CODEWITHSUNDEEP
pythontensorflowconvolution
koryakinp
koryakinp

Reputation: 4135

tensorflow compute tf.nn.conv2d

I have calculated a convolution between a 3x3 image and two 2x2 filters in Excel manually:

enter image description here

I want to reproduce the same result using tensorflow tf.nn.conv2d:

x_raw = np.array([
    [2,5,3],
    [3,4,2],
    [4,1,1]
])

f_raw = np.array(
[[
    [2,1],
    [3,4]
],[
    [4,1],
    [1,2]   
]])

f = tf.constant(f_raw, dtype=tf.float32)
x = tf.constant(x_raw, dtype=tf.float32)

filter = tf.reshape(f, [2, 2, 1, 2])
image  = tf.reshape(x, [1, 3, 3, 1])

tf.nn.conv2d(image, filter, [1, 1, 1, 1], "VALID").eval()

But the output I have from tensorflow is off:

array([[[[35.,33.],[37.,25.]],[[35.,25.],[19.,15.]]]], dtype=float32)

What I am doing wrong ?

Upvotes: 2

Views: 371

Answers (1)

Tim
Tim

Reputation: 10719

To get identical results as in your excel example, you need to make the following changes:

  1. create two seperate weights
  2. calculate convolutions for each weight separately

Code example: 

x_raw = np.array([
    [2,5,3],
    [3,4,2],
    [4,1,1]
])
#created two seperate weights 
weight1 = np.array(
[[
    [2,1],
    [3,4]
]])

weight2 = np.array(
[[
    [4,1],
    [1,2]
]]
)
weight1 = tf.constant(weight1, dtype=tf.float32)
weight2 = tf.constant(weight2, dtype=tf.float32)
x = tf.constant(x_raw, dtype=tf.float32)

#change out_channels to 1 
filter1 = tf.reshape(weight1, [2, 2, 1, 1])
filter2 = tf.reshape(weight2, [2, 2, 1, 1])
image = tf.reshape(x, [1, 3, 3, 1])

with tf.Session() as sess:
  print(tf.nn.conv2d(image, filter1, [1, 1, 1, 1], "VALID").eval())
  print(tf.nn.conv2d(image, filter2, [1, 1, 1, 1], "VALID").eval())

Upvotes: 2

Related Questions