tf.nn.conv2d用法
tf.nn.conv2d：给定input和4D filters张量计算2D卷积。
tf.nn.conv2d(
input, filters, strides, padding, data_format='NHWC', dilations=None, name=None
)输入input张量可以具有秩4或更高，其中，形状维度[:-3] 被认为是批量维度（batch_shape）。
给定shape的输入张量 batch_shape + [in_height, in_width, in_channels]和shape的过滤器/内核张量[filter_height, filter_width, in_channels, out_channels]，batch_shape 为图片的数量，in_height为图片高度，in_width为图片宽度，in_channels为图片的通道数，灰度图该值为1，彩色图为3；其中 filter_height为卷积核高度，filter_width为卷积核宽度，in_channels是输入通道数 ，和 input的 in_channels要保持一致，out_channels是输出通道数，即卷积核数量。此op执行以下操作：
- 将过滤器展平为形状为
[filter_height * filter_width * in_channels, output_channels]的二维矩阵 。 - 从输入张量中提取图像补丁以形成形状为
[batch, out_height, out_width, filter_height * filter_width * in_channels]的虚拟 张量。 - 对于每个补丁，将滤波器矩阵和图像补丁向量右乘。
详细而言，使用默认的NHWC格式，
output[b, i, j, k] =
sum_{di, dj, q} input[b, strides[1] * i + di, strides[2] * j + dj, q] *
filter[di, dj, q, k]必须有strides[0] = strides[3] = 1。对于最常见的水平和垂直步幅相同的情况，strides = [1, stride, stride, 1]。
Args |
|
|---|---|
input |
一个Tensor。必须是下列类型之一： half，bfloat16，float32，float64。等级至少为4的张量。维度顺序根据data_format的值进行解释。所有内部的3个维度都用作批次维度。有关详情，请参见下文。 |
filters |
一个Tensor。必须具有与input相同的类型。形状为[filter_height, filter_width, in_channels, out_channels]的4D张量 |
strides |
一个int或ints列表具有长度1，2或4。input的每个维度的滑动窗口的步幅。如果给出单个值，则将其复制到H和W维度中。默认情况下，N和C维度设置为1。维度顺序由data_format的值确定，有关详细信息，请参见下文。 |
padding |
任一string "SAME"或"VALID"表示使用的填充算法来类型，或一个列表表示在每个维度的开始和结束的明确填充。如果使用显式填充，而data_format是"NHWC"，则应采用形式 [[0, 0], [pad_top,pad_bottom], [pad_left, pad_right], [0, 0]]。如果使用显式填充，并且data_format是"NCHW"，则应采用形式[[0, 0], [0, 0],[pad_top, pad_bottom], [pad_left, pad_right]]。表示的是卷积的形式，是否考虑边界。"SAME"是考虑边界，不足的时候用0去填充周围，"VALID"则不考虑 |
data_format |
string来自"NHWC","NCHW"的可选内容。默认为 "NHWC"。指定输入和输出数据的数据格式。使用默认格式“ NHWC”，数据按以下顺序存储：batch_shape + [height, width, channels] 。或者，格式可以是“ NCHW”，数据存储顺序为：batch_shape + [channels, height, width] 。 |
dilations |
一个int或ints列表具有长度1，2或者4，默认为1。对input的各个维度的扩张因子。如果给出单个值，则将其复制到H和W维度中。默认情况下，N和C维度设置为1。如果设置为k> 1，则该维度上每个过滤器元素之间将又k-1个跳过单元格。维度顺序由data_format的值确定，有关详细信息，请参见上文。如果4-d张量必须为1，则批量和深度维度的膨胀。 |
name |
操作的名称（可选）。 |
Returns |
|
|---|---|
一个Tensor。具有和input相同的类型和相同的外部批处理形状。 |
示例程序：
import tensorflow as tf
import numpy as np
x_in = np.array([[
[[2], [1], [2], [0], [1]],
[[1], [3], [2], [2], [3]],
[[1], [1], [3], [3], [0]],
[[2], [2], [0], [1], [1]],
[[0], [0], [3], [1], [2]], ]])
kernel_in = np.array([
[[[2, 0.1]], [[3, 0.2]]],
[[[0, 0.3]], [[1, 0.4]]], ])
x = tf.constant(x_in, dtype=tf.float32)
kernel = tf.constant(kernel_in, dtype=tf.float32)
# out = tf.nn.conv2d(x, kernel, strides=[1, 1, 1, 1], padding='VALID')
# 输出结果是一个[1, 4, 4, 2]的张量：
# Tensor("Conv2D:0", shape=(1, 4, 4, 2), dtype=float32)
# [[[[10. 1.9 ]
# [10. 2.2 ]
# [ 6. 1.6000001]
# [ 6. 2. ]]
#
# [[12. 1.4000001]
# [15. 2.2 ]
# [13. 2.7000003]
# [13. 1.7 ]]
#
# [[ 7. 1.7 ]
# [11. 1.3000001]
# [16. 1.3000001]
# [ 7. 1. ]]
#
# [[10. 0.6 ]
# [ 7. 1.4000001]
# [ 4. 1.5000001]
# [ 7. 1.4000001]]]]
out = tf.nn.conv2d(x, kernel, strides=[1, 1, 1, 1], padding='SAME')
# 输出结果是一个[1, 5, 5, 2]的张量：
# Tensor("Conv2D:0", shape=(1, 5, 5, 2), dtype=float32)
# [[[[10. 1.9 ]
# [10. 2.2 ]
# [ 6. 1.6000001 ]
# [ 6. 2. ]
# [ 2. 1. ]]
#
# [[12. 1.4000001 ]
# [15. 2.2 ]
# [13. 2.7000003 ]
# [13. 1.7 ]
# [ 6. 0.3 ]]
#
# [[ 7. 1.7 ]
# [11. 1.3000001 ]
# [16. 1.3000001 ]
# [ 7. 1. ]
# [ 0. 0.3 ]]
#
# [[10. 0.6 ]
# [ 7. 1.4000001 ]
# [ 4. 1.5000001 ]
# [ 7. 1.4000001 ]
# [ 2. 0.70000005]]
#
# [[ 0. 0. ]
# [ 9. 0.6 ]
# [ 9. 0.5 ]
# [ 8. 0.5 ]
# [ 4. 0.2 ]]]]
with tf.Session() as sess:
print(out)
print(sess.run(out))
import tensorflow as tf
# case 1
# 输入是1个 3*3 大小的张量，输入张量通道数是5，卷积核是 1*1 大小，数量是1
# 步长是[1,1,1,1]最后得到一个 3*3 的feature map
# 最后输出就是一个 shape为[1,3,3,1] 的张量
input = tf.Variable(tf.random_normal([1, 3, 3, 5]))
filter = tf.Variable(tf.random_normal([1, 1, 5, 1]))
op1 = tf.nn.conv2d(input, filter, strides=[1, 1, 1, 1], padding='SAME')
# 输出结果：
# ******************** op2 ********************
# [[[[ 2.1436303 ]
# [ 2.0738544 ]
# [-0.17929938]]
#
# [[-2.0726924 ]
# [-1.5222 ]
# [ 1.3480619 ]]
#
# [[ 0.59065 ]
# [ 4.6908927 ]
# [-2.7029667 ]]]]
# case 2
# 输入是1个 3*3 大小的张量，输入张量通道数是5，卷积核是 2*2 大小，数量是1
# 步长是[1,1,1,1]最后得到一个 3*3 的feature map
# 最后输出就是一个 shape为[1,3,3,1] 的张量
input = tf.Variable(tf.random_normal([1, 3, 3, 5]))
filter = tf.Variable(tf.random_normal([2, 2, 5, 1]))
op2 = tf.nn.conv2d(input, filter, strides=[1, 1, 1, 1], padding='SAME')
# 输出结果：
# ******************** op2 ********************
# [[[[-2.726498 ]
# [ 4.578858 ]
# [-4.247362 ]]
#
# [[-4.46906 ]
# [ 2.0851839 ]
# [ 0.12348425]]
#
# [[-0.16621417]
# [-3.6783643 ]
# [ 1.403814 ]]]]
# case 3
# 输入是1个 3*3 大小的张量，输入张量通道数是5，卷积核是 3*3 大小，数量是1
# 步长是[1,1,1,1]最后得到一个 1*1 的feature map (不考虑边界)
# 最后输出就是一个 shape为[1,1,1,1] 的张量
input = tf.Variable(tf.random_normal([1, 3, 3, 5]))
filter = tf.Variable(tf.random_normal([3, 3, 5, 1]))
op3 = tf.nn.conv2d(input, filter, strides=[1, 1, 1, 1], padding='VALID')
# 输出结果：
# ******************** op3 ********************
# [[[[-14.692696]]]]
# case 4
# 输入是1个 5*5 大小的张量，输入张量通道数是5，卷积核是 3*3 大小，数量是1
# 步长是[1,1,1,1]最后得到一个 3*3 的feature map (不考虑边界)
# 最后输出就是一个 shape为[1,3,3,1] 的张量
input = tf.Variable(tf.random_normal([1, 5, 5, 5]))
filter = tf.Variable(tf.random_normal([3, 3, 5, 1]))
op4 = tf.nn.conv2d(input, filter, strides=[1, 1, 1, 1], padding='VALID')
# 输出结果：
# ******************** op4 ********************
# [[[[ 7.8252077]
# [ -6.8241425]
# [ -4.5439234]]
#
# [[ 5.2957616]
# [ -2.3116708]
# [-10.82259 ]]
#
# [[ -2.663844 ]
# [ 2.8965201]
# [ 3.1595905]]]]
# case 5
# 输入是1个 5*5 大小的张量，输入张量通道数是5，卷积核是 3*3 大小，数量是1
# 步长是[1,1,1,1]最后得到一个 5*5 的feature map (考虑边界)
# 最后输出就是一个 shape为[1,5,5,1] 的张量
input = tf.Variable(tf.random_normal([1, 5, 5, 5]))
filter = tf.Variable(tf.random_normal([3, 3, 5, 1]))
op5 = tf.nn.conv2d(input, filter, strides=[1, 1, 1, 1], padding='SAME')
# 输出结果：
# ******************** op5 ********************
# [[[[ 3.446119 ]
# [ 6.889248 ]
# [-0.74263513]
# [-5.5886226 ]
# [-4.0011005 ]]
#
# [[-7.598516 ]
# [-2.5811138 ]
# [-5.9887714 ]
# [-2.423468 ]
# [ 3.302377 ]]
#
# [[-4.1364555 ]
# [-1.5923785 ]
# [-0.2860764 ]
# [ 1.7675163 ]
# [ 7.9709225 ]]
#
# [[ 4.975916 ]
# [ 3.2795277 ]
# [-1.8256065 ]
# [-4.5424066 ]
# [ 3.1844525 ]]
#
# [[-0.63261276]
# [-3.9194276 ]
# [-6.0611115 ]
# [-6.9680305 ]
# [-2.8860292 ]]]]
# case 6
# 输入是1个 5*5 大小的张量，输入张量通道数是5，卷积核是 3*3 大小，数量是7
# 步长是[1,1,1,1]最后得到一个 5*5 的feature map (考虑边界)
# 最后输出就是一个 shape为[1,5,5,7] 的张量
input = tf.Variable(tf.random_normal([1, 5, 5, 5]))
filter = tf.Variable(tf.random_normal([3, 3, 5, 7]))
op6 = tf.nn.conv2d(input, filter, strides=[1, 1, 1, 1], padding='SAME')
# 输出结果：
# ******************** op6 ********************
# [[[[-3.0840011e+00 1.2335920e-01 -4.5059223e+00 -2.1750627e+00
# -2.7304339e+00 -5.1604855e-01 1.3690400e+00]
# [ 4.7401037e+00 -6.0819392e+00 -3.0767975e+00 2.6486878e+00
# -6.2960935e+00 -1.0540801e+01 3.0819988e+00]
# [ 1.7505054e+00 5.4914937e+00 -1.0307097e+00 -4.4227147e+00
# 4.4376817e-01 3.7327261e+00 -2.4275784e+00]
# [-4.3396387e+00 3.7363658e+00 -1.0957268e+00 4.5229203e-01
# 2.0793560e+00 9.6609974e-01 -5.0022750e+00]
# [ 1.1345127e+00 1.9144246e-01 -3.1545307e-02 -4.1797557e-01
# -1.0004331e+00 6.3729405e-01 4.8672795e-02]]
#
# [[ 4.9904194e+00 -2.6558199e+00 -3.3613873e-01 3.5926490e+00
# 2.0895982e+00 2.3332577e+00 5.3478036e+00]
# [-4.6558094e+00 3.3834887e+00 -2.0608113e+00 -7.6377983e+00
# 5.3955240e+00 7.0295696e+00 -7.2131968e+00]
# [-1.7864041e+00 -7.0538968e-01 1.5037076e+01 2.4281028e-01
# -4.6876159e+00 -1.1438875e+00 2.5164890e+00]
# [ 2.8490059e+00 -3.0830842e-01 1.1872043e-01 6.8919549e+00
# 1.7080204e+00 1.7034137e+00 2.3760687e-01]
# [ 2.2561960e+00 3.4825525e+00 -1.4772111e+01 -1.0685290e+01
# -6.5574465e+00 7.2971749e+00 -1.4437375e+00]]
#
# [[ 4.2684340e+00 -4.3361559e+00 1.0803232e+00 3.5441415e+00
# 8.9037914e+00 7.2594132e+00 7.2638750e+00]
# [-6.8646002e-01 -8.9739448e-01 5.8036337e+00 7.4364100e+00
# 1.1704000e+01 4.4205704e+00 3.0446523e-01]
# [-2.8831110e+00 -5.2022009e+00 7.9378448e+00 -3.2306526e+00
# 7.3811979e+00 6.2627382e+00 -4.6261063e+00]
# [-6.3880229e-01 -1.0363024e+01 7.0129766e+00 8.6191149e+00
# 1.2183407e+01 7.0540304e+00 -2.1471415e+00]
# [ 1.6519511e-01 9.4482350e-01 3.3437994e+00 1.7440913e+00
# 1.0958947e+01 -6.3905849e+00 7.7824841e+00]]
#
# [[-5.9248290e+00 7.9262257e-03 -3.4675770e+00 2.5212393e+00
# 4.6964250e+00 -1.2980299e+01 6.6482816e+00]
# [ 1.6509128e-01 1.3398497e+01 -1.6516523e+01 -5.0966377e+00
# 3.5735250e+00 4.7559333e+00 -1.3719524e+00]
# [ 5.6323524e+00 -1.7552935e+01 6.5047069e+00 2.4060578e+00
# -6.7384124e+00 -4.1142387e+00 8.3369007e+00]
# [ 1.1545516e+01 4.2000041e+00 3.6326356e+00 3.3518002e+00
# -5.8180132e+00 -8.6836344e-01 6.1310697e+00]
# [-6.1742949e+00 2.8111699e+00 4.4471732e-01 -1.3193527e+00
# -5.4388676e+00 -1.3804424e+00 -4.0689611e-01]]
#
# [[-1.9953527e+00 3.3288798e+00 -4.4504986e+00 6.2108145e+00
# 4.2928514e+00 -1.0003451e+01 -3.9065666e+00]
# [ 1.3419192e+00 7.2219830e+00 -6.9047012e+00 3.3905752e+00
# -5.0909300e+00 1.3293836e+01 -7.6556101e+00]
# [-7.3027096e+00 4.5334468e+00 -2.6074278e-01 -1.0191909e+01
# 3.3845627e+00 -7.1441424e-01 -1.0066732e+00]
# [ 2.2974596e+00 -4.5104642e+00 9.2605171e+00 3.9580934e+00
# 3.9759703e+00 -7.4361342e-01 1.2278461e+00]
# [ 3.2552500e+00 -7.5348072e+00 -9.5567679e-01 2.1697264e+00
# -3.3733122e+00 2.1065104e+00 -2.6742821e+00]]]]
# case 7
# 输入是1个 5*5 大小的张量，输入张量通道数是5，卷积核是 3*3 大小，数量是7
# 步长是[1,2,2,1]最后得到7个 3*3 的feature map (考虑边界)
# 最后输出就是一个 shape为[1,3,3,7] 的张量
input = tf.Variable(tf.random_normal([1, 5, 5, 5]))
filter = tf.Variable(tf.random_normal([3, 3, 5, 7]))
op7 = tf.nn.conv2d(input, filter, strides=[1, 2, 2, 1], padding='SAME')
# 输出结果：
# ******************** op7 ********************
# [[[[ 0.293367 -6.8127875 0.72176373 -4.609955 0.29551882
# 4.215855 -0.9913437 ]
# [ 0.36416888 11.682545 4.806132 -7.0474143 4.167954
# 2.4354155 -2.5111024 ]
# [ -0.55445254 -6.615763 -2.0677602 -1.5091647 -7.727282
# -11.306014 2.890195 ]]
#
# [[ -0.28828835 -0.80325335 1.7683787 6.4272923 -4.341469
# 0.8893982 0.39250994]
# [ 3.596493 -0.8119854 -3.97007 -2.0908122 -1.1287336
# -0.3859545 -7.195946 ]
# [ -5.793783 6.1878567 15.109923 0.6127636 -3.6183724
# 1.4538065 -10.928911 ]]
#
# [[ 7.3411665 -4.1733685 -3.3012686 -3.3117685 -0.5984683
# 0.8338492 -1.9214034 ]
# [ -0.5107512 -12.7436285 1.8505284 -1.0571606 0.7512574
# 4.2915535 9.583618 ]
# [ 2.5063615 5.879011 4.90127 -0.07179046 2.8028073
# -0.05491097 -5.3217688 ]]]]
# case 8
# 输入是10 个 5*5 大小的张量，输入张量通道数是5，卷积核是 3*3 大小，数量是7
# 步长是[1,2,2,1]最后每张图得到7个 3*3 的feature map (考虑边界)
# 最后输出就是一个 shape为[10,3,3,7] 的张量
input = tf.Variable(tf.random_normal([10, 5, 5, 5]))
filter = tf.Variable(tf.random_normal([3, 3, 5, 7]))
op8 = tf.nn.conv2d(input, filter, strides=[1, 2, 2, 1], padding='SAME')
# 输出结果：
# ******************** op8 ********************
# [[[[-1.50972009e+00 3.28846717e+00 2.56372422e-01 -2.70064521e+00
# 4.40395069e+00 8.26167464e-01 -7.68952465e+00]
# [-1.46069360e+00 1.66591525e+00 4.58909464e+00 2.48840261e+00
# 1.63206520e+01 -5.99188614e+00 1.35770631e+00]
# [ 4.75995207e+00 7.77311897e+00 -2.12295747e+00 4.07630301e+00
# -3.84346151e+00 -1.19016916e-01 1.87570047e+00]]
#
# [[ 8.78928661e+00 8.35301638e-01 -7.33191681e+00 4.09586620e+00
# 8.46557617e-01 8.00785637e+00 -2.03714895e+00]
# [ 5.73713017e+00 -1.49388247e+01 3.70609355e+00 9.50882626e+00
# 1.17435322e+01 5.17303848e+00 -1.73004842e+00]
# [-4.42001724e+00 -3.54912615e+00 7.96449041e+00 -3.40062141e-01
# 1.23415983e+00 -6.74519539e-02 -7.09112930e+00]]
#
# [[ 6.89682722e-01 1.01104822e+01 -2.62892461e+00 6.52079821e+00
# -1.58322823e+00 -6.27171707e+00 3.94091272e+00]
# [ 5.16624117e+00 2.95724177e+00 -3.22336292e+00 -3.29455805e+00
# 7.71984863e+00 3.12455177e-01 1.52014565e+00]
# [ 3.40628886e+00 -7.10564077e-01 7.25713825e+00 -2.20483243e-01
# 1.51156664e-01 -2.01525712e+00 -8.09532547e+00]]]
#
#
# [[[ 2.34457016e+00 -2.26403069e+00 3.01330864e-01 -4.59966421e+00
# -5.70768118e-01 3.28723526e+00 1.81965542e+00]
# [-2.02637410e+00 -2.98199654e+00 -2.61881232e+00 -7.22932386e+00
# 3.42993617e+00 2.40504479e+00 -5.41764259e+00]
# [ 2.87678647e+00 1.45507336e+00 -1.83620012e+00 2.49595642e+00
# -5.10922480e+00 2.40312648e+00 1.78781736e+00]]
#
# [[-1.33651924e+01 3.00249600e+00 9.26462364e+00 2.58863783e+00
# 3.29223633e+00 -5.51135063e+00 4.08538818e+00]
# [ 6.96175003e+00 6.51313400e+00 -2.43331528e+00 3.99290252e+00
# 3.79648089e+00 -3.50887942e+00 3.35554075e+00]
# [-2.33746815e+00 2.42192769e+00 3.05796742e+00 -2.99886394e+00
# -4.59325254e-01 -9.66714096e+00 -3.06445456e+00]]
#
# [[-4.99330187e+00 -3.59256339e+00 3.12972355e+00 -1.05485499e+00
# -2.55326033e+00 4.25835180e+00 -6.87904644e+00]
# [-6.78382349e+00 1.99727798e+00 -5.68853617e+00 1.16159916e+00
# -6.48815918e+00 1.27502060e+00 -1.44755840e-02]
# [-6.61731911e+00 -2.98836422e+00 -1.89204836e+00 -5.07599735e+00
# 1.98823047e+00 4.22331476e+00 -3.81494570e+00]]]
#
#
# [[[ 6.55350590e+00 1.52602494e+00 2.64053488e+00 3.48391485e+00
# 1.15554178e+00 5.93114519e+00 8.92291641e+00]
# [-7.22469378e+00 1.97034907e+00 2.99531627e+00 -4.34610367e+00
# -3.67712796e-01 -3.64595604e+00 -7.84149694e+00]
# [-7.76820278e+00 5.99960327e+00 -8.41420841e+00 2.29520798e+00
# 4.24104357e+00 9.19977474e+00 -4.61686659e+00]]
#
# [[ 8.27398205e+00 2.88631511e+00 -5.95147419e+00 7.13959980e+00
# -2.54925585e+00 -6.65015578e-01 -7.24622011e+00]
# [ 5.36609936e+00 -2.25370550e+00 -1.00714719e+00 6.40559196e+00
# -3.44093418e+00 -1.10686178e+01 -1.93581736e+00]
# [ 6.08330536e+00 1.16270723e+01 -6.12922239e+00 1.02644575e+00
# -8.62347031e+00 -4.47479963e+00 6.06679058e+00]]
#
# [[-6.74393320e+00 -2.22659349e-01 -3.77906322e-01 -4.90764093e+00
# 2.15560746e+00 6.62665665e-01 -8.02735519e+00]
# [ 6.33290350e-01 1.03738439e+00 1.54830575e-01 -1.13407755e+00
# 5.91823578e+00 1.70822430e+01 -4.38055634e-01]
# [-2.77781868e+00 -5.48576641e+00 2.56255603e+00 6.71352196e+00
# 6.03837252e+00 3.98823690e+00 8.05403328e+00]]]
#
#
# [[[-6.27049685e-01 1.15633488e+01 -6.36027515e-01 5.66315889e+00
# 2.28375840e+00 9.08275545e-01 1.16598463e+00]
# [ 2.53564405e+00 8.72430134e+00 -1.99402618e+00 8.17446709e+00
# 2.74033046e+00 6.30941105e+00 9.17519188e+00]
# [-3.74645138e+00 -8.57678294e-01 -2.83845067e+00 -3.27615452e+00
# 2.56140089e+00 5.98653507e+00 -1.09049511e+01]]
#
# [[ 3.24537659e+00 -3.93446064e+00 -1.52721763e+00 5.86781311e+00
# -1.97805035e+00 1.28334188e+00 -2.98240376e+00]
# [-1.21900101e+01 1.25004463e+01 4.92444181e+00 -2.70350194e+00
# 7.81507492e+00 -9.05650043e+00 6.47177219e+00]
# [-2.17206478e+00 2.71480703e+00 -9.16565228e+00 4.69869089e+00
# -4.38394928e+00 5.74612093e+00 3.59565282e+00]]
#
# [[-3.55511594e+00 2.76028395e+00 -3.60184312e+00 7.99856138e+00
# -4.38458920e+00 -8.98869276e-01 -8.10075521e-01]
# [ 8.20520210e+00 -2.68970108e+00 -9.29989719e+00 -1.22854757e+00
# -5.82088470e+00 -5.44573784e+00 -8.23349571e+00]
# [ 3.67080784e+00 1.18026161e+00 -2.92942858e+00 -7.73343849e+00
# 2.77626085e+00 -4.28008842e+00 -2.14514184e+00]]]
#
#
# [[[ 2.12814093e+00 -7.68777609e-01 -8.84735203e+00 1.01067667e+01
# 4.76778126e+00 4.78147745e+00 -1.67699254e+00]
# [ 4.97244644e+00 1.94252181e+00 -2.71501064e+00 -4.14783287e+00
# -7.91691661e-01 3.18883848e+00 -7.51171255e+00]
# [-2.92162299e+00 -2.16129565e+00 3.80590534e+00 5.97961855e+00
# 4.99147081e+00 -5.57082367e+00 -1.89473879e+00]]
#
# [[ 2.64872313e+00 -6.40528393e+00 -1.62670207e+00 2.38367653e+00
# 3.88705778e+00 1.39000535e-01 1.55138087e+00]
# [ 1.25824089e+01 4.76570749e+00 -7.49335003e+00 6.50031090e+00
# 1.96111858e+00 -4.04969454e+00 7.39542055e+00]
# [ 1.88336353e+01 3.54755116e+00 7.74576569e+00 2.64947295e+00
# -3.44345450e+00 7.32074857e-01 1.50061584e+00]]
#
# [[-3.47981930e+00 -2.59158301e+00 -3.34343934e+00 8.46221507e-01
# 2.33314896e+00 3.76661897e+00 -2.47455430e+00]
# [ 7.20970154e-01 -5.69754934e+00 -3.22284698e+00 -2.73109198e+00
# -3.16789818e+00 -2.03302526e+00 -2.95625877e+00]
# [-5.52691269e+00 4.82660437e+00 1.01005960e+00 2.21910906e+00
# 1.57971847e+00 3.68322039e+00 3.75630736e+00]]]
#
#
# [[[ 5.41193104e+00 -5.60384941e+00 2.29562593e+00 -9.02732134e-01
# 5.91344452e+00 -2.84974337e+00 -8.05228114e-01]
# [ 3.65591240e+00 -4.20973110e+00 -2.51142955e+00 3.83922720e+00
# -4.78865433e+00 1.13573515e+00 4.27588224e+00]
# [-7.49597371e-01 -8.19201112e-01 -2.69533873e+00 -2.92425513e-01
# -9.61732864e-01 -1.60479331e+00 -2.03666830e+00]]
#
# [[-1.37315083e+00 -9.93513107e+00 -9.46694183e+00 -4.27486229e+00
# 1.70914292e-01 4.13093662e+00 -1.33161507e+01]
# [ 1.42857313e+00 -4.43079710e+00 1.07908468e+01 3.77378225e+00
# 3.78649712e+00 -6.18113041e+00 -4.32539082e+00]
# [-3.15826833e-01 -3.38590169e+00 -6.04250371e-01 -4.27673054e+00
# -2.64095235e+00 2.44551945e+00 -7.75257301e+00]]
#
# [[ 1.18270099e+00 -8.82114220e+00 -1.65353990e+00 -7.77074385e+00
# 5.09965706e+00 1.26115513e+00 5.37919807e+00]
# [-1.93771791e+00 -3.69820595e+00 9.32795525e+00 5.11645746e+00
# 7.01560497e+00 8.65833187e+00 -5.40101194e+00]
# [-3.98796463e+00 -2.34203637e-02 -2.60023904e+00 1.02625144e+00
# 4.13492680e+00 1.92184746e+00 4.07373190e+00]]]
#
#
# [[[-6.84364319e-01 -4.76772118e+00 -5.59659576e+00 2.59804058e+00
# -4.45792150e+00 2.85404205e+00 4.37403154e+00]
# [ 9.87957764e+00 4.24151230e+00 -5.65920019e+00 3.64386129e+00
# 6.02909446e-01 1.11303949e+01 -3.01928926e+00]
# [-7.46744573e-01 -5.60856485e+00 1.99480343e+00 -9.92031753e-01
# 1.71692371e+00 -5.59550583e-01 -1.56595182e+00]]
#
# [[ 9.75834179e+00 8.77900183e-01 -4.63732815e+00 4.03338528e+00
# 1.47323322e+00 1.23197660e+01 3.10974097e+00]
# [-4.69103575e-01 -1.42769079e+01 4.73733187e+00 9.42406559e+00
# -7.56288624e+00 -1.26075115e+01 -2.31584907e+00]
# [ 6.94194674e-01 -3.53304219e+00 8.35210919e-01 -5.16879797e+00
# 5.44092989e+00 -5.92492199e+00 5.38006544e+00]]
#
# [[-5.09272289e+00 5.69651365e-01 9.49377251e+00 4.91886139e+00
# -4.38747263e+00 -1.27646389e+01 5.52901745e+00]
# [-7.93955231e+00 -4.50405884e+00 1.59659004e+00 -8.24533939e+00
# -1.65851283e+00 -5.41553068e+00 -8.03626060e+00]
# [-3.18145752e+00 2.66282582e+00 9.36366320e-01 -1.10870361e+00
# -4.80783844e+00 -2.30249023e+00 -5.55814934e+00]]]
#
#
# [[[ 5.35861683e+00 2.81997848e+00 -2.70675421e-01 -1.49736440e+00
# -2.72707272e+00 9.93852735e-01 2.10917950e+00]
# [ 9.54258060e+00 -3.62865114e+00 -7.98306227e-01 -1.77635348e+00
# -4.55513382e+00 6.12996149e+00 -7.45813084e+00]
# [-3.70084548e+00 2.80985504e-01 -1.28691113e+00 5.09892893e+00
# 2.69848919e+00 4.83566880e-01 4.12019444e+00]]
#
# [[-3.95883560e-01 1.04449043e+01 5.16901970e+00 5.41939878e+00
# 1.82462847e+00 -8.01707745e+00 2.00431228e+00]
# [-1.88117683e-01 2.92954755e+00 4.14290667e+00 8.43845749e+00
# -3.81140685e+00 -3.14104676e+00 -7.85914612e+00]
# [-8.65340996e+00 1.11263824e+00 5.96140480e+00 1.93812943e+00
# 4.93798304e+00 -1.07395992e+01 -6.05933189e-01]]
#
# [[ 1.68662715e+00 1.91131592e-01 -5.49581814e+00 -2.99100828e+00
# 3.26321900e-01 2.22558951e+00 -7.21933961e-01]
# [-3.62809539e+00 -1.14094102e+00 7.86778629e-01 -3.37758040e+00
# 9.88929462e+00 1.51389647e+00 -7.54234600e+00]
# [ 3.16409206e+00 -5.93904066e+00 5.72136593e+00 2.06540251e+00
# 1.19954090e+01 4.82101727e+00 -2.20609307e+00]]]
#
#
# [[[ 1.04300761e+00 1.63646793e+00 6.86399341e-01 1.34852600e+00
# -4.03470707e+00 -2.25256681e+00 3.33122182e+00]
# [-1.46150219e+00 6.01696110e+00 -8.95973873e+00 -5.05144358e+00
# -5.28098297e+00 6.84933281e+00 -3.63140082e+00]
# [ 3.74540508e-01 -7.87882507e-01 3.23215270e+00 -2.20066023e+00
# -4.33330297e-01 5.46882105e+00 -5.77409935e+00]]
#
# [[ 6.64719820e-01 1.83429480e+00 -2.47371340e+00 -5.44579744e+00
# -2.51808548e+00 -3.36716509e+00 7.91948617e-01]
# [ 1.18216543e+01 7.04637825e-01 3.57614374e+00 3.26182246e-02
# -7.08487892e+00 -1.53310990e+00 3.27570367e+00]
# [-1.55577135e+00 2.11997604e+00 -4.86663532e+00 1.29599333e+00
# -4.90606833e+00 -1.16319454e+00 2.90553045e+00]]
#
# [[ 6.70130253e+00 3.28602600e+00 5.35130310e+00 1.09373245e+01
# -1.68153787e+00 -5.31269217e+00 1.27796793e+01]
# [ 2.71093917e+00 4.41750097e+00 -2.81792474e+00 -5.20114708e+00
# -4.55729961e+00 -7.18441391e+00 -5.07527828e+00]
# [ 1.90964675e+00 3.77511740e+00 2.33712316e+00 -7.08522856e-01
# 3.72151232e+00 -1.05427623e-01 -1.03531027e+00]]]
#
#
# [[[ 2.04985380e+00 -6.48390770e+00 1.95798242e+00 6.57322884e+00
# 4.12863731e-01 -4.32278967e+00 4.89580631e+00]
# [ 9.95611787e-01 -2.08505154e+00 5.77722263e+00 -4.75273275e+00
# 4.74745226e+00 2.42263794e+00 2.28105402e+00]
# [ 4.28825855e+00 -8.90439606e+00 1.48928604e+01 -8.79860461e-01
# 7.60016537e+00 -3.02103829e+00 3.51416063e+00]]
#
# [[-4.18446255e+00 -6.44563854e-01 -1.60029554e+00 -9.40447903e+00
# 2.96056366e+00 4.05623341e+00 -3.08270931e+00]
# [ 5.90721655e+00 6.26168966e+00 -1.17052495e+00 1.20180407e+01
# -2.22426128e+00 1.84180009e+00 5.88260651e+00]
# [ 1.56474435e+00 5.05135679e+00 -3.33587337e+00 2.66880488e+00
# -4.89722109e+00 7.00591683e-01 7.88860321e-01]]
#
# [[ 3.55771160e+00 -8.34005165e+00 6.57881498e+00 -5.94304085e+00
# -1.29626989e-02 -7.68445730e+00 -8.56293917e-01]
# [-7.78887987e+00 2.45494366e+00 5.37120914e+00 -2.09734797e-01
# -7.03242493e+00 -2.25485846e-01 3.39174938e+00]
# [-5.31516409e+00 -4.23801708e+00 2.69230103e+00 -6.37099981e+00
# -2.49476051e+00 8.04008389e+00 -7.84046292e-01]]]]
init = tf.initialize_all_variables()
with tf.Session() as sess:
sess.run(init)
print('*' * 20 + ' op1 ' + '*' * 20)
print(sess.run(op1))
print('*' * 20 + ' op2 ' + '*' * 20)
print(sess.run(op2))
print('*' * 20 + ' op3 ' + '*' * 20)
print(sess.run(op3))
print('*' * 20 + ' op4 ' + '*' * 20)
print(sess.run(op4))
print('*' * 20 + ' op5 ' + '*' * 20)
print(sess.run(op5))
print('*' * 20 + ' op6 ' + '*' * 20)
print(sess.run(op6))
print('*' * 20 + ' op7 ' + '*' * 20)
print(sess.run(op7))
print('*' * 20 + ' op8 ' + '*' * 20)
print(sess.run(op8))
import tensorflow as tf
x = tf.constant([[1., 2., 3.],
[4., 5., 6.]])
x = tf.reshape(x, [1, 2, 3, 1]) # give a shape accepted by tf.nn.max_pool
valid_pad = tf.nn.max_pool(x, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID')
same_pad = tf.nn.max_pool(x, [1, 2, 2, 1], [1, 2, 2, 1], padding='SAME')
with tf.Session() as sess:
print(valid_pad)
print(same_pad)
# Tensor("MaxPool:0", shape=(1, 1, 1, 1), dtype=float32)
# Tensor("MaxPool_1:0", shape=(1, 1, 2, 1), dtype=float32)
print(valid_pad.get_shape())
print(same_pad.get_shape())
# (1, 1, 1, 1)
# (1, 1, 2, 1)
print(sess.run(valid_pad))
print(sess.run(same_pad))
# [[[[5.]]]]
# [[[[5.]
# [6.]]]]
max pooling窗口为2×2，两个维度的步长strides=2。第一次由于窗口可以覆盖，橙色区域做max pooling，没什么问题，如下：
接下来就是SAME和VALID的区别所在：由于步长为2，当向右滑动两步之后，VALID方式发现余下的窗口不到2×2，所以直接将第三列舍弃，而SAME方式并不会把多出的一列丢弃，但是只有一列了不够2×2，即进行填充！
如上图所示，SAME会增加第四列以保证可以达到2×2，但为了不影响原始信息，一般以0来填充。
按原理来说，还有一种填充方式，即FULL填充方式，该方式为卷积后输出特征大小与输入特征大小相同，但是填充方式多种多样，故没有准确的填充方法，了解即可。