\n\uc21c\ud658 \uc2e0\uacbd\ub9dd(RNN)\uc740 \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\ub294 \ub370 \uac15\ub825\ud55c \ub2a5\ub825\uc744 \uac00\uc9c4 \ub525\ub7ec\ub2dd \ubaa8\ub378\uc785\ub2c8\ub2e4. \uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 RNN\uc758 \uae30\ucd08 \uac1c\ub150\ubd80\ud130 \uc2dc\uc791\ud558\uc5ec, \ud30c\uc774\ud1a0\uce58\ub85c RNN\uc744 \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc744 \uc790\uc138\ud788 \uc124\uba85\ud569\ub2c8\ub2e4.<\/p>\n<\/header>\n\n1. \uc21c\ud658 \uc2e0\uacbd\ub9dd(RNN)\uc758 \uac1c\uc694<\/h2>\n
\n \uc21c\ud658 \uc2e0\uacbd\ub9dd(RNN)\uc740 \uc774\uc804\uc758 \uc815\ubcf4\uac00 \ud604\uc7ac\uc758 \uc815\ubcf4\uc5d0 \uc601\ud5a5\uc744 \ubbf8\uce60 \uc218 \uc788\ub3c4\ub85d \uc124\uacc4\ub41c \uc2e0\uacbd\ub9dd \uad6c\uc870\uc785\ub2c8\ub2e4. \uc774\ub294 \uc8fc\ub85c \uc2dc\ud000\uc2a4 \ub370\uc774\ud130(\uc608: \uc790\uc5f0\uc5b4 \ud14d\uc2a4\ud2b8, \uc2dc\uacc4\uc5f4 \ub370\uc774\ud130) \ucc98\ub9ac\uc5d0 \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \uc804\ud1b5\uc801\uc778 \uc2e0\uacbd\ub9dd\uc740 \uc785\ub825 \ub370\uc774\ud130\uc758 \ub3c5\ub9bd\uc131\uc744 \uac00\uc815\ud558\uc9c0\ub9cc, RNN\uc740 \uc2dc\uac04\uc5d0 \ub530\ub978 \uc885\uc18d\uc131\uc744 \ud559\uc2b5\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n RNN\uc758 \uae30\ubcf8 \uad6c\uc870\uc5d0\uc11c \uac01 \uc2dc\uc810\uc744 \ub098\ud0c0\ub0b4\ub294 \uc785\ub825\uc740 \uc774\uc804 \uc2dc\uc810\uc758 \uc740\ub2c9 \uc0c1\ud0dc\uc640 \ud568\uaed8 \ubaa8\ub378\uc5d0 \uc785\ub825\ub429\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uc5f0\uacb0\ub85c \uc778\ud574 RNN\uc740 \uc815\ubcf4\uc758 \ud750\ub984\uc744 \uc2dc\ud000\uc2a4\uc5d0 \ub530\ub77c \ucc98\ub9ac\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n<\/section>\n\n2. RNN\uc758 \uad6c\uc870<\/h2>\n
\n RNN\uc758 \uae30\ubcf8 \uad6c\uc870\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:\n <\/p>\n
\n- \uc785\ub825\uce35: \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\ub97c \uc785\ub825\ubc1b\uc2b5\ub2c8\ub2e4.<\/li>\n
- \uc740\ub2c9\uce35: \uc2dc\uac04\uc801\uc73c\ub85c \uc5f0\uacb0\ub41c \uc5ec\ub7ec \uce35\uc744 \uac00\uc9d1\ub2c8\ub2e4.<\/li>\n
- \ucd9c\ub825\uce35: \ucd5c\uc885 \uc608\uce21 \uacb0\uacfc\ub97c \uc81c\uacf5\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n
![\"RNN](\"rnn_structure.png\")
<\/p>\n
\n\uc218\ud559\uc801 \ud45c\ud604:<\/strong> RNN\uc758 \uc5c5\ub370\uc774\ud2b8\ub294 \ub2e4\uc74c\uacfc \uac19\uc774 \ud45c\ud604\ub429\ub2c8\ub2e4:
\n
\n ht<\/sub> = f(Whh<\/sub>ht-1<\/sub> + Wxh<\/sub>xt<\/sub> + bh<\/sub>)
\n
\n yt<\/sub> = Why<\/sub>ht<\/sub> + by<\/sub>\n<\/p>\n<\/section>\n\n3. RNN\uc758 \ud55c\uacc4<\/h2>\n
\n \uc804\ud1b5\uc801\uc778 RNN\uc740 \uae34 \uc2dc\ud000\uc2a4\uc5d0 \ub300\ud55c \uc758\uc874\uc131\uc744 \ud559\uc2b5\ud558\ub294 \ub370 \ud55c\uacc4\uac00 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub294 \uae30\uc6b8\uae30 \uc18c\uc2e4(vanishing gradient) \ubb38\uc81c\ub85c \uc774\uc5b4\uc9c0\uba70, \uc774\uc5d0 \ub300\ud55c \ud574\uacb0\ucc45\uc73c\ub85c LSTM(\uc7a5\uae30 \ub2e8\uae30 \uba54\ubaa8\ub9ac)\uc640 GRU(\uac8c\uc774\ud2f0\ub4dc \uc21c\ud658 \uc720\ub2db) \uac19\uc740 \ub2e4\uc591\ud55c RNN \ubcc0\ud615\uc774 \uc81c\uc548\ub418\uc5c8\uc2b5\ub2c8\ub2e4.\n <\/p>\n<\/section>\n\n4. PyTorch\ub85c RNN \uad6c\ud604\ud558\uae30<\/h2>\n
\n \uc774\ubc88 \uc139\uc158\uc5d0\uc11c\ub294 \ud30c\uc774\ud1a0\uce58\ub97c \uc0ac\uc6a9\ud558\uc5ec \uac04\ub2e8\ud55c RNN \ubaa8\ub378\uc744 \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc6b0\ub9ac\ub294 \uc720\uba85\ud55c IMDB \uc601\ud654 \ub9ac\ubdf0 \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc601\ud654 \ub9ac\ubdf0\uc758 \uae0d\uc815\/\ubd80\uc815 \uac10\uc815\uc744 \ubd84\ub958\ud558\ub294 \uc791\uc5c5\uc744 \uc218\ud589\ud560 \uac83\uc785\ub2c8\ub2e4.\n <\/p>\n<\/section>\n\n4.1 \ub370\uc774\ud130 \ub85c\ub4dc \ubc0f \uc804\ucc98\ub9ac<\/h3>\n
\n \ud30c\uc774\ud1a0\uce58\uc758 torchtext \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud558\uc5ec IMDB \ub370\uc774\ud130\ub97c \ub85c\ub4dc\ud558\uace0 \uc804\ucc98\ub9ac\ud569\ub2c8\ub2e4.\n <\/p>\n
\nimport torch\nfrom torchtext.datasets import IMDB\nfrom torchtext.data import Field, BucketIterator\n\nTEXT = Field(tokenize='spacy', include_lengths=True)\nLABEL = Field(dtype=torch.float)\n\ntrain_data, test_data = IMDB.splits(TEXT, LABEL)\nTEXT.build_vocab(train_data, max_size=25000)\nLABEL.build_vocab(train_data)\n\ntrain_iterator, test_iterator = BucketIterator.splits(\n (train_data, test_data), \n batch_size=64, \n sort_within_batch=True)\n <\/code><\/pre>\n\n \uc704 \ucf54\ub4dc\ub294 IMDB \ub370\uc774\ud130\uc14b\uc744 \ub85c\ub4dc\ud558\uace0, \uc774\ub97c \uc704\ud55c \ud14d\uc2a4\ud2b8\uc640 \ub808\uc774\ube14 \ud544\ub4dc\ub97c \uc815\uc758\ud558\uc5ec \ub370\uc774\ud130\uc14b\uc744 \uc804\ucc98\ub9ac\ud558\ub294 \uacfc\uc815\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.\n <\/p>\n<\/section>\n\n4.2 RNN \ubaa8\ub378 \uc815\uc758\ud558\uae30<\/h3>\n
\n RNN \ubaa8\ub378\uc744 \uc815\uc758\ud569\ub2c8\ub2e4. \ud30c\uc774\ud1a0\uce58\uc758 nn.Module\uc744 \uc0c1\uc18d\ubc1b\uc544 \uae30\ubcf8 \ubaa8\ub378\uc744 \uad6c\ud604\ud558\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\nimport torch.nn as nn\n\nclass RNN(nn.Module):\n def __init__(self, input_dim, emb_dim, hidden_dim, output_dim):\n super().__init__()\n \n self.embedding = nn.Embedding(input_dim, emb_dim)\n self.rnn = nn.RNN(emb_dim, hidden_dim)\n self.fc = nn.Linear(hidden_dim, output_dim)\n self.dropout = nn.Dropout(0.5)\n \n def forward(self, text, text_length):\n embedded = self.dropout(self.embedding(text))\n packed_embedded = nn.utils.rnn.pack_padded_sequence(embedded, text_length)\n packed_output, hidden = self.rnn(packed_embedded)\n output, output_length = nn.utils.rnn.pad_packed_sequence(packed_output)\n return self.fc(hidden.squeeze(0))\n <\/code><\/pre>\n\n \uc774 \ucf54\ub4dc\uc5d0\uc11c\ub294 \uc785\ub825 \ucc28\uc6d0, \uc784\ubca0\ub529 \ucc28\uc6d0, \uc740\ub2c9 \ucc28\uc6d0, \ucd9c\ub825 \ucc28\uc6d0\uc744 \uc778\uc790\ub85c \ubc1b\uc544 RNN \ubaa8\ub378\uc744 \uad6c\uc131\ud569\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc740 \uc784\ubca0\ub529 \ub808\uc774\uc5b4, RNN \ub808\uc774\uc5b4, \uadf8\ub9ac\uace0 \ucd9c\ub825 \ub808\uc774\uc5b4\ub85c \uc774\ub8e8\uc5b4\uc838 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n<\/section>\n\n4.3 \ubaa8\ub378 \ud6c8\ub828\ud558\uae30<\/h3>\n
\n \uc774\uc81c \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0a4\uae30 \uc704\ud55c \uacfc\uc815\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc190\uc2e4 \ud568\uc218\ub85c\ub294 \uc774\uc9c4 \uad50\ucc28 \uc5d4\ud2b8\ub85c\ud53c\ub97c \uc0ac\uc6a9\ud558\uace0, \ucd5c\uc801\ud654 \uae30\ubc95\uc73c\ub85c Adam\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4.\n <\/p>\n
\nimport torch.optim as optim\n\ndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n\nmodel = RNN(len(TEXT.vocab), 100, 256, 1)\nmodel = model.to(device)\n\noptimizer = optim.Adam(model.parameters())\ncriterion = nn.BCEWithLogitsLoss()\ncriterion = criterion.to(device)\n\ndef train(model, iterator, optimizer, criterion):\n model.train()\n epoch_loss = 0\n \n for batch in iterator:\n text, text_length = batch.text\n labels = batch.label\n \n optimizer.zero_grad()\n predictions = model(text, text_length).squeeze(1)\n \n loss = criterion(predictions, labels)\n loss.backward()\n optimizer.step()\n \n epoch_loss += loss.item()\n \n return epoch_loss \/ len(iterator)\n <\/code><\/pre>\n\n train \ud568\uc218\ub294 \uc8fc\uc5b4\uc9c4 \ub370\uc774\ud130 \ubc30\uce58\uc5d0 \ub300\ud574 \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0a4\uace0 \uc190\uc2e4\uc744 \ubc18\ud658\ud569\ub2c8\ub2e4.\n <\/p>\n<\/section>\n\n4.4 \ubaa8\ub378 \ud3c9\uac00\ud558\uae30<\/h3>\n
\n \ubaa8\ub378\uc744 \ud3c9\uac00\ud558\ub294 \ud568\uc218\ub3c4 \uc815\uc758\ud560 \ud544\uc694\uac00 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c \ucf54\ub4dc\ub97c \ud1b5\ud574 \ud3c9\uac00\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\ndef evaluate(model, iterator, criterion):\n model.eval()\n epoch_loss = 0\n \n with torch.no_grad():\n for batch in iterator:\n text, text_length = batch.text\n labels = batch.label\n \n predictions = model(text, text_length).squeeze(1)\n loss = criterion(predictions, labels)\n epoch_loss += loss.item()\n \n return epoch_loss \/ len(iterator)\n <\/code><\/pre>\n\n evaluate \ud568\uc218\ub294 \ud3c9\uac00 \ub370\uc774\ud130\uc5d0 \ub300\ud574 \ubaa8\ub378\uc744 \ud3c9\uac00\ud558\uace0 \uc190\uc2e4 \uac12\uc744 \ubc18\ud658\ud569\ub2c8\ub2e4.\n <\/p>\n<\/section>\n\n4.5 \ud6c8\ub828 \ubc0f \ud3c9\uac00 \ub8e8\ud504<\/h3>\n
\n \ub9c8\uc9c0\ub9c9\uc73c\ub85c, \ud6c8\ub828 \ubc0f \ud3c9\uac00 \ub8e8\ud504\ub97c \uc791\uc131\ud558\uc5ec \ubaa8\ub378\uc758 \ud559\uc2b5\uc744 \uc218\ud589\ud569\ub2c8\ub2e4.\n <\/p>\n
\nN_EPOCHS = 5\n\nfor epoch in range(N_EPOCHS):\n train_loss = train(model, train_iterator, optimizer, criterion)\n valid_loss = evaluate(model, test_iterator, criterion)\n\n print(f'Epoch: {epoch+1:02}, Train Loss: {train_loss:.3f}, Valid Loss: {valid_loss:.3f}')\n <\/code><\/pre>\n\n \uc774 \ub8e8\ud504\ub294 \uc8fc\uc5b4\uc9c4 \uc5d0\ud3ec\ud06c \uc218\uc5d0 \ub530\ub77c \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0a4\uace0 \ub9e4 \uc5d0\ud3ec\ud06c\ub9c8\ub2e4 \ud6c8\ub828 \uc190\uc2e4\uacfc \uac80\uc99d \uc190\uc2e4\uc744 \ucd9c\ub825\ud569\ub2c8\ub2e4.\n <\/p>\n<\/section>\n\n5. \uacb0\ub860<\/h2>\n
\n \uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \uc21c\ud658 \uc2e0\uacbd\ub9dd(RNN)\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \ud30c\uc774\ud1a0\uce58\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc774 \ubaa8\ub378\uc744 \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc744 \ubc30\uc6e0\uc2b5\ub2c8\ub2e4. RNN\uc740 \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\ub294 \ub370 \ud6a8\uacfc\uc801\uc774\uc9c0\ub9cc, \uae34 \uc2dc\ud000\uc2a4\uc5d0 \ub300\ud55c \ud55c\uacc4\uac00 \uc874\uc7ac\ud569\ub2c8\ub2e4. \uc774\uc5d0 \ub530\ub77c LSTM\uacfc GRU\uc640 \uac19\uc740 \ubcc0\ud615 \ubaa8\ub378\uc744 \uace0\ub824\ud560 \ud544\uc694\uac00 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \ub0b4\uc6a9\uc744 \uae30\ubc18\uc73c\ub85c \ub354 \ub098\uc544\uac00 \ub2e4\uc591\ud55c \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\uc5d0 \ub300\ud55c \uc2e4\ud5d8\uc744 \ud574\ubcf4\ub294 \uac83\ub3c4 \uc88b\uc740 \ud559\uc2b5\uc774 \ub420 \uac83\uc785\ub2c8\ub2e4.\n <\/p>\n<\/section>\n