaa8\uc774\ub4dc \ud568\uc218\uc785\ub2c8\ub2e4.<\/p>\nh_t<\/code>\ub294 \ub2e4\uc74c\uacfc \uac19\uc774 \uacc4\uc0b0\ub429\ub2c8\ub2e4.<\/p>\nh_t = (1 - z_t) * h_{t-1} + z_t * tanh(W_h * x_t + U_h * (r_t * h_{t-1}))<\/code><\/pre>\nW_h<\/code>, U_h<\/code>\ub294 \ub610 \ub2e4\ub978 \uac00\uc911\uce58\uc785\ub2c8\ub2e4.<\/p>\n2. GRU\uc758 \uc7a5\uc810<\/h2>\n
\n
3. GRU \uc140 \uad6c\ud604\ud558\uae30<\/h2>\n
3.1 GRU \uc140 \uad6c\ud604<\/h3>\n
import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nclass GRUSimple(nn.Module):\n def __init__(self, input_size, hidden_size):\n super(GRUSimple, self).__init__()\n self.input_size = input_size\n self.hidden_size = hidden_size\n \n # \uac00\uc911\uce58 \ucd08\uae30\ud654\n self.Wz = nn.Parameter(torch.Tensor(hidden_size, input_size))\n self.Uz = nn.Parameter(torch.Tensor(hidden_size, hidden_size))\n self.Wr = nn.Parameter(torch.Tensor(hidden_size, input_size))\n self.Ur = nn.Parameter(torch.Tensor(hidden_size, hidden_size))\n self.Wh = nn.Parameter(torch.Tensor(hidden_size, input_size))\n self.Uh = nn.Parameter(torch.Tensor(hidden_size, hidden_size))\n\n self.reset_parameters()\n\n def reset_parameters(self):\n for param in self.parameters():\n stdv = 1.0 \/ param.size(0) ** 0.5\n param.data.uniform_(-stdv, stdv)\n\n def forward(self, x_t, h_prev):\n r_t = torch.sigmoid(self.Wr @ x_t + self.Ur @ h_prev)\n z_t = torch.sigmoid(self.Wz @ x_t + self.Uz @ h_prev)\n h_hat_t = torch.tanh(self.Wh @ x_t + self.Uh @ (r_t * h_prev))\n \n h_t = (1 - z_t) * h_prev + z_t * h_hat_t\n return h_t\n<\/code><\/pre>\n__init__<\/code> \uba54\uc11c\ub4dc\uc5d0\uc11c \uc785\ub825 \ud06c\uae30\uc640 \uc228\uaca8\uc9c4 \uc0c1\ud0dc \ud06c\uae30\ub97c \ucd08\uae30\ud654\ud558\uba70, \uac00\uc911\uce58 \ub9e4\uac1c\ubcc0\uc218\ub97c \uc815\uc758\ud569\ub2c8\ub2e4. reset_parameters<\/code> \uba54\uc11c\ub4dc\ub294 \uac00\uc911\uce58\ub97c \ucd08\uae30\ud654\ud569\ub2c8\ub2e4. forward<\/code> \uba54\uc11c\ub4dc\uc5d0\uc11c\ub294 \uc785\ub825\uacfc \uc774\uc804 \uc0c1\ud0dc\ub97c \uae30\ubc18\uc73c\ub85c \uc0c8\ub85c\uc6b4 \uc740\ub2c9 \uc0c1\ud0dc\ub97c \uacc4\uc0b0\ud569\ub2c8\ub2e4.<\/p>\n3.2 GRU \uc140 \ud14c\uc2a4\ud2b8<\/h3>\n
input_size = 5\nhidden_size = 3\nx_t = torch.randn(input_size) # \uc784\uc758\uc758 \uc785\ub825 \uc0dd\uc131\nh_prev = torch.zeros(hidden_size) # \ucd08\uae30 \uc740\ub2c9 \uc0c1\ud0dc\n\ngru_cell = GRUSimple(input_size, hidden_size)\nh_t = gru_cell(x_t, h_prev)\n\nprint(\"\ud604\uc7ac \uc740\ub2c9 \uc0c1\ud0dc h_t:\", h_t)\n<\/code><\/pre>\nh_t<\/code>\ub97c \ucd9c\ub825\ud569\ub2c8\ub2e4.<\/p>\n4. GRU\ub97c \uc774\uc6a9\ud55c RNN \ubaa8\ub378<\/h2>\n
class GRUModel(nn.Module):\n def __init__(self, input_size, hidden_size, output_size):\n super(GRUModel, self).__init__()\n self.gru = GRUSimple(input_size, hidden_size)\n self.fc = nn.Linear(hidden_size, output_size)\n\n def forward(self, x):\n h_t = torch.zeros(self.gru.hidden_size) # \ucd08\uae30 \uc740\ub2c9 \uc0c1\ud0dc\n\n for t in range(x.size(0)):\n h_t = self.gru(x[t], h_t) # \uac01 \ud0c0\uc784\uc2a4\ud15d\ub9c8\ub2e4 GRU \uc0ac\uc6a9\n output = self.fc(h_t) # \ub9c8\uc9c0\ub9c9 \uc740\ub2c9 \uc0c1\ud0dc\ub97c \ucd9c\ub825\uc73c\ub85c \ubcc0\ud658\n return output\n<\/code><\/pre>\nGRUModel<\/code> \ud074\ub798\uc2a4\ub294 GRU \uc140\uc744 \uc774\uc6a9\ud558\uc5ec \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\ub294 \ubaa8\ub378\uc744 \uad6c\uc131\ud569\ub2c8\ub2e4. forward<\/code> \uba54\uc11c\ub4dc\ub294 \uc785\ub825 \uc2dc\ud000\uc2a4\ub97c \ubc18\ubcf5\ud558\uba70 GRU \uc140\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc740\ub2c9 \uc0c1\ud0dc\ub97c \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4. \ub9c8\uc9c0\ub9c9 \uc740\ub2c9 \uc0c1\ud0dc\ub294 \uc120\ud615 \uc870\ud569\uc744 \ud1b5\ud574 \ucd5c\uc885 \ucd9c\ub825\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4.<\/p>\n4.1 RNN \ubaa8\ub378 \ud14c\uc2a4\ud2b8<\/h3>\n
input_size = 5\nhidden_size = 3\noutput_size = 2\nseq_length = 10\n\nx = torch.randn(seq_length, input_size) # \uc784\uc758\uc758 \uc2dc\ud000\uc2a4 \ub370\uc774\ud130 \uc0dd\uc131\n\nmodel = GRUModel(input_size, hidden_size, output_size)\noutput = model(x)\n\nprint(\"\ubaa8\ub378\uc758 \ucd9c\ub825:\", output)\n<\/code><\/pre>\n5. GRU\uc758 \ud65c\uc6a9<\/h2>\n
6. \uacb0\ub860<\/h2>\n