padding卷积的两种方式“SAME”和“VALID”

conv2d是常用的实现卷积的，Tensorflow调用Conv的api时，常用代码如下：

查看：https://tensorflow.google.cn/api_docs/python/tf/nn/conv2d

tf.nn.conv2d(
    input,
    filter,
    strides,
    padding,
    use_cudnn_on_gpu=True,
    data_format='NHWC',
    dilations=[1, 1, 1, 1],
    name=None
)

对于其中一个参数padding的理解为：

padding: A `string` from: `"SAME", "VALID"`.
      The type of padding algorithm to use.

padding有两个方式可以选择：“SAME” and “VALID”

举两个栗子：

One：padding='SAME'

import tensorflow as tf

input = tf.Variable(tf.random_normal([1,5,5,3]))
filter = tf.Variable(tf.random_normal([3,3,3,7]))

result = tf.nn.conv2d(input, filter, strides=[1,2,2,1],padding='SAME')
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
print(sess.run(result))
print(result.shape)
sess.close()

结果为： 

[[[[  6.88815355e-01  -1.58929396e+00  -8.13352680e+00   3.47248018e-01
     -2.10637522e+00  -2.47548366e+00  -3.29180861e+00]
   [ -1.50164223e+00  -2.82424307e+00  -2.40781856e+00  -2.55665493e+00
     -3.89841533e+00  -6.71445191e-01   3.10867667e+00]
   [ -3.39479542e+00  -1.40321875e+00   2.29996824e+00  -3.98842275e-01
      7.90905952e-03  -1.71421432e+00  -5.47636747e-01]]

  [[ -1.07995415e+00  -2.21969414e+00  -1.43076777e-01   2.65041399e+00
     -4.38491011e+00  -4.83550358e+00   8.30997753e+00]
   [  1.35791779e+00  -1.38357902e+00  -4.50581169e+00   1.22106361e+00
     -1.36877072e+00  -1.19497585e+00  -3.64005876e+00]
   [ -3.07881045e+00   1.33630781e+01  -4.33032846e+00   1.98507690e+00
     -1.34837186e+00  -3.44964921e-01  -5.76371312e-01]]

  [[ -4.02724743e-01  -3.08082283e-01   1.51205099e+00  -2.11967897e+00
      8.77675891e-01  -3.89271736e-01   1.28933489e+00]
   [  1.05681574e+00   3.83993292e+00   1.46158600e+00   5.12251711e+00
     -4.37659168e+00  -5.88564873e-02   8.72927666e-01]
   [  3.13625002e+00  -2.52725768e+00  -1.89247894e+00  -2.89734745e+00
      2.49475980e+00  -7.85117006e+00   4.73596001e+00]]]]
(1, 3, 3, 7)

Two：padding='VALID'

import tensorflow as tf

input = tf.Variable(tf.random_normal([1,5,5,3]))
filter = tf.Variable(tf.random_normal([3,3,3,7]))

result = tf.nn.conv2d(input, filter, strides=[1,2,2,1],padding='VALID')
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
print(sess.run(result))
print(result.shape)
sess.close()

结果为：

[[[[  3.30246162e+00   1.00174313e+01   1.02988682e+01  -3.38870287e+00
      3.57620907e+00   9.25950432e+00   1.40226996e+00]
   [  2.39865661e+00   4.90117121e+00   6.27546692e+00  -7.14295626e+00
     -1.87810266e+00   4.73461962e+00  -8.87438393e+00]]

  [[  5.66498578e-01   1.21167574e+01  -2.98488545e+0