如何根据相应卷积层的训练值权重计算反卷积层的权重

Question

这可能吗？ 反卷积层的对应层是tf.conv2d_transpose（），但是文档指出它只是一个转置转换层，而不是真正的反卷积层。 那么如何计算去反层的权重呢？ 像下面的代码一样，如何使y == x（通过从W计算W2）？这可能吗？还是唯一的方法是训练去转化层？

# [batch, height, width, depth]
x_image = tf.placeholder(tf.float32,shape=[3,2])
x = tf.reshape(x_image,[1,3,2,1])

#Filter: W  [kernel_height, kernel_width, output_depth, input_depth]
W_cpu = np.array([[-1,1]],dtype=np.float32)
W = tf.Variable(W_cpu)
W = tf.reshape(W, [1,2,1,1])

W_cpu2 = np.array([[-1,1]],dtype=np.float32)
W2 = tf.Variable(W_cpu2)
W2 = tf.reshape(W2, [1,2,1,1])

strides=[1, 1, 1, 1]
padding='VALID'

z = tf.nn.conv2d(x, W, strides=strides, padding=padding)

y = tf.nn.conv2d_transpose(z, W2, [1,3,2,1],strides, padding)

x_data = np.array([[1,-1],[2,2],[1,2]],dtype=np.float32)
with tf.Session() as sess:
    init = tf.initialize_all_variables()
    sess.run(init)

    x = (sess.run(x, feed_dict={x_image: x_data}))
    W = (sess.run(W, feed_dict={x_image: x_data}))
    z = (sess.run(z, feed_dict={x_image: x_data}))
    y = (sess.run(y, feed_dict={x_image: x_data}))

    print("The shape of x:\t", x.shape, ",\t and the x.reshape(3,2) is :")
    print(x.reshape(3,2))
    print()

    print ("The shape of x:\t", W.shape, ",\t and the W.reshape(1,2) is :")
    print (W.reshape(1,2))
    print ("")


    print ("The shape of z:\t", W.shape, ",\t and the W.reshape(1,2) is :")
    print (z.reshape(3))
    print ("")

    print ("The shape of y:\t", y.shape, ",\t and the y.reshape(3,3) is :")
    print (y.reshape(3,2))
    print ("")

但是在Matthew D. Zeiler和Rob Fergus 的可视化和理解卷积网络中，他们提出了一种对“转置”进行反变换的方法。 相同过滤器*的版本，这是否意味着形状已移置或包含权重？