zoukankan      html  css  js  c++  java
  • 卷积神经网络---padding、 pool、 Activation layer

    #coding:utf-8
    import tensorflow as tf
    tf.reset_default_graph()
    image = tf.random_normal([1, 112, 96, 3])
    in_channels = 3
    out_channels = 32
    kernel_size = 5
    conv_weight = tf.Variable(tf.truncated_normal([kernel_size, kernel_size, in_channels, out_channels], stddev=0.1,
                                                  dtype=tf.float32))
    
    print 'image shape', image.get_shape()
    print 'conv weight shape', conv_weight.get_shape()
    bias = tf.Variable(tf.zeros([out_channels], dtype=tf.float32))
    conv = tf.nn.conv2d(image, conv_weight, strides=[1, 3, 3, 1], padding='SAME')
    conv = tf.nn.bias_add(conv, bias)
    print 'conv output shape with SAME padded', conv.get_shape()
    
    conv = tf.nn.conv2d(image, conv_weight, strides=[1, 3, 3, 1], padding='VALID')
    conv = tf.nn.bias_add(conv, bias)
    print 'conv output shape with VALID padded', conv.get_shape()
    
    
    '''
    两种padding方式的不同
    SAME 简而言之就是丢弃,像素不够的时候对那部分不进行卷积,输出图像的宽高计算公式如下(向上取整,进1):
    HEIGHT = ceil(float(in_height)/float(strides[1]))
    WIDTH = ceil(float(in_width)/float(strides[2]))
    
    VALID 简而言之就是补全,像素不够的时候补0,输出图像的宽高计算公式如下
    HEIGHT = ceil(float(in_height - filter_height + 1)/float(strides[1]))
    WIDTH = ceil(float(in_width - filter_width + 1)/float(strides[2]))
    '''

     打印结果

     image shape (1, 112, 96, 3)
     conv weight shape (5, 5, 3, 32)
     conv output shape with SAME padded (1, 38, 32, 32)
     conv output shape with VALID padded (1, 36, 31, 32)

    pool_size = 3
    pool = tf.nn.max_pool(conv, ksize=[1, pool_size, pool_size, 1], strides=[1, 2, 2, 1], padding='SAME')
    print pool.get_shape()
    pool = tf.nn.max_pool(conv, ksize=[1, pool_size, pool_size, 1], strides=[1, 2, 2, 1], padding='VALID')
    print pool.get_shape()

    结果

    (1, 18, 16, 32)
    (1, 17, 15, 32)

    #激活层
    relu = tf.nn.relu(pool)
    print relu.get_shape()
    l2_regularizer = tf.contrib.layers.l2_regularizer(1.0)
    def prelu(x, name = 'prelu'):
        with tf.variable_scope(name):
            alphas = tf.get_variable('alpha', x.get_shape()[-1], initializer=tf.constant_initializer(0.25), regularizer=l2_regularizer, dtype=
                                     tf.float32)
        pos = tf.nn.relu(x)
        neg = tf.multiply(alphas, (x - abs(x)) * 0.5)
        return pos + neg
    prelu_out = prelu(pool)
    print prelu_out.get_shape()
  • 相关阅读:
    hdu4549(费马小定理 + 快速幂)
    java 四种实现延迟加载的方法
    浅谈java缓存
    JavaWeb实现文件上传下载功能实例解析
    java分页之页面分页
    浅析Java开发模式—Model1、Model2和三层
    构建一个简单的基于MVC模式的JavaWeb
    Java中MVC详解以及优缺点总结
    Servlet
    通过JDBC进行简单的增删改查(以MySQL为例)
  • 原文地址:https://www.cnblogs.com/cnugis/p/9309113.html
Copyright © 2011-2022 走看看