I found total number of neurons of ResNet-50 model is 26,560 and 94,059 in two different papers. Their titles are below:

1: DeepXplore: Automated Whitebox Testing of Deep Learning Systems 2: Testing Deep Neural Network based Image Classifiers

I found another article (link) that clearly describes how many neurons a CNN does have. If I follow this technique then the neuron count does not match with the count mentioned in those papers.

Could anybody please help me how I can calculate number of neurons in a DL model? Also which layers have neurons and which don't have.

  • $\begingroup$ Do you mind providing links ot where you found that ResNet-50 has 26k and 94k neurons, that seems very far off from the actual number of neurons they have. Secondly, are you asking for neurons using a framework or theoretically? $\endgroup$
    – PSub
    Mar 13 at 6:14
  • $\begingroup$ In my post I have mentioned about two papers and if you go through these papers you will find two tables and in those tables the authors mentioned that number of total neurons. I want to know the theoretical concept to calculate number of neurons a DL model (specially CNN) does have. Also in my post I have mentioned about a link and if you have a look in that link then you will find out how we can calculate number of neurons a model does have. I also appreciate your effort if you can help me how can I count number of neurons through a framework like Tensorflow or Keras. $\endgroup$
    – Abdul Awal
    Mar 13 at 14:55

I hope you are talking about the parameter count in a convolutional neural network.

Below is basic understanding about calculating the parameters. After getting this, it would be self explanatory.

kernel count = k
kernel size = (a, a)
depth of input = d 
Total number of parameters: k*(a*a)*d + k

Now, for below scenario which has one input layer with one con2d layer.

enter image description here

To cross check the number of parameters, we need k,d,a:

 k = 64, d = 3, a*a = 3*3

Total parameters: 64*(3*3)*3 + 64 = 1792

Hope its clear now.

  • $\begingroup$ Thanks for your comment. I know how to calculate number of trainable parameters. But my concern is how to calculate number of neurons a DL model (specially CNN) does have. Also I have mentioned about two papers in my post and the authors mentioned total number of neurons in those papers. If you have a look to those papers then you can see there are two tables in those papers and the number of neurons are mentioned along with their DL model. $\endgroup$
    – Abdul Awal
    Mar 13 at 15:13
  • $\begingroup$ Selecting hyperparameters totally depends on how developer would like to see it. It is selected after hyperparameter tunings like kerastuner, GridCVSearch, RandomCVSearch . $\endgroup$ Mar 13 at 15:23

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.