keras自定义层

当你在复现别人的代码或者有个新奇的点子需要自定义一个网络层时，希望这篇博客可以帮助到你！

主要有两种方法来自定义网络层：

• keras.core.lambda()：简单实现，不包含可训练参数；
• 编写Layer继承类：可实现复杂网络层，可自定义可训练参数；

1. keras.core.lambda()

如果只是对输入进行一些变换，并不包含可训练参数权重，可以自定义个函数，并用lanbda封装成keras支持的网络层。

keras.layers.core.Lambda(function, output_shape=None, mask=None, arguments=None)
# function：要实现的函数，该函数仅接受一个变量，即上一层的输出
# output_shape：函数应该返回的值的shape，可以是一个tuple，也可以是一个根据输入shape计算输出shape的函数
# mask: 
# arguments：可选，字典，用来记录向函数中传递的其他关键字参数

import numpy as np
from keras.layers import Reshape
from keras.layers import merge
from keras.layers import Input, Lambda
from keras.models import Model

# 简单版, 取平方
model.add(Lambda(lambda x: x ** 2))

# 复杂版：数据切片
def slice(x,index):
　　return x[:,:,index]
 
a = Input(shape=(4,2))
x1 = Lambda(slice,output_shape=(4,1),arguments={‘index‘:0})(a)
x2 = Lambda(slice,output_shape=(4,1),arguments={‘index‘:1})(a)
x1 = Reshape((4,1,1))(x1)
x2 = Reshape((4,1,1))(x2)
output = merge([x1,x2],mode=‘concat‘)
 
model = Model(a, output)
x_test = np.array([[[1,2],[2,3],[3,4],[4,5]]])
print(model.predict(x_test))

注意Lambda 是可以进行参数传递的,传递的方式如:x1 = Lambda(slice,output_shape=(4,1),arguments={‘index‘:0})(a)

2. 编写Layer继承类

对于简单、无状态的自定义操作，你也许可以通过 layers.core.Lambda 层来实现。但是对于那些包含了可训练权重的自定义层，你应该自己实现这种层。

在keras中，你只需要实现三个方法即可：

• build(input_shape): 定义权重的地方。这个方法必须设 self.built = True。
• call(x): 这里是编写层的功能的地方。只需要关注传入 call 的第一个参数：输入张量，除非你希望你的层支持masking。
• compute_output_shape(input_shape): 如果你的层更改了输入张量的形状，你应该在这里定义形状变化的逻辑，这让Keras能够自动推断各层的形状。

2.1 简单示例

from keras import backend as K
from keras.engine.topology import Layer

class MyLayer(Layer):

    def __init__(self, output_dim, **kwargs):
        self.output_dim = output_dim
        super(MyLayer, self).__init__(**kwargs)

    def build(self, input_shape):
        # 为该层创建一个可训练的权重
        self.kernel = self.add_weight(name='kernel', 
                                      shape=(input_shape[1], self.output_dim),
                                      initializer='uniform',
                                      trainable=True)
        super(MyLayer, self).build(input_shape)  # 一定要在最后调用它

    def call(self, x):
        return K.dot(x, self.kernel)

    def compute_output_shape(self, input_shape):
        return (input_shape[0], self.output_dim)

2.2 自己实现keras实现DANet中的通道注意力

from keras.layers import Activation, Conv2D
import keras.backend as K
import tensorflow as tf
from keras.layers import Layer


# DANet（语义分割）中的通道注意力模块
class PAM(Layer):
    def __init__(self,
                 gamma_initializer=tf.zeros_initializer(),
                 gamma_regularizer=None,
                 gamma_constraint=None,
                 **kwargs):
        super(PAM, self).__init__(**kwargs)
        self.gamma_initializer = gamma_initializer
        self.gamma_regularizer = gamma_regularizer
        self.gamma_constraint = gamma_constraint

    # 添加权重
    def build(self, input_shape):
        self.gamma = self.add_weight(shape=(1, ),
                                     initializer=self.gamma_initializer,
                                     name='gamma',
                                     regularizer=self.gamma_regularizer,
                                     constraint=self.gamma_constraint)

        self.built = True
	
    def compute_output_shape(self, input_shape):
        return input_shape
	
    # 层功能实现处
    def call(self, input):
        input_shape = input.get_shape().as_list()
        _, h, w, filters = input_shape

        b = Conv2D(filters // 8, 1, use_bias=False, kernel_initializer='he_normal')(input)
        c = Conv2D(filters // 8, 1, use_bias=False, kernel_initializer='he_normal')(input)
        d = Conv2D(filters, 1, use_bias=False, kernel_initializer='he_normal')(input)

        vec_b = K.reshape(b, (-1, h * w, filters // 8))
        vec_cT = tf.transpose(K.reshape(c, (-1, h * w, filters // 8)), (0, 2, 1))
        bcT = K.batch_dot(vec_b, vec_cT)
        softmax_bcT = Activation('softmax')(bcT)
        vec_d = K.reshape(d, (-1, h * w, filters))
        bcTd = K.batch_dot(softmax_bcT, vec_d)
        bcTd = K.reshape(bcTd, (-1, h, w, filters))

        out = self.gamma*bcTd + input
        return out

2.3 keras源码中的LocallyConnected2D

class LocallyConnected2D(Layer):
    """Locally-connected layer for 2D inputs.

    The `LocallyConnected2D` layer works similarly
    to the `Conv2D` layer, except that weights are unshared,
    that is, a different set of filters is applied at each
    different patch of the input.

    # Examples
    ```python
        # apply a 3x3 unshared weights convolution with 64 output filters
        # on a 32x32 image with `data_format="channels_last"`:
        model = Sequential()
        model.add(LocallyConnected2D(64, (3, 3), input_shape=(32, 32, 3)))
        # now model.output_shape == (None, 30, 30, 64)
        # notice that this layer will consume (30*30)*(3*3*3*64)
        # + (30*30)*64 parameters

        # add a 3x3 unshared weights convolution on top, with 32 output filters:
        model.add(LocallyConnected2D(32, (3, 3)))
        # now model.output_shape == (None, 28, 28, 32)
    ```

    # Arguments
        filters: Integer, the dimensionality of the output space
            (i.e. the number of output filters in the convolution).
        kernel_size: An integer or tuple/list of 2 integers, specifying the
            width and height of the 2D convolution window.
            Can be a single integer to specify the same value for
            all spatial dimensions.
        strides: An integer or tuple/list of 2 integers,
            specifying the strides of the convolution along the width and height.
            Can be a single integer to specify the same value for
            all spatial dimensions.
        padding: Currently only support `"valid"` (case-insensitive).
            `"same"` will be supported in future.
        data_format: A string,
            one of `channels_last` (default) or `channels_first`.
            The ordering of the dimensions in the inputs.
            `channels_last` corresponds to inputs with shape
            `(batch, height, width, channels)` while `channels_first`
            corresponds to inputs with shape
            `(batch, channels, height, width)`.
            It defaults to the `image_data_format` value found in your
            Keras config file at `~/.keras/keras.json`.
            If you never set it, then it will be "channels_last".
        activation: Activation function to use
            (see [activations](../activations.md)).
            If you don't specify anything, no activation is applied
            (ie. "linear" activation: `a(x) = x`).
        use_bias: Boolean, whether the layer uses a bias vector.
        kernel_initializer: Initializer for the `kernel` weights matrix
            (see [initializers](../initializers.md)).
        bias_initializer: Initializer for the bias vector
            (see [initializers](../initializers.md)).
        kernel_regularizer: Regularizer function applied to
            the `kernel` weights matrix
            (see [regularizer](../regularizers.md)).
        bias_regularizer: Regularizer function applied to the bias vector
            (see [regularizer](../regularizers.md)).
        activity_regularizer: Regularizer function applied to
            the output of the layer (its "activation").
            (see [regularizer](../regularizers.md)).
        kernel_constraint: Constraint function applied to the kernel matrix
            (see [constraints](../constraints.md)).
        bias_constraint: Constraint function applied to the bias vector
            (see [constraints](../constraints.md)).

    # Input shape
        4D tensor with shape:
        `(samples, channels, rows, cols)` if data_format='channels_first'
        or 4D tensor with shape:
        `(samples, rows, cols, channels)` if data_format='channels_last'.

    # Output shape
        4D tensor with shape:
        `(samples, filters, new_rows, new_cols)` if data_format='channels_first'
        or 4D tensor with shape:
        `(samples, new_rows, new_cols, filters)` if data_format='channels_last'.
        `rows` and `cols` values might have changed due to padding.
    """

    @interfaces.legacy_conv2d_support
    def __init__(self, filters,
                 kernel_size,
                 strides=(1, 1),
                 padding='valid',
                 data_format=None,
                 activation=None,
                 use_bias=True,
                 kernel_initializer='glorot_uniform',
                 bias_initializer='zeros',
                 kernel_regularizer=None,
                 bias_regularizer=None,
                 activity_regularizer=None,
                 kernel_constraint=None,
                 bias_constraint=None,
                 **kwargs):
        super(LocallyConnected2D, self).__init__(**kwargs)
        self.filters = filters
        self.kernel_size = conv_utils.normalize_tuple(kernel_size, 2, 'kernel_size')
        self.strides = conv_utils.normalize_tuple(strides, 2, 'strides')
        self.padding = conv_utils.normalize_padding(padding)
        if self.padding != 'valid':
            raise ValueError('Invalid border mode for LocallyConnected2D '
                             '(only "valid" is supported): ' + padding)
        self.data_format = K.normalize_data_format(data_format)
        self.activation = activations.get(activation)
        self.use_bias = use_bias
        self.kernel_initializer = initializers.get(kernel_initializer)
        self.bias_initializer = initializers.get(bias_initializer)
        self.kernel_regularizer = regularizers.get(kernel_regularizer)
        self.bias_regularizer = regularizers.get(bias_regularizer)
        self.activity_regularizer = regularizers.get(activity_regularizer)
        self.kernel_constraint = constraints.get(kernel_constraint)
        self.bias_constraint = constraints.get(bias_constraint)
        self.input_spec = InputSpec(ndim=4)

    def build(self, input_shape):
        if self.data_format == 'channels_last':
            input_row, input_col = input_shape[1:-1]
            input_filter = input_shape[3]
        else:
            input_row, input_col = input_shape[2:]
            input_filter = input_shape[1]
        if input_row is None or input_col is None:
            raise ValueError('The spatial dimensions of the inputs to '
                             ' a LocallyConnected2D layer '
                             'should be fully-defined, but layer received '
                             'the inputs shape ' + str(input_shape))
        output_row = conv_utils.conv_output_length(input_row, self.kernel_size[0],
                                                   self.padding, self.strides[0])
        output_col = conv_utils.conv_output_length(input_col, self.kernel_size[1],
                                                   self.padding, self.strides[1])
        self.output_row = output_row
        self.output_col = output_col
        self.kernel_shape = (
            output_row * output_col,
            self.kernel_size[0] * self.kernel_size[1] * input_filter,
            self.filters)
        self.kernel = self.add_weight(shape=self.kernel_shape,
                                      initializer=self.kernel_initializer,
                                      name='kernel',
                                      regularizer=self.kernel_regularizer,
                                      constraint=self.kernel_constraint)
        if self.use_bias:
            self.bias = self.add_weight(shape=(output_row, output_col, self.filters),
                                        initializer=self.bias_initializer,
                                        name='bias',
                                        regularizer=self.bias_regularizer,
                                        constraint=self.bias_constraint)
        else:
            self.bias = None
        if self.data_format == 'channels_first':
            self.input_spec = InputSpec(ndim=4, axes={1: input_filter})
        else:
            self.input_spec = InputSpec(ndim=4, axes={-1: input_filter})
        self.built = True

    def compute_output_shape(self, input_shape):
        if self.data_format == 'channels_first':
            rows = input_shape[2]
            cols = input_shape[3]
        elif self.data_format == 'channels_last':
            rows = input_shape[1]
            cols = input_shape[2]

        rows = conv_utils.conv_output_length(rows, self.kernel_size[0],
                                             self.padding, self.strides[0])
        cols = conv_utils.conv_output_length(cols, self.kernel_size[1],
                                             self.padding, self.strides[1])

        if self.data_format == 'channels_first':
            return (input_shape[0], self.filters, rows, cols)
        elif self.data_format == 'channels_last':
            return (input_shape[0], rows, cols, self.filters)

    def call(self, inputs):
        output = K.local_conv2d(inputs,
                                self.kernel,
                                self.kernel_size,
                                self.strides,
                                (self.output_row, self.output_col),
                                self.data_format)

        if self.use_bias:
            output = K.bias_add(output, self.bias, data_format=self.data_format)

        output = self.activation(output)
        return output

    def get_config(self):
        config = {
            'filters': self.filters,
            'kernel_size': self.kernel_size,
            'strides': self.strides,
            'padding': self.padding,
            'data_format': self.data_format,
            'activation': activations.serialize(self.activation),
            'use_bias': self.use_bias,
            'kernel_initializer': initializers.serialize(self.kernel_initializer),
            'bias_initializer': initializers.serialize(self.bias_initializer),
            'kernel_regularizer': regularizers.serialize(self.kernel_regularizer),
            'bias_regularizer': regularizers.serialize(self.bias_regularizer),
            'activity_regularizer':
                regularizers.serialize(self.activity_regularizer),
            'kernel_constraint': constraints.serialize(self.kernel_constraint),
            'bias_constraint': constraints.serialize(self.bias_constraint)
        }
        base_config = super(LocallyConnected2D, self).get_config()
        return dict(list(base_config.items()) + list(config.items()))

2.4 别人的例子

from keras.layers.core import Layer
from keras.engine import InputSpec
from keras import backend as K
try:
    from keras import initializations
except ImportError:
    from keras import initializers as initializations
# 继承父类Layer
class Scale(Layer):
    '''
    该层功能：
        通过向量元素依次相乘（Element wise multiplication）调整上层输出的形状。
        out = in * gamma + beta,
        gamma代表权重weights，beta代表偏置bias
    参数列表：
        axis: int型，代表需要做scale的轴方向，axis=-1 代表选取默认方向（横行）。
        momentum: 对数据方差和标准差做指数平均时的动量.
        weights: 初始权重，是一个包含两个numpy array的list, shapes:[(input_shape,), (input_shape,)]
        beta_init: 偏置量的初始化方法名。(参考Keras.initializers.只有weights未传参时才会使用.
        gamma_init: 权重量的初始化方法名。(参考Keras.initializers.只有weights未传参时才会使用.
    '''
    def __init__(self, weights=None, axis=-1, beta_init = 'zero', gamma_init = 'one', momentum = 0.9, **kwargs):
        # 参数**kwargs代表按字典方式继承父类
        self.momentum = momentum
        self.axis = axis
        self.beta_init = initializers.Zeros()
        self.gamma_init = initializers.Ones()
        self.initial_weights = weights
        super(Scale, self).__init__(**kwargs)

    def build(self, input_shape):
        self.input_spec = [InputSpec(shape=input_shape)]
        # 1：InputSpec(dtype=None, shape=None, ndim=None, max_ndim=None, min_ndim=None, axes=None)
        #Docstring:     
        #Specifies the ndim, dtype and shape of every input to a layer.
        #Every layer should expose (if appropriate) an `input_spec` attribute:a list of instances of InputSpec (one per input tensor).
        #A None entry in a shape is compatible with any dimension
        #A None shape is compatible with any shape.

        # 2:self.input_spec: List of InputSpec class instances
        # each entry describes one required input:
        #     - ndim
        #     - dtype
        # A layer with `n` input tensors must have
        # an `input_spec` of length `n`.

        shape = (int(input_shape[self.axis]),)

        # Compatibility with TensorFlow >= 1.0.0
        self.gamma = K.variable(self.gamma_init(shape), name='{}_gamma'.format(self.name))
        self.beta = K.variable(self.beta_init(shape), name='{}_beta'.format(self.name))

        self.trainable_weights = [self.gamma, self.beta]

        if self.initial_weights is not None:
            self.set_weights(self.initial_weights)
            del self.initial_weights

    def call(self, x, mask=None):
        input_shape = self.input_spec[0].shape
        broadcast_shape = [1] * len(input_shape)
        broadcast_shape[self.axis] = input_shape[self.axis]

        out = K.reshape(self.gamma, broadcast_shape) * x + K.reshape(self.beta, broadcast_shape)
        return out

    def get_config(self):
        config = {"momentum": self.momentum, "axis": self.axis}
        base_config = super(Scale, self).get_config()
        return dict(list(base_config.items()) + list(config.items()))