Tensorflow:空洞卷积(Atrous convolution) atrous_conv2d的代码实现

12 篇文章 1 订阅
11 篇文章 0 订阅

一、参考文献

https://blog.csdn.net/mao_xiao_feng/article/details/78003730

https://blog.csdn.net/davincil/article/details/80902366

二、概念

空洞卷积的基本概念则是扩大卷积核,例如下图(作图为标准卷积核,有图为空洞卷积核):

                  

这里使用空洞卷积的时候padding一定要注意,因为卷积核可能比输入还要打,所以尽量使用padding=‘SAME’

这篇博客中对空洞卷积有一个详细的流程图解释:https://blog.csdn.net/mao_xiao_feng/article/details/78003730

上图中输入是一个4*4的一个双通道的图片,在padding=same中,卷积核相乘的时候进行了填充,相反valid是不对输出进行天填充,所以最后的输出的维度是偏小的

上图的卷积就是空洞卷积,原本3*3*2的卷积核在空洞步长rate=2的情况下变成了5*5*2的卷积核,这就是为什么要使用padding=same的原因了,因为卷积核5*5已经大于4*4的输入了,不然会报错!!

三、Tensorflow中的空洞卷积

函数的调用为:

atrous_conv2d(value, filters, rate, padding, name=None)

value: 
指需要做卷积的输入图像,要求是一个4维Tensor,具有[batch, height, width, channels]这样的shape,具体含义是[训练时一个batch的图片数量, 图片高度, 图片宽度, 图像通道数]

filters: 
相当于CNN中的卷积核,要求是一个4维Tensor,具有[filter_height, filter_width, channels, out_channels]这样的shape,具体含义是[卷积核的高度,卷积核的宽度,图像通道数,卷积核个数],同理这里第三维channels,就是参数value的第四维

rate: 
要求是一个int型的正数,正常的卷积操作应该会有stride(即卷积核的滑动步长),但是空洞卷积是没有stride参数的,这一点尤其要注意。取而代之,它使用了新的rate参数,那么rate参数有什么用呢?它定义为我们在输入图像上卷积时的采样间隔,你可以理解为卷积核当中穿插了(rate-1)数量的“0”,把原来的卷积核插出了很多“洞洞”,这样做卷积时就相当于对原图像的采样间隔变大了。具体怎么插得,可以看后面更加详细的描述。此时我们很容易得出rate=1时,就没有0插入,此时这个函数就变成了普通卷积。如果rate=2,就把3*3的卷积核变成5*5了

padding: 
string类型的量,只能是”SAME”,”VALID”其中之一,这个值决定了不同边缘填充方式。
 

四、完整代码实现

数据集+相关原版代码链接:https://blog.csdn.net/qq_28626909/article/details/88866707

我的空洞卷积是对上文博客中的CNN代码进行了修改实现的

#!D:/workplace/python
# -*- coding: utf-8 -*-
# @File  : cifar10_atrous_conv2d.py
# @Author: WangYe
# @Date  : 2019/3/26
# @Software: PyCharm

# coding:utf-8
# 导入官方cifar10模块
#from tensorflow.image.cifar10 import cifar10
# import cifar10
# import tensorflow as tf
#
# # tf.app.flags.FLAGS是tensorflow的一个内部全局变量存储器
# FLAGS = tf.app.flags.FLAGS
# # cifar10模块中预定义下载路径的变量data_dir为'/tmp/cifar10_eval',预定义如下:
# # tf.app.flags.DEFINE_string('data_dir', './cifar10_data',
# #                           """Path to the CIFAR-10 data directory.""")
# # 为了方便,我们将这个路径改为当前位置
# FLAGS.data_dir = './cifar10_data'
#
# # 如果不存在数据文件则下载,并且解压
# cifar10.maybe_download_and_extract()
import cifar10,cifar10_input
import tensorflow as tf
import numpy as np
import time

#cifar10.maybe_download_and_extract()
max_steps = 3000
batch_size = 128
data_dir = r'./cifar10_data/cifar-10-batches-bin'


def variable_with_weight_loss(shape, stddev, w1):
    var = tf.Variable(tf.truncated_normal(shape, stddev=stddev))
    if w1 is not None:
        weight_loss = tf.multiply(tf.nn.l2_loss(var), w1, name='weight_loss')
        tf.add_to_collection('losses', weight_loss)
    return var


images_train, labels_train = cifar10_input.distorted_inputs(data_dir=data_dir, batch_size=batch_size)

images_test, labels_test = cifar10_input.inputs(eval_data=True,
                                                data_dir=data_dir,
                                                batch_size=batch_size)

image_holder = tf.placeholder(tf.float32, [batch_size, 24, 24, 3])
label_holder = tf.placeholder(tf.float32, [batch_size])
#第一层


#输入为   batch*24*24*3   卷积后为b*24*24*64   池化后为b*14*14*64   b 是batch的简写
weight1 = variable_with_weight_loss(shape=[5, 5, 3, 64], stddev=0.05, w1=0)
kernel1 = tf.nn.conv2d(image_holder, weight1, [1, 1, 1, 1], padding='SAME')
bias1 = tf.Variable(tf.constant(0.0, shape=[64]))
# conv1 = tf.nn.relu(tf.nn.bias_add(kernel1,bias1))
conv1 = tf.nn.relu(kernel1 + bias1)
pool1 = tf.nn.max_pool(conv1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME')
norm1 = tf.nn.lrn(pool1, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)



#第二层  输入b*14*14*64  卷积后卫 b*14*14*64    池化后卫为   b*7*7*64
weight2 = variable_with_weight_loss(shape=[5, 5, 64, 64], stddev=0.05, w1=0.0)
kernel2 = tf.nn.conv2d(norm1, weight2, [1, 1, 1, 1], padding='SAME')
bias2 = tf.Variable(tf.constant(0.1, shape=[64]))
conv2 = tf.nn.relu(kernel2 + bias1)
norm2 = tf.nn.lrn(conv2, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
pool2 = tf.nn.max_pool(norm2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME')



#第三层   自己添加一层   Atrous Convolution卷积层  空洞卷积(膨胀卷积)
weight3 = variable_with_weight_loss(shape=[5, 5, 64, 128], stddev=0.05, w1=0.0)
"""
rate: 
要求是一个int型的正数,正常的卷积操作应该会有stride(即卷积核的滑动步长),
但是空洞卷积是没有stride参数的,这一点尤其要注意。取而代之,它使用了新的rate参数,
那么rate参数有什么用呢?它定义为我们在输入图像上卷积时的采样间隔,
你可以理解为卷积核当中穿插了(rate-1)数量的“0”,把原来的卷积核插出了很多“洞洞”,
这样做卷积时就相当于对原图像的采样间隔变大了。具体怎么插得,可以看后面更加详细的描述。
此时我们很容易得出rate=1时,就没有0插入,此时这个函数就变成了普通卷积。
"""
#输入为  b*7*7*64  空洞卷积缺少strides参数,所以仍然卷积之后是  b*7*7*128
kernel3 = tf.nn.atrous_conv2d(pool2,weight3,rate=2,padding='SAME')   #空洞卷积会把上面的卷积核编程  10*10的大小
bias3 = tf.Variable(tf.constant(0.1, shape=[128]))     #输出维度128
conv3 = tf.nn.relu(kernel3 + bias3)    #   输出为b*7*7*128
norm3 = tf.nn.lrn(conv3, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
pool3 = tf.nn.max_pool(norm3, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME') #strides的步长上下都是2,输出为b*4*4*128


#全链接
reshape = tf.reshape(pool3, [batch_size, -1])  #分开成为
dim = reshape.get_shape()[1].value       #
weight3 = variable_with_weight_loss([dim, 384], stddev=0.04, w1=0.004)
bias3 = tf.Variable(tf.constant(0.1, shape=[384]))
local3 = tf.nn.relu(tf.matmul(reshape, weight3) + bias3)
#全连接层
weight4 = variable_with_weight_loss([384, 192], stddev=0.04, w1=0.004)
bias4 = tf.Variable(tf.constant(0.1, shape=[192]))
local4 = tf.nn.relu(tf.matmul(local3, weight4) + bias4)
#输出层
weight5 = variable_with_weight_loss([192, 10], stddev=1 / 192, w1=0.0)
bias5 = tf.Variable(tf.constant(0.0, shape=[10]))
logits = tf.matmul(local4, weight5) + bias5


def loss(logits, labels):
    labels = tf.cast(labels, tf.int64)
    cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=labels,
                                                                   name='cross_entropy_per_example')
    cross_entropy_mean = tf.reduce_mean(cross_entropy, name='cross_entropy')
    tf.add_to_collection('losses', cross_entropy_mean)
    return tf.add_n(tf.get_collection('losses'), name='total_loss')


loss = loss(logits, label_holder)

train_op = tf.train.AdamOptimizer(1e-3).minimize(loss)
top_k_op = tf.nn.in_top_k(logits, tf.cast(label_holder, tf.int64), 1)

# num_examples = 10000
# import math
# num_iter = int(math.ceil(num_examples/ batch_size))
# true_count = 0
# total_sample_count = num_iter * batch_size
# step = 0
# with tf.Session() as sess:
#     sess.run(tf.global_variables_initializer())
#     tf.train.start_queue_runners()
#     for step in range(max_steps):
#         start_time = time.time()
#         image_batch,label_batch = sess.run([images_train,labels_train])
#         _,loss_value = sess.run([train_op,loss],feed_dict={image_holder: image_batch,label_holder:label_batch})
#         duration = time.time()-start_time
#         if step % 10 == 0:
#             examples_per_sec = batch_size /duration
#             sec_per_batch = float(duration)

#             format_str = ('step %d,loss=%.2f (%.1f examples/sec;%.3f sec/batch)')
#             print(format_str % (step,loss_value,examples_per_sec,sec_per_batch))

#     while step< num_iter:
#         image_batch,label_batch = sess.run([images_test,labels_test])
#         predictions = sess.run([top_k_op],feed_dict = {image_holder:image_batch,
#                                                        label_holder:label_batch})
#         true_count += np.sum(predictions)
#         step+=1
#         if step % 10 ==0:
#             print true_count
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
tf.train.start_queue_runners()
for step in range(max_steps):
    start_time = time.time()
    image_batch, label_batch = sess.run([images_train, labels_train])
    _, loss_value = sess.run([train_op, loss], feed_dict={image_holder: image_batch, label_holder: label_batch})
    duration = time.time() - start_time
    if step % 10 == 0:
        examples_per_sec = batch_size / duration
        sec_per_batch = float(duration)

        format_str = ('step %d,loss=%.2f (%.1f examples/sec;%.3f sec/batch)')
        print(format_str % (step, loss_value, examples_per_sec, sec_per_batch))

num_examples = 10000
import math

num_iter = int(math.ceil(num_examples / batch_size))
true_count = 0
total_sample_count = num_iter * batch_size
step = 0
# with tf.Session() as sess:
while step < num_iter:
    image_batch, label_batch = sess.run([images_test, labels_test])
    predictions = sess.run([top_k_op], feed_dict={image_holder: image_batch,
                                                  label_holder: label_batch})
    true_count += np.sum(predictions)
    step += 1
    if step % 10 == 0:
        print(true_count)

precision = float(true_count) / total_sample_count
print('precision @ 1 =%.3f' % precision)

 

截图可以看出代码是没问题的

五、说明

代码只是为了实现空洞卷积,所以并没有在意准确率什么的,请谅解

  • 4
    点赞
  • 32
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值