\uc624\ub298\ub0a0 \uc778\uacf5\uc9c0\ub2a5(AI) \ubd84\uc57c\uc5d0\uc11c \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP)\uc640 \ucef4\ud4e8\ud130 \ube44\uc804(CV)\uc740 \ub9e4\uc6b0 \uc911\uc694\ud55c \uc601\uc5ed\uc785\ub2c8\ub2e4. \ud2b9\ud788, \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38(Transformers)<\/strong> \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \ucd5c\uadfc \uba87 \ub144\uac04 \ub9ce\uc740 \uad00\uc2ec\uc744 \ubc1b\uace0 \uc788\uc73c\uba70, \uc0ac\uc6a9\uc790\uac00 \uc190\uc27d\uac8c \ucd5c\uc2e0 \ub525\ub7ec\ub2dd \ubaa8\ub378\uc744 \ud65c\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d \ub3d5\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc774\uc6a9\ud55c CLIP \ubaa8\ub378 \uc124\uce58 \ubc0f \ud65c\uc6a9\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 NLP \ubc0f CV \ubaa8\ub378\uc744 \uc190\uc27d\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d \uc124\uacc4\ub41c \ud30c\uc774\uc36c \ub77c\uc774\ube0c\ub7ec\ub9ac\uc785\ub2c8\ub2e4. \uc774 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 BERT, GPT-2, T5 \ub4f1 \ucd5c\uc2e0 \ubaa8\ub378\ub4e4\uc744 \uc9c0\uc6d0\ud558\uba70, \uc0ac\uc6a9\uc790\uac00 \uc774\ub7ec\ud55c \ubaa8\ub378\uc744 \uc190\uc27d\uac8c \ub2e4\uc6b4\ub85c\ub4dc\ud558\uace0 \ud6c8\ub828\ud560 \uc218 \uc788\ub3c4\ub85d \ud569\ub2c8\ub2e4.<\/p>\n \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub294 \uc8fc\uc758(attention) \uba54\ucee4\ub2c8\uc998\uc744 \uae30\ubc18\uc73c\ub85c \ud55c \ubaa8\ub378\ub85c, \uc790\uc5f0\uc5b4 \ucc98\ub9ac \uc791\uc5c5\uc5d0 \ub9e4\uc6b0 \uac15\ub825\ud55c \uc131\ub2a5\uc744 \ubc1c\ud718\ud569\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc740 \uc785\ub825 \uc2dc\ud000\uc2a4\uc758 \ubaa8\ub4e0 \ub2e8\uc5b4\uc5d0 \ub300\ud574 \ub3d9\uc2dc \ucc98\ub9ac\ud560 \uc218 \uc788\ub294 \ub2a5\ub825 \ub355\ubd84\uc5d0, \ubcd1\ub82c \ucc98\ub9ac\uac00 \uac00\ub2a5\ud558\uc5ec \ud6c8\ub828 \uc18d\ub3c4\uac00 \ub9e4\uc6b0 \ube60\ub985\ub2c8\ub2e4.<\/p>\n CLIP \ubaa8\ub378\uc740 OpenAI\uc5d0\uc11c \uac1c\ubc1c\ud55c \ubaa8\ub378\ub85c, \uc774\ubbf8\uc9c0\ub97c \uc124\uba85\ud558\ub294 \uc790\uc5f0\uc5b4 \ud14d\uc2a4\ud2b8\uc640 \uc774\ubbf8\uc9c0\ub97c \ud568\uaed8 \ud559\uc2b5\ud569\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc740 \uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8 \uac04\uc758 \uad00\uacc4\ub97c \uc774\ud574\ud558\uace0, \uc8fc\uc5b4\uc9c4 \ud14d\uc2a4\ud2b8 \uc124\uba85\uc5d0 \ub9de\ub294 \uc774\ubbf8\uc9c0\ub97c \ucc3e\uac70\ub098 \uc8fc\uc5b4\uc9c4 \uc774\ubbf8\uc9c0\uc5d0 \ub300\ud55c \uc124\uba85\uc744 \uc0dd\uc131\ud558\ub294 \ub4f1\uc758 \uc791\uc5c5\uc744 \uc218\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\uc640 CLIP \ubaa8\ub378\uc744 \uc124\uce58\ud558\ub294 \uacfc\uc815\uc740 \ub9e4\uc6b0 \uac04\ub2e8\ud569\ub2c8\ub2e4. \uc544\ub798\uc758 \ub2e8\uacc4\uc5d0 \ub530\ub77c \uc9c4\ud589\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n \uba3c\uc800, Python \ubc0f pip\uac00 \uc124\uce58\ub418\uc5b4 \uc788\uc5b4\uc57c \ud569\ub2c8\ub2e4. \uc544\ub798\uc758 \uba85\ub839\uc5b4\ub97c \uc0ac\uc6a9\ud558\uc5ec \uceec\ub7ec\uc640 \ud074\ub9bd\uc744 \uc124\uce58\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n \uc704\uc758 \uba85\ub839\uc5b4\ub97c \uc2e4\ud589\ud558\uba74 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\uc640 PyTorch, TorchVision\uc774 \uc124\uce58\ub429\ub2c8\ub2e4.<\/p>\n CLIP \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uae30 \uc704\ud574 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud1b5\ud574 \ubaa8\ub378\uc744 \ubd88\ub7ec\uc62c \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc544\ub798\ub294 \uc774\ub97c \uc704\ud55c \uc608\uc81c \ucf54\ub4dc\uc785\ub2c8\ub2e4:<\/p>\n \ubaa8\ub378\uc774 \uc124\uce58\ub418\uc5c8\uc73c\ub2c8, \uc774\uc81c CLIP \ubaa8\ub378\uc744 \uc2e4\uc81c\ub85c \ud65c\uc6a9\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc544\ub798\uc758 \uc608\uc81c \ucf54\ub4dc\ub97c \ud1b5\ud574 \uc8fc\uc5b4\uc9c4 \ud14d\uc2a4\ud2b8\uc5d0 \ub9de\ub294 \uc774\ubbf8\uc9c0\ub97c \ucc3e\ub294 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n CLIP\uc740 \ud14d\uc2a4\ud2b8\uc640 \uc774\ubbf8\uc9c0\ub97c \uc784\ubca0\ub529\ud558\uc5ec \uc720\uc0ac\ub3c4 \uc810\uc218\ub97c \uacc4\uc0b0\ud569\ub2c8\ub2e4. \uc544\ub798\uc758 \ucf54\ub4dc\ub97c \ud1b5\ud574 \uc774\ub97c \uc2e4\uc81c\ub85c \uc218\ud589\ud574 \ubcfc \uc218 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\uc640 CLIP \ubaa8\ub378\uc744 \uc124\uce58\ud558\uace0 \ud65c\uc6a9\ud558\ub294 \ubc29\ubc95\uc744 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. CLIP \ubaa8\ub378\uc740 \uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8 \uac04\uc758 \ubcf5\uc7a1\ud55c \uad00\uacc4\ub97c \uc798 \ud559\uc2b5\ud558\uc5ec \ub2e4\uc591\ud55c \uc791\uc5c5\uc5d0 \ud6a8\uacfc\uc801\uc73c\ub85c \ud65c\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ucd94\ud6c4 \ub354 \ub9ce\uc740 \uc608\uc81c\uc640 \uc751\uc6a9 \uc0ac\ub840\ub97c \ud1b5\ud574 CLIP \ubaa8\ub378\uc758 \uac00\ub2a5\uc131\uc744 \ud0d0\uad6c\ud574 \ubcfc \uac83\uc744 \uad8c\uc7a5\ud569\ub2c8\ub2e4.<\/p>\n \ucc38\uace0:<\/strong> \uc704\uc758 \ucf54\ub4dc \uc608\uc81c\ub294 CLIP \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uae30 \uc704\ud55c \uae30\ubcf8\uc801\uc778 \uc0ac\uc6a9\ubc95\uc744 \uc81c\uacf5\ud558\uba70, \uc2e4\uc804\uc5d0\uc11c\ub294 \uc801\uc808\ud55c \ub370\uc774\ud130\uc14b\uacfc \uc870\uac74\uc5d0 \ub9de\ucdb0 \ucd94\uac00\uc801\uc778 \uc804\ucc98\ub9ac \ubc0f \ud6c8\ub828\uc774 \ud544\uc694\ud569\ub2c8\ub2e4.<\/p>\n<\/div>\n <\/body><\/p>\n","protected":false},"excerpt":{"rendered":"1. \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub780?<\/h2>\n
1.1 \uae30\ubcf8 \uac1c\ub150<\/h3>\n
2. CLIP(Contrastive Language-Image Pretraining) \ubaa8\ub378 \uc18c\uac1c<\/h2>\n
2.1 CLIP\uc758 \uc8fc\uc694 \ud2b9\uc9d5<\/h3>\n
\n
3. CLIP \ubaa8\ub378 \uc124\uce58\ud558\uae30<\/h2>\n
3.1 \ud658\uacbd \uc124\uc815<\/h3>\n
pip install transformers torch torchvision<\/code><\/pre>\n3.2 CLIP \ubaa8\ub378 \ubd88\ub7ec\uc624\uae30<\/h3>\n
import torch\nfrom transformers import CLIPProcessor, CLIPModel\n\n# CLIP \ubaa8\ub378\uacfc \ud504\ub85c\uc138\uc11c \ubd88\ub7ec\uc624\uae30\nmodel = CLIPModel.from_pretrained(\"openai\/clip-vit-base-patch16\")\nprocessor = CLIPProcessor.from_pretrained(\"openai\/clip-vit-base-patch16\")<\/code><\/pre>\n4. CLIP \ubaa8\ub378 \ud65c\uc6a9\ud558\uae30<\/h2>\n
4.1 \ud14d\uc2a4\ud2b8\uc640 \uc774\ubbf8\uc9c0 \ubd88\ub7ec\uc624\uae30<\/h3>\n
# \uc608\uc81c \uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8 \uc785\ub825\ntexts = [\"a photo of a cat\", \"a photo of a dog\"]\nimages = [\"path\/to\/cat_image.jpg\", \"path\/to\/dog_image.jpg\"]\n\n# \uc774\ubbf8\uc9c0 \uc804\ucc98\ub9ac\ninputs = processor(text=texts, images=images, return_tensors=\"pt\", padding=True)<\/code><\/pre>\n4.2 \uc720\uc0ac\ub3c4 \uacc4\uc0b0\ud558\uae30<\/h3>\n
# \uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8\uc758 \uc784\ubca0\ub529 \uacc4\uc0b0\nwith torch.no_grad():\n outputs = model(**inputs)\n\n# \uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8 \uc720\uc0ac\ub3c4 \uacc4\uc0b0\nlogits_per_image = outputs.logits_per_image # \uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8 \uc720\uc0ac\ub3c4\nprobs = logits_per_image.softmax(dim=1) # \ud655\ub960\ub85c \ubcc0\ud658\n\nprint(\"\uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8\uc758 \uc720\uc0ac\ub3c4 \ud655\ub960:\", probs.numpy())<\/code><\/pre>\n5. \uacb0\ub860<\/h2>\n