tf.nn.conv2d 和 tf.nn.max_pool 中 padding 分别为 'VALID' 和 'SAME' 的直觉上的经验和测试代码

这个地方一开始是迷糊的，写代码做比较分析，总结出直觉上的经验．

某人若想看精准的解释，移步这个网址(http://blog.csdn.net/fireflychh/article/details/73743849)，但我觉得直觉上的经验更有用，如下:

直觉上的经验：

1. 一件确定的事: padding　无论取 'SAME' 还是取 'VALID', 它在 conv2d　和　max_pool 上的表现是一致的;
2. padding = 'SAME' 时，输出并不一定和原图size一致，但会保证覆盖原图所有像素，不会舍弃边上的莫些元素;
3. padding = 'VALID' 时，输出的size总比原图的size小，有时不会覆盖原图所有元素(既，可能舍弃边上的某些元素).

# -*- coding: utf-8 -*-
import tensorflow as tf
import numpy as np

def pooling_show():
    a = tf.Variable(tf.random_normal(X))
    pooling = tf.nn.max_pool(a, pooling_filter, pooling_strides, padding=pad)
    # VALID (1, 2, 2, 7)
    # SAME (1, 3, 3, 7)

    init = tf.global_variables_initializer()

    with tf.Session() as sess:
        sess.run(init)

        print 'image: '
        image = sess.run(a)
        print image.shape

        print 'pooling result: '
        res = sess.run(pooling)
        print res.shape

def conv2d_padding_show():
    # [1, 13, 13, 2] ---> [m, height, width, channel]
    input = tf.Variable(tf.random_normal(X))
    # [6, 6, 2, 7] ---> [height, width, prev_channel, output_channel]
    filter = tf.Variable(tf.random_normal(conv2d_filter))

    op = tf.nn.conv2d(input, filter, strides=conv2d_strides, padding=pad)
    # VALID (1, 2, 2, 7)
    # SAME  (1, 3, 3, 7)

    init = tf.global_variables_initializer()

    with tf.Session() as sess:
        sess.run(init)

        print 'image: '
        image = sess.run(input)
        print image.shape

        print 'result: '
        res = sess.run(op)
        print res.shape

pad = 'VALID'

# X ---> [m, height, width, channel]
# X = [1, 13, 13, 7]
X = [1, 8, 8, 3]

# ---> [1, f, f, 1]
# pooling_filter = [1, 6, 6, 1]
pooling_filter = [1, 2, 2, 1]

# ---> [1, s, s, 1]
# pooling_strides = [1, 5, 5, 1]
pooling_strides = [1, 2, 2, 1]

# ---> [height, width, prev_channel, output_channel]
# conv2d_filter = [6, 6, 7, 7]
conv2d_filter = [2, 2, 3, 3]

# ---> [1, s, s, 1]
# conv2d_strides = [1, 5, 5, 1]
conv2d_strides = [1, 2, 2, 1]

# 自己改改 X, fileter, strides 的值，配合直觉经验，会有更好的理解
conv2d_padding_show()
pooling_show()