facenet_keras.h5能否适配TensorFlow2.16.1与Keras3.3.3？遇EOFError求解

我尝试结合FaceNet Keras、TensorFlow与MTCNN实现人脸识别，但运行代码时抛出EOFError: EOF read where object expected错误，想确认facenet_keras.h5模型是否还能正常使用。

我的环境版本：

• TensorFlow 2.16.1
• Keras 3.3.3
• Python 3.11.9

此前无需加载Keras模型的代码在该环境下可正常运行，相关完整代码如下：

import os
from os import listdir
from PIL import Image as Img
from numpy import asarray
from numpy import expand_dims
from matplotlib import pyplot
from keras.models import load_model  #type: ignore
import numpy as np
import tensorflow as tf
from keras_facenet import FaceNet # type: ignore
import pickle
import cv2

HaarCascade = cv2.CascadeClassifier(cv2.samples.findFile(cv2.data.haarcascades + 'haarcascade_frontalface_default.xml'))

MyFaceNet = load_model("F:\\Andre\\FaceRecognitiontest/facenet_keras.h5")

folder='faces/'
database = {}

for filename in listdir(folder):

    path = folder + filename
    gbr1 = cv2.imread(folder + filename)
    
    wajah = HaarCascade.detectMultiScale(gbr1,1.1,4)
    
    if len(wajah)>0:
        x1, y1, width, height = wajah[0]         
    else:
        x1, y1, width, height = 1, 1, 10, 10
        
    x1, y1 = abs(x1), abs(y1)
    x2, y2 = x1 + width, y1 + height
    
    gbr = cv2.cvtColor(gbr1, cv2.COLOR_BGR2RGB)
    gbr = Img.fromarray(gbr)                  # 从OpenCV格式转PIL格式
    gbr_array = asarray(gbr)
    
    face = gbr_array[y1:y2, x1:x2]                        
    
    face = Img.fromarray(face)                       
    face = face.resize((160,160))
    face = asarray(face)
    
    face = face.astype('float32')
    mean, std = face.mean(), face.std()
    face = (face - mean) / std
    
    face = expand_dims(face, axis=0)
    signature = MyFaceNet.predict(face)
    
    database[os.path.splitext(filename)[0]]=signature

myfile = open("data.pkl", "wb")
pickle.dump(database, myfile)
myfile.close()

myfile = open("data.pkl", "rb")
database = pickle.load(myfile)
myfile.close()

def findFaces(frame):
    gbr1 = frame
    gbr = cv2.cvtColor(gbr1, cv2.COLOR_BGR2RGB)
    gbr = Img.fromarray(gbr)                  # 从OpenCV格式转PIL格式
    gbr_array = asarray(gbr)

    wajah = HaarCascade.detectMultiScale(gbr1,1.1,4)

    for (x1,y1,w,h) in wajah:
        x1, y1 = abs(x1), abs(y1)
        x2, y2 = x1 + w, y1 + h

        face = gbr_array[y1:y2, x1:x2]                        

        face = Img.fromarray(face)                       
        face = face.resize((160,160))
        face = asarray(face)

        face = face.astype('float32')
        mean, std = face.mean(), face.std()
        face = (face - mean) / std

        face = expand_dims(face, axis=0)
        signature = MyFaceNet.predict(face)

        min_dist=100
        identity=' '
        for key, value in database.items() :
            dist = np.linalg.norm(value-signature)
            if dist < min_dist:
                min_dist = dist
                identity = key

        cv2.putText(gbr1,identity, (x1,y1),cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 0), 1, cv2.LINE_AA)
        cv2.rectangle(gbr1,(x1,y1),(x2,y2), (0,255,0), 2)

    return gbr1

cap = cv2.VideoCapture(0)

while True:
    ret, frame = cap.read()
    if not ret:
        break

    frame = findFaces(frame)
    cv2.imshow('Live Stream', frame)

    if cv2.waitKey(1) & 0xFF == ord('q'):
        break

cap.release()
cv2.destroyAllWindows()

错误核心原因 #

EOFError通常意味着程序读取文件时提前遇到了文件末尾，针对你的场景，大概率是模型文件损坏或Keras版本兼容性问题导致加载失败。

验证模型可用性的步骤 #

1. 检查模型文件完整性

   • 对比文件大小：官方发布的facenet_keras.h5约为100MB左右，如果你的文件远小于这个值，说明下载过程中断，需重新下载完整模型。
   • 用HDF5工具验证文件结构：

     import h5py
     try:
         with h5py.File("F:\\Andre\\FaceRecognitiontest/facenet_keras.h5", 'r') as f:
             print("模型文件结构正常：", list(f.keys()))
     except Exception as e:
         print("模型文件损坏：", str(e))
     

     若报错则说明文件已损坏，需重新获取模型。

2. 适配Keras 3.x兼容性
   旧版FaceNet模型是用Keras 2.x保存的，直接用Keras 3加载可能存在格式兼容问题，修改加载代码如下：

   # 改用TensorFlow集成的Keras加载，并关闭编译（避免旧编译配置冲突）
   from tensorflow.keras.models import load_model
   MyFaceNet = load_model("F:\\Andre\\FaceRecognitiontest/facenet_keras.h5", compile=False)
   

3. 替换为keras_facenet库的自动加载方式
   你已经导入了keras_facenet.FaceNet，可以直接使用该库的预训练模型，无需手动管理h5文件，兼容问题由库自动处理：

   from keras_facenet import FaceNet
   MyFaceNet = FaceNet()  # 自动加载官方预训练模型
   

4. 排除pickle环节干扰
   暂时注释掉代码中pickle.dump和pickle.load相关逻辑，单独测试模型加载和单张图片预测，确认错误是否来自模型加载环节。

内容的提问来源于stack exchange，提问作者André Pimentel