续

上集说到语义搜索，这集接着玩一下图搜图，这种场景在电商中很常见——拍照搜商品。图搜图实现非常类似语义搜索，代码逻辑结构都很类似…

开搞

还是老地方modelscope找个Vision Transformer模型，这里选用vit-base-patch16-224,如果还想玩玩文搜图，可以选用支持多模态的multi-modal_clip-vit-base-patch16_zh

D:\python2023>modelscope download --model AI-ModelScope/vit-base-patch16-224

准备测试数据

运行代码

from PIL import Image
from transformers import ViTFeatureExtractor, ViTModel
import torch
import time
from elasticsearch import Elasticsearch

# 初始化模型和图片特征提取器
MODEL_PATH = 'C:\\Users\\Administrator\\.cache\\modelscope\\hub\\AI-ModelScope\\vit-base-patch16-224'
feature_extractor = ViTFeatureExtractor.from_pretrained(MODEL_PATH)
model = ViTModel.from_pretrained(MODEL_PATH)


def extract_features(image_path):
    image = Image.open(image_path)
    inputs = feature_extractor(images=image, return_tensors="pt")
    with torch.no_grad():
        outputs = model(**inputs)
        last_hidden_states = outputs.last_hidden_state
        # 取出CLS token的输出作为图片的特征向量
        features = last_hidden_states[:, 0].squeeze()
    return features.numpy()

def store_image_features(image_id, features):
    doc = {
        'image_id': image_id,
        'features': features.tolist()
    }
    res = es.index(index=index_name, id=image_id, body=doc)
    print(res['result'])

def search_similar_images(query_image_path, top_k=5):
    query_features = extract_features(query_image_path)
    body = {
        "size": top_k,
        "query": {
            "script_score": {
                "query": {"match_all": {}},
                "script": {
                    "source": "cosineSimilarity(params.query_vector, 'features') + 1.0",
                    "params": {"query_vector": query_features.tolist()}
                }
            }
        }
    }

    response = es.search(index=index_name, body=body)
    similar_images = [hit['_id'] for hit in response['hits']['hits']]
    return similar_images

# 假设有一个图片ID和路径的字典
images = {
    'img_dog': 'D:\\1\\dog.jpg',
    'img_wolf': 'D:\\1\\wolf.jpg',
    'img_person': 'D:\\1\\person.jpg',
    'img_montain': 'D:\\1\\montain.jpg',
}
# 调用ES api创建索引
es = Elasticsearch([{'scheme':'http','host':'192.168.72.128','port':9200}])
index_name = 'image_search'
body = {
    "mappings": {
        "properties": {
            "features": {
                "type": "dense_vector",
                "dims": 768  # ViT base模型的特征向量维度
            }
        }
    }
}
if not es.indices.exists(index=index_name):
    es.indices.create(index=index_name, body=body)
# 存储图片特征到ES
for img_id, img_path in images.items():
    img_features = extract_features(img_path)
    store_image_features(img_id, img_features)

# ES向量搜索找到某张图片最相似的图片集
time.sleep(3)
similar_ids = search_similar_images("D:\\1\\test_dog.jpg")
print(f"找到的相似图片ID: {similar_ids}")

可以看到最相似的还是狗，狼也像狗所以次之，然后是人，其实相关度已经很低了，最不相关的是风景图

看看索引情况

图片向量ES内存使用还不大，和文本数据基本一样，主要是因为图片特征向量维度都使用了768，如果不搜索不不准确，可以调高维度，但ES内存使用会增加。反之图片干扰特征少一点，特征向量维度小一些也会有较好的搜索效果。