基于Milvus的图像相似性搜索无结果问题求助

我参考图像反向搜索的Notebook开发相似性搜索功能，目前遇到搜索无结果的问题，但验证同一张图的向量余弦相似度接近1，说明向量存储是正确的，不清楚问题出在哪。

我的实现代码 #

初始化参数与索引创建 #

import csv
from glob import glob
from pathlib import Path
from statistics import mean

from towhee import pipe, ops, DataCollection
from pymilvus import connections, FieldSchema, CollectionSchema, DataType, Collection, utility


# Towhee parameters
MODEL = 'resnet50'

# Milvus parameters
HOST = [MY_HOST]
PORT = [MY_PORT]
TOPK = 5
DIM = 2048 
COLLECTION_NAME = 'images'
INDEX_TYPE = 'IVF_FLAT'
METRIC_TYPE = 'L2'

index_params = {
    'metric_type': METRIC_TYPE,
    'index_type': INDEX_TYPE,
    'params': {"nlist": 2048}
}

collection.create_index(field_name='image', index_params=index_params, index_name = "image_index")

图像向量化与数据插入 #

# Load image path
def load_image(x):
    if x.endswith('csv'):
        with open(x) as f:
            reader = csv.reader(f)
            next(reader)
            for item in reader:
                yield item[1]
    else:
        for item in glob(x):
            yield item
            
# Embedding pipeline
p_embed = (
    pipe.input('src')
        .flat_map('src', 'img_path', load_image)
        .map('img_path', 'img', ops.image_decode())
        .map('img', 'vec', ops.image_embedding.timm(model_name=MODEL))
)


image_save_dir = [MY_IMAGE_PATH]
p_display = p_embed.output('img_path', 'img', 'vec') 
result = DataCollection(p_display(image_save_dir))

# check result
result.show()
print(result[0]['img_path'])
print(result[0]['vec'])    

connections.connect(alias='default', host=HOST, port=PORT)
collection_name = "clothes"
collection = Collection(name = collection_name)

for i, r in enumerate(result):
    vector = r['vec']
    collection.insert([
        {
            "clothes_id" : i,
            "category" : "top",
            "color" : "black",
            "image" : vector,
            "gender" : ["F"],
            "style" : ["casual"],
            "thickness" : [],
            "season" : ["spring"]
        }
    ])

搜索代码（无结果） #

p_search_pre = (
        p_embed.map('vec', ('search_res'), ops.ann_search.milvus_client(
                    host=HOST, port=PORT, limit=5,
                    collection_name="clothes"))
               .map('search_res', 'pred', lambda x: [y[0] for y in x]) # get id
)

p_search = p_search_pre.output('img_path', 'pred') 

# Search for example query image(s)
collection.load()
dc = p_search('[MY_IMAGE_PATH]/test37.png')

# Display search results with image paths
DataCollection(dc).show()

验证代码（余弦相似度接近1） #

from numpy import dot
from numpy.linalg import norm
import numpy as np

def cos_similarity(A, B):
    return dot(A, B) / (norm(A) * norm(B))

test_image = '[MY_IMAGE_PATH]/test38.png'
p_display = p_embed.output('img_path', 'img', 'vec') 
result = DataCollection(p_display(image_save_dir))

collection_name = 'clothes' 
collection = Collection(name=collection_name)

# get image vector where clothes_id=38
save_results = collection.query(
    expr="clothes_id == 38",
    output_fields=["clothes_id", "category", "color", "gender", "style", "thickness", "season", "image"]
)

if save_results:
    saved_image_vector = save_results[0]["image"]
    result_vector = np.array(result[0]['vec'])
    saved_image_vector = np.array(saved_image_vector)

    # get similarity
    similarity = cos_similarity(result_vector, saved_image_vector)
    print(f"cosine similarity : {similarity}")
else:
    print("there is no images")

问题点 #

• 执行搜索时无任何结果返回，但验证同一张图像的向量余弦相似度约为1，说明向量存储正确。
• 不清楚搜索无结果的原因。

可能的原因与解决方案 #

1. 索引创建时机错误 #

代码开头直接调用collection.create_index时，collection还未连接到Milvus或指定目标clothes集合，这行代码并未对实际使用的集合生效。
解决：在数据插入完成后，对clothes集合创建索引并加载：

# 插入数据后添加以下代码
collection.create_index(field_name='image', index_params=index_params, index_name="image_index")
collection.load()

2. IVF_FLAT索引缺少搜索参数 #

IVF_FLAT索引搜索时需要指定nprobe参数（默认值过小会导致无法命中数据），当前搜索算子未配置该参数。
解决：在搜索算子中添加search_params={"nprobe": 10}（可根据nlist调整，一般设为nlist的1/10~1/20）：

p_search_pre = (
        p_embed.map('vec', ('search_res'), ops.ann_search.milvus_client(
                    host=HOST, port=PORT, limit=5,
                    collection_name="clothes",
                    search_params={"nprobe": 10}))
               .map('search_res', 'pred', lambda x: [y[0] for y in x])
)

3. 搜索结果解析逻辑错误 #

Milvus搜索结果格式可能因版本不同有差异，当前lambda x: [y[0] for y in x]的解析逻辑可能不正确。
解决：先打印search_res查看实际结构，再调整解析逻辑：

p_search_pre = (
        p_embed.map('vec', ('search_res'), ops.ann_search.milvus_client(
                    host=HOST, port=PORT, limit=5,
                    collection_name="clothes"))
               .map('search_res', 'pred', lambda x: print(x)) # 先打印结果结构
)

4. 集合字段维度不匹配 #

确认clothes集合中image字段的维度是否为2048（resnet50输出的向量维度），维度不匹配会导致搜索无结果。
解决：通过print(collection.schema)查看字段定义，确保image字段的dim参数为2048。

内容的提问来源于stack exchange，提问作者Rashad Tockey