How to choose a pre-trained neural network for your image classification

Hello,

I'm looking into the pre-trained neural networks such as : mobilenet , googlenet, resnet 50, VGG16 and alex net

and I'm wondering how to choose one of them to transfer learning?

my image classification program is simple and it's about detecting traffic signs, but the question is how should i choose one of the above networks?

how do I know which one of them will perform better than the others ?

Answer 1

[Nagasai Bharat]

Hi,

From my understanding you are trying to use a pretrained model for an image classification task. In order to do that you can find the available pretrained networks from this doc. The doc provides a section on Transfer learning . And an example on how to use this network can be accessed from here. The example also shows how to evaluate the network too.

[Abdulaziz Alotaibi]

Thank you, I know that , but the question is what differentiate different networks from each other?

for example, should i pick googlenet or Alexnet ? because I don't know what to choose with my classification task

[Nagasai Bharat]

The choice would be on the requirement of the project you are working on. The time the model takes to predict, memory contraint of the model parameters, the harware you are using to train(more layers implies more time taken by the learning algorithm) and finally the accuracy you are getting without overfiting.

An option would be to try each model and cross-check their acceracy metric and choose the best one also taking into consideration of the above requirements.

Machine Learning Challenges: Choosing the Best Classification Model and Avoiding Overfitting - MATLAB & Simulink (mathworks.com) - This could help you

[Abdulaziz Alotaibi]

Thank you!