\n \ub525\ub7ec\ub2dd\uc740 \uc778\uacf5\uc9c0\ub2a5\uc758 \ud55c \ubd84\uc57c\ub85c, \uc778\uacf5\uc2e0\uacbd\ub9dd\uc744 \uc774\uc6a9\ud558\uc5ec \ub370\uc774\ud130\uc5d0\uc11c \ud328\ud134\uc744 \ud559\uc2b5\ud558\uace0 \uc608\uce21\ud558\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4. \uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \uc21c\ud658 \uc2e0\uacbd\ub9dd(Recurrent Neural Network, RNN)\uc758 \uac1c\ub150\uacfc PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec RNN \ubaa8\ub378\uc744 \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \ud55c\ubc88 \uc790\uc138\ud788 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n \uc21c\ud658 \uc2e0\uacbd\ub9dd(RNN)\uc740 \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\uae30 \uc704\ud574 \uc124\uacc4\ub41c \uc2e0\uacbd\ub9dd\uc758 \uc77c\uc885\uc785\ub2c8\ub2e4. \uc77c\ubc18\uc801\uc778 \uc778\uacf5\uc2e0\uacbd\ub9dd\uc740 \uace0\uc815\ub41c \uc785\ub825 \ud06c\uae30\ub97c \uac00\uc9c0\uba70, \ub370\uc774\ud130\ub97c \ud55c \ubc88\uc5d0 \ucc98\ub9ac\ud558\ub294 \ub370 \ubc18\ud574, RNN\uc740 \ub0b4\ubd80 \uc0c1\ud0dc\ub97c \uae30\uc5b5\ud558\uace0 \uacfc\uac70\uc758 \uc815\ubcf4\ub97c \uc774\uc6a9\ud558\uc5ec \ud604\uc7ac\uc758 \ucd9c\ub825\uc5d0 \uc601\ud5a5\uc744 \ubbf8\uce69\ub2c8\ub2e4. \uc774\ub294 \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP)\uc640 \uac19\uc740 \ubd84\uc57c\uc5d0\uc11c \ub9e4\uc6b0 \uc720\uc6a9\ud569\ub2c8\ub2e4.\n <\/p>\n
\n RNN\uc758 \uae30\ubcf8 \uad6c\uc870\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4. \uac01 \uc2dc\uac04 \ub2e8\uacc4\uc5d0\uc11c\uc758 \uc785\ub825 \\( x_t \\)\ub294 \uc774\uc804 \ub2e8\uacc4\uc758 \uc740\ub2c9 \uc0c1\ud0dc \\( h_{t-1} \\)\uc640 \ud568\uaed8 \ucc98\ub9ac\ub418\uc5b4 \uc0c8\ub85c\uc6b4 \uc740\ub2c9 \uc0c1\ud0dc \\( h_t \\)\ub97c \uc0dd\uc131\ud569\ub2c8\ub2e4. \uc774\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uc218\uc2dd\uc73c\ub85c \ub098\ud0c0\ub0bc \uc218 \uc788\uc2b5\ub2c8\ub2e4:\n <\/p>\n
\n h_t = f(W_h * h_{t-1} + W_x * x_t)\n <\/pre>\n\n \uc5ec\uae30\uc11c \\( f \\)\ub294 \ud65c\uc131\ud654 \ud568\uc218, \\( W_h \\)\ub294 \uc740\ub2c9 \uc0c1\ud0dc\uc758 \uac00\uc911\uce58, \\( W_x \\)\ub294 \uc785\ub825\uc758 \uac00\uc911\uce58\uc785\ub2c8\ub2e4.\n <\/p>\n
2.2 RNN\uc758 \uc7a5\ub2e8\uc810<\/h3>\n
\n RNN\uc740 \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\ub294 \ub370 \uac15\uc810\uc774 \uc788\uc9c0\ub9cc, \uae34 \uc2dc\ud000\uc2a4\uc758 \uacbd\uc6b0 \uae30\uc6b8\uae30\uac00 \uc18c\uc2e4(vanishing gradient) \ub610\ub294 \ud3ed\uc8fc(exploding gradient)\ud558\ub294 \ubb38\uc81c\ub85c \uc778\ud574 \ud559\uc2b5\uc774 \uc5b4\ub824\uc6cc\uc9c0\ub294 \ub2e8\uc810\uc774 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ubb38\uc81c\ub97c \uadf9\ubcf5\ud558\uae30 \uc704\ud574 LSTM(Long Short-Term Memory)\uc774\ub098 GRU(Gated Recurrent Unit) \uac19\uc740 \uac1c\uc120\ub41c \uad6c\uc870\uac00 \uc0ac\uc6a9\ub429\ub2c8\ub2e4.\n <\/p>\n
3. PyTorch\ub97c \ud1b5\ud55c RNN \uad6c\ud604<\/h2>\n
\n \uc774\uc81c PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec \uae30\ubcf8\uc801\uc778 RNN \ubaa8\ub378\uc744 \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc774 \uc608\uc81c\uc5d0\uc11c\ub294 \uac04\ub2e8\ud55c \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubb38\uc81c, \uc989 \ubb38\uc7a5\uc5d0\uc11c \uac01 \ub2e8\uc5b4\uc758 \ub2e4\uc74c \ub2e8\uc5b4\ub97c \uc608\uce21\ud558\ub294 \ubb38\uc81c\ub97c \ub2e4\ub8f0 \uac83\uc785\ub2c8\ub2e4.\n <\/p>\n
3.1 \ub370\uc774\ud130 \uc900\ube44<\/h3>\n
\n \uba3c\uc800 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ubd88\ub7ec\uc624\uace0 \ub370\uc774\ud130\ub97c \uc900\ube44\ud558\uaca0\uc2b5\ub2c8\ub2e4. \uc774 \uc608\uc81c\uc5d0\uc11c\ub294 \uac04\ub2e8\ud55c \ubb38\uc7a5\uc744 \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4.\n <\/p>\n
\n import torch\n import torch.nn as nn\n import numpy as np\n from sklearn.preprocessing import OneHotEncoder\n\n # \ub370\uc774\ud130 \uc900\ube44\n sentences = ['\ub098\ub294 \ubc25\uc744 \uba39\uc5c8\ub2e4', '\ub098\ub294 \uc0ac\uacfc\ub97c \uc88b\uc544\ud55c\ub2e4', '\ub098\ub294 \ucf54\ub529\uc744 \ud55c\ub2e4']\n words = set(' '.join(sentences).split())\n word_to_index = {word: i for i, word in enumerate(words)}\n index_to_word = {i: word for i, word in enumerate(words)}\n <\/pre>\n\n \uc704\uc758 \ucf54\ub4dc\ub294 \ubb38\uc7a5\uc5d0\uc11c \ub2e8\uc5b4\ub97c \ucd94\ucd9c\ud558\uace0, \uac01 \ub2e8\uc5b4\uc5d0 \uc778\ub371\uc2a4\ub97c \ubd80\uc5ec\ud558\ub294 \uacfc\uc815\uc785\ub2c8\ub2e4. \uc774\uc81c \ub2e4\uc74c \ub2e8\uacc4\ub85c \ub118\uc5b4\uac00\uc11c \ub2e8\uc5b4\ub97c \uc6d0-\ud56b \uc778\ucf54\ub529\uc73c\ub85c \ubcc0\ud658\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n # \uc6d0-\ud56b \uc778\ucf54\ub529\n ohe = OneHotEncoder(sparse=False)\n X = []\n y = []\n\n for sentence in sentences:\n words = sentence.split()\n for i in range(len(words) - 1):\n X.append(word_to_index[words[i]])\n y.append(word_to_index[words[i + 1]])\n\n X = np.array(X).reshape(-1, 1)\n y = np.array(y).reshape(-1, 1)\n\n X_onehot = ohe.fit_transform(X)\n y_onehot = ohe.fit_transform(y)\n <\/pre>\n3.2 RNN \ubaa8\ub378 \uad6c\ucd95<\/h3>\n
\n \uc774\uc81c RNN \ubaa8\ub378\uc744 \uad6c\ucd95\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. PyTorch\uc5d0\uc11c RNN\uc740 nn.RNN<\/code> \ud074\ub798\uc2a4\ub97c \ud1b5\ud574 \uad6c\ud604\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n\n class 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)\n out, _ = self.rnn(x, h0)\n out = self.fc(out[:, -1, :])\n return out\n <\/pre>\n3.3 \ubaa8\ub378 \ud559\uc2b5<\/h3>\n
\n \ubaa8\ub378\uc744 \uc0dd\uc131\ud55c \ud6c4, \uc190\uc2e4 \ud568\uc218\uc640 \ucd5c\uc801\ud654 \uae30\ubc95\uc744 \uc124\uc815\ud558\uace0, \ud559\uc2b5\uc744 \uc9c4\ud589\ud558\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n input_size = len(words)\n hidden_size = 5\n output_size = len(words)\n\n model = RNNModel(input_size, hidden_size, output_size)\n criterion = nn.CrossEntropyLoss()\n optimizer = torch.optim.Adam(model.parameters(), lr=0.01)\n\n num_epochs = 1000\n for epoch in range(num_epochs):\n model.train()\n optimizer.zero_grad()\n\n X_tensor = torch.Tensor(X_onehot).view(-1, 1, input_size)\n y_tensor = torch.Tensor(y).long().view(-1)\n\n outputs = model(X_tensor)\n loss = criterion(outputs, y_tensor)\n loss.backward()\n optimizer.step()\n\n if (epoch + 1) % 100 == 0:\n print(f'Epoch [{epoch + 1}\/{num_epochs}], Loss: {loss.item():.4f}')\n <\/pre>\n3.4 \ubaa8\ub378 \ud3c9\uac00<\/h3>\n
\n \ud559\uc2b5\uc774 \uc644\ub8cc\ub41c \ud6c4, \ubaa8\ub378\uc744 \ud3c9\uac00\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc0c8\ub85c\uc6b4 \uc785\ub825\uc5d0 \ub300\ud574 \ub2e4\uc74c \ub2e8\uc5b4\ub97c \uc608\uce21\ud558\ub294 \uacfc\uc815\uc744 \uc124\uba85\ud569\ub2c8\ub2e4.\n <\/p>\n
\n def predict_next_word(model, current_word):\n model.eval()\n with torch.no_grad():\n input_index = word_to_index[current_word]\n input_onehot = ohe.transform([[input_index]])\n input_tensor = torch.Tensor(input_onehot).view(-1, 1, input_size)\n output = model(input_tensor)\n next_word_index = torch.argmax(output).item()\n return index_to_word[next_word_index]\n\n # \uc608\uce21\n next_word = predict_next_word(model, '\ub098\ub294')\n print(f\"\ub2e4\uc74c \ub2e8\uc5b4 \uc608\uce21: {next_word}\")\n <\/pre>\n4. \uacb0\ub860<\/h2>\n
\n \uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \uc21c\ud658 \uc2e0\uacbd\ub9dd(RNN)\uc758 \uac1c\ub150\uacfc PyTorch\ub97c \ud1b5\ud55c \uae30\ubcf8\uc801\uc778 RNN \ubaa8\ub378 \uad6c\ud604 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. RNN\uc740 \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\ub294 \ub370 \uac15\ub825\ud55c \ub3c4\uad6c\uc774\uc9c0\ub9cc \uae34 \uc2dc\ud000\uc2a4\uc5d0\uc11c\ub294 LSTM\uc774\ub098 GRU\uc640 \uac19\uc740 \ubcc0\ud615\uc774 \ud544\uc694\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
4.1 RNN\uc758 \ubc1c\uc804 \ubc29\ud5a5<\/h3>\n
\n RNN\uc740 \uae30\ubcf8\uc801\uc778 \ud615\ud0dc\uc77c \ubfd0\uc774\uba70, \ucd5c\uadfc\uc5d0\ub294 Transformer\uc640 \uac19\uc740 \ub354 \ubc1c\uc804\ub41c \ubaa8\ub378\ub4e4\uc774 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubd84\uc57c\uc5d0\uc11c \uc8fc\ubaa9\ubc1b\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c \ub354 \ub098\uc544\uac00 \uac15\ub825\ud55c \ubaa8\ub378\uc744 \ud559\uc2b5\ud558\ub824\uba74 \ub2e4\uc591\ud55c \ub525\ub7ec\ub2dd \uae30\ubc95\uacfc \uc544\ud0a4\ud14d\ucc98\uc5d0 \ub300\ud55c \uc774\ud574\uac00 \ud544\uc694\ud569\ub2c8\ub2e4.\n <\/p>\n
4.2 \ucd94\uac00 \ud559\uc2b5 \uc790\ub8cc<\/h3>\n
\n \uc21c\ud658 \uc2e0\uacbd\ub9dd\uc5d0 \ub300\ud55c \ub354 \uae4a\uc774 \uc788\ub294 \ud559\uc2b5\uc744 \uc6d0\ud558\uc2e0\ub2e4\uba74 \ub2e4\uc74c \uc790\ub8cc\ub97c \ucd94\ucc9c\ud569\ub2c8\ub2e4:\n <\/p>\n
\n- \ub525\ub7ec\ub2dd \ucc45: Ian Goodfellow, Yoshua Bengio, Aaron Courville\uc758 “Deep Learning”<\/li>\n
- PyTorch \uacf5\uc2dd \ubb38\uc11c<\/li>\n
- Coursera\uc758 \ub525\ub7ec\ub2dd \uac15\uc758<\/li>\n<\/ul>\n
5. \ucc38\uace0\ubb38\ud5cc<\/h2>\n