CODEWITHSUNDEEP
pythontensorflowconv-neural-networkconvolution
mamafoku
mamafoku

Reputation: 1139

Tensorflow Convolution with Different Filter Sizes

I would like to convolve over my data feed with filters of different sizes and was wondering how I can achieve the following setup using Tensorflow enter image description here

In other words, I would like to have two parallel convolutions and connect them in the flattening layer, right before feeding into the fully connected layer but am unsure about the connections.

Any code snippet or sources on approach would tremendously help!

Upvotes: 1

Views: 1079

Answers (1)

jkschin
jkschin

Reputation: 5844

Assume batch size 100 and image data of size 28x28x1. 

import tensorflow as tf

inp = tf.placeholder(tf.float32, shape=[100, 28, 28, 1])
left_branch = tf.layers.conv2d(input=inp, filters=N, kernel_size=[L, M])
right_branch = tf.layers.conv2d(input=inp, filters=P, kernel_size=[R, S])

left_reshape = tf.reshape(left_branch, [100, num_outputs_in_left_branch])
right_reshape = tf.reshape(right_branch, [100, num_outputs_in_right_branch])

combined_branch = tf.concat([left_reshape, right_reshape], axis=1)
combined_branch = tf.layers.dense(combined_branch, num_units_in_dense)

Upvotes: 2

Related Questions