使用Tensorflow过滤Tensor

Question

I'm attempting to filter a matrix that represents a point cloud in tensorflow. It is an nx 3 matrix.

I only want to keep rows with z > eps . This corresponds to column index 2 of the matrix.

I have the following code:

import numpy as np
import tensorflow as tf

point_cloud = tf.placeholder(tf.float32, shape=[None,3])
eps = tf.placeholder(tf.float32)

mask = tf.greater(point_cloud[:,2], eps)
reduced_cloud = tf.boolean_mask(point_cloud, mask)


init = tf.global_variables_initializer()

with tf.Session() as sess:
    sess.run(init)
    _cloud = np.random.rand(5000,3)
    feed = {point_cloud:_cloud, eps:0.0025}
    _filtered = sess.run(reduced_cloud, feed_dict=feed)

When I run the above code I get this:

ValueError: Number of mask dimensions must be specified, even if some dimensions are None.  E.g. shape=[None] is ok, but shape=None is not.

I don't understand the error message, having tried to specify shape in a number of places with no success, and the documentation seems to suggest the boolean_mask only works with np.array s. Is there any way to do this entirely on the tensorflow graph?

Answer 1

You haven't specified the shape of eps which needs to be a 1D-tensor:

import numpy as np
import tensorflow as tf

point_cloud = tf.placeholder(tf.float32, shape=[None,3])
eps = tf.placeholder(tf.float32, shape=[None])

mask = tf.greater(point_cloud[:,2], eps)
reduced_cloud = tf.boolean_mask(point_cloud, mask)

init = tf.global_variables_initializer()
with tf.Session() as sess:
     sess.run(init)
     _cloud = np.random.rand(5000,3)
     feed = {point_cloud:_cloud, eps:[0.0025]}
     _filtered = sess.run(reduced_cloud, feed_dict=feed)