CODEWITHSUNDEEP
pythontensorflow
YW P Kwon
YW P Kwon

Reputation: 2168

How to apply interp1d to each element of a tensor in Tensorflow

Let say, I have a interpolation function.

def mymap():
    x = np.arange(256)
    y = np.random.rand(x.size)*255.0
    return interp1d(x, y)

This guy maps a number in [0,255] to a number following the profile given by x and y (now y is random, though). When I do following, each value in image gets mapped nicely.

x = imread('...')
x_ = mymap()(x)

However, how can I do this in Tensorflow? I want to do something like

img = tf.placeholder(tf.float32, [64, 64, 1], name="img")
distorted_image = tf.map_fn(mymap(), img)

But it results in an error saying

ValueError: setting an array element with a sequence.

For information, I checked if a function map is simple as below, it works well

mymap2 = lambda x: x+10
distorted_image = tf.map_fn(mymap2, img)

How can I map each number in a tensor? Could anyone help?

Upvotes: 0

Views: 633

Answers (1)

Olivier Moindrot
Olivier Moindrot

Reputation: 28198

The function input of tf.map_fn needs to be a function written with Tensorflow ops. For instance, this one will work:

def this_will_work(x):
    return tf.square(x)

img = tf.placeholder(tf.float32, [64, 64, 1])
res = tf.map_fn(this_will_work, img)

This one will not work:

def this_will_not_work(x):
    return np.sinh(x)

img = tf.placeholder(tf.float32, [64, 64, 1])
res = tf.map_fn(this_will_not_work, img)

Because np.sinh cannot be applied to a TensorFlow tensor (np.sinh(tf.constant(1)) returns an error).

Solutions

You can write your interpolation function in TensorFlow, and maybe ask for help in another StackOverflow question.

If you absolutely want to use scipy.interpolate.interp1d, you will need to keep the code encapsulated in python. For that, you can use tf.py_func, and use your scipy function inside.

Upvotes: 1

Related Questions