\ub525\ub7ec\ub2dd\uc740 \ud604\uc7ac \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \ud65c\uc6a9\ub418\uba70, \uadf8 \uc911 \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP)\ub294 \ud2b9\ud788 \ube60\ub974\uac8c \ubc1c\uc804\ud558\uace0 \uc788\ub294 \ubd84\uc57c\uc785\ub2c8\ub2e4. \ud5c8\uae45\ud398\uc774\uc2a4(Hugging Face)\ub294 \uc774\ub7ec\ud55c \ub525\ub7ec\ub2dd \ubaa8\ub378\uc744 \uc27d\uac8c \ub2e4\ub8f0 \uc218 \uc788\ub3c4\ub85d \ub2e4\uc591\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc81c\uacf5\ud558\ub294 \ud50c\ub7ab\ud3fc\uc73c\ub85c \uc720\uba85\ud569\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 Transformers \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud65c\uc6a9\ud558\uc5ec \uc0ac\uc804\ud559\uc2b5(pre-training) \ubc0f Trainer \ud074\ub798\uc2a4\ub97c \uc0ac\uc6a9\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc790\uc138\ud788 \uc124\uba85\ud558\uace0, \uc2e4\uc2b5\ud560 \uc218 \uc788\ub294 \ud30c\uc774\uc36c \ucf54\ub4dc \uc608\uc81c\ub97c \uc81c\uacf5\ud569\ub2c8\ub2e4.<\/p>\n
\ud5c8\uae45\ud398\uc774\uc2a4 Transformers\ub294 \ubcc0\ud658\uae30(Transformer) \uc544\ud0a4\ud14d\ucc98\ub97c \uae30\ubc18\uc73c\ub85c \ud55c \ub2e4\uc591\ud55c \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubaa8\ub378\uc744 \uc27d\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d \ud574\uc8fc\ub294 \ub77c\uc774\ube0c\ub7ec\ub9ac\uc785\ub2c8\ub2e4. \uc774 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 BERT, GPT-2, RoBERTa, T5 \ub4f1\uacfc \uac19\uc740 \ub2e4\uc591\ud55c \uc0ac\uc804\ud559\uc2b5 \ubaa8\ub378\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \uc6b0\ub9ac\ub294 \ubcf5\uc7a1\ud55c \ubaa8\ub378 \uad6c\ud604 \uc5c6\uc774\ub3c4 \uac04\ud3b8\ud558\uac8c \uc790\uc5f0\uc5b4 \ucc98\ub9ac \uc791\uc5c5\uc744 \uc218\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
Trainer \ud074\ub798\uc2a4\ub294 \ud5c8\uae45\ud398\uc774\uc2a4 Transformers \ub77c\uc774\ube0c\ub7ec\ub9ac\uc5d0\uc11c \uc81c\uacf5\ud558\ub294 \uace0\uc218\uc900 API\ub85c, \ubaa8\ub378 \ud6c8\ub828\uacfc \ud3c9\uac00\ub97c \uac04\ud3b8\ud558\uac8c \uc218\ud589\ud560 \uc218 \uc788\ub3c4\ub85d \ub3d5\ub294 \ud074\ub798\uc2a4\uc785\ub2c8\ub2e4. Trainer \ud074\ub798\uc2a4\ub97c \uc0ac\uc6a9\ud558\uba74 \uc0ac\uc6a9\uc790 \uc815\uc758 \ud6c8\ub828 \ub8e8\ud504\ub97c \uc791\uc131\ud558\uc9c0 \uc54a\uace0\ub3c4 \uc6d0\ud558\ub294 \ubaa8\ub378\uc744 \ud559\uc2b5\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. Trainer \ud074\ub798\uc2a4\ub97c \uc0ac\uc6a9\ud560 \ub54c\ub294 \ub370\uc774\ud130\uc14b, \ubaa8\ub378, \ud6c8\ub828 \uc778\uc790(training arguments), \uadf8\ub9ac\uace0 \ud3c9\uac00 \uc778\uc790\ub97c \uc9c0\uc815\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
\uba3c\uc800 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud574\uc57c \ud569\ub2c8\ub2e4. \uc544\ub798\uc758 \uba85\ub839\uc5b4\ub97c \uc2e4\ud589\ud558\uc5ec Transformers\uc640 \ud568\uaed8 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
!pip install transformers datasets<\/code><\/pre>\n2.2. Trainer \ud074\ub798\uc2a4 \uc0ac\uc6a9 \uc900\ube44<\/h3>\n
Trainer \ud074\ub798\uc2a4\ub97c \uc0ac\uc6a9\ud558\uae30 \uc704\ud574 \ud544\uc694\ud55c \uc900\ube44 \uc791\uc5c5\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- \ubaa8\ub378 \ub85c\ub529: \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ubaa8\ub378 \ud5c8\ube0c\uc5d0\uc11c \uc6d0\ud558\ub294 \ubaa8\ub378\uc744 \ub85c\ub529\ud569\ub2c8\ub2e4.<\/li>\n
- \ud1a0\ud06c\ub098\uc774\uc800 \uc124\uc815: \uc785\ub825 \ub370\uc774\ud130\ub97c \ubca1\ud130\ub85c \ubcc0\ud658\ud558\uae30 \uc704\ud574 \ud1a0\ud06c\ub098\uc774\uc800\ub97c \uc124\uc815\ud569\ub2c8\ub2e4.<\/li>\n
- \ub370\uc774\ud130\uc14b \uc900\ube44: \ud6c8\ub828 \ubc0f \ud3c9\uac00\uc5d0 \uc0ac\uc6a9\ud560 \ub370\uc774\ud130\uc14b\uc744 \uc900\ube44\ud569\ub2c8\ub2e4.<\/li>\n
- \ud6c8\ub828 \uc778\uc790 \uc124\uc815: \ud6c8\ub828 \uacfc\uc815\uc5d0\uc11c \uc0ac\uc6a9\ud560 \ub2e4\uc591\ud55c \uc778\uc790\ub4e4\uc744 \uc124\uc815\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n
3. \ub370\uc774\ud130\uc14b \uc900\ube44\ud558\uae30<\/h2>\n
\uc5ec\uae30\uc11c\ub294 IMDb \uc601\ud654 \ub9ac\ubdf0 \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uae0d\uc815 \ubc0f \ubd80\uc815 \ub9ac\ubdf0\ub97c \ubd84\ub958\ud558\ub294 \ubaa8\ub378\uc744 \ud559\uc2b5\ud560 \uac83\uc785\ub2c8\ub2e4. \uc774\ub97c \uc704\ud574 Hugging Face\uc758 datasets \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud65c\uc6a9\ud558\uc5ec IMDb \ub370\uc774\ud130\uc14b\uc744 \ub2e4\uc6b4\ub85c\ub4dc\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
from datasets import load_dataset\n\ndataset = load_dataset(\"imdb\")\ntrain_dataset = dataset[\"train\"]\ntest_dataset = dataset[\"test\"]<\/code><\/pre>\n4. \ubaa8\ub378\uacfc \ud1a0\ud06c\ub098\uc774\uc800 \uc124\uc815\ud558\uae30<\/h2>\n
\uc6b0\ub9ac\ub294 BERT \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud560 \uac83\uc774\uba70, Hugging Face\uc5d0\uc11c \uc81c\uacf5\ud558\ub294 ‘bert-base-uncased’ \ubaa8\ub378\uc744 \ub85c\ub4dc\ud558\uaca0\uc2b5\ub2c8\ub2e4. \ub3d9\uc2dc\uc5d0 \ud574\ub2f9 \ubaa8\ub378\uc5d0 \ub9de\ub294 \ud1a0\ud06c\ub098\uc774\uc800\ub3c4 \uc124\uc815\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
from transformers import BertTokenizer, BertForSequenceClassification\n\nmodel_name = \"bert-base-uncased\"\ntokenizer = BertTokenizer.from_pretrained(model_name)\nmodel = BertForSequenceClassification.from_pretrained(model_name)<\/code><\/pre>\n5. \ub370\uc774\ud130\uc14b \uc804\ucc98\ub9ac<\/h2>\n
\ub370\uc774\ud130\uc14b\uc744 \ubaa8\ub378\uc5d0 \ub9de\uac8c \uc804\ucc98\ub9ac\ud574\uc57c \ud569\ub2c8\ub2e4. \ud14d\uc2a4\ud2b8 \ub370\uc774\ud130\ub97c \ud1a0\ud070\ud654\ud558\uace0, \ud544\uc694\ud558\uba74 \ud328\ub529\uc744 \ucd94\uac00\ud558\uc5ec \uace0\uc815\ub41c \uae38\uc774\ub85c \ub9de\ucd94\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
def preprocess_function(examples):\n return tokenizer(examples[\"text\"], truncation=True, padding=\"max_length\", max_length=512)\n\ntrain_tokenized = train_dataset.map(preprocess_function, batched=True)\ntest_tokenized = test_dataset.map(preprocess_function, batched=True)<\/code><\/pre>\n6. Trainer \ud074\ub798\uc2a4 \uc124\uc815<\/h2>\n
\uc774\uc81c Trainer \ud074\ub798\uc2a4\ub97c \uc124\uc815\ud558\uae30 \uc704\ud574 \ud6c8\ub828 \uc778\uc790\ub97c \uc815\uc758\ud574\uc57c \ud569\ub2c8\ub2e4. \ud6c8\ub828 \uc778\uc790\ub294 \ud6c8\ub828 \uacfc\uc815\uc758 \ud558\uc774\ud37c\ud30c\ub77c\ubbf8\ud130\ub97c \uc815\uc758\ud569\ub2c8\ub2e4.<\/p>\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=8,\n per_device_eval_batch_size=8,\n num_train_epochs=3,\n weight_decay=0.01,\n)\n\ntrainer = Trainer(\n model=model,\n args=training_args,\n train_dataset=train_tokenized,\n eval_dataset=test_tokenized,\n)<\/code><\/pre>\n7. \ubaa8\ub378 \ud6c8\ub828<\/h2>\n
\ubaa8\ub378 \ud6c8\ub828\uc744 \uc2dc\uc791\ud569\ub2c8\ub2e4. \uc544\ub798 \ucf54\ub4dc\ub85c \ud6c8\ub828\uc744 \uc218\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
trainer.train()<\/code><\/pre>\n8. \ubaa8\ub378 \ud3c9\uac00<\/h2>\n
\ud6c8\ub828 \ud6c4 \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud3c9\uac00 \uba54\ud2b8\ub9ad\uc744 \ud655\uc778\ud558\uc5ec \ubaa8\ub378\uc774 \uc5bc\ub9c8\ub098 \uc798 \uc791\ub3d9\ud558\ub294\uc9c0 \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
trainer.evaluate()<\/code><\/pre>\n9. \ubaa8\ub378 \uc608\uce21<\/h2>\n
\uc774\uc81c \ud6c8\ub828\ub41c \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\uc5d0 \ub300\ud55c \uc608\uce21\uc744 \uc218\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
def predict(texts):\n inputs = tokenizer(texts, padding=True, truncation=True, return_tensors=\"pt\")\n with torch.no_grad():\n outputs = model(**inputs)\n predictions = torch.argmax(outputs.logits, dim=-1)\n return predictions\n\nsample_texts = [\"I love this movie!\", \"This is the worst film ever.\"]\npredictions = predict(sample_texts)\nprint(predictions)<\/code><\/pre>\n10. \ub9c8\ubb34\ub9ac<\/h2>\n
\ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 Transformers \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud65c\uc6a9\ud558\uc5ec Trainer \ud074\ub798\uc2a4\ub97c \uc774\uc6a9\ud55c \ubaa8\ub378 \uc0ac\uc804\ud559\uc2b5\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \ud5c8\uae45\ud398\uc774\uc2a4\ub294 \ub2e4\uc591\ud55c \uc0ac\uc804\ud559\uc2b5\ub41c \ubaa8\ub378\uc744 \uc81c\uacf5\ud558\uc5ec, \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc5d0\uc11c\uc758 \uc5ec\ub7ec \uc791\uc5c5\uc744 \uc27d\uac8c \uc218\ud589\ud560 \uc218 \uc788\ub3c4\ub85d \ub3d5\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ubc88 \uc608\uc81c\ub97c \ud1b5\ud574 \uc190\uc27d\uac8c \ubaa8\ub378\uc744 \ud6c8\ub828\ud558\uace0 \ud3c9\uac00\ud558\ub294 \ubc29\ubc95\uc744 \uc775\ud614\uae30\ub97c \ubc14\ub78d\ub2c8\ub2e4. \uc55e\uc73c\ub85c\ub3c4 \ud5c8\uae45\ud398\uc774\uc2a4\uc640 \ub525\ub7ec\ub2dd\uc758 \ub2e4\uc591\ud55c \uac00\ub2a5\uc131\uc744 \ud0d0\uc0c9\ud574 \ub098\uac00\uc2dc\uae38 \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n
\uac10\uc0ac\ud569\ub2c8\ub2e4!<\/p>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"