\ub525\ub7ec\ub2dd(Deep Learning)\uc740 \ube44\uc120\ud615\uc801\uc778 \ud568\uc218\ub97c \ud1b5\ud574 \ubcf5\uc7a1\ud55c \ud328\ud134\uc744 \ud559\uc2b5\ud558\ub294 \uae30\ubc95\uc73c\ub85c, \uc778\uacf5\uc2e0\uacbd\ub9dd(Artificial Neural Networks)\uc744 \uae30\ubc18\uc73c\ub85c \ud569\ub2c8\ub2e4. \uc774 \uae00\uc5d0\uc11c\ub294 \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\ub294 \ub370 \ud2b9\ud654\ub41c Recurrent Neural Networks(RNN)\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc PyTorch\ub97c \uc774\uc6a9\ud55c \uad6c\ud604 \ubc29\ubc95\uc744 \uc790\uc138\ud788 \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
RNN\uc740 \uc21c\ud658 \uc2e0\uacbd\ub9dd(Recurrent Neural Network)\uc758 \uc57d\uc790\ub85c, \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\ub294 \ub370 \uc801\ud569\ud55c \uad6c\uc870\ub97c \uac00\uc9c4 \uc2e0\uacbd\ub9dd\uc785\ub2c8\ub2e4. \uc77c\ubc18\uc801\uc778 \uc2e0\uacbd\ub9dd\uc740 \uc785\ub825 \ub370\uc774\ud130\uc758 \ubaa8\ub4e0 \uc694\uc18c\ub97c \ub3c5\ub9bd\uc801\uc73c\ub85c \ucc98\ub9ac\ud558\uc9c0\ub9cc, RNN\uc740 \uc774\uc804 \uc0c1\ud0dc\uc758 \ucd9c\ub825\uc744 \ud604\uc7ac \uc0c1\ud0dc\uc758 \uc785\ub825\uc73c\ub85c \ub2e4\uc2dc \uc0ac\uc6a9\ud558\ub294 \ubc29\uc2dd\uc73c\ub85c \uc2dc\ud000\uc2a4 \uac04\uc758 \uc5f0\uad00\uc131\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4.<\/p>\n
RNN\uc758 \uae30\ubcf8 \uad6c\uc870\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \ud2b9\uc9d5\uc744 \uac00\uc9d1\ub2c8\ub2e4:<\/p>\n
RNN\uc740 \uc5ec\ub7ec \uc7a5\uc810\uc744 \uac00\uc9d1\ub2c8\ub2e4:<\/p>\n
\ud558\uc9c0\ub9cc RNN\uc740 \ub2e4\uc74c\uacfc \uac19\uc740 \ub2e8\uc810\ub3c4 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
RNN\uc758 \ub3d9\uc791 \ubc29\uc2dd\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4. \uc785\ub825 \uc2dc\ud000\uc2a4\uc758 \uac01 \uc694\uc18c\ub294 \uc7ac\uadc0\uc801\uc73c\ub85c \ucc98\ub9ac\ub418\uba70, \uc774\uc804 \uc0c1\ud0dc\uc758 \ucd9c\ub825\uc740 \ud604\uc7ac \uc0c1\ud0dc\uc758 \uc785\ub825\uc73c\ub85c \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \uc774\ub97c \uc218\uc2dd\uc73c\ub85c \ud45c\ud604\ud558\uba74 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\n h_t = f(W_xh * x_t + W_hh * h_{t-1} + b_h)\n y_t = W_hy * h_t + b_y\n <\/code><\/pre>\n\uc5ec\uae30\uc11c:<\/p>\n
\nh_t<\/code>: \ud604\uc7ac \uc2dc\uc810 t<\/code>\uc758 \uc740\ub2c9 \uc0c1\ud0dc(hidden state)<\/li>\nx_t<\/code>: \ud604\uc7ac \uc2dc\uc810 t<\/code>\uc758 \uc785\ub825<\/li>\nW_xh<\/code>, W_hh<\/code>, W_hy<\/code>: \uac00\uc911\uce58 \ud589\ub82c<\/li>\nb_h<\/code>, b_y<\/code>: \ud3b8\ud5a5(bias) \ubca1\ud130<\/li>\nf<\/code>: \ud65c\uc131\ud654 \ud568\uc218 (\uc608: tanh, ReLU \ub4f1)<\/li>\n<\/ul>\n3. PyTorch\uc5d0\uc11c\uc758 RNN \uad6c\ud604<\/h2>\n
\uc774\uc81c PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec RNN\uc744 \uad6c\ud604\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c\uc740 RNN \uacc4\uce35\uc744 \ub9cc\ub4e4\uc5b4 \uac04\ub2e8\ud55c \uc2dc\ud000\uc2a4 \ud559\uc2b5\uc744 \uc9c4\ud589\ud558\ub294 \uc608\uc81c\uc785\ub2c8\ub2e4.<\/p>\n
3.1 RNN \ubaa8\ub378 \uc815\uc758<\/h3>\n\nimport torch\nimport torch.nn as nn\n\nclass RNNModel(nn.Module):\n def __init__(self, input_size, hidden_size, output_size):\n super(RNNModel, self).__init__()\n self.hidden_size = hidden_size\n self.rnn = nn.RNN(input_size, hidden_size, batch_first=True)\n self.fc = nn.Linear(hidden_size, output_size)\n\n def forward(self, x):\n h0 = torch.zeros(1, x.size(0), self.hidden_size).to(x.device) # \ucd08\uae30 \uc740\ub2c9 \uc0c1\ud0dc\n out, _ = self.rnn(x, h0)\n out = self.fc(out[:, -1, :]) # \ub9c8\uc9c0\ub9c9 \ud0c0\uc784\uc2a4\ud15d\uc758 \ucd9c\ub825\n return out\n <\/code><\/pre>\n3.2 \ub370\uc774\ud130 \uc900\ube44<\/h3>\n
\uc774\uc81c RNN \ubaa8\ub378\uc744 \ud559\uc2b5\ud560 \ub370\uc774\ud130\ub97c \uc900\ube44\ud569\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \uac04\ub2e8\ud55c \uc2dc\uacc4\uc5f4 \uc608\uce21\uc744 \uc704\ud574 \uc0ac\uc778 \ud568\uc218\ub97c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\nimport numpy as np\n\n# \ub370\uc774\ud130 \uc0dd\uc131\ndef create_dataset(seq_length):\n x = np.linspace(0, 100, seq_length)\n y = np.sin(x)\n return x, y\n\n# \ub370\uc774\ud130 \ubcc0\ud658\ndef transform_data(x, y, seq_length):\n x_data = []\n y_data = []\n for i in range(len(x) - seq_length):\n x_data.append(x[i:i + seq_length])\n y_data.append(y[i + seq_length])\n return np.array(x_data), np.array(y_data)\n\nseq_length = 10\nx, y = create_dataset(200)\nx_data, y_data = transform_data(x, y, seq_length)\n\n# PyTorch \ud150\uc11c\ub85c \ubcc0\ud658\nx_data = torch.FloatTensor(x_data).view(-1, seq_length, 1)\ny_data = torch.FloatTensor(y_data).view(-1, 1)\n <\/code><\/pre>\n3.3 \ubaa8\ub378 \ud6c8\ub828<\/h3>\n
\ubaa8\ub378 \ud6c8\ub828\uc744 \uc704\ud574 \uc190\uc2e4 \ud568\uc218\uc640 \ucd5c\uc801\ud654 \uc54c\uace0\ub9ac\uc998\uc744 \uc815\uc758\ud558\uace0, \uc5d0\ud3ed(epoch)\ub9c8\ub2e4 \ubaa8\ub378\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4.<\/p>\n
\n# \ubaa8\ub378 \ucd08\uae30\ud654\ninput_size = 1\nhidden_size = 16\noutput_size = 1\nmodel = RNNModel(input_size, hidden_size, output_size)\n\n# \uc190\uc2e4 \ud568\uc218 \ubc0f \ucd5c\uc801\ud654 \uc54c\uace0\ub9ac\uc998 \uc124\uc815\ncriterion = nn.MSELoss()\noptimizer = torch.optim.Adam(model.parameters(), lr=0.01)\n\n# \ubaa8\ub378 \ud6c8\ub828\nnum_epochs = 100\nfor epoch in range(num_epochs):\n model.train()\n optimizer.zero_grad() # \uae30\uc6b8\uae30 \ucd08\uae30\ud654\n\n outputs = model(x_data)\n loss = criterion(outputs, y_data)\n \n loss.backward() # \uae30\uc6b8\uae30 \uacc4\uc0b0\n optimizer.step() # \uac00\uc911\uce58 \uc5c5\ub370\uc774\ud2b8\n\n if (epoch+1) % 10 == 0:\n print(f'Epoch [{epoch+1}\/{num_epochs}], Loss: {loss.item():.4f}')\n <\/code><\/pre>\n4. RNN\uc758 \ubcc0\ud615<\/h2>\n
RNN\uc758 \uc5ec\ub7ec \ubcc0\ud615\ub4e4\uc774 \uc874\uc7ac\ud569\ub2c8\ub2e4. \ub300\ud45c\uc801\uc778 \uac83\ub4e4\uc740 Long Short-Term Memory(LSTM)\uc640 Gated Recurrent Unit(GRU)\uc785\ub2c8\ub2e4.<\/p>\n
4.1 LSTM<\/h3>\n
LSTM\uc740 RNN\uc758 \uae30\uc6b8\uae30 \uc18c\uc2e4 \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574 \uace0\uc548\ub41c \uad6c\uc870\uc785\ub2c8\ub2e4. LSTM\uc740 \uc140 \uc0c1\ud0dc(cell state)\uc640 \uc5ec\ub7ec \uac8c\uc774\ud2b8(gate)\ub97c \ud1b5\ud574 \uc815\ubcf4\ub97c \uc120\ud0dd\uc801\uc73c\ub85c \uae30\uc5b5\ud558\uac70\ub098 \uc78a\uc5b4\ubc84\ub9b4 \uc218 \uc788\ub294 \ub2a5\ub825\uc744 \uac00\uc9d1\ub2c8\ub2e4. \uc774\ub85c \uc778\ud574 \uc7a5\uae30\uc801\uc778 \uc758\uc874\uc131\uc744 \ucc98\ub9ac\ud558\ub294 \ub370 \ub354 \ud6a8\uacfc\uc801\uc785\ub2c8\ub2e4.<\/p>\n
4.2 GRU<\/h3>\n
GRU\ub294 LSTM\ubcf4\ub2e4 \uad6c\uc870\uac00 \uac04\ub2e8\ud558\uba70, \ube44\uc2b7\ud55c \uc131\ub2a5\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. GRU\ub294 \ub450 \uac1c\uc758 \uac8c\uc774\ud2b8(\ub9ac\uc14b \uac8c\uc774\ud2b8\uc640 \uc5c5\ub370\uc774\ud2b8 \uac8c\uc774\ud2b8)\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc815\ubcf4 \ud750\ub984\uc744 \uc870\uc808\ud569\ub2c8\ub2e4.<\/p>\n
5. RNN\uc758 \uc751\uc6a9 \ubd84\uc57c<\/h2>\n
RNN\uc740 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \uc751\uc6a9\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- \uc74c\uc131 \uc778\uc2dd<\/strong>: \uc5f0\uc18d\uc801\uc778 \uc74c\uc131 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\uc5ec \ubb38\uc7a5\uc744 \uc774\ud574\ud569\ub2c8\ub2e4.<\/li>\n
- \uc790\uc5f0\uc5b4 \ucc98\ub9ac<\/strong>: \uae30\uacc4 \ubc88\uc5ed, \uac10\uc815 \ubd84\uc11d \ub4f1\uc5d0\uc11c \ubb38\uc7a5\uc758 \uc758\ubbf8\ub97c \ubd84\uc11d\ud569\ub2c8\ub2e4.<\/li>\n
- \uc2dc\uacc4\uc5f4 \uc608\uce21<\/strong>: \uae08\uc735 \ub370\uc774\ud130\ub098 \ub0a0\uc528 \uc608\uce21 \ub4f1\uc758 \uc2dc\uacc4\uc5f4 \ub370\uc774\ud130\ub97c \ubaa8\ub378\ub9c1\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n
6. \uacb0\ub860<\/h2>\n
\ubcf8 \uae00\uc5d0\uc11c\ub294 RNN\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc PyTorch\ub97c \uc774\uc6a9\ud55c \uad6c\ud604 \ubc29\ubc95, \ubcc0\ud615 \ubaa8\ub378 \ubc0f \uc751\uc6a9 \ubd84\uc57c\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. RNN\uc740 \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\uc758 \ud2b9\uc131\uc744 \uc798 \ubc18\uc601\ud558\uba70, \ub525\ub7ec\ub2dd \ubd84\uc57c\uc5d0\uc11c \uc911\uc694\ud55c \uc5ed\ud560\uc744 \ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ub525\ub7ec\ub2dd\uc744 \uacf5\ubd80\ud558\uba74\uc11c RNN\uc758 \ub2e4\uc591\ud55c \ubcc0\ud615\uc744 \uc775\ud788\uace0, \ud2b9\uc815 \ubb38\uc81c\uc5d0 \uc801\ud569\ud55c \ubaa8\ub378\uc744 \uc120\ud0dd\ud558\ub294 \uac83\uc774 \uc911\uc694\ud569\ub2c8\ub2e4.<\/p>\n
\ucc38\uace0 \uc790\ub8cc<\/h3>\n