CNN 3 LeNet5, AlexNet, MyCNN,ResNet 模型重载

  model 	accuracy
{'AlexNet.h5': 0.3725, 
'LeNet5.h5': 0.2192, 
'MiniVGG_CNN_model.h5': 0.1748, 
'MiniVGG_CNN_model_augu.h5': 0.1752, 
'MyCNN.h5': 0.2073}

对比文章二的准确率结果来看，重新加载模型准确率大大降低，但是AlexNet反而性能得到提升。
有知道原因的可以留言。
GitHub 上同样情况的讨论
模型重载，预测

def model_prediction(model_name, test_size):
    (X_train_full, y_train_full), (X_test, y_test) = datasets.cifar10.load_data()
    class_names = ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']
    # X_train_full, X_test = X_train_full / 255.0, X_test / 255.0
    # X_train, y_train = X_train_full[0:1000], y_train_full[0:1000]
    # X_valid, y_valid = X_train_full[1000:2000], y_train_full[1000:2000]
    # print(y_test[:10])
    idxs = np.arange(X_test.shape[0])
    np.random.shuffle(idxs)
    X_test, y_test = X_test[idxs], y_test[idxs]
    # print(y_test[:10])


    model = keras.models.load_model(model_name)
    model.summary()
    r = model.predict(X_test[:test_size], batch_size=32, verbose=1)
    count = 0
    wrong_names = []
    for i in range(test_size):
        if np.argmax(r[i]) == y_test[i][0]:
            count += 1
        else:
            wrong_names.append(class_names[y_test[i][0]])
    # print(count, wrong_names)

    # plot_images(X_test[:test_size], y_train=y_test[:test_size])
    # plot_images(X_test[:test_size], prediction=r, flag=True)

    return count/test_size

运行每个程序，得到预测准确率

if __name__ == '__main__':

    model_percent = {}
    name = ["AlexNet.h5", "LeNet5.h5", "MiniVGG_CNN_model.h5", "MiniVGG_CNN_model_augu.h5", "MyCNN.h5"]

    for n in name:
        K.clear_session()  # Some memory clean-up
        my_per = model_prediction(n, 10000)
        model_percent[n] = my_per

    print(model_percent)