Passing dlarrays to Alexnet

Dear community,

As a part of training a custom DL model, I want to pass an 'image' to Alexnet and extract an intermediate layer's activation. The issue that I have is that the image is dlarray, i.e., a matrix with learnable parameters of my custom network - therefore, it needs to stay dlarray in order to be traceable. Alexnet is not accepting dlarrays, so activation(), predict() and forward() are useless here. Any advice on how to best approach this?

Answer 1

[Himanshu]

Hello,

To pass a "dlarray" to AlexNet and extract intermediate activations, convert the "dlarray" using "extractdata", process it through AlexNet, and convert the output back if needed.

Deep learning array for customization: https://www.mathworks.com/help/deeplearning/ref/dlarray.html

Extract data from "dlarray": https://www.mathworks.com/help/deeplearning/ref/dlarray.extractdata.html

Alternatively, convert AlexNet to a "dlnetwork" using MATLAB's "dlnetwork" function, which accepts "dlarray" inputs directly. This allows you to extract activations using the "forward" method while maintaining parameter traceability.

Compute deep learning network output for training: https://www.mathworks.com/help/deeplearning/ref/dlnetwork.forward.html

I hope this helps.