Keras.predict为预训练模型给出不同的结果

Question

我有一个经过预先训练的分类器，当应用程序启动时我正在加载其权重。

TF版本1.14.0 Keras版本2.2.4

this.props.children[0]()

我正在加载具有正确ID且形状正确的给定图像，并且每次从分类器获得不同的预测

    def create_discriminators_256(self):


        inp_256 = Input(shape = [256, 256, 3])


        conv256_filters=32


        x=Conv2D(filters = conv256_filters, kernel_size = 3, padding = 'same', activation = 'relu', kernel_initializer = 'he_normal')(inp_256)

        x = d_block(x, 64) #128
        x = d_block(x, 128) #64
        x = d_block(x, 192) #32
        x = d_block(x, 256) #16
        x = d_block(x, 320) #8
        x = d_block(x, 320)  #4

        x = Conv2D(filters = 128, kernel_size = 3, padding = 'same',activation="relu", kernel_initializer = 'he_normal')(x)
        x = Dropout(0.4)(x)

        x = Conv2D(filters = 1, kernel_size = 3, padding = 'same', kernel_initializer = 'he_normal')(x)
        x = Dropout(0.4)(x)

        x = GlobalAveragePooling2D()(x)

        model_D_256 = Model(inputs = [inp_256], outputs = x)



def d_block(inp, fil):

    r = Conv2D(filters = fil, kernel_size = 3, padding = 'same', kernel_initializer = 'he_normal')(inp)
    r = LeakyReLU(0.01)(r)
    r = AveragePooling2D()(r)
    r = Conv2D(filters = fil, kernel_size = 3, padding = 'same', kernel_initializer = 'he_normal')(r)
    out = LeakyReLU(0.01)(r)

    return out

我检查了加载后分类器网络的权重是否始终相同，因此问题似乎来自于预测函数。

这里没有进行任何培训，因为我正在加载预先训练的模型的重量。

有什么想法吗？