CODEWITHSUNDEEP
pythontensorflowscipy
J Agustin Barrachina
J Agustin Barrachina

Reputation: 4100

Scipy convolve2d different from tensorflow conv2d

Here is my code:

import tensorflow as tf
import numpy as np
from scipy import signal

img2 = np.array([
    [10, 10, 10, 0, 0, 0],
    [10, 10, 10, 0, 0, 0],
    [10, 10, 10, 0, 0, 0],
    [10, 10, 10, 0, 0, 0],
    [10, 10, 10, 0, 0, 0],
    [10, 10, 10, 0, 0, 0]
]).astype(np.float32)
k = np.array([
    [1., 0., -1.],
    [1., 0., -1.],
    [1., 0., -1.]
]).astype(np.float32)

img_tf = tf.constant(tf.reshape(img2, (1, 6, 6, 1)), dtype=tf.float32)
k_tf = tf.constant(tf.reshape(k, (3, 3, 1, 1)), dtype=tf.float32)
conv_tf = tf.nn.conv2d(img_tf, k_tf, strides=[1, 1], padding="SAME")[0, ..., 0]
print("conv_tf: " + str(conv_tf))

np_conv = np.array(signal.convolve2d(img2 , k, "same"), np.int32)
print("sp_conv:\n" + str(np_conv))

The output is:

conv_tf: tf.Tensor(
[[-20.   0.  20.  20.   0.   0.]
 [-30.   0.  30.  30.   0.   0.]
 [-30.   0.  30.  30.   0.   0.]
 [-30.   0.  30.  30.   0.   0.]
 [-30.   0.  30.  30.   0.   0.]
 [-20.   0.  20.  20.   0.   0.]], shape=(6, 6), dtype=float32)
sp_conv:
[[ 20   0 -20 -20   0   0]
 [ 30   0 -30 -30   0   0]
 [ 30   0 -30 -30   0   0]
 [ 30   0 -30 -30   0   0]
 [ 30   0 -30 -30   0   0]
 [ 20   0 -20 -20   0   0]]

Now one seam as the opposite of the other one, why is there this difference?

Upvotes: 0

Views: 865

Answers (1)

J Agustin Barrachina
J Agustin Barrachina

Reputation: 4100

In the end, I found the response.

The issue is scipy does the mathematically 'correct' convolution, whereas tensorflow does the convolution oriented to a Convolutional Neural Network (CNN) application.

Therefore, scipy inverts the kernel before applying the convolution (as explained here) whereas tensorflow does not.

Upvotes: 2

Related Questions