{"id":29570,"date":"2024-10-28T01:55:30","date_gmt":"2024-10-28T01:55:30","guid":{"rendered":"http:\/\/atmokpo.com\/w\/?p=29570"},"modified":"2024-11-26T06:52:14","modified_gmt":"2024-11-26T06:52:14","slug":"%ed%97%88%ea%b9%85%ed%8e%98%ec%9d%b4%ec%8a%a4-%ed%8a%b8%eb%a0%8c%ec%8a%a4%ed%8f%ac%eb%a8%b8-%ed%99%9c%ec%9a%a9%ea%b0%95%ec%a2%8c-clip-%ec%a0%84%ec%b2%98%eb%a6%ac","status":"publish","type":"post","link":"https:\/\/atmokpo.com\/w\/29570\/","title":{"rendered":"\ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub80c\uc2a4\ud3ec\uba38 \ud65c\uc6a9\uac15\uc88c, CLIP \uc804\ucc98\ub9ac"},"content":{"rendered":"

<\/p>\n

\ub525\ub7ec\ub2dd\uc758 \ucd5c\uc2e0 \ud2b8\ub80c\ub4dc \uc911 \ud558\ub098\ub294 \ub2e4\uc591\ud55c \uba40\ud2f0\ubaa8\ub2ec(multi-modal) \ubaa8\ub378\uc758 \ucd9c\ud604\uc785\ub2c8\ub2e4. \ud2b9\ud788 OpenAI\uc758 CLIP(Contrastive Language\u2013Image Pretraining) \ubaa8\ub378\uc740 \uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8 \uac04\uc758 \uad00\uacc4\ub97c \ud559\uc2b5\ud558\uc5ec \ub2e4\uc591\ud55c \uc77c\uc744 \uc218\ud589\ud560 \uc218 \uc788\ub294 \ub9e4\uc6b0 \uac15\ub825\ud55c \ubc29\ubc95\ub860\uc785\ub2c8\ub2e4. \uc774\ubc88 \uae00\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4(Hugging Face) \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud1b5\ud574 CLIP \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\ub294 \ubc29\ubc95\uacfc \uadf8 \uc804\ucc98\ub9ac \uacfc\uc815\uc744 \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n

1. CLIP \ubaa8\ub378 \uc18c\uac1c<\/h2>\n

CLIP \ubaa8\ub378\uc740 \uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8 \uc30d\uc744 \ub3d9\uc2dc\uc5d0 \ud559\uc2b5\ud558\uc5ec, \uc774\ubbf8\uc9c0\uac00 \uc5b4\ub5a4 \ub0b4\uc6a9\uc778\uc9c0\ub97c \uc774\ud574\ud558\uace0, \uc8fc\uc5b4\uc9c4 \ud14d\uc2a4\ud2b8 \uc124\uba85\uacfc \uadf8 \uc774\ubbf8\uc9c0 \uac04\uc758 \uc720\uc0ac\uc131\uc744 \uce21\uc815\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uc811\uadfc \ubc29\uc2dd\uc740 \ub370\uc774\ud130\uc14b\uc744 \uc120\ud0dd\ud560 \ud544\uc694 \uc5c6\uc774 \ube44\uc9c0\ub3c4 \ud559\uc2b5\uc744 \ud1b5\ud574 \ub2e4\uc591\ud55c \uc791\uc5c5\uc5d0 \uc720\uc5f0\ud558\uac8c \uc801\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n

2. CLIP \uc804\ucc98\ub9ac \uacfc\uc815<\/h2>\n

CLIP \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uae30 \uc704\ud574\uc11c\ub294 \uc785\ub825 \uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8\uc5d0 \ub300\ud55c \uc801\uc808\ud55c \uc804\ucc98\ub9ac\uac00 \ud544\uc694\ud569\ub2c8\ub2e4. \uc804\ucc98\ub9ac \uacfc\uc815\uc740 \ub2e4\uc74c\uacfc \uac19\uc740 \ub2e8\uacc4\ub85c \uc774\ub8e8\uc5b4\uc9d1\ub2c8\ub2e4:<\/p>\n

    \n
  1. \uc774\ubbf8\uc9c0 \ub85c\ub4dc \ubc0f \ud06c\uae30 \uc870\uc815<\/li>\n
  2. \uc774\ubbf8\uc9c0\uc758 \uc815\uaddc\ud654<\/li>\n
  3. \ud14d\uc2a4\ud2b8 \ud1a0\ud070\ud654<\/li>\n<\/ol>\n

    2.1 \uc774\ubbf8\uc9c0 \ub85c\ub4dc \ubc0f \ud06c\uae30 \uc870\uc815<\/h3>\n

    \ubaa8\ub378\uc5d0 \uc785\ub825\ud560 \uc774\ubbf8\uc9c0\ub294 \uc77c\uc815\ud55c \ud06c\uae30\ub85c \uc870\uc815\ud574\uc57c \ud569\ub2c8\ub2e4. \uc77c\ubc18\uc801\uc73c\ub85c CLIP \ubaa8\ub378\uc740 224×224 \ud06c\uae30\uc758 \uc774\ubbf8\uc9c0\ub97c \uc694\uad6c\ud569\ub2c8\ub2e4. \uc774\ub97c \uc704\ud574 \ud30c\uc774\uc36c\uc758 PIL \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n

    2.2 \uc774\ubbf8\uc9c0 \uc815\uaddc\ud654<\/h3>\n

    \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ub192\uc774\uae30 \uc704\ud574\uc11c\ub294 \uc774\ubbf8\uc9c0\uc758 \ud53d\uc140 \uac12\uc744 \uc815\uaddc\ud654\ud574\uc57c \ud569\ub2c8\ub2e4. CLIP \ubaa8\ub378\uc740 \uc77c\ubc18\uc801\uc73c\ub85c mean=[0.48145466, 0.4578275, 0.40821073]<\/code>\uc640 std=[0.26862954, 0.26130258, 0.27577711]<\/code>\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc815\uaddc\ud654\ud569\ub2c8\ub2e4.<\/p>\n

    2.3 \ud14d\uc2a4\ud2b8 \ud1a0\ud070\ud654<\/h3>\n

    \ud14d\uc2a4\ud2b8\ub294 \ubbf8\ub9ac \uc815\uc758\ub41c \ud1a0\ud06c\ub098\uc774\uc800\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc778\ucf54\ub529\ub418\uc5b4\uc57c \ud569\ub2c8\ub2e4. CLIP\uc740 BPE(Byte Pair Encoding) \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\ubbc0\ub85c, \uc774\ub97c \ud1b5\ud574 \ud14d\uc2a4\ud2b8\ub97c \uc815\uc218 \uc778\ub371\uc2a4\ub85c \ubcc0\ud658\ud569\ub2c8\ub2e4.<\/p>\n

    3. \ucf54\ub4dc \uc608\uc81c<\/h2>\n

    \uc774\uc81c \uc704\uc758 \uc804\ucc98\ub9ac \uacfc\uc815\uc744 \ud30c\uc774\uc36c \ucf54\ub4dc\ub85c \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc774 \uc608\uc81c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 transformers<\/code> \ub77c\uc774\ube0c\ub7ec\ub9ac\uc640 PIL<\/code> \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \uba3c\uc800 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud569\ub2c8\ub2e4.<\/p>\n

    pip install transformers torch torchvision pillow<\/code><\/pre>\n
    3.1 \uc774\ubbf8\uc9c0 \uc804\ucc98\ub9ac \ucf54\ub4dc<\/h3>\n
    \nfrom PIL import Image\nimport requests\nfrom transformers import CLIPProcessor\n\n# CLIPProcessor \ubd88\ub7ec\uc624\uae30\nprocessor = CLIPProcessor.from_pretrained(\"openai\/clip-vit-base-patch16\")\n\n# \uc774\ubbf8\uc9c0 URL\nimage_url = \"https:\/\/example.com\/image.jpg\"  # \uc801\uc808\ud55c \uc774\ubbf8\uc9c0 URL\ub85c \ubcc0\uacbd\ud558\uc138\uc694.\nimage = Image.open(requests.get(image_url, stream=True).raw)\n\n# \uc774\ubbf8\uc9c0\ub97c \uc804\ucc98\ub9ac\ninputs = processor(images=image, return_tensors=\"pt\", padding=True)\nprint(inputs)\n    <\/code><\/pre>\n
    3.2 \ud14d\uc2a4\ud2b8 \uc804\ucc98\ub9ac \ucf54\ub4dc<\/h3>\n
    \n# \ud14d\uc2a4\ud2b8 \uc785\ub825\ntext = \"A label describing the image\"\ntext_inputs = processor(text=[text], return_tensors=\"pt\", padding=True)\nprint(text_inputs)\n    <\/code><\/pre>\n
    4. \ubaa8\ub378 \uc608\uce21<\/h2>\n
    \uc804\ucc98\ub9ac\uac00 \uc644\ub8cc\ub41c \uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8\ub97c \ubaa8\ub378\uc5d0 \uc785\ub825\ud558\uc5ec \uc608\uce21\uc744 \uc218\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud5c8\uae45\ud398\uc774\uc2a4\uc758 CLIPModel<\/code>\uc744 \uc0ac\uc6a9\ud558\uba74 \ub2e4\uc74c\uacfc \uac19\uc774 \uc608\uce21\uc744 \uc9c4\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
    \nfrom transformers import CLIPModel\n\n# CLIP \ubaa8\ub378 \ubd88\ub7ec\uc624\uae30\nmodel = CLIPModel.from_pretrained(\"openai\/clip-vit-base-patch16\")\n\n# \uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8\uc758 \ud2b9\uc9d5 \ucd94\ucd9c\nwith torch.no_grad():\n    outputs = model(**inputs, **text_inputs)\n\n# \uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8 \uac04\uc758 \uc720\uc0ac\ub3c4 \uacc4\uc0b0\nlogits_per_image = outputs.logits_per_image  # (batch_size, text_length)\nlogits_per_text = outputs.logits_per_text      # (batch_size, image_length)\nprint(logits_per_image)\nprint(logits_per_text)\n    <\/code><\/pre>\n
    5. \uacb0\ub860<\/h2>\n
    \uc774\ubc88 \ud3ec\uc2a4\ud2b8\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub97c \ud65c\uc6a9\ud558\uc5ec CLIP \ubaa8\ub378\uc758 \uc804\ucc98\ub9ac \uacfc\uc815\uc744 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \uc774\ubbf8\uc9c0 \ubc0f \ud14d\uc2a4\ud2b8\uc758 \uc804\ucc98\ub9ac\ub294 \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \uadf9\ub300\ud654\ud558\ub294 \ub370 \ub9e4\uc6b0 \uc911\uc694\ud55c \ub2e8\uacc4\uc785\ub2c8\ub2e4. \uc774\uc81c \uc5ec\ub7ec\ubd84\uc740 CLIP \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\uc5ec \uc774\ubbf8\uc9c0 \ubc0f \ud14d\uc2a4\ud2b8 \uac04\uc758 \uad00\uacc4\ub97c \ubcf4\ub2e4 \uc798 \uc774\ud574\ud560 \uc218 \uc788\uae30\ub97c \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n
    6. \ucc38\uace0\uc790\ub8cc<\/h2>\n