I have two arrays and I can't figure out how to simply convolve them in tensorflow. In numpy I would do:
import numpy as np
from numpy import convolve
a=np.array([1,0,3])
b=np.array([-1,2])
convolve(a,b)
and the output is array([-1, 2, -3, 6]).
How can I do this in tensorflow? I tried this:
a=tf.constant([1,0,3])
b=tf.constant([-1,2])
tf.nn.convolution(a,b,strides=1)
But I get ValueError: `num_spatial_dims` must be 1, 2, or 3. Received: num_spatial_dims=-1.
Tensorflow does not have a full padding option like in numpy, so you'll have to add padding yourself:
a=tf.constant([1,0,3])
b=tf.constant([-1,2])
b=tf.reverse(b,axis=[0])
pad_add = len(b)-1
paddings = tf.constant([[pad_add,pad_add]])
a = tf.pad(a,paddings)
a=tf.reshape(a,(1,len(a),1))
b=tf.reshape(b,(len(b),1,1))
output =tf.nn.convolution(input = a, filters = b)
print(output.numpy().flatten())
This gives:[-1 2 -3 6]