Keras之SqueezeNet

keras实现SqueezeNet 网络，具体网络结构参照https://blog.csdn.net/xbinworld/article/details/50897870博客

之前也这样压缩过自己的网络，识别效果还不错。现在根据模板简单构建下网络。

虽然说参数能够减少很多，其实卷积层数增加了，就是利用时间来换取空间，看个人如何取舍了吧。

#coding:utf-8

import os
os.environ['KERAS_BACKEND'] = 'tensorflow'
import numpy as np
from keras.utils.vis_utils import plot_model
from keras import backend as K
from keras.layers.convolutional import Conv2D, MaxPooling2D
from keras.layers.pooling import GlobalAvgPool2D
from keras.layers import Input,Concatenate
from keras.models import Model
#np.random.seed(55)
'''Fire Module是本文的核心构件，思想非常简单，就是将原来简单的一层conv层变成两层：
squeeze层+expand层，各自带上Relu激活层。在squeeze层里面全是1x1的卷积kernel，数量记为S11；
在expand层里面有1x1和3x3的卷积kernel，数量分别记为E11和E33，
要求S11 < input map number即满足上面的设计原则（2）。
expand层之后将1x1和3x3的卷积output feature maps在channel维度拼接起来。'''
###padding = 'same'   output = int{((w,h)-kernel_size+pading_num)/2}+1 ,pading_num = int(kernel_size/2)
###pading = 'valid'   output = int{((w,h)-kernel_size)/2}+1
def SqueezeNet(img_w,img_h,n_channels):
    x = Input(shape=(img_w,img_h,n_channels))
    conv1 = Conv2D(filters=96,kernel_size=(7,7),strides=(2,2),padding='same',activation='relu')(x)
    maxpool1 = MaxPooling2D(pool_size=(3,3),strides=(2,2))(conv1)
    fire1_squee = Conv2D(filters=16, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(maxpool1)
    fire1_expan1 = Conv2D(filters=64, kernel_size=(3, 3), strides=(1, 1), padding='same', activation='relu')(fire1_squee)
    fire1_expan2 = Conv2D(filters=64, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire1_squee)
    fire1_out = Concatenate(axis=-1)([fire1_expan1, fire1_expan2])
    fire2_squee = Conv2D(filters=16, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire1_out)
    fire2_expan1 = Conv2D(filters=64, kernel_size=(3, 3), strides=(1, 1), padding='same', activation='relu')(fire2_squee)
    fire2_expan2 = Conv2D(filters=64, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire2_squee)
    fire2_out = Concatenate(axis=-1)([fire2_expan1, fire2_expan2])
    fire3_squee = Conv2D(filters=32, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire2_out)
    fire3_expan1 = Conv2D(filters=128, kernel_size=(3, 3), strides=(1, 1), padding='same', activation='relu')(fire3_squee)
    fire3_expan2 = Conv2D(filters=128, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire3_squee)
    fire3_out = Concatenate(axis=-1)([fire3_expan1, fire3_expan2])
    maxpool2 = MaxPooling2D(pool_size=(3,3),strides=(2,2))(fire3_out)
    fire4_squee = Conv2D(filters=32, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(maxpool2)
    fire4_expan1 = Conv2D(filters=128, kernel_size=(3, 3), strides=(1, 1), padding='same', activation='relu')(fire4_squee)
    fire4_expan2 = Conv2D(filters=128, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire4_squee)
    fire4_out = Concatenate(axis=-1)([fire4_expan1, fire4_expan2])
    fire5_squee = Conv2D(filters=48, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire4_out)
    fire5_expan1 = Conv2D(filters=192, kernel_size=(3, 3), strides=(1, 1), padding='same', activation='relu')(fire5_squee)
    fire5_expan2 = Conv2D(filters=192, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire5_squee)
    fire5_out = Concatenate(axis=-1)([fire5_expan1, fire5_expan2])
    fire6_squee = Conv2D(filters=48, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire5_out)
    fire6_expan1 = Conv2D(filters=192, kernel_size=(3, 3), strides=(1, 1), padding='same', activation='relu')(fire6_squee)
    fire6_expan2 = Conv2D(filters=192, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire6_squee)
    fire6_out = Concatenate(axis=-1)([fire6_expan1, fire6_expan2])
    fire7_squee = Conv2D(filters=64, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire6_out)
    fire7_expan1 = Conv2D(filters=256, kernel_size=(3, 3), strides=(1, 1), padding='same', activation='relu')(fire7_squee)
    fire7_expan2 = Conv2D(filters=256, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire7_squee)
    fire7_out = Concatenate(axis=-1)([fire7_expan1, fire7_expan2])
    maxpool3 = MaxPooling2D(pool_size=(3,3),strides=(2,2))(fire7_out)
    fire8_squee = Conv2D(filters=64, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(maxpool3)
    fire8_expan1 = Conv2D(filters=256, kernel_size=(3, 3), strides=(1, 1), padding='same', activation='relu')(fire8_squee)
    fire8_expan2 = Conv2D(filters=256, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire8_squee)
    fire8_out = Concatenate(axis=-1)([fire8_expan1, fire8_expan2])
    conv2 = Conv2D(filters=1000, kernel_size=(1, 1), strides=(1, 1), padding='same', activation='relu')(fire8_out)
    Gap = GlobalAvgPool2D(data_format='channels_last')(conv2)
    model = Model(inputs=x,outputs=Gap)
    return model

if __name__ == '__main__':
    model = SqueezeNet(224,224,3)
    model.summary()
    plot_model(model, to_file='model.png', show_shapes=True, show_layer_names=False)