在Tensorflow中,如何根据索引在Tensor中分配值？

Question

我想根据索引在张量中分配值.

例如,根据池值和tf.nn.max_pool_with_argmax的相应索引输出,我想将这些池值重新放回到具有索引的原始拆分Tensor中.

我发现输出指数tf.nn.max_pool_with_argmax是扁平的.一个问题:如何将它们解开回Tensorflow中的坐标？

另一个问题:如果给出索引,如何将池化张量的每个值分配给Tensorflow中原始解析张量的位置？

非常感谢你.

我试图制作代码来实现这一点,但我可以使用numpy.我不知道如何tf.nn.max_pool_with_argmax在Tensorflow中分配到unpooling张量之后获得平坦的索引.

ksize = 3
stride = 1

input_image = tf.placeholder(tf.float32, name='input_image')

#conv1
kernel = tf.Variable(tf.truncated_normal([ksize, ksize, 3, 16],stddev=0.1),
                    name='kernel')
conv = tf.nn.conv2d(input_image, kernel, [1,stride,stride,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0, shape = [16]), name = 'biases')
bias = tf.nn.bias_add(conv, biases)
conv1 = tf.nn.relu(bias, name='conv1')

#pool1
pool1, pool1_indices = tf.nn.max_pool_with_argmax(conv1, ksize=[1, 2, 2, 1], 
                                                  strides=[1, 2, 2, 1], 
                                                  padding='SAME', name='pool1')

#upsample by assigning the values of pool1 to the position in unpooling Tensor according to pool1_indices                                                
indices = pool1_indices
unravel_pool1_indices = np.unravel_index(indices,[4,32,32,16])
unravel_pool1_coordinates = np.array(unravel_pool1_indices)
coor_shape = np.shape(unravel_pool1_coordinates)
unravel_pool1_coordinates = np.reshape(unravel_pool1_coordinates,(coor_shape[0],coor_shape[1]*coor_shape[2]*coor_shape[3]*coor_shape[4]))
unravel_pool1_coordinates = unravel_pool1_coordinates.T

values = pool1
values = np.reshape(values,(np.size(values)))

up1 = tf.constant(0.0, shape = [4,32,32,16])
delta = tf.SparseTensor(unravel_pool1_coordinates, values, shape = [4,32,32,16])

result = up1 + tf.sparse_tensor_to_dense(delta)


with tf.Session() as session:
    session.run(tf.initialize_all_variables())
    test_image = np.random.rand(4,32,32,3)
    sess_outputs = session.run([pool1, pool1_indices],
                               {input_image.name: test_image})

Answer 1

有一个待处理的 PR 应该可以解决这个问题：

https://github.com/tensorflow/tensorflow/issues/1793