Tensorflow conditional throwing value error

Question

I am trying to use conditionals with tensorflow and I am getting the error:

ValueError: Shapes (1,) and () are not compatible

Below is the code I use that is throwing the error. It is saying the error is in the conditional

import tensorflow as tf
import numpy as np

X = tf.constant([1, 0])
Y = tf.constant([0, 1])
BOTH = tf.constant([1, 1])
WORKING = tf.constant(1)

def create_mult_func(tf, amount, list):
    def f1():
        return tf.scalar_mul(amount, list)
    return f1

def create_no_op_func(tensor):
    def f1():
        return tensor
    return f1

def stretch(tf, points, dim, amount):
    """points is a 2 by ??? tensor, dim is a 1 by 2 tensor, amount is tensor scalor"""
    x_list, y_list = tf.split(0, 2, points)
    x_stretch, y_stretch = tf.split(1, 2, dim)
    is_stretch_X = tf.equal(x_stretch, WORKING, name="is_stretch_x")
    is_stretch_Y = tf.equal(y_stretch, WORKING, name="is_stretch_Y")
    x_list_stretched = tf.cond(is_stretch_X,
                               create_mult_func(tf, amount, x_list), create_no_op_func(x_list))
    y_list_stretched = tf.cond(is_stretch_Y,
                               create_mult_func(tf, amount, y_list), create_no_op_func(y_list))
    return tf.concat(1, [x_list_stretched, y_list_stretched])

example_points = np.array([[1, 1], [2, 2], [3, 3]], dtype=np.float32)
example_point_list = tf.placeholder(tf.float32)

result = stretch(tf, example_point_list, X, 1)
sess = tf.Session()

with tf.Session() as sess:
    result = sess.run(result, feed_dict={example_point_list: example_points})
    print(result)

Stack trace:

  File "/path/test2.py", line 36, in <module>
    result = stretch(tf, example_point_list, X, 1)
  File "/path/test2.py", line 28, in stretch
    create_mult_func(tf, amount, x_list), create_no_op_func(x_list))
  File "/path/tensorflow/python/ops/control_flow_ops.py", line 1142, in cond
    p_2, p_1 = switch(pred, pred)
  File "/path/tensorflow/python/ops/control_flow_ops.py", line 203, in switch
    return gen_control_flow_ops._switch(data, pred, name=name)
  File "/path/tensorflow/python/ops/gen_control_flow_ops.py", line 297, in _switch
    return _op_def_lib.apply_op("Switch", data=data, pred=pred, name=name)
  File "/path/tensorflow/python/ops/op_def_library.py", line 655, in apply_op
    op_def=op_def)
  File "/path/tensorflow/python/framework/ops.py", line 2156, in create_op
    set_shapes_for_outputs(ret)
  File "/path/tensorflow/python/framework/ops.py", line 1612, in set_shapes_for_outputs
    shapes = shape_func(op)
  File "/path/tensorflow/python/ops/control_flow_ops.py", line 2032, in _SwitchShape
    unused_pred_shape = op.inputs[1].get_shape().merge_with(tensor_shape.scalar())
  File "/path/tensorflow/python/framework/tensor_shape.py", line 554, in merge_with
    (self, other))
ValueError: Shapes (1,) and () are not compatible

I have tried changing the WORKING to be an array instead of a scalar.

I believe that the problem is that tf.equal is returning an int32 instead of the bool that it is supposed to return according to the documentation

Answer 1

The problem is in the first argument to tf.cond. From the documentation here, about the type of the first argument to tf.cond :

pred: A scalar determining whether to return the result of fn1 or fn2.

Note that it has to be a scalar. You are using the result of comparing a tensor and a tensor, which gives you a (1,) tensor, NOT a scalar. You can convert it to a scalar using the tf.reshape operator as follows :

t = tf.equal(x_stretch, WORKING, name="is_stretch_x")
x_list_stretched = tf.cond(tf.reshape(t, []),
                           create_mult_func(tf, amount, x_list), create_no_op_func(x_list))

Complete working program :

import tensorflow as tf
import numpy as np

X = tf.constant([1, 0])
Y = tf.constant([0, 1])
BOTH = tf.constant([1, 1])
WORKING = tf.constant(1)

def create_mult_func(tf, amount, list):
    def f1():
        return tf.scalar_mul(amount, list)
    return f1

def create_no_op_func(tensor):
    def f1():
        return tensor
    return f1

def stretch(tf, points, dim, amount):
    """points is a 2 by ??? tensor, dim is a 1 by 2 tensor, amount is tensor scalor"""
    x_list, y_list = tf.split(0, 2, points)
    x_stretch, y_stretch = tf.split(0, 2, dim)
    is_stretch_X = tf.equal(x_stretch, WORKING, name="is_stretch_x")
    is_stretch_Y = tf.equal(y_stretch, WORKING, name="is_stretch_Y")
    x_list_stretched = tf.cond(tf.reshape(is_stretch_X, []),
                               create_mult_func(tf, amount, x_list), create_no_op_func(x_list))
    y_list_stretched = tf.cond(tf.reshape(is_stretch_Y, []),
                               create_mult_func(tf, amount, y_list), create_no_op_func(y_list))
    return tf.pack([x_list_stretched, y_list_stretched])

example_points = np.array([[1, 1], [2, 2]], dtype=np.float32)
example_point_list = tf.placeholder(tf.float32)

result = stretch(tf, example_point_list, X, 1)
sess = tf.Session()

with tf.Session() as sess:
    result = sess.run(result, feed_dict={example_point_list: example_points})
    print(result)