MilvusによるマルチモーダルRAG

このチュートリアルでは、Milvus、可視化BGEモデル、GPT-4oを用いたマルチモーダルRAGを紹介します。このシステムでは、ユーザは画像をアップロードし、テキスト指示を編集することができます。このテキスト指示はBGEの構成検索モデルによって処理され、候補画像を検索します。その後、GPT-4oがリランカーとして機能し、最適な画像を選択し、選択の根拠を提供します。この強力な組み合わせにより、Milvusによる効率的な検索、BGEモデルによる正確な画像処理とマッチング、GPT-4oによる高度なリランキングを活用した、シームレスで直感的な画像検索が可能になります。

準備

依存関係のインストール

$ pip install --upgrade pymilvus openai datasets opencv-python timm einops ftfy peft tqdm

$ git clone https://github.com/FlagOpen/FlagEmbedding.git
$ pip install -e FlagEmbedding

データのダウンロード

以下のコマンドは、サンプルデータをダウンロードし、ローカルフォルダー"./images_folder "に展開します：

• 画像：カテゴリー "Appliance"、"Cell_Phones_and_Accessories"、"Electronics "から約900の画像を含むAmazon Reviews 2023のサブセット。

• leopard.jpg: クエリー画像の例。

$ wget https://github.com/milvus-io/bootcamp/releases/download/data/amazon_reviews_2023_subset.tar.gz
$ tar -xzf amazon_reviews_2023_subset.tar.gz

ロード埋め込みモデル

ビジュアライズドBGEモデル "bge-visualized-base-en-v1.5 "を使用して、画像とテキストの埋め込みを生成します。

1.重みのダウンロード

$ wget https://huggingface.co/BAAI/bge-visualized/resolve/main/Visualized_base_en_v1.5.pth

2.エンコーダの構築

import torch
from FlagEmbedding.visual.modeling import Visualized_BGE


class Encoder:
    def __init__(self, model_name: str, model_path: str):
        self.model = Visualized_BGE(model_name_bge=model_name, model_weight=model_path)
        self.model.eval()

    def encode_query(self, image_path: str, text: str) -> list[float]:
        with torch.no_grad():
            query_emb = self.model.encode(image=image_path, text=text)
        return query_emb.tolist()[0]

    def encode_image(self, image_path: str) -> list[float]:
        with torch.no_grad():
            query_emb = self.model.encode(image=image_path)
        return query_emb.tolist()[0]


model_name = "BAAI/bge-base-en-v1.5"
model_path = "./Visualized_base_en_v1.5.pth"  # Change to your own value if using a different model path
encoder = Encoder(model_name, model_path)

データのロード

このセクションでは、エンベッディングに対応する画像をデータベースに読み込みます。

埋め込み画像の生成

データディレクトリからすべてのjpeg画像を読み込み、エンコーダを適用して画像を埋め込みに変換します。

import os
from tqdm import tqdm
from glob import glob


# Generate embeddings for the image dataset
data_dir = (
    "./images_folder"  # Change to your own value if using a different data directory
)
image_list = glob(
    os.path.join(data_dir, "images", "*.jpg")
)  # We will only use images ending with ".jpg"
image_dict = {}
for image_path in tqdm(image_list, desc="Generating image embeddings: "):
    try:
        image_dict[image_path] = encoder.encode_image(image_path)
    except Exception as e:
        print(f"Failed to generate embedding for {image_path}. Skipped.")
        continue
print("Number of encoded images:", len(image_dict))

Generating image embeddings: 100%|██████████| 900/900 [00:20<00:00, 44.08it/s]

Number of encoded images: 900

Milvusへの挿入

Milvusに画像を挿入します。

from pymilvus import MilvusClient


dim = len(list(image_dict.values())[0])
collection_name = "multimodal_rag_demo"

# Connect to Milvus client given URI
milvus_client = MilvusClient(uri="./milvus_demo.db")

# Create Milvus Collection
# By default, vector field name is "vector"
milvus_client.create_collection(
    collection_name=collection_name,
    auto_id=True,
    dimension=dim,
    enable_dynamic_field=True,
)

# Insert data into collection
milvus_client.insert(
    collection_name=collection_name,
    data=[{"image_path": k, "vector": v} for k, v in image_dict.items()],
)

{'insert_count': 900,
 'ids': [451537887696781312, 451537887696781313, ..., 451537887696782211],
 'cost': 0}

生成リランカーによるマルチモーダル検索

このセクションでは、まずマルチモーダルなクエリで関連する画像を検索し、LLMサービスを使って結果をランク付けし、説明付きで最適な画像を見つけます。

検索の実行

画像とテキストで構成されたクエリデータを使って、高度な画像検索を実行します。

query_image = os.path.join(
    data_dir, "leopard.jpg"
)  # Change to your own query image path
query_text = "phone case with this image theme"

# Generate query embedding given image and text instructions
query_vec = encoder.encode_query(image_path=query_image, text=query_text)

search_results = milvus_client.search(
    collection_name=collection_name,
    data=[query_vec],
    output_fields=["image_path"],
    limit=9,  # Max number of search results to return
    search_params={"metric_type": "COSINE", "params": {}},  # Search parameters
)[0]

retrieved_images = [hit.get("entity").get("image_path") for hit in search_results]
print(retrieved_images)

['./images_folder/images/518Gj1WQ-RL._AC_.jpg', './images_folder/images/41n00AOfWhL._AC_.jpg', './images_folder/images/51Wqge9HySL._AC_.jpg', './images_folder/images/51R2SZiywnL._AC_.jpg', './images_folder/images/516PebbMAcL._AC_.jpg', './images_folder/images/51RrgfYKUfL._AC_.jpg', './images_folder/images/515DzQVKKwL._AC_.jpg', './images_folder/images/51BsgVw6RhL._AC_.jpg', './images_folder/images/51INtcXu9FL._AC_.jpg']

GPT-4oによる再ランク

LLMを使って画像のランク付けを行い、ユーザのクエリと検索結果に基づいて最適な結果の説明を生成します。

1.パノラマ画像の作成

import numpy as np
import cv2

img_height = 300
img_width = 300
row_count = 3


def create_panoramic_view(query_image_path: str, retrieved_images: list) -> np.ndarray:
    """
    creates a 5x5 panoramic view image from a list of images

    args:
        images: list of images to be combined

    returns:
        np.ndarray: the panoramic view image
    """
    panoramic_width = img_width * row_count
    panoramic_height = img_height * row_count
    panoramic_image = np.full(
        (panoramic_height, panoramic_width, 3), 255, dtype=np.uint8
    )

    # create and resize the query image with a blue border
    query_image_null = np.full((panoramic_height, img_width, 3), 255, dtype=np.uint8)
    query_image = Image.open(query_image_path).convert("RGB")
    query_array = np.array(query_image)[:, :, ::-1]
    resized_image = cv2.resize(query_array, (img_width, img_height))

    border_size = 10
    blue = (255, 0, 0)  # blue color in BGR
    bordered_query_image = cv2.copyMakeBorder(
        resized_image,
        border_size,
        border_size,
        border_size,
        border_size,
        cv2.BORDER_CONSTANT,
        value=blue,
    )

    query_image_null[img_height * 2 : img_height * 3, 0:img_width] = cv2.resize(
        bordered_query_image, (img_width, img_height)
    )

    # add text "query" below the query image
    text = "query"
    font_scale = 1
    font_thickness = 2
    text_org = (10, img_height * 3 + 30)
    cv2.putText(
        query_image_null,
        text,
        text_org,
        cv2.FONT_HERSHEY_SIMPLEX,
        font_scale,
        blue,
        font_thickness,
        cv2.LINE_AA,
    )

    # combine the rest of the images into the panoramic view
    retrieved_imgs = [
        np.array(Image.open(img).convert("RGB"))[:, :, ::-1] for img in retrieved_images
    ]
    for i, image in enumerate(retrieved_imgs):
        image = cv2.resize(image, (img_width - 4, img_height - 4))
        row = i // row_count
        col = i % row_count
        start_row = row * img_height
        start_col = col * img_width

        border_size = 2
        bordered_image = cv2.copyMakeBorder(
            image,
            border_size,
            border_size,
            border_size,
            border_size,
            cv2.BORDER_CONSTANT,
            value=(0, 0, 0),
        )
        panoramic_image[
            start_row : start_row + img_height, start_col : start_col + img_width
        ] = bordered_image

        # add red index numbers to each image
        text = str(i)
        org = (start_col + 50, start_row + 30)
        (font_width, font_height), baseline = cv2.getTextSize(
            text, cv2.FONT_HERSHEY_SIMPLEX, 1, 2
        )

        top_left = (org[0] - 48, start_row + 2)
        bottom_right = (org[0] - 48 + font_width + 5, org[1] + baseline + 5)

        cv2.rectangle(
            panoramic_image, top_left, bottom_right, (255, 255, 255), cv2.FILLED
        )
        cv2.putText(
            panoramic_image,
            text,
            (start_col + 10, start_row + 30),
            cv2.FONT_HERSHEY_SIMPLEX,
            1,
            (0, 0, 255),
            2,
            cv2.LINE_AA,
        )

    # combine the query image with the panoramic view
    panoramic_image = np.hstack([query_image_null, panoramic_image])
    return panoramic_image

クエリ画像と検索された画像をインデックスと組み合わせてパノラマビューにします。

from PIL import Image

combined_image_path = os.path.join(data_dir, "combined_image.jpg")
panoramic_image = create_panoramic_view(query_image, retrieved_images)
cv2.imwrite(combined_image_path, panoramic_image)

combined_image = Image.open(combined_image_path)
show_combined_image = combined_image.resize((300, 300))
show_combined_image.show()

パノラマビューの作成

2.再ランク付けと説明

合成された画像を、適切なプロンプトとともにマルチモーダルLLMサービスに送信し、検索された結果を説明付きでランク付けします。GPT-4oをLLMとして有効にするには、OpenAIのAPI Keyを用意する必要があります。

import requests
import base64

openai_api_key = "sk-***"  # Change to your OpenAI API Key


def generate_ranking_explanation(
    combined_image_path: str, caption: str, infos: dict = None
) -> tuple[list[int], str]:
    with open(combined_image_path, "rb") as image_file:
        base64_image = base64.b64encode(image_file.read()).decode("utf-8")

    information = (
        "You are responsible for ranking results for a Composed Image Retrieval. "
        "The user retrieves an image with an 'instruction' indicating their retrieval intent. "
        "For example, if the user queries a red car with the instruction 'change this car to blue,' a similar type of car in blue would be ranked higher in the results. "
        "Now you would receive instruction and query image with blue border. Every item has its red index number in its top left. Do not misunderstand it. "
        f"User instruction: {caption} \n\n"
    )

    # add additional information for each image
    if infos:
        for i, info in enumerate(infos["product"]):
            information += f"{i}. {info}\n"

    information += (
        "Provide a new ranked list of indices from most suitable to least suitable, followed by an explanation for the top 1 most suitable item only. "
        "The format of the response has to be 'Ranked list: []' with the indices in brackets as integers, followed by 'Reasons:' plus the explanation why this most fit user's query intent."
    )

    headers = {
        "Content-Type": "application/json",
        "Authorization": f"Bearer {openai_api_key}",
    }

    payload = {
        "model": "gpt-4o",
        "messages": [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": information},
                    {
                        "type": "image_url",
                        "image_url": {"url": f"data:image/jpeg;base64,{base64_image}"},
                    },
                ],
            }
        ],
        "max_tokens": 300,
    }

    response = requests.post(
        "https://api.openai.com/v1/chat/completions", headers=headers, json=payload
    )
    result = response.json()["choices"][0]["message"]["content"]

    # parse the ranked indices from the response
    start_idx = result.find("[")
    end_idx = result.find("]")
    ranked_indices_str = result[start_idx + 1 : end_idx].split(",")
    ranked_indices = [int(index.strip()) for index in ranked_indices_str]

    # extract explanation
    explanation = result[end_idx + 1 :].strip()

    return ranked_indices, explanation

ランク付け後の画像インデックスと最も良い結果の理由を取得します：

ranked_indices, explanation = generate_ranking_explanation(
    combined_image_path, query_text
)

3.最良結果を説明付きで表示する

print(explanation)

best_index = ranked_indices[0]
best_img = Image.open(retrieved_images[best_index])
best_img = best_img.resize((150, 150))
best_img.show()

Reasons: The most suitable item for the user's query intent is index 6 because the instruction specifies a phone case with the theme of the image, which is a leopard. The phone case with index 6 has a thematic design resembling the leopard pattern, making it the closest match to the user's request for a phone case with the image theme.

ベスト結果

クイックデプロイ

このチュートリアルでオンラインデモを開始する方法については、サンプルアプリケーションを参照してください。