fasterRcnn实现（一）

特征提取主干网络采用resnet50

resnet50结构图

其中，conv2d_block和identity_block结构图为

主干网络实现采用tf.contrib.slim模块

# -*- coding: UTF-8 -*-

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt 
import tensorflow.contrib.slim as slim

def conv_block(input_tensor,kernal_size,filters,stage,block,strides=2,is_training=True):
   
    filters1,filters2,filters3 = filters
    
    conv_name_base = 'res' + str(stage) + block + '_branch'
    bn_name_base = 'bn' + str(stage) + block + '_branch'
    
    with slim.arg_scope([slim.conv2d],
                        activation_fn=tf.nn.relu,
                        weights_initializer=tf.truncated_normal_initializer(0.0, 0.01),
                        weights_regularizer=slim.l2_regularizer(0.0005),
                        normalizer_fn=slim.batch_norm,
                        normalizer_params={'is_training': is_training,
                                           'zero_debias_moving_mean': True,
                                           'decay': 0.99, 
                                           'epsilon': 0.0001,         
                                           'scale': 0.2,
                                           'updates_collections': tf.GraphKeys.UPDATE_OPS}):   
        
        x = slim.conv2d(input_tensor,filters1,[1,1],stride=strides,scope=conv_name_base + '2a')

        
        x = slim.conv2d(x,filters2,kernal_size,scope=conv_name_base + '2b')

        
        x = slim.conv2d(x,filters3,[1,1],activation_fn=None,scope=conv_name_base + '2c')

        
        shortcut = slim.conv2d(input_tensor,filters3,[1,1],stride=strides,activation_fn=None,scope=conv_name_base+'1')
        
        x = x + shortcut
        x = tf.nn.relu(x)
    return x

def identity_block(input_tensor,kernal_size,filters,stage,block,is_training=True):
    filters1,filters2,filters3 = filters
    
    conv_name_base = 'res' + str(stage) + block + '_branch'
    bn_name_base = 'bn' + str(stage) + block + '_branch'      
    
    with slim.arg_scope([slim.conv2d],
                        activation_fn=tf.nn.relu,
                        weights_initializer=tf.truncated_normal_initializer(0.0, 0.01),
                        weights_regularizer=slim.l2_regularizer(0.0005),
                        normalizer_fn=slim.batch_norm,
                        normalizer_params={'is_training': is_training,
                                           'zero_debias_moving_mean': True,
                                           'decay': 0.99, 
                                           'epsilon': 0.0001,         
                                           'scale': 0.2,
                                           'updates_collections': tf.GraphKeys.UPDATE_OPS}):   
        
        x = slim.conv2d(input_tensor,filters1,[1,1],scope= conv_name_base + '2a')
        x = slim.conv2d(x,filters2,kernal_size,scope=conv_name_base + '2b')
        x = slim.conv2d(x,filters3,[1,1],scope=conv_name_base + '2c')
        
        x = x + input_tensor
        x = tf.nn.relu(x)
    return x
   
def ResNet50(inputs,is_training = True):
    #建立两个瓶颈结构，conv_block有可能改变特征图的大小，identity_block不会改变特征图的大小
    
    img_input = inputs
    
    with slim.arg_scope([slim.conv2d],
                        activation_fn=tf.nn.relu,
                        weights_initializer=tf.truncated_normal_initializer(0.0, 0.01),
                        weights_regularizer=slim.l2_regularizer(0.0005),
                        normalizer_fn=slim.batch_norm,
                        #加入bn层后，要定义传入batch_norm的参数，是字典形式
                        normalizer_params={'is_training': is_training,
                                           'zero_debias_moving_mean': True,
                                           'decay': 0.99, 
                                           'epsilon': 0.0001,         
                                           'scale': 0.2,
                                           'updates_collections': tf.GraphKeys.UPDATE_OPS}):    
        #stage1 --压缩4倍
        paddings = [[0,0],[3,3],[3,3],[0,0]]
        x = tf.pad(img_input,paddings,"CONSTANT")
        x = slim.conv2d(x, 64, (7,7),stride=2,padding='VALID',scope='conv1') # 上面补0了，这里就不用SAME了，特征图计算公式是针对padding=valid的（自己手动补0），如果padding=SAME，stride=2，特征图为输入的1/2，如果stride=1，特征图和输入大小一致。
       
        x = slim.max_pool2d(x,[3,3],padding="SAME")
        
        #stage2 --不压缩
        x = conv_block(x,[3,3],[64,64,256],stage=2,block='a',strides=1)
        x = identity_block(x,[3,3],[64,64,256],stage=2,block='b')
        x = identity_block(x,[3,3],[64,64,256],stage=2,block='c')  
        
        #stage3 --压缩2倍
        x = conv_block(x,[3,3],[128,128,512],stage=3,block='a')
        x = identity_block(x,[3,3],[128,128,512],stage=3,block='b')
        x = identity_block(x,[3,3],[128,128,512],stage=3,block='c')
        x = identity_block(x,[3,3],[128,128,512],stage=3,block='d')
        
        #stage4 --压缩2倍
        x = conv_block(x,[3,3],[256,256,1024],stage=4,block='a')
        x = identity_block(x,[3,3],[256,256,1024],stage=4,block='b')
        x = identity_block(x,[3,3],[256,256,1024],stage=4,block='c')
        x = identity_block(x,[3,3],[256,256,1024],stage=4,block='d')
        x = identity_block(x,[3,3],[256,256,1024],stage=4,block='e')
        x = identity_block(x,[3,3],[256,256,1024],stage=4,block='f')
    
    #输入如果是600*600的图像，返回的时38X38的特征
    return x

if __name__ == '__main__':
    
    img = tf.constant(1,dtype=tf.float32,shape=[1,600,600,3])
    x = ResNet50(img)
    print(x)

这样就得到了共享特征图。