如何在损失函数中使用 tf.contrib.image.rotate？

Question

我在损失函数中使用了 tf.conbrib.image.rotate，出现错误：

No gradients provided for any variable, check your graph for ops that do not support gradients,

我的代码是：

import tensorflow as tf

image_tensor = tf.placeholder(dtype=tf.float32, shape=[None,320,320,1])
target_tensor = tf.placeholder(dtype=tf.float32, shape=[None,320,320,1])
s = tf.concat([image_tensor, target_tensor],axis=3)
s = tf.layers.flatten(s)
w = tf.get_variable(initializer=tf.truncated_normal([204800,1], stddev=0.1),name='w')
b = tf.get_variable(initializer=tf.truncated_normal([1], stddev=0.1),name='b')
a = tf.matmul(s,w)+b
diff = tf.contrib.image.rotate(image_tensor, a[:,0], interpolation='BILINEAR') - target_tensor
loss = tf.reduce_sum(tf.square(diff))

optimizer = tf.train.GradientDescentOptimizer(0.001)
train = optimizer.minimize(loss)

我的tensorflow是：1.4.0，我的电脑是Win10。

顺便问一下，如何在 tensorflow 中旋转 3D 图像？ tf.conbrib.image.rotate 仅适用于 2D 图像。

Answer 1

tf.contrib.image.rotate 无法优化。我找到了空间变换网络的有用代码 (https://github.com/kevinzakka/spatial-transformer-network)。