update func

imgsearchimg/api/app.py

@@ -4,26 +4,18 @@ from image_search import ImageSearchEngine
 import magic
 from urllib.request import urlretrieve
 import time
+from . import config
 
 app = Flask(__name__)
 
-# 配置常量
-UPLOAD_FOLDER = 'static/images'
-ALLOWED_EXTENSIONS = {'png', 'jpg', 'jpeg', 'gif', 'bmp'}
-
-# 搜索参数默认值
-TOP_K = 5  # 默认返回的最大结果数
-MIN_SCORE = 0.0  # 默认最小相似度分数
-MAX_SCORE = 100.0  # 默认最大相似度分数
-
-os.makedirs(UPLOAD_FOLDER, exist_ok=True)
+os.makedirs(config.UPLOAD_FOLDER, exist_ok=True)
 
 # 初始化图像搜索引擎
 search_engine = ImageSearchEngine()
 
 def allowed_file(filename):
     """检查文件是否允许上传"""
-    return '.' in filename and filename.rsplit('.', 1)[1].lower() in ALLOWED_EXTENSIONS
+    return '.' in filename and filename.rsplit('.', 1)[1].lower() in config.ALLOWED_EXTENSIONS
 
 def is_valid_image(file_path):
     """检查文件是否为有效的图片文件"""
@@ -59,7 +51,7 @@ def upload_file():
             return jsonify({'error': '无效的图片URL'}), 400
         
         time_base = str(time.time() * 1000)
-        image_path = os.path.join(UPLOAD_FOLDER, time_base + os.path.basename(image_url))
+        image_path = os.path.join(config.UPLOAD_FOLDER, time_base + os.path.basename(image_url))
         urlretrieve(image_url, image_path)
 
         if allowed_file(image_path) and is_valid_image(image_path):
@@ -92,9 +84,9 @@ def search():
             return jsonify({'error': '无效的图片URL'}), 400
             
         # 获取可选参数
-        limit = data.get('limit', TOP_K)
-        min_score = data.get('min_score', MIN_SCORE)
-        max_score = data.get('max_score', MAX_SCORE)
+        limit = data.get('limit', config.TOP_K)
+        min_score = data.get('min_score', config.MIN_SCORE)
+        max_score = data.get('max_score', config.MAX_SCORE)
         
         # 验证参数类型和范围
         try:
@@ -115,7 +107,7 @@ def search():
         
         start_download_time = time.time()
         time_base = str(time.time() * 1000)
-        image_path = os.path.join(UPLOAD_FOLDER, time_base + os.path.basename(image_url))
+        image_path = os.path.join(config.UPLOAD_FOLDER, time_base + os.path.basename(image_url))
         urlretrieve(image_url, image_path)
         end_download_time = time.time()
         print(f"下载耗时: { end_download_time - start_download_time } s",)

imgsearchimg/api/config.py

@@ -0,0 +1,7 @@
+UPLOAD_FOLDER = 'static/images'
+ALLOWED_EXTENSIONS = {'png', 'jpg', 'jpeg', 'gif', 'bmp'}
+ALIYUN_FUNCTION_URL = "http://img-fun-alnilgdbfc.cn-hangzhou.fcapp.run"
+
+TOP_K = 5
+MIN_SCORE = 0.0
+MAX_SCORE = 100.0

imgsearchimg/api/image_search.py

@@ -72,6 +72,11 @@ class ImageSearchEngine:
 
 
     def _process_image(self, image_path: str) -> Optional[torch.Tensor]:
+
+        # 这里调用阿里云云函数处理图片
+
+
+
         """处理单张图片并提取特征。
         
         Args:

imgsearchimg/func/Dockerfile

@@ -1,6 +1,3 @@
-# 使用Python 3.11.11作为基础镜像
-# FROM python:3.11.11-slim
-# FROM swr.cn-north-4.myhuaweicloud.com/ddn-k8s/docker.io/wallies/python-cuda:3.11-cuda11.8-runtime
 FROM registry.cn-hangzhou.aliyuncs.com/serverless_devs/pytorch:22.12-py3
 
 RUN pip config set global.index-url https://mirrors.aliyun.com/pypi/simple/
@@ -32,4 +29,4 @@ RUN pip install --no-cache-dir -r requirements.txt
 EXPOSE 5001
 
 # 设置默认命令
-CMD ["python", "func_test.py"]
+CMD ["python", "vector_func.py"]

imgsearchimg/func/image_search.py

@@ -1,25 +1,20 @@
-import faiss
+
 import numpy as np
 from PIL import Image
-import io
-import os
-from typing import List, Tuple, Optional, Union
+from typing import  Optional
 import torch
 import torchvision.transforms as transforms
 import torchvision.models as models
 from torchvision.models import ResNet50_Weights
-from scipy import ndimage
 
 import torch.nn.functional as F
-from pymongo import MongoClient
-import datetime
 import time
 
 class ImageSearchEngine:
+
     def __init__(self):
-     
-        # 强制使用CPU设备
-        self.device = torch.device("cpu")
+        # 检查GPU是否可用（仅用于PyTorch模型）
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         print(f"使用设备: {self.device}")
         
         # 定义基础预处理转换
@@ -30,28 +25,21 @@ class ImageSearchEngine:
         ])
         
     
-        # 加载预训练的ResNet模型并保持全精度
+        # 加载预训练的ResNet模型
         self.model = models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V2)
         # 移除最后的全连接层
         self.model = torch.nn.Sequential(*list(self.model.children())[:-1])
-        self.model = self.model.float().to(self.device).eval()
+        self.model = self.model.to(self.device)
+        self.model.eval()
         
         # 初始化FAISS索引（2048是ResNet50的特征维度）
         self.dimension = 2048
         # self.index = faiss.IndexFlatL2(self.dimension)
 
-
-    def _batch_generator(self, cursor, batch_size):
-        """从MongoDB游标中分批生成数据"""
-        batch = []
-        for doc in cursor:
-            batch.append(doc)
-            if len(batch) == batch_size:
-                yield batch
-                batch = []
-        if batch:
-            yield batch
-
+        # 改为支持删除的索引
+        # base_index = faiss.IndexFlatL2(self.dimension)
+        # self.index = faiss.IndexIDMap(base_index) 
+        
 
     def _process_image(self, image_path: str) -> Optional[torch.Tensor]:
         """处理单张图片并提取特征。
@@ -109,9 +97,6 @@ class ImageSearchEngine:
             多尺度特征向量，处理失败返回None
         """
         try:
-            # 三重精度保障
-            self.model = self.model.float()
-            
             # 获取原图信息
             orig_w, orig_h = image.size
             max_edge = max(orig_w, orig_h)
@@ -119,14 +104,14 @@ class ImageSearchEngine:
 
             # 动态调整策略 -------------------------------------------
             # 策略1：根据最大边长确定基准尺寸
-            base_size = min(max_edge, 2048)  # 最大尺寸限制
+            base_size = min(max_edge, 3000)  # 不超过模型支持的最大尺寸
             
             # 策略2：自动生成窗口尺寸（等比数列）
             min_size = 224  # 最小特征尺寸
-            num_scales = 3  # 采样点数
+            num_scales = 4  # 固定采样点数
             scale_factors = np.logspace(0, 1, num_scales, base=2)
             window_sizes = [int(base_size * f) for f in scale_factors]
-            window_sizes = sorted({min(max(s, min_size), 2048) for s in window_sizes})
+            window_sizes = sorted({min(max(s, min_size), 3000) for s in window_sizes})
             
             # 策略3：根据长宽比调整尺寸组合
             if aspect_ratio > 1.5:  # 宽幅图像
@@ -143,8 +128,9 @@ class ImageSearchEngine:
                 self.base_transform
             ])
             
-            # 
-            img_base = base_transform(image).unsqueeze(0).to(torch.float32).to(self.device)
+            # 半精度加速
+            self.model.half()
+            img_base = base_transform(image).unsqueeze(0).to(self.device).half()
 
             # 动态特征提取 ------------------------------------------
             features = []
@@ -152,21 +138,21 @@ class ImageSearchEngine:
                 # 保持长宽比的重采样
                 target_size = (int(size*aspect_ratio), size) if aspect_ratio > 1 else (size, int(size/aspect_ratio))
                 
-                # CPU兼容的插值
+                # GPU加速的智能插值
                 img_tensor = torch.nn.functional.interpolate(
                     img_base, 
                     size=target_size,
-                    mode='bilinear',
+                    mode= 'area' if size < base_size else 'bicubic',  # 下采样用area，上采样用bicubic
                     align_corners=False
-                ).to(torch.float32)
+                )
 
                 # 自适应归一化（保持原图统计特性）
                 if hasattr(self, 'adaptive_normalize'):
                     img_tensor = self.adaptive_normalize(img_tensor)
 
                 # 混合精度推理
-                with torch.no_grad():
-                    feature = self.model(img_tensor).to(torch.float32)
+                with torch.no_grad(), torch.cuda.amp.autocast():
+                    feature = self.model(img_tensor)
                 
                 features.append(feature.squeeze().float())
 
@@ -185,9 +171,6 @@ class ImageSearchEngine:
             print(f"智能特征提取失败: {e}")
             return None
 
-
-
-
     def _extract_sliding_window_features(self, image: Image.Image) -> Optional[torch.Tensor]:
         """优化版滑动窗口特征提取（动态调整+批量处理）
         
@@ -198,9 +181,6 @@ class ImageSearchEngine:
             滑动窗口特征向量，处理失败返回None
         """
         try:
-            # 三重精度保障
-            self.model = self.model.float()
-            
             # 获取原图信息
             orig_w, orig_h = image.size
             aspect_ratio = orig_w / orig_h
@@ -212,10 +192,10 @@ class ImageSearchEngine:
                 int(2 ** np.round(np.log2(max_dim * 0.1))),  # 约10%尺寸
                 int(2 ** np.floor(np.log2(max_dim * 0.5))),  # 约50%尺寸
                 int(2 ** np.ceil(np.log2(max_dim)))          # 接近原图尺寸
-            } & {256, 512, 1024, 2048})  # 与预设尺寸取交集
+            } & {256, 512, 1024, 2048, 3000})  # 与预设尺寸取交集
             
             # 智能步长调整（窗口尺寸越大步长越大）
-            stride_ratios = {256:0.5, 512:0.4, 1024:0.3, 2048:0.2}
+            stride_ratios = {256:0.5, 512:0.4, 1024:0.3, 2048:0.2, 3000:0.15}
             
             # 预处理优化 --------------------------------------------
             # 生成基准图像（最大窗口尺寸）
@@ -224,11 +204,15 @@ class ImageSearchEngine:
                         (max_win_size, int(max_win_size / aspect_ratio))
             
             transform = transforms.Compose([
-                transforms.Resize(base_size[::-1], interpolation=transforms.InterpolationMode.BILINEAR),
+                transforms.Resize(base_size[::-1], interpolation=transforms.InterpolationMode.LANCZOS),
                 self.base_transform
             ])
-            base_img = transform(image).to(torch.float32).to(self.device)
+            base_img = transform(image).to(self.device)
             
+            # 半精度加速
+            self.model.half()
+            base_img = base_img.half()
+
             # 批量特征提取 ------------------------------------------
             all_features = []
             for win_size in window_sizes:
@@ -241,7 +225,7 @@ class ImageSearchEngine:
                 num_w = (w - win_size) // stride + 1
                 
                 # 调整窗口数量上限（防止显存溢出）
-                MAX_WINDOWS = 16  # 最大窗口数
+                MAX_WINDOWS = 32  # 根据显存调整
                 if num_h * num_w > MAX_WINDOWS:
                     stride = int(np.sqrt(h * w * win_size**2 / MAX_WINDOWS))
                     num_h = (h - win_size) // stride + 1
@@ -260,11 +244,11 @@ class ImageSearchEngine:
                     continue
 
                 # 批量处理（自动分块防止OOM）
-                BATCH_SIZE = 4  # 批处理大小
-                with torch.no_grad():
+                BATCH_SIZE = 8  # 根据显存调整
+                with torch.no_grad(), torch.cuda.amp.autocast():
                     for i in range(0, len(windows), BATCH_SIZE):
-                        batch = torch.stack(windows[i:i+BATCH_SIZE]).to(torch.float32)
-                        features = self.model(batch).to(torch.float32)
+                        batch = torch.stack(windows[i:i+BATCH_SIZE])
+                        features = self.model(batch)
                         all_features.append(features.cpu().float())  # 转移至CPU释放显存
 
             # 特征融合 ---------------------------------------------
@@ -274,10 +258,9 @@ class ImageSearchEngine:
             final_feature = torch.cat([f.view(-1, f.shape[-1]) for f in all_features], dim=0)
             final_feature = final_feature.mean(dim=0).to(self.device)
 
-            return final_feature.float()
+            return final_feature
 
         except Exception as e:
             print(f"滑动窗口特征提取失败: {e}")
             return None
 
-

imgsearchimg/func/requirements.txt

@@ -1,8 +1,5 @@
-# torch==2.0.1
-# torchvision==0.15.2
-# faiss-cpu==1.7.4
+
 Flask==2.3.3
 Pillow==10.0.0
 numpy==1.24.3
-# scipy==1.11.2
 python-magic==0.4.27

imgsearchimg/func/func_test.py → imgsearchimg/func/vector_func.py

@@ -20,23 +20,6 @@ os.makedirs(UPLOAD_FOLDER, exist_ok=True)
 # 初始化图像搜索引擎
 search_engine = ImageSearchEngine()
 
-
-# @app.route("/initialize", methods=["POST"])
-# def initialize():
-#     # See FC docs for all the HTTP headers: https://www.alibabacloud.com/help/doc-detail/132044.htm#common-headers
-#     request_id = request.headers.get("x-fc-request-id", "")
-#     print("FC Initialize Start RequestId: " + request_id)
-
-#     # do your things
-#     # Use the following code to get temporary credentials
-#     # access_key_id = request.headers['x-fc-access-key-id']
-#     # access_key_secret = request.headers['x-fc-access-key-secret']
-#     # access_security_token = request.headers['x-fc-security-token']
-
-#     print("FC Initialize End RequestId: " + request_id)
-#     return "Function is initialized, request_id: " + request_id + "\n"
-
-
 @app.route("/", methods=["POST"])
 def invoke():
     """处理图片上传请求"""
@@ -54,16 +37,13 @@ def invoke():
         urlretrieve(image_url, image_path)
         
         start_time = time.time()
-        search_engine._process_image(image_path)
+        vector = search_engine._process_image(image_path)
         os.remove(image_path)
-        return jsonify({"vector_spend_time": time.time() - start_time}),200
+        return jsonify({"vector_spend_time": time.time() - start_time,"vector": vector.tolist()}),200
             
     except Exception as e:
         return jsonify({'error': str(e)}), 500
 
-
-
-
 if __name__ == '__main__':
     app.run(host='0.0.0.0', port=5001, debug=False)
 

imgsearchimg/readme.md

@@ -0,0 +1,32 @@
+# 以图搜图api
+
+## 搜图api
+
+1. 加载数据库存储的图片向量特征
+2. 通过图片url调用云函数提取向量
+3. 获取url对应图片特征进行增删查操作
+4. 将product_id集合返回给应用服务
+
+## func云函数提取图片向量特征
+
+1. 通过url内网下载图片然后在容器函数中利用gpu资源提取图片向量特征
+2. 返回图片向量给api服务
+
+
+开发:
+
+## 云函数
+
+1. 整理云函数提取图片向量特征,完善日志和向量返回
+2. 配置集成对象存储内网获取图片
+3. 思考如何优化冷启动的长时间等待
+
+## api
+
+1. 改造api集成云函数处理
+2. 应用服务器与api服务的调用
+
+## 应用服务
+
+1. 确定与搜图api的调用接口(参数/返回)
+2. 测试搜图的实际效果优化云函数图片向量特征算法