[英]What is the padding strategy of tensorflow conv2d?
i follow the question and answer from stackoverflow 我遵循stackoverflow的问题和答案
however i am still confused of the start index and padding strategy of tf.nn.conv2d after i have the following tests, hopefully someone can give me a clue here, especially on odd and even strides 但我仍然困惑tf.nn.conv2d的开始索引和填充策略后,我有以下的测试,希望有人能在这里给我一个线索,特别是奇数和偶数的进步
array height(h),kernel size(f), stride number(s)
数组高度(h),内核大小(f),步幅数
h,f,s = 4,3,2
padding number on left column (pl) padding on right column (pr) of matrix x
矩阵x左列(pl)的填充数右列(pr)的填充数
pl = int((f-1)/2)
pr = int(np.ceil((f-1)/2))
tf.reset_default_graph()
x = np.arange(1*h*h*1).reshape(1,h,h,1)
w = np.ones((f,f,1,1))
xc = tf.constant(x,np.float32)
wc = tf.constant(w,np.float32)
xp = np.pad(x,((0,0),(pl,pr),(pl,pr),(0,0)),'constant',constant_values = 0)
xcp = tf.constant(xp,np.float32)
zs = tf.nn.conv2d(xc,wc,strides=[1,s,s,1],padding='SAME')
zv = tf.nn.conv2d(xc,wc,strides=[1,s,s,1],padding='VALID')
zp = tf.nn.conv2d(xcp,wc,strides=[1,s,s,1],padding='VALID')
with tf.Session() as sess:
os = sess.run(zs)
ov = sess.run(zv)
op = sess.run(zp)
print('x shape: ', x.shape,' kernel: ',f,' stride: ',s,'\n',x[0,:,:,0])
print(' 'SAME' os shape: ', os.shape,'\n',os[0,:,:,0])
print(' 'VALID' ov shape: ', ov.shape,'\n',ov[0,:,:,0])
print(' 'VALID' op shape: ', op.shape,' pl: ',pl,' pr: ', pr,'\n',op[0,:,:,0])
in case of pooling in convolution, the zero padding should pad around the array x like how i define xp , however, i cannot figure out the how conv2d the start index of it 在卷积池的情况下,零填充应该像我定义xp一样在数组x周围填充,但是,我无法弄清楚conv2d的起始索引如何
origin matrix x 原始矩阵x
x shape: (1, 4, 4, 1) kernel: 3 stride: 2
[[ 0 1 2 3]
[ 4 5 6 7]
[ 8 9 10 11]
[12 13 14 15]]
In 'same' type convolution, why tf.nn.conv2d does not pad zero on the left in this case ? 在“相同”类型的卷积中,为什么在这种情况下tf.nn.conv2d不在左侧填充零?
'SAME' os shape: (1, 2, 2, 1)
[[45. 39.]
[66. 50.]]
valid convolution on matrix x 在矩阵x上的有效卷积
'VALID' ov shape: (1, 1, 1, 1)
[[45.]]
valid type convolution after zero padding from xp (as my expected result) 从xp填充零后的有效类型卷积(作为我的预期结果)
'VALID' op shape: (1, 2, 2, 1) pl: 1 pr: 1
[[10. 24.]
[51. 90.]]
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.