\uc774\ubc88 \uae00\uc5d0\uc11c\ub294 \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP)\uc5d0\uc11c \uc790\uc8fc \uc0ac\uc6a9\ub418\ub294 \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38(Hugging Face Transformers)<\/strong> \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud65c\uc6a9\ud558\uc5ec \uac10\uc131 \ubd84\uc11d\uc744 \uc218\ud589\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uac10\uc131 \ubd84\uc11d\uc740 \ud14d\uc2a4\ud2b8 \ub370\uc774\ud130\ub97c \ud1b5\ud574 \uac10\uc815\uc774\ub098 \uac10\uc131\uc744 \ucd94\ucd9c\ud558\ub294 \uae30\uc220\ub85c, \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \uc720\uc6a9\ud558\uac8c \uc0ac\uc6a9\ub429\ub2c8\ub2e4.<\/p>\n \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \ub2e4\uc591\ud55c \uc0ac\uc804 \ud559\uc2b5\ub41c \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubaa8\ub378\uc744 \uc27d\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\uac8c \ud574\uc8fc\ub294 \ud30c\uc774\uc36c \ub77c\uc774\ube0c\ub7ec\ub9ac\uc785\ub2c8\ub2e4. BERT, GPT-2, T5 \ub4f1 \uc5ec\ub7ec \uc885\ub958\uc758 \ubaa8\ub378\uc744 \uc9c0\uc6d0\ud558\uba70, \ud2b9\ud788 Fine-tuning\uc774 \uc6a9\uc774\ud558\uc5ec \ub2e4\uc591\ud55c \uc791\uc5c5\uc5d0 \ub9de\uac8c \ubaa8\ub378\uc744 \uc870\uc815\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n \uac10\uc131 \ubd84\uc11d\uc740 \uc8fc\ub85c \uc544\ub798\uc640 \uac19\uc740 \uc791\uc5c5\uc744 \ud3ec\ud568\ud569\ub2c8\ub2e4:<\/p>\n \uac10\uc131 \ubd84\uc11d\uc740 \uba38\uc2e0\ub7ec\ub2dd, \ub525\ub7ec\ub2dd\uc758 \uae30\ubc95\uc744 \ud1b5\ud574 \uad6c\ud604\ud560 \uc218 \uc788\uc73c\uba70, \ud559\uc2b5 \ub370\uc774\ud130\uc758 \ud488\uc9c8\uacfc \uc591\uc774 \uacb0\uacfc\uc5d0 \ud070 \uc601\ud5a5\uc744 \ubbf8\uce69\ub2c8\ub2e4.<\/p>\n \uc774 \uac15\uc88c\ub97c \uc9c4\ud589\ud558\uae30 \uc704\ud574 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud558\uaca0\uc2b5\ub2c8\ub2e4. \uc544\ub798\uc758 \uba85\ub839\uc5b4\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38\uc640 \ud1a0\ud070\ud654 \ub77c\uc774\ube0c\ub7ec\ub9ac\uc778 \uac10\uc131 \ubd84\uc11d\uc744 \uc704\ud55c \ub370\uc774\ud130\uc14b\uc73c\ub85c\ub294 \uc720\uba85\ud55c IMDb \uc601\ud654 \ub9ac\ubdf0 \ub370\uc774\ud130\uc14b<\/strong>\uc744 \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4. \uc774 \ub370\uc774\ud130\uc14b\uc740 \uc601\ud654\uc5d0 \ub300\ud55c \uae0d\uc815\uc801 \ub610\ub294 \ubd80\uc815\uc801\uc778 \ub9ac\ubdf0\ub97c \ud3ec\ud568\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n \ub370\uc774\ud130\ub97c \ubaa8\ub378\uc5d0 \uc785\ub825\ud560 \uc218 \uc788\ub3c4\ub85d \uc804\ucc98\ub9ac\ud558\uaca0\uc2b5\ub2c8\ub2e4. \ud14d\uc2a4\ud2b8\ub97c \uc815\uc81c\ud558\uace0, \ub808\uc774\ube14\uc744 \uc22b\uc790\ub85c \ubcc0\ud658\ud558\ub294 \uacfc\uc815\uc774 \ud544\uc694\ud569\ub2c8\ub2e4.<\/p>\n \uac10\uc131 \ubd84\uc11d\uc744 \uc704\ud55c \uc0ac\uc804 \ud559\uc2b5\ub41c BERT \ubaa8\ub378\uc744 \ubd88\ub7ec\uc624\uaca0\uc2b5\ub2c8\ub2e4. \ub610\ud55c, \ud14d\uc2a4\ud2b8\ub97c \ud1a0\ud070\ud654\ud558\uace0 \ubaa8\ub378\uc5d0 \uc785\ub825\ud560 \uc218 \uc788\ub294 \ud615\uc2dd\uc73c\ub85c \ubcc0\ud658\ud569\ub2c8\ub2e4.<\/p>\n \ubaa8\ub378\uc5d0 \uc785\ub825\ud560 \uc218 \uc788\ub3c4\ub85d \ud14d\uc2a4\ud2b8\ub97c \ud1a0\ud070\ud654\ud569\ub2c8\ub2e4. \uac01 \ub9ac\ubdf0\ub97c \ubaa8\ub378\uc5d0 \uc785\ub825\ud558\uae30 \uc704\ud574\uc11c \uc801\uc808\ud55c \ud615\ud0dc\ub85c \ubcc0\ud658\ud558\ub294 \uacfc\uc815\uc785\ub2c8\ub2e4.<\/p>\n \ubaa8\ub378\uc744 \ud559\uc2b5\uc2dc\ud0a4\uae30 \uc704\ud574\uc11c\ub294 \uc8fc\uc5b4\uc9c4 \ub370\uc774\ud130\ub85c Fine-tuning\uc744 \uc218\ud589\ud574\uc57c \ud569\ub2c8\ub2e4. \uc774\uc81c PyTorch\uc758 \ub370\uc774\ud130 \ub85c\ub354\ub97c \uc0ac\uc6a9\ud558\uc5ec \ub370\uc774\ud130\uc14b\uc744 \uad6c\uc131\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n \ubaa8\ub378\uc744 \ud6c8\ub828\ud558\uae30 \uc704\ud574 \uc190\uc2e4 \ud568\uc218\uc640 \uc635\ud2f0\ub9c8\uc774\uc800\ub97c \uc815\uc758\ud558\uace0, \uc5d0\ud3ec\ud06c\uc5d0 \uac78\uccd0 \ud6c8\ub828\uc744 \uc9c4\ud589\ud569\ub2c8\ub2e4.<\/p>\n \ubaa8\ub378\uc744 \ud3c9\uac00\ud558\uc5ec \uc131\ub2a5\uc744 \ud655\uc778\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uac80\uc99d \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc815\ud655\ub3c4\uc640 \uc190\uc2e4\uc744 \uce21\uc815\ud569\ub2c8\ub2e4.<\/p>\n \ubaa8\ub378\uc774 \ud559\uc2b5\uc774 \uc644\ub8cc\ub418\uba74 \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\uc5d0 \ub300\ud55c \uc608\uce21\uc744 \uc9c4\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc544\ub798\ub294 \uc2e4\uc81c \uc608\uce21\uc744 \ud558\ub294 \uc608\uc81c \ucf54\ub4dc\uc785\ub2c8\ub2e4.<\/p>\n \uc774 \uae00\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud65c\uc6a9\ud558\uc5ec \uac10\uc131 \ubd84\uc11d\uc744 \uc218\ud589\ud558\ub294 \uc804 \uacfc\uc815\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \ubaa8\ub378\uc744 Fine-tuning\ud558\uace0 \uc2e4\uc81c \ub370\uc774\ud130\ub97c \uc608\uce21\ud558\ub294 \uacfc\uc815\uc744 \ud1b5\ud574 \ub525\ub7ec\ub2dd \ubaa8\ub378\uc758 \ud65c\uc6a9 \uac00\ub2a5\uc131\uc744 \ud655\uc778\ud560 \uc218 \uc788\uc5c8\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c\ub3c4 \ub2e4\uc591\ud55c \uc790\uc5f0\uc5b4 \ucc98\ub9ac \uc791\uc5c5\uc5d0 \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub97c \uc751\uc6a9\ud560 \uc218 \uc788\uc744 \uac83\uc785\ub2c8\ub2e4.<\/p>\n <\/body><\/p>\n","protected":false},"excerpt":{"rendered":"1. \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub780?<\/h2>\n
2. \uac10\uc131 \ubd84\uc11d\uc758 \uac1c\uc694<\/h2>\n
\n
3. \ud658\uacbd \uc124\uc815<\/h2>\n
transformers<\/code> \ubc0f torch<\/code>\ub97c \uc124\uce58\ud558\uc138\uc694.<\/p>\npip install transformers torch<\/code><\/pre>\n4. \ub370\uc774\ud130\uc14b \uc900\ube44\ud558\uae30<\/h2>\n
from sklearn.datasets import fetch_openml\n\ndata = fetch_openml('IMDb', version=1)\ntexts, labels = data['data'], data['target']\n<\/code><\/pre>\n5. \ub370\uc774\ud130 \uc804\ucc98\ub9ac<\/h2>\n
import pandas as pd\n\ndf = pd.DataFrame({'text': texts, 'label': labels})\ndf['label'] = df['label'].apply(lambda x: 1 if x == 'pos' else 0)\ntexts = df['text'].tolist()\nlabels = df['label'].tolist()\n<\/code><\/pre>\n6. \ubaa8\ub378 \ubd88\ub7ec\uc624\uae30<\/h2>\n
from transformers import AutoTokenizer, AutoModelForSequenceClassification\n\nmodel_name = 'nlptown\/bert-base-multilingual-uncased-sentiment'\ntokenizer = AutoTokenizer.from_pretrained(model_name)\nmodel = AutoModelForSequenceClassification.from_pretrained(model_name)\n<\/code><\/pre>\n7. \ud14d\uc2a4\ud2b8 \ud1a0\ud070\ud654<\/h2>\n
encodings = tokenizer(texts, truncation=True, padding=True, max_length=128, return_tensors=\"pt\")\n<\/code><\/pre>\n8. \ubaa8\ub378 \ud559\uc2b5<\/h2>\n
import torch\nfrom torch.utils.data import DataLoader, Dataset\n\nclass SentimentDataset(Dataset):\n def __init__(self, encodings, labels):\n self.encodings = encodings\n self.labels = labels\n\n def __getitem__(self, idx):\n item = {key: val[idx] for key, val in self.encodings.items()}\n item['labels'] = torch.tensor(self.labels[idx])\n return item\n\n def __len__(self):\n return len(self.labels)\n\ndataset = SentimentDataset(encodings, labels)\ntrain_loader = DataLoader(dataset, batch_size=16, shuffle=True)\n<\/code><\/pre>\n9. \ubaa8\ub378 \ud6c8\ub828<\/h2>\n
from transformers import AdamW\n\noptimizer = AdamW(model.parameters(), lr=5e-5)\n\nmodel.train()\nfor epoch in range(3):\n for batch in train_loader:\n optimizer.zero_grad()\n outputs = model(**batch)\n loss = outputs.loss\n loss.backward()\n optimizer.step()\n print(f'Epoch: {epoch}, Loss: {loss.item()}')\n<\/code><\/pre>\n10. \ubaa8\ub378 \ud3c9\uac00<\/h2>\n
model.eval()\ncorrect = 0\ntotal = 0\n\nwith torch.no_grad():\n for batch in train_loader:\n outputs = model(**batch)\n predictions = outputs.logits.argmax(dim=-1)\n correct += (predictions == batch['labels']).sum().item()\n total += batch['labels'].size(0)\n\naccuracy = correct \/ total\nprint(f'Accuracy: {accuracy}')\n<\/code><\/pre>\n11. \uc608\uce21\ud558\uae30<\/h2>\n
def predict_sentiment(text):\n inputs = tokenizer(text, return_tensors='pt', truncation=True, padding=True, max_length=128)\n with torch.no_grad():\n outputs = model(**inputs)\n prediction = outputs.logits.argmax(dim=-1)\n return '\uae0d\uc815' if prediction.item() == 1 else '\ubd80\uc815'\n\ntest_text = \"\uc774 \uc601\ud654\ub294 \uc815\ub9d0 \uc7ac\ubbf8\uc788\uc5c8\uc2b5\ub2c8\ub2e4!\"\nprint(f'\uc608\uce21: {predict_sentiment(test_text)}')\n<\/code><\/pre>\n12. \uacb0\ub860<\/h2>\n
13. \ucc38\uace0 \ubb38\ud5cc<\/h2>\n
\n