tensorflow联合学习resnet失败

Question

我对联合学习API进行了一些实验。实际上，我尝试对10个客户培训简单的重新设备。根据数据和指标，培训似乎是成功的。但是评估以及本地和联合的失败。

有人有建议吗？

模型：

def create_keras_resnet_model(): 

    inputs = tf.keras.layers.Input(shape=(28,28,1))
    bn0 = tf.keras.layers.BatchNormalization(scale=True)(inputs)

    conv1 = tf.keras.layers.Conv2D(filters=32, 
                               kernel_size=(7,7), 
                               padding='same', 
                               activation='relu', 
                               kernel_initializer='uniform')(bn0)
    conv1 = tf.keras.layers.Conv2D(filters=32, 
                               kernel_size=(7,7), 
                               padding='same', 
                               activation='relu', 
                               kernel_initializer='uniform')(conv1)
    bn1 = tf.keras.layers.BatchNormalization(scale=True)(conv1)
    max_pool1 = tf.keras.layers.MaxPooling2D(pool_size=(2,2), padding = 'same')(bn1)


    conv2 = tf.keras.layers.Conv2D(filters=32, 
                               kernel_size=(5,5), 
                               padding='same', 
                               activation='relu', 
                               kernel_initializer='uniform')(max_pool1)
    conv2 = tf.keras.layers.Conv2D(filters=32, 
                               kernel_size=(5,5), 
                               padding='same', 
                               activation='relu', 
                               kernel_initializer='uniform')(conv2)
    conv2 = tf.keras.layers.Conv2D(filters=32, 
                               kernel_size=(5,5), 
                               padding='same', 
                               activation='relu', 
                               kernel_initializer='uniform')(conv2)
    bn2 = tf.keras.layers.BatchNormalization(scale=True)(conv2)

    res1_conv = tf.keras.layers.Conv2D(filters = 32,
                                  kernel_size = (3,3),
                                  padding = 'same',
                                  kernel_initializer='uniform')(max_pool1)
    res1_bn = tf.keras.layers.BatchNormalization(scale=True)(res1_conv)

    add1 = tf.keras.layers.Add()([res1_bn, bn2])


    act1 = tf.keras.layers.Activation('relu')(add1)
    max_pool2 = tf.keras.layers.MaxPooling2D(pool_size=(2,2), padding = 'same')(act1)

    conv3 = tf.keras.layers.Conv2D(filters=32, 
                               kernel_size=(5,5), 
                               padding='same', 
                               activation='relu', 
                               kernel_initializer='uniform')(max_pool2)
    conv3 = tf.keras.layers.Conv2D(filters=32, 
                               kernel_size=(5,5), 
                               padding='same', 
                               activation='relu', 
                               kernel_initializer='uniform')(conv3)
    conv3 = tf.keras.layers.Conv2D(filters=32, 
                               kernel_size=(5,5), 
                               padding='same', 
                               activation='relu', 
                               kernel_initializer='uniform')(conv3)
    bn2 = tf.keras.layers.BatchNormalization(scale=True)(conv3)

    res2_conv = tf.keras.layers.Conv2D(filters = 32,
                                  kernel_size = (3,3),
                                  padding = 'same',
                                  kernel_initializer='uniform')(max_pool2)
    res2_bn = tf.keras.layers.BatchNormalization(scale=True)(res2_conv)

    add2 = tf.keras.layers.Add()([res2_bn, bn2])

    act2 = tf.keras.layers.Activation('relu')(add2)
    max_pool3 = tf.keras.layers.MaxPooling2D(pool_size=(2,2), padding = 'same')(act2)

    flatten = tf.keras.layers.Flatten()(max_pool3)

    dense1 = tf.keras.layers.Dense(128, activation='relu')(flatten)

    do = tf.keras.layers.Dropout(0.20)(dense1)
    dense2 = tf.keras.layers.Dense(10, activation='softmax')(do)

    model = tf.keras.models.Model(inputs=[inputs], outputs=[dense2])

    return model

模型只是一个简单的重新设备。 在培训中，我为EMNIST使用TensorFlow联合模拟数据集，在这里为10个时代使用10个客户。

到目前为止一切都很好...

我已经调整了准备数据的提供的功能。我已经用一个简单的CNN测试了整个过程，所有这些都非常安静。

def preprocess(dataset):

def batch_format_fn(element):
    return collections.OrderedDict(
        x=tf.reshape(element['pixels'], [-1, 28, 28, 1]),
        y=tf.reshape(element['label'], [-1, 1])
    )
    
return dataset.repeat(NUM_EPOCHS).shuffle(SHUFFLE_BUFFER, seed=42).batch(BATCH_SIZE).map(batch_format_fn).prefetch(PREFETCH_BUFFER)

使用TensorFlow进行评估过程显示出一个奇怪的结果。准确性将达到11％左右，损失的含义在7到8之间。

如果我将权重复制到本地模型并在本地进行评估，则相同的结果。如果我尝试从测试数据中预测单个图像，则会引发异常：

ValueError: Input 0 of layer dense_10 is incompatible with the layer: expected axis -1 of input shape to have value 512 but received input with shape (None, 128)

此处的模型summaray：

Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_13 (InputLayer)           [(None, 28, 28, 1)]  0                                            
__________________________________________________________________________________________________
batch_normalization_72 (BatchNo (None, 28, 28, 1)    4           input_13[0][0]                   
__________________________________________________________________________________________________
conv2d_120 (Conv2D)             (None, 28, 28, 32)   1600        batch_normalization_72[0][0]     
__________________________________________________________________________________________________
conv2d_121 (Conv2D)             (None, 28, 28, 32)   50208       conv2d_120[0][0]                 
__________________________________________________________________________________________________
batch_normalization_73 (BatchNo (None, 28, 28, 32)   128         conv2d_121[0][0]                 
__________________________________________________________________________________________________
max_pooling2d_36 (MaxPooling2D) (None, 14, 14, 32)   0           batch_normalization_73[0][0]     
__________________________________________________________________________________________________
conv2d_122 (Conv2D)             (None, 14, 14, 32)   25632       max_pooling2d_36[0][0]           
__________________________________________________________________________________________________
conv2d_123 (Conv2D)             (None, 14, 14, 32)   25632       conv2d_122[0][0]                 
__________________________________________________________________________________________________
conv2d_125 (Conv2D)             (None, 14, 14, 32)   9248        max_pooling2d_36[0][0]           
__________________________________________________________________________________________________
conv2d_124 (Conv2D)             (None, 14, 14, 32)   25632       conv2d_123[0][0]                 
__________________________________________________________________________________________________
batch_normalization_75 (BatchNo (None, 14, 14, 32)   128         conv2d_125[0][0]                 
__________________________________________________________________________________________________
batch_normalization_74 (BatchNo (None, 14, 14, 32)   128         conv2d_124[0][0]                 
__________________________________________________________________________________________________
add_24 (Add)                    (None, 14, 14, 32)   0           batch_normalization_75[0][0]     
                                                                 batch_normalization_74[0][0]     
__________________________________________________________________________________________________
activation_24 (Activation)      (None, 14, 14, 32)   0           add_24[0][0]                     
__________________________________________________________________________________________________
max_pooling2d_37 (MaxPooling2D) (None, 7, 7, 32)     0           activation_24[0][0]              
__________________________________________________________________________________________________
conv2d_126 (Conv2D)             (None, 7, 7, 32)     25632       max_pooling2d_37[0][0]           
__________________________________________________________________________________________________
conv2d_127 (Conv2D)             (None, 7, 7, 32)     25632       conv2d_126[0][0]                 
__________________________________________________________________________________________________
conv2d_129 (Conv2D)             (None, 7, 7, 32)     9248        max_pooling2d_37[0][0]           
__________________________________________________________________________________________________
conv2d_128 (Conv2D)             (None, 7, 7, 32)     25632       conv2d_127[0][0]                 
__________________________________________________________________________________________________
batch_normalization_77 (BatchNo (None, 7, 7, 32)     128         conv2d_129[0][0]                 
__________________________________________________________________________________________________
batch_normalization_76 (BatchNo (None, 7, 7, 32)     128         conv2d_128[0][0]                 
__________________________________________________________________________________________________
add_25 (Add)                    (None, 7, 7, 32)     0           batch_normalization_77[0][0]     
                                                                 batch_normalization_76[0][0]     
__________________________________________________________________________________________________
activation_25 (Activation)      (None, 7, 7, 32)     0           add_25[0][0]                     
__________________________________________________________________________________________________
max_pooling2d_38 (MaxPooling2D) (None, 4, 4, 32)     0           activation_25[0][0]              
__________________________________________________________________________________________________
flatten_12 (Flatten)            (None, 512)          0           max_pooling2d_38[0][0]           
__________________________________________________________________________________________________
dense_24 (Dense)                (None, 128)          65664       flatten_12[0][0]                 
__________________________________________________________________________________________________
dropout_12 (Dropout)            (None, 128)          0           dense_24[0][0]                   
__________________________________________________________________________________________________
dense_25 (Dense)                (None, 10)           1290        dropout_12[0][0]                 
==================================================================================================
Total params: 291,694
Trainable params: 291,372
Non-trainable params: 322
__________________________________________________________________________________________________

我没有从Karas Util包中带有to_categorical函数的标签。但是，为什么例外，密集层的输入是错误的？为什么培训工作？

Answer 1

问题是批处理层。在执行下一个（）函数在Tensorflow联合框架中，只有可训练的权重处理。批处理层的值是不可训练的。但是，每个客户都会有自己的均值和标准偏差。对于此层，需要禁用培训。

    for layer in model.layers:
        if isinstance(layer, tf.keras.layers.BatchNormalization):
            layer.trainable = False
            layer._per_input_updates = {}