CODEWITHSUNDEEP
tensorflow
eof
eof

Reputation: 349

TensorFlow convolution of 2D array

I have some data where I have a 2D array A and a kernel K. I would like to compute a convolution of these. How do I do this in TensorFlow. It looks like the built-in convolution operations take a 4D tensor as input, since they assume that we're computing a batch with multiple color channels. How do I use these ops with my data?

Example:

a = np.random.randint(2, size=(10,10))
k = [[1,1,1],[1,1,1],[1,1,1]]
# Compute convolution of these??
c = ...

Upvotes: 2

Views: 3706

Answers (3)

Umar Karim
Umar Karim

Reputation: 41

In addition to kaufmanu's answer, you can use tf.squeeze if you need a 2D result. I.e.,

c=tf.squeeze(tf.nn.conv2d(a_tensor, k_weight,padding='VALID',strides=[1, 1, 1, 1]))

to perform the convolution.

Upvotes: 1

kafman
kafman

Reputation: 2860

The answers posted so far all miss one important point: Tensorflow does not compute a convolution, but a cross-correlation as is stated in the doc:

Note that although these ops are called "convolution", they are strictly speaking "cross-correlation" since the filter is combined with an input window without reversing the filter.

If you really want to compute a convolution, you will have to reverse the kernel before passing it into conv2d, i.e. flip it once on the horizontal then on the vertical axis. Using Miriam's answer, this could look like this:

import tensorflow as tf
import numpy as np

a = np.random.randint(2, size=(10,10))
k = np.array([[1,1,1],[1,1,1],[1,1,1]],dtype=np.float32)
flip = [slice(None, None, -1), slice(None, None, -1)]
k = k[flip]

a=a.astype(np.float32)
a_tensor = tf.reshape(a, [1, 10, 10, 1])
k_weight = tf.reshape(np.array(k), [3,3,1,1])
c=tf.nn.conv2d(a_tensor, k_weight,padding='VALID',strides=[1, 1, 1, 1])
sess=tf.Session()
c.eval(session=sess)

Note that in this specific example flipping the kernel is technically in vain, because for symmetric kernels convolution and cross-correlation is the same thing. However, as soon as you have a non-symmetric kernels, you must flip it if you want Tensorflow to actually compute a convolution.

Upvotes: 1

Yu Zeng
Yu Zeng

Reputation: 33

import tensorflow as tf
import numpy as np

a = np.random.randint(2, size=(10,10))
k = [[1,1,1],[1,1,1],[1,1,1]]

tensor_a = tf.constant(a, tf.float32)
tensor_k = tf.constant(k, tf.float32)

tensor_res = tf.nn.convolution(tf.reshape(tensor_a, [1, 10, 10, 1]), tf.reshape(tensor_k, [3, 3, 1, 1]), padding='VALID')

sess = tf.Session()
print(sess.run(tensor_res))

The Computational Graph tutorial is here

Convolution helper doc

Upvotes: 3

Related Questions