TF版网络模型搭建常用代码备忘

本文主要介绍如何搭建一个网络并训练

最近，我在写代码时经常碰到这样的情况，明明记得代码应该怎么写，在写出来的代码调试时，总是有些小错误。原因不是接口参数个数不对，就是位置不对。为了节约上网查找时间，现记录下常用操作，以备需要时快速查看。

根据网络结构不同功能，主要分这几大块：网络基本结构元组件，网络常用结构，Tensorboard调试接口，数据预处理常用操作，后处理常用操作。

1、搭建一个基础网络所需的元组件：

import tensorflow as tf
import glog as log

class basenet(object):
'''
base model for other specific cnn
'''
def __init__(self):
      pass

def weight_variable(self,shape,sdtdev=0.1,name):
     initial=tf.truncated_normal(shape=shape,mean=0.0,stddev=sdtdev)
     if name is None:
         return tf.Variable(initial)
     else:
         return tf.get_variable(name=name,initial=initialdef bias_variable(self,shape,name):     initial=tf.constant(.01,shape=shape)

     if name is None:
         return tf.Variable(initial,name=name)
     else:
         return tf.get_variable(name=name,initial=initial)
@staticmethod
def conv2d(self,x,w,s=1,name=None,padding='SAME'):
    #with tf.variable_scope(name):
    if s == 1:
         x = tf.nn.conv2d(x,w,strides=[1,s,s,1],padding=padding)
    else:
         x = tf.nn.conv2d(x,w,strides=[1,s,s,1],padding=padding)
         #log.info('basenet conv2d x:{:}'.format(x.get_shape().as_list()))
    return x

def conv2d_transpose(x,w,b,output_shape,stride=2):
    if output_shape is None:
        output_shape =x.get_shape().as_list()
        output_shape[1]*=2
        output_shape[2]*=2
        output_shape[3]=w.get_shape().as_list()[2]
    conv = tf.nn.conv2d_transpose(x,w,output_shape,strides=[1,stride,stride,1],padding='SAME')
    return tf.nn.bias_add(conv,b)
def maxpool(self,x,k=2,s=2,padding='SAME'):
    x= tf.nn.max_pool(x,ksize=[1,k,k,1],strides=[1,s,s,1],padding=padding)
    return x

def avgpool(self,x,k,s,padding='SAME'):
    x= tf.nn.avg_pool(x,ksize=[1,k,k,1],strides=[1,s,s,1],padding=padding)

def local_response_norm(x):
    return tf.nn.lrn(x,depth_radius=5,bias=2,alpha=1e-4,beta=0.75)
def relu(self,x,name):
    x = tf.nn.relu(x)
    return x

def leaky_relu(x,alpha=0.0,name=""):
    return tf.maximum(alpha*x,x,name)

def relu6(self,x):
    x= min(max(0,x), 6)
    x = tf.nn.relu(x)
    return x

def batch_norm(x,output,phrase,scope='bn',decay=0.9,eps=1e-5):
    with tf.variable_scope(scope):
        beta=tf.get_variable(name='beta',shape=[output],initializer=tf.constant_initializer(0.05))
        gamma=tf.get_variable(name='gamma',shape=[output],initialzer=tf.random_normal_initializer(1.0,0.02)
        batch_mean,batch_var=tf.nn.moment(x,[0,1,2],name='moment')

        def mean_var_2_update():
            ema_apply_op = ema.apply([batch_mean,batch_var])
            with tf.control_dependencies([ema_apply_op]):
                return tf.identity(batch_mean),tf.identity(batch_var)
        mean,var = tf.cond(phrase,mean_var_2_update,lambda:(ema.average(batch_mean),ema.average(batch_var))
        normed = tf.nn.batch_normalization(x,mean,var,beta,gamma,eps)
    return normed
def wx_b(self,x,w,b):
    x = tf.matmul(x,w)+b
    log.info('basenet wx_b x:{:}'.format(x.get_shape().as_list()))
    return x

def fc(self,x,w,b):
    x = tf.add(tf.matmul(x,w),b)
    return x        

2、常用网络结构

网络结构1：

def bottleneck_unit(x,out_chan1,out_chan2,down_stride=False,up_stride=False,name=None):

3、数据预处理常用操作

def save_img(image,save_dir,name,mean=None):
    if mean:
        image=unprocess_image(image,mean)
    misc.imsave(os.path.join(save_dir,name+'.png'),image)

def process_image(image,mean_pixel):
    return image-mean_pixel

4、Tensorboard常用接口

def add_regular_to_summary(var):
    if var is not None:
        tf.summary.histogram(var.op.name,var)
        tf.add_to_collection('reg_loss',tf.nn.l2_loss(var))

def add_activation_to_summary(var):
    if var is not None:
        tf.summary.histogram(var.op.name+'/activation',var)
        tf.summary.scalar(var.op.name+'/sparsity',tf.nn.zero_fraction(var))

def add_gradient_to_summary(grad,var):
    if grad is not None:
        tf.summary.histogram(var.op.name+'/gradient',grad)