0
$\begingroup$

I have created 4 x 4 2d images from a signal. Now, I want to feed this data to a Convolutional neural Network. How I can choose the nubmber of filter, Kernel Size for this small, shape and size of dataset (4 x 4, 320 images). Should I use maxpooling layer ?

Note: I have already implemented a model for this dataset and accuracy is 78.125%. The model is,

enter image description here

$\endgroup$
5
  • $\begingroup$ Are you asking how to input the values? or Which values to feed? also do add tensorflow in your tags for better exposure to the question. $\endgroup$
    – Academic
    Nov 23, 2020 at 14:35
  • $\begingroup$ @SoumyaKundu thanks. No, I am looking for suggestion how I can choose Kernel size & depth of model for this data. $\endgroup$ Nov 23, 2020 at 14:44
  • $\begingroup$ towardsdatascience.com/… - refer here! $\endgroup$
    – Academic
    Nov 23, 2020 at 14:49
  • $\begingroup$ Why you need CNN for 4x4 i.e. 16 dimensions data? Do you believe that CNN will find something special but a simple model will not even if data is very small? $\endgroup$
    – 10xAI
    Nov 24, 2020 at 9:51
  • $\begingroup$ @10xAI Thanks for your response. I want to compare the best performance between ML and CNN. So, I did this. Is there any good way for doing this? $\endgroup$ Nov 25, 2020 at 11:36

1 Answer 1

0
$\begingroup$

Since the images are just 4x4 in size, you can do the following :

  • You can resize the image to a much larger dimension like 28x28 and then use sharpen or histogram equalization to bring out the contrast. Then use a 3x3x16, 3x3x 32 kernel arrays in 2 convolutional layers. The rest is fully connected.

  • The images can be input as 4x4 images but use a shorter neural net. One convolutional layer with padding and 3x3x 32 shape would do.

  • Use something similar to google nets, inception module, in which output from multiple convolutions are combined.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.