\ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38(Transformers)<\/strong> \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud65c\uc6a9\ud558\uc5ec DistilGPT-2<\/strong> \ubaa8\ub378\uc744 \uc774\uc6a9\ud55c \ubb38\uc7a5 \uc791\uc131\uc744 \uc2e4\uc2b5\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. DistilGPT-2\ub294 OpenAI\uc758 GPT-2 \ubaa8\ub378\uc744 \uae30\ubc18\uc73c\ub85c \ud558\uc5ec \uacbd\ub7c9\ud654\ub41c \ubaa8\ub378\ub85c, \ube60\ub974\uace0 \ud6a8\uc728\uc801\uc73c\ub85c \ud14d\uc2a4\ud2b8 \uc0dd\uc131\uc744 \uc218\ud589\ud558\ub294 \ub370 \ucd5c\uc801\ud654\ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc740 \ub2e4\uc591\ud55c \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP) \ud0dc\uc2a4\ud06c\uc5d0\uc11c \uc0ac\uc6a9\ub420 \uc218 \uc788\uc73c\uba70, \ud2b9\ud788 \uc790\ub3d9 \uc791\ubb38\uc5d0 \ub6f0\uc5b4\ub09c \uc131\ub2a5\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<\/p>\n DistilGPT-2 \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uae30 \uc704\ud574 \uc544\ub798\uc758 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n \uc704\uc758 \uba85\ub839\uc5b4\ub97c \ud1b5\ud574 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ubaa8\ub450 \uc124\uce58\ud569\ub2c8\ub2e4.<\/p>\n \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uae30 \uc704\ud574 \uba3c\uc800 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc784\ud3ec\ud2b8\ud558\uace0, DistilGPT-2 \ubaa8\ub378\uacfc \ud1a0\ud06c\ub098\uc774\uc800\ub97c \ub85c\ub4dc\ud569\ub2c8\ub2e4.<\/p>\n \uc774 \ucf54\ub4dc\ub294 distilgpt2<\/strong>\ub77c\ub294 \uc0ac\uc804 \ud559\uc2b5\ub41c \ubaa8\ub378\uc744 \ub85c\ub4dc\ud569\ub2c8\ub2e4. \uc774\uc81c \uc6b0\ub9ac\ub294 \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud574 \ud14d\uc2a4\ud2b8\ub97c \uc0dd\uc131\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n \uc774\uc81c \uc900\ube44\ub41c \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud574 \ud14d\uc2a4\ud2b8\ub97c \uc0dd\uc131\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \ubb38\uc7a5\uc758 \uc2dc\uc791 \ubd80\ubd84\uc744 \uc8fc\uba74, \ubaa8\ub378\uc774 \uadf8\uc5d0 \uae30\ubc18\ud558\uc5ec \ub2e4\uc74c \ub2e8\uc5b4\ub4e4\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4.<\/p>\n \uc704 \ucf54\ub4dc\uc5d0\uc11c \ud14d\uc2a4\ud2b8 \uc0dd\uc131 \uc2dc \uc5ec\ub7ec \uac00\uc9c0 \ud30c\ub77c\ubbf8\ud130\ub97c \uc870\uc815\ud558\uc5ec \uc0dd\uc131\ub41c \uacb0\uacfc\uc5d0 \ucc28\uc774\ub97c \uc904 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc5ec\uae30\uc5d0 \uc8fc\uc694 \ud30c\ub77c\ubbf8\ud130\ub97c \uc18c\uac1c\ud569\ub2c8\ub2e4:<\/p>\n \uc774\uc81c \uc774\ub7ec\ud55c \ud30c\ub77c\ubbf8\ud130\ub97c \uc0ac\uc6a9\ud574 \ud14d\uc2a4\ud2b8\ub97c \uc0dd\uc131\ud558\ub294 \uc608\uc81c\ub97c \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4:<\/p>\n \uc704\uc758 \uc774\ubc88\uc5d0\ub294 \uc704\uc5d0\uc11c \ud559\uc2b5\ud55c \ub0b4\uc6a9\uc744 \ubc14\ud0d5\uc73c\ub85c \uc9e7\uc740 \uc18c\uc124\uc744 \uc791\uc131\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc0ac\uc6a9\uc790\uac00 \uc81c\uacf5\ud55c \ud504\ub86c\ud504\ud2b8\uc5d0 \ub530\ub77c \uc774\uc57c\uae30\uc758 \uc804\uac1c\ub97c \uc0dd\uc131\ud574\ubcf4\uc138\uc694.<\/p>\n \uc774 \ucf54\ub4dc\ub294 \uc8fc\uc5b4\uc9c4 \ud504\ub86c\ud504\ud2b8\uc5d0\uc11c \uc2dc\uc791\ud558\uc5ec \ucd5c\ub300 150\uc790\uc758 \ud14d\uc2a4\ud2b8\ub97c \uc0dd\uc131\ud569\ub2c8\ub2e4. \ud15c\ud37c\ub7ec\uccd0\ub97c \ub192\uac8c \uc124\uc815\ud558\uc600\uae30\uc5d0 \ubcf4\ub2e4 \ucc3d\uc758\uc801\uc778 \ub0b4\uc6a9\uc774 \uc0dd\uc131\ub420 \uac83\uc785\ub2c8\ub2e4. \uc0dd\uc131\ub41c \uc18c\uc124\uc744 \ud1b5\ud574 DistilGPT-2\uc758 \ud14d\uc2a4\ud2b8 \uc0dd\uc131 \ub2a5\ub825\uc744 \uccb4\ud5d8\ud574\ubcfc \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n \uc704\uc5d0\uc11c \uc0dd\uc131\ud55c \ubb38\uc7a5\uc744 \uc77d\uc5b4\ubcf4\uace0, \uc0dd\uc131\ub41c \ud14d\uc2a4\ud2b8\uac00 \uc5b4\ub5a4 \ub290\ub08c\uc744 \uc8fc\ub294\uc9c0, \uadf8\ub9ac\uace0 \ubaa8\ub378\uc774 \uc5b4\ub5a4 \ubd80\ubd84\uc5d0\uc11c \uae30\ub300\ub97c \ub6f0\uc5b4\ub118\uac70\ub098 \ubd80\uc871\ud588\ub294\uc9c0 \ud53c\ub4dc\ubc31\uc744 \uc81c\uacf5\ud558\uc138\uc694. \uc774 \ud53c\ub4dc\ubc31\uc740 \ubaa8\ub378\uc758 \uac1c\uc120\uc810 \ubc0f \ud65c\uc6a9 \ubc29\uc548\uc744 \ucc3e\ub294 \ub370 \ub3c4\uc6c0\uc774 \ub420 \uac83\uc785\ub2c8\ub2e4.<\/p>\n \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc774\uc6a9\ud558\uc5ec DistilGPT-2 \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \uc790\uc5f0\uc5b4 \uc0dd\uc131(NLG) \ubd84\uc57c\uc5d0\uc11c \uc774 \ubaa8\ub378\uc740 \ub180\ub77c\uc6b4 \uac00\ub2a5\uc131\uc744 \ubcf4\uc5ec\uc8fc\uba70, \ub2e4\uc591\ud55c \ud14d\uc2a4\ud2b8 \uc0dd\uc131\uc5d0 \ud65c\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \uc778\uacf5\uc9c0\ub2a5\uc758 \ucc3d\uc758\uc801\uc778 \uac00\ub2a5\uc131\uc744 \ub354\uc6b1 \ubc1c\uc804\uc2dc\ucf1c\uac08 \uc218 \uc788\uc744 \uac83\uc785\ub2c8\ub2e4.<\/p>\n \uc774\ucc98\ub7fc \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\uc5ec \ud544\uc694\ud55c \uc5ec\ub7ec \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ud14c\uc2a4\ud06c\ub97c \uad6c\ud604\ud574\ubcf4\uc138\uc694. \uc2e4\uc81c\ub85c \uc5ec\ub7ec\ubd84\uc758 \uc544\uc774\ub514\uc5b4\uc640 \uacb0\ud569\ud558\uc5ec \ub9e4\uc6b0 \ud765\ubbf8\ub85c\uc6b4 \uacb0\uacfc\ub97c \ub9cc\ub4e4 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n1. \uc0ac\uc804 \uc900\ube44<\/h2>\n
pip install transformers torch<\/code><\/pre>\n\n
2. \ubaa8\ub378 \ub85c\ub4dc\ud558\uae30<\/h2>\n
from transformers import DistilGPT2LMHeadModel, DistilGPT2Tokenizer\n\n# DistilGPT-2 \ubaa8\ub378\uacfc \ud1a0\ud06c\ub098\uc774\uc800 \ub85c\ub4dc\ntokenizer = DistilGPT2Tokenizer.from_pretrained('distilgpt2')\nmodel = DistilGPT2LMHeadModel.from_pretrained('distilgpt2')<\/code><\/pre>\n3. \ud14d\uc2a4\ud2b8 \uc0dd\uc131\ud558\uae30<\/h2>\n
# \ud14d\uc2a4\ud2b8 \uc0dd\uc131 \ud568\uc218 \uc815\uc758\ndef generate_text(prompt, max_length=50):\n # \uc785\ub825 \ud14d\uc2a4\ud2b8\ub97c \ud1a0\ud070\ud654\n inputs = tokenizer.encode(prompt, return_tensors='pt')\n \n # \ubaa8\ub378\uc744 \ud1b5\ud574 \ud14d\uc2a4\ud2b8 \uc0dd\uc131\n outputs = model.generate(inputs, max_length=max_length, num_return_sequences=1)\n \n # \uc0dd\uc131\ub41c \ud14d\uc2a4\ud2b8 \ub514\ucf54\ub529\n return tokenizer.decode(outputs[0], skip_special_tokens=True)\n\n# \uc0ac\uc6a9 \uc608\nprompt = \"\ub525\ub7ec\ub2dd\uc740\"\ngenerated_text = generate_text(prompt)\nprint(generated_text)<\/code><\/pre>\ngenerate_text<\/code> \ud568\uc218\ub294 \uc8fc\uc5b4\uc9c4 \ud504\ub86c\ud504\ud2b8(prompt) \ub2e4\uc74c\uc5d0 \ucd5c\ub300 max_length<\/code> \ub9cc\ud07c\uc758 \ud14d\uc2a4\ud2b8\ub97c \uc0dd\uc131\ud569\ub2c8\ub2e4. tokenizer.encode<\/code>\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc785\ub825 \ud14d\uc2a4\ud2b8\ub97c \ud1a0\ud070\ud654\ud558\uace0, model.generate<\/code>\ub97c \ud1b5\ud574 \ud14d\uc2a4\ud2b8\ub97c \uc0dd\uc131\ud55c \ud6c4\uc5d0, \uc0dd\uc131\ub41c \ud14d\uc2a4\ud2b8\ub97c \ub2e4\uc2dc \ub514\ucf54\ub529\ud558\uc5ec \ubb38\uc790\uc5f4 \ud615\ud0dc\ub85c \ubc18\ud658\ud569\ub2c8\ub2e4.<\/p>\n4. \ub2e4\uc591\ud55c \ud14d\uc2a4\ud2b8 \uc0dd\uc131 \ud30c\ub77c\ubbf8\ud130 \uc870\uc815\ud558\uae30<\/h2>\n
\n
# \ub2e4\uc591\ud55c \ud30c\ub77c\ubbf8\ud130\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud14d\uc2a4\ud2b8 \uc0dd\uc131\ndef generate_text_with_params(prompt, max_length=50, temperature=1.0, top_k=50, top_p=0.9):\n inputs = tokenizer.encode(prompt, return_tensors='pt')\n outputs = model.generate(\n inputs,\n max_length=max_length,\n num_return_sequences=1,\n temperature=temperature,\n top_k=top_k,\n top_p=top_p,\n do_sample=True\n )\n return tokenizer.decode(outputs[0], skip_special_tokens=True)\n\n# \uc0ac\uc6a9 \uc608\nprompt = \"\uc778\uacf5\uc9c0\ub2a5\uc758 \ubc1c\uc804\uc740\"\ngenerated_text = generate_text_with_params(prompt, temperature=0.7, top_k=30, top_p=0.85)\nprint(generated_text)<\/code><\/pre>\ngenerate_text_with_params<\/code> \ud568\uc218\uc5d0\uc11c\ub294 \ucd94\uac00\uc801\uc778 \ud30c\ub77c\ubbf8\ud130\ub97c \ud1b5\ud574 \ud14d\uc2a4\ud2b8 \uc0dd\uc131 \ubc29\uc2dd\uc744 \uc870\uc815\ud569\ub2c8\ub2e4. \uc774\ub807\uac8c \ub2e4\uc591\ud55c \ubc29\ubc95\uc73c\ub85c \ud14d\uc2a4\ud2b8\ub97c \uc0dd\uc131\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n5. \uc2e4\uc2b5: \uc18c\uc124 \uc791\uc131\ud558\uae30<\/h2>\n
# \uc18c\uc124 \uc791\uc131 \uc608\uc81c\nstory_prompt = \"\ud55c\ub54c \ud3c9\ud654\ub86d\ub358 \ub9c8\uc744\uc5d0 \uac11\uc791\uc2a4\ub7ec\uc6b4 \ud3ed\ud48d\uc774 \ubab0\uc544\ucce4\ub2e4. \uc8fc\ubbfc\ub4e4\uc740 \ubaa8\ub450\"\nstory = generate_text_with_params(story_prompt, max_length=150, temperature=0.9)\nprint(story)<\/code><\/pre>\n6. \uacb0\uacfc \ubc0f \ud53c\ub4dc\ubc31<\/h2>\n
7. \uacb0\ub860<\/h2>\n
8. \ucc38\uace0 \uc790\ub8cc<\/h2>\n