GhostNet学习

GhostNet

GhostNet: More Features from Cheap Operations

code

该论文为CVPR 2020的一篇论文，作者来自华为诺亚方舟实验室。

在这篇论文中提出了一个ghost模块，可以通过廉价操作来生成更多的特征图。作者认为，在现有的cnn模型中，会包含丰富的冗余信息，即一些特征图可以由其它特征图廉价操作（线性变换）得到。下面具体介绍下GhostNet模块的实现方式。

ghost module

首先论文中介绍了Ghost Module，通过普通卷积得到m个通道的特征，通过对这m个通道做可分离的线性变换，我们可以得到$n=m\times s$个通道的特征。

kernel_initializer = tf.contrib.layers.variance_scaling_initializer(2.0)

@layer_register(log_shape=True)
def MyDepthConv(x, kernel_shape, channel_mult=1, padding='SAME', stride=1, rate=1, data_format='NHWC',
              W_init=None, activation=tf.identity):
    in_shape = x.get_shape().as_list()
    if data_format=='NHWC':
        in_channel = in_shape[3]
        stride_shape = [1, stride, stride, 1]
    elif data_format=='NCHW':
        in_channel = in_shape[1]
        stride_shape = [1, 1, stride, stride]
    out_channel = in_channel * channel_mult

    if W_init is None:
        W_init = kernel_initializer
    kernel_shape = shape2d(kernel_shape) #[kernel_shape, kernel_shape]
    filter_shape = kernel_shape + [in_channel, channel_mult]

    W = tf.get_variable('DW', filter_shape, initializer=W_init)
    conv = tf.nn.depthwise_conv2d(x, W, stride_shape, padding=padding, rate=[rate,rate], data_format=data_format)
    if activation is None:
        return conv
    else:
        return activation(conv, name='output')

    
def GhostModule(name, x, filters, kernel_size, dw_size, ratio, padding='SAME', strides=1, data_format='NHWC', use_bias=False,
                activation=tf.identity):
    with tf.variable_scope(name):
        init_channels = math.ceil(filters / ratio)
        x = Conv2D('conv1', x, init_channels, kernel_size, strides=strides, activation=activation, data_format=data_format,
                   kernel_initializer=kernel_initializer, use_bias=use_bias)
        if ratio == 1:
            return x #activation(x, name='output')
        dw1 = MyDepthConv('dw1', x, [dw_size,dw_size], channel_mult=ratio-1, stride=1, data_format=data_format, activation=activation)
        dw1 = dw1[:,:,:,:filters-init_channels] if data_format=='NHWC' else dw1[:,:filters-init_channels,:,:]
        x = tf.concat([x, dw1], 3 if data_format=='NHWC' else 1)
        return x

ghost block

在本文中有两种ghost block结构，一种为步长为1，一种为步长为2.两种block的结构分别如下如所示。

其中在步长等于2时，shortcut通过一个步长为2的深度可分离卷积进行降维计算。代码如下：

res = DepthConv(end_point, net, conv_def.kernel, stride=layer_stride, 
                                        data_format='NHWC', activation=BNNoReLU)
                       
res = MyConv2D(end_point, res, depth(conv_def.depth), [1, 1], strides=1, data_format='NHWC',activation=BNNoReLU, use_bias=False)



def MyConv2D(
        inputs,
        filters,
        kernel_size,
        strides=(1, 1),
        padding='same',
        data_format='channels_last',
        dilation_rate=(1, 1),
        activation=None,
        use_bias=True,
        kernel_initializer=kernel_initializer,#tf.contrib.layers.variance_scaling_initializer(2.0),
        bias_initializer=tf.zeros_initializer(),
        kernel_regularizer=None,
        bias_regularizer=None,
        activity_regularizer=None,
        ):
    """
    A wrapper around `tf.layers.Conv2D`.
    Some differences to maintain backward-compatibility:
    1. Default kernel initializer is variance_scaling_initializer(2.0).
    2. Default padding is 'same'.
    Variable Names:
    * ``W``: weights
    * ``b``: bias
    """
    with rename_get_variable({'kernel': 'W', 'bias': 'b'}):
        layer = tf.layers.Conv2D(
            filters,
            kernel_size,
            strides=strides,
            padding=padding,
            data_format=data_format,
            dilation_rate=dilation_rate,
            activation=activation,
            use_bias=use_bias,
            kernel_initializer=kernel_initializer,
            bias_initializer=bias_initializer,
            kernel_regularizer=kernel_regularizer,
            bias_regularizer=bias_regularizer,
            activity_regularizer=activity_regularizer,
            _reuse=tf.get_variable_scope().reuse)
        ret = layer.apply(inputs, scope=tf.get_variable_scope())
        ret = tf.identity(ret, name='output')

    ret.variables = VariableHolder(W=layer.kernel)
    if use_bias:
        ret.variables.b = layer.bias
return ret

def DepthConv(x, kernel_shape, padding='SAME', stride=1, data_format='NHWC',
              W_init=None, activation=tf.identity):
    in_shape = x.get_shape().as_list()
    if data_format=='NHWC':
        in_channel = in_shape[3]
        stride_shape = [1, stride, stride, 1]
    elif data_format=='NCHW':
        in_channel = in_shape[1]
        stride_shape = [1, 1, stride, stride]
    out_channel = in_channel
    channel_mult = out_channel // in_channel

    if W_init is None:
        W_init = kernel_initializer
    kernel_shape = shape2d(kernel_shape) #[kernel_shape, kernel_shape]
    filter_shape = kernel_shape + [in_channel, channel_mult]

    W = tf.get_variable('W', filter_shape, initializer=W_init)
    conv = tf.nn.depthwise_conv2d(x, W, stride_shape, padding=padding, data_format=data_format)
return activation(conv, name='output')