\ub525\ub7ec\ub2dd\uacfc \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP)\uc758 \ubc1c\uc804\uc740 \uc624\ub298\ub0a0\uc758 \uae30\uc220 \ud601\uc2e0\uc744 \uc774\ub044\ub294 \uc911\uc694\ud55c \uc694\uc18c \uc911 \ud558\ub098\uc785\ub2c8\ub2e4. \uc5ec\ub7ec \ub77c\uc774\ube0c\ub7ec\ub9ac\uc640 \ud504\ub808\uc784\uc6cc\ud06c\uac00 \uc874\uc7ac\ud558\uc9c0\ub9cc, \ud5c8\uae45\ud398\uc774\uc2a4(Hugging Face)<\/strong>\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38(transformers)<\/strong> \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \ud2b9\ud788 \uc9c1\uad00\uc801\uc774\uace0 \uc0ac\uc6a9\ud558\uae30 \uc27d\uac8c \uc124\uacc4\ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \uae00\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub97c \ud65c\uc6a9\ud558\uc5ec \ubb38\uc11c \ubd84\ub958(Classification) \ubaa8\ub378\uc744 \uad6c\ucd95\ud558\uace0, \ubaa8\ub378 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \ub2e4\ub8f0 \uac83\uc785\ub2c8\ub2e4.<\/p>\n \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \ub2e4\uc591\ud55c \ubaa8\ub378 \uc544\ud0a4\ud14d\ucc98\ub97c \uc9c0\uc6d0\ud558\uba70, Pre-trained \ubaa8\ub378\uc744 \uc27d\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d \ud574\uc90d\ub2c8\ub2e4. \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub294 \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc5d0 \ud601\uc2e0\uc744 \uc77c\uc73c\ud0a8 \ubaa8\ub378\ub85c, Attention Mechanism<\/strong>\uc744 \uae30\ubc18\uc73c\ub85c \ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \ubaa8\ub378\ub4e4\uc740 \ub300\uaddc\ubaa8 \ub370\uc774\ud130\uc14b\uc73c\ub85c \uc0ac\uc804 \ud559\uc2b5\ub418\uc5b4, \ud2b9\uc815 \uc791\uc5c5\uc5d0 \ub300\ud574 \ubbf8\uc138 \uc870\uc815(Fine-tuning)\ud558\uc5ec \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub97c \uc0ac\uc6a9\ud558\uae30 \uc704\ud574 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud558\uaca0\uc2b5\ub2c8\ub2e4. \uc8fc\ub85c \uc0ac\uc6a9\ud560 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 transformers<\/strong>, torch<\/strong>, datasets<\/strong>\uc785\ub2c8\ub2e4. \ub2e4\uc74c\uc758 \uba85\ub839\uc5b4\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc124\uce58\ud558\uc138\uc694:<\/p>\n \ubb38\uc11c \ubd84\ub958\ub97c \uc704\ud574 \uc0ac\uc6a9\ud560 \ub370\uc774\ud130\uc14b\uc744 \uc900\ube44\ud569\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 AG News<\/strong> \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4. AG News\ub294 \ub274\uc2a4 \uae30\uc0ac \ubd84\ub958\ub97c \uc704\ud55c \ub370\uc774\ud130\uc14b\uc73c\ub85c, 4\uac1c\uc758 \ud074\ub798\uc2a4\ub97c \uac16\uace0 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n \ub2e4\uc74c\uc758 \ucf54\ub4dc\ub97c \uc2e4\ud589\ud558\uba74 \ub370\uc774\ud130\uc14b\uc744 \ub2e4\uc6b4\ub85c\ub4dc\ubc1b\uace0, \ud559\uc2b5 \ubc0f \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\ub85c \ub098\ub215\ub2c8\ub2e4.<\/p>\n \ub370\uc774\ud130\ub97c \ubd88\ub7ec\uc628 \ud6c4\uc5d0\ub294 \ud14d\uc2a4\ud2b8\uc640 \ub808\uc774\ube14\uc744 \ubd84\ub9ac\ud558\uace0, \ud544\uc694\ud55c \uc804\ucc98\ub9ac\ub97c \uc9c4\ud589\ud574\uc57c \ud569\ub2c8\ub2e4. \ub2e4\uc74c \ucf54\ub4dc\ub294 \uc0d8\ud50c \ub370\uc774\ud130\uc640 \ub77c\ubca8\uc744 \ud655\uc778\ud558\ub294 \uacfc\uc815\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<\/p>\n \uc774\uc81c transformers<\/strong> \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud1b5\ud574 \uc0ac\uc804 \ud559\uc2b5\ub41c \ubaa8\ub378\uacfc \ud1a0\ud06c\ub098\uc774\uc800\ub97c \ubd88\ub7ec\uc635\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 BertForSequenceClassification<\/strong> \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4.<\/p>\n \ubb38\uc11c \ubd84\ub958\ub97c \uc704\ud574 \uac01 \ud14d\uc2a4\ud2b8\ub97c BERT \ubaa8\ub378\uc5d0 \ub9de\uac8c \ud1a0\ud070\ud654\ud569\ub2c8\ub2e4. \ub2e4\uc74c \ucf54\ub4dc\uc5d0\uc11c\ub294 Padding\uc744 \ucd94\uac00\ud574 \ubc30\uce58 \ucc98\ub9ac\uc5d0 \uc6a9\uc774\ud558\ub3c4\ub85d \ub9cc\ub4ed\ub2c8\ub2e4.<\/p>\n \ubaa8\ub378\uc744 \ud559\uc2b5\ud558\uae30 \uc704\ud574 Trainer<\/strong> \ud074\ub798\uc2a4\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4. Trainer\ub294 \ud559\uc2b5 \ubc0f \ud3c9\uac00\ub97c \uc790\ub3d9\uc73c\ub85c \ucc98\ub9ac\ud574 \uc90d\ub2c8\ub2e4. \ub2e4\uc74c \ucf54\ub4dc\ub294 \uc124\uc815 \ubc0f \ud559\uc2b5\uc744 \uc704\ud55c \uc900\ube44 \uacfc\uc815\uc744 \ud3ec\ud568\ud569\ub2c8\ub2e4.<\/p>\n \ubaa8\ub378 \ud559\uc2b5\uc774 \uc644\ub8cc\ub418\uba74, \ud3c9\uac00 \ud568\uc218\ub97c \ud1b5\ud574 \uc131\ub2a5\uc744 \uce21\uc815\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. metrics<\/strong> \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud65c\uc6a9\ud558\uc5ec \uc815\ud655\ub3c4\ub97c \uacc4\uc0b0\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub97c \uc0ac\uc6a9\ud558\uc5ec \ub370\uc774\ud130\uc14b\uc744 \ubd88\ub7ec\uc624\uace0, \uc0ac\uc804 \ud559\uc2b5\ub41c \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\uc5ec \ud14d\uc2a4\ud2b8 \ubd84\ub958\ub97c \uc218\ud589\ud558\ub294 \ubc29\ubc95\uc744 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \uc774 \uacfc\uc815\uc744 \ud1b5\ud574 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \uc791\uc5c5\uc5d0\uc11c \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ubaa8\ub378\uc758 \uc720\uc6a9\uc131\uc744 \uc54c \uc218 \uc788\uc5c8\uc2b5\ub2c8\ub2e4. \ucd94\uac00\uc801\uc73c\ub85c Hyperparameter \ud29c\ub2dd\uc774\ub098 \ub2e4\uc591\ud55c \ubaa8\ub378\uc744 \uc2dc\ub3c4\ud574\ubcf4\uba74 \uc131\ub2a5\uc744 \ub354\uc6b1 \ud5a5\uc0c1\uc2dc\ud0ac \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n \ub354 \ub9ce\uc740 \uc815\ubcf4\uc640 \uc608\uc81c\ub97c \uc6d0\ud558\ub294 \ub3c5\uc790\ub97c \uc704\ud574 \ub2e4\uc74c\uc758 \uc790\ub8cc\ub97c \ucd94\ucc9c\ud569\ub2c8\ub2e4:<\/p>\n1. \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub780?<\/h2>\n
2. \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac \uc124\uce58\ud558\uae30<\/h2>\n
!pip install transformers torch datasets<\/code><\/pre>\n3. \ub370\uc774\ud130\uc14b \uc900\ube44\ud558\uae30<\/h2>\n
\n
from datasets import load_dataset\n\ndataset = load_dataset(\"ag_news\")<\/code><\/pre>\n4. \ub370\uc774\ud130 \uc804\ucc98\ub9ac<\/h2>\n
train_texts = dataset['train']['text']\ntrain_labels = dataset['train']['label']\n\ntest_texts = dataset['test']['text']\ntest_labels = dataset['test']['label']\n\nprint(\"\uc0d8\ud50c \ub274\uc2a4 \uae30\uc0ac:\", train_texts[0])\nprint(\"\ub77c\ubca8:\", train_labels[0])<\/code><\/pre>\n5. \ubaa8\ub378 \ubc0f \ud1a0\ud06c\ub098\uc774\uc800 \uc900\ube44\ud558\uae30<\/h2>\n
from transformers import BertTokenizer, BertForSequenceClassification\n\nmodel_name = \"bert-base-uncased\"\ntokenizer = BertTokenizer.from_pretrained(model_name)\nmodel = BertForSequenceClassification.from_pretrained(model_name, num_labels=4)<\/code><\/pre>\n6. \ub370\uc774\ud130 \ud1a0\ud070\ud654<\/h2>\n
def tokenize_function(examples):\n return tokenizer(examples[\"text\"], padding=\"max_length\", truncation=True)\n\ntokenized_datasets = dataset.map(tokenize_function, batched=True)<\/code><\/pre>\n7. \ubaa8\ub378 \ud559\uc2b5\ud558\uae30<\/h2>\n
from transformers import Trainer, TrainingArguments\n\ntraining_args = TrainingArguments(\n output_dir=\".\/results\",\n evaluation_strategy=\"epoch\",\n learning_rate=2e-5,\n per_device_train_batch_size=16,\n per_device_eval_batch_size=16,\n num_train_epochs=3,\n weight_decay=0.01,\n)\n\ntrainer = Trainer(\n model=model,\n args=training_args,\n train_dataset=tokenized_datasets['train'],\n eval_dataset=tokenized_datasets['test']\n)\n\ntrainer.train()<\/code><\/pre>\n8. \ubaa8\ub378 \ud3c9\uac00\ud558\uae30<\/h2>\n
import numpy as np\nfrom sklearn.metrics import accuracy_score\n\npredictions, label_ids, _ = trainer.predict(tokenized_datasets['test'])\npreds = np.argmax(predictions, axis=1)\n\naccuracy = accuracy_score(label_ids, preds)\nprint(\"\ubd84\ub958 \uc815\ud655\ub3c4:\", accuracy)<\/code><\/pre>\n9. \uacb0\ub860<\/h2>\n
10. \ucc38\uace0 \uc790\ub8cc<\/h2>\n