tesnorflow conv deconv,padding

1.padding test

input = tf.placeholder(tf.float32, shape=(1,2, 2,1))
simpleconv=slim.conv2d(input,1,[3,3],stride = 1,activation_fn = None,scope = 'simpleconv3')
sess.run(tf.global_variables_initializer())
weights=graph.get_tensor_by_name("simpleconv3/weights:0")
sess.run(tf.assign(weights,tf.constant(1.0,shape=weights.shape)))
a=np.ndarray(shape=(1,2,2,1),dtype='float',buffer=np.array([1.0,2,3,4]))
simpleconvout=sess.run(simpleconv,feed_dict={input:a.astype('float32')})
print simpleconvout
[[[[ 10.000000]
[ 10.000000]]

[[ 10.000000]
[ 10.000000]]]]

input1 = tf.placeholder(tf.float32, shape=(1,4, 4,1))
simpleconv=slim.conv2d(input1,1,[3,3],stride = 2,activation_fn = None,scope = 'simpleconv3')
sess.run(tf.global_variables_initializer())
weights=graph.get_tensor_by_name("simpleconv3/weights:0")
sess.run(tf.assign(weights,tf.constant(1.0,shape=weights.shape)))
a=np.ndarray(shape=(1,4,4,1),dtype='float',buffer=np.array([1.0,2,3,4,2,3,4,5,3,4,5,6,4,5,6,7]))
simpleconvout=sess.run(simpleconv,feed_dict={input1:a.astype('float32')})

print simpleconvout

[[[[ 27.]
[ 27.]]

[[ 27.]
[ 24.]]]]

simpledeconv=slim.conv2d_transpose(input,1,[3,3],stride = 2,activation_fn = None,scope = 'simpledeconv')
sess.run(tf.global_variables_initializer())
weights=graph.get_tensor_by_name("simpledeconv/weights:0")
sess.run(tf.assign(weights,tf.constant(1.0,shape=weights.shape)))
a=np.ndarray(shape=(1,2,2,1),dtype='float',buffer=np.array([1.0,2,3,4]))
simpleconvout=sess.run(simpledeconv,feed_dict={input:a.astype('float32')})
print simpleconvout

[[[[ 1.000000]
[ 1.000000]
[ 3.000000]
[ 2.000000]]

[[ 1.000000]
[ 1.000000]
[ 3.000000]
[ 2.000000]]

[[ 4.000000]
[ 4.000000]
[ 10.000000]
[ 6.000000]]

[[ 3.000000]
[ 3.000000]
[ 7.000000]
[ 4.000000]]]]

 

conv stride=1是四周padding 0,stride=2是down right padding 0

deconv是top left各插了两行0

而torch中的deconv是四周padding一圈0

 

参考http://blog.csdn.net/lujiandong1/article/details/53728053 

'SAME' padding方式时，如果padding的数目是奇数，则多的padding在右边（下边）

 

2.实现custom-padding

https://stackoverflow.com/questions/37659538/custom-padding-for-convolutions-in-tensorflow 

实现custom conv decon
def conv(input,num_outputs,kernel_size,stride=1,padW=0,padH=0,activation_fn=None,scope=None):
    padded_input = tf.pad(input, [[0, 0], [padH, padH], [padW, padW], [0, 0]], "CONSTANT")
    return slim.conv2d(padded_input,num_outputs,kernel_size,stride = stride,padding="VALID",activation_fn = activation_fn ,scope = scope)
input1 = tf.placeholder(tf.float32, shape=(1,4, 4,1))
a=np.ndarray(shape=(1,4,4,1),dtype='float',buffer=np.array([1.0,2,3,4,2,3,4,5,3,4,5,6,4,5,6,7]))
simpleconv=conv(input1,1,[3,3],stride = 2,padW=1,padH=1,activation_fn = None,scope = 'conv')
sess.run(tf.global_variables_initializer())
weights=graph.get_tensor_by_name("conv/weights:0")
sess.run(tf.assign(weights,tf.constant(1.0,shape=weights.shape)))
simpleconvout=sess.run(simpleconv,feed_dict={input1:a.astype('float32')})
print simpleconvout
[[[[ 8.]
[ 21.]]

[[ 21.]
[ 45.]]]]

 

 

 

def deconv(input,num_outputs,kernel_size,stride=2,activation_fn=None,scope=None):
    N,H,W,C = [i.value for i in input.get_shape()]
    out = slim.conv2d_transpose(input,num_outputs,kernel_size,stride = stride,padding="VALID",activation_fn = activation_fn ,scope = scope)
    return tf.slice(out, [0, kernel_size[0]/2,kernel_size[1]/2, 0], [N, H*stride, W*stride,num_outputs])

input = tf.placeholder(tf.float32, shape=(1,2, 2,1))
a=np.ndarray(shape=(1,2,2,1),dtype='float',buffer=np.array([1.0,2,3,4]))
simpledeconv=deconv(input,1,[3,3],stride = 2,activation_fn = None,scope = 'simpledeconv1')
sess.run(tf.global_variables_initializer())
weights=graph.get_tensor_by_name("simpledeconv1/weights:0")
sess.run(tf.assign(weights,tf.constant(1.0,shape=weights.shape)))
out=sess.run(simpledeconv,feed_dict={input:a.astype('float32')})
print out

[[[[ 1.]
[ 3.]
[ 2.]
[ 2.]]

[[ 4.]
[ 10.]
[ 6.]
[ 6.]]

[[ 3.]
[ 7.]
[ 4.]
[ 4.]]

[[ 3.]
[ 7.]
[ 4.]
[ 4.]]]]