tensorflow - 如何在Keras模型中替换(或插入)中间层？

Question

我有一个训练有素的 Keras 模型，我想要:

1)用相同但没有偏差的Con2D层替换Con2D层。

2) 在第一次激活之前添加 BatchNormalization 层

我该怎么做？

def keras_simple_model():
    from keras.models import Model
    from keras.layers import Input, Dense,  GlobalAveragePooling2D
    from keras.layers import Conv2D, MaxPooling2D, Activation

    inputs1 = Input((28, 28, 1))
    x = Conv2D(4, (3, 3), activation=None, padding='same', name='conv1')(inputs1)
    x = Activation('relu')(x)
    x = Conv2D(4, (3, 3), activation=None, padding='same', name='conv2')(x)
    x = Activation('relu')(x)
    x = MaxPooling2D((2, 2), strides=(2, 2), name='pool1')(x)

    x = Conv2D(8, (3, 3), activation=None, padding='same', name='conv3')(x)
    x = Activation('relu')(x)
    x = Conv2D(8, (3, 3), activation=None, padding='same', name='conv4')(x)
    x = Activation('relu')(x)
    x = MaxPooling2D((2, 2), strides=(2, 2), name='pool2')(x)

    x = GlobalAveragePooling2D()(x)
    x = Dense(10, activation=None)(x)
    x = Activation('softmax')(x)

    model = Model(inputs=inputs1, outputs=x)
    return model


if __name__ == '__main__':
    model = keras_simple_model()
    print(model.summary())

Answer 1

以下函数允许您在之前、之后插入新层，或者替换原始模型中名称与正则表达式，包括非序列模型，例如 DenseNet 或 ResNet。

import re
from keras.models import Model

def insert_layer_nonseq(model, layer_regex, insert_layer_factory,
                        insert_layer_name=None, position='after'):

    # Auxiliary dictionary to describe the network graph
    network_dict = {'input_layers_of': {}, 'new_output_tensor_of': {}}

    # Set the input layers of each layer
    for layer in model.layers:
        for node in layer._outbound_nodes:
            layer_name = node.outbound_layer.name
            if layer_name not in network_dict['input_layers_of']:
                network_dict['input_layers_of'].update(
                        {layer_name: [layer.name]})
            else:
                network_dict['input_layers_of'][layer_name].append(layer.name)

    # Set the output tensor of the input layer
    network_dict['new_output_tensor_of'].update(
            {model.layers[0].name: model.input})

    # Iterate over all layers after the input
    model_outputs = []
    for layer in model.layers[1:]:

        # Determine input tensors
        layer_input = [network_dict['new_output_tensor_of'][layer_aux] 
                for layer_aux in network_dict['input_layers_of'][layer.name]]
        if len(layer_input) == 1:
            layer_input = layer_input[0]

        # Insert layer if name matches the regular expression
        if re.match(layer_regex, layer.name):
            if position == 'replace':
                x = layer_input
            elif position == 'after':
                x = layer(layer_input)
            elif position == 'before':
                pass
            else:
                raise ValueError('position must be: before, after or replace')

            new_layer = insert_layer_factory()
            if insert_layer_name:
                new_layer.name = insert_layer_name
            else:
                new_layer.name = '{}_{}'.format(layer.name, 
                                                new_layer.name)
            x = new_layer(x)
            print('New layer: {} Old layer: {} Type: {}'.format(new_layer.name,
                                                            layer.name, position))
            if position == 'before':
                x = layer(x)
        else:
            x = layer(layer_input)

        # Set new output tensor (the original one, or the one of the inserted
        # layer)
        network_dict['new_output_tensor_of'].update({layer.name: x})

        # Save tensor in output list if it is output in initial model
        if layer_name in model.output_names:
            model_outputs.append(x)

    return Model(inputs=model.inputs, outputs=model_outputs)

与纯顺序模型的简单情况的区别在于，在迭代各层以查找关键层之前，您首先解析图形并将每个层的输入层存储在辅助字典中。然后，当您迭代各层时，您还存储每层的新输出张量，该张量用于在构建新模型时确定每层的输入层。

一个用例如下，其中在 ResNet50 的每个激活层之后插入 Dropout 层:

from keras.applications.resnet50 import ResNet50
from keras.models import load_model

model = ResNet50()
def dropout_layer_factory():
    return Dropout(rate=0.2, name='dropout')
model = insert_layer_nonseq(model, '.*activation.*', dropout_layer_factory)

# Fix possible problems with new model
model.save('temp.h5')
model = load_model('temp.h5')

model.summary()