alexnet_params.h not found

[k0nze]

In the imagenet directory is a alexnet.c which I would like to compile. However, the ? alexnet_params.h and alexnet_images.h are missing from that directory, and due to that, the compilation of AlexNet fails. Is it possible to get the header files for AlexNet?

[hngenc]

We generated one internally a long time ago, but we didn't end up pushing it because it was pretty large. I've taken a look, and I can't find it anymore. If I end up finding the file later, I'll update you.

In the meantime, if you want to run alexnet.c, you could ust images.h instead of alexnet_images.h, and perhaps generate dummy parameters for all the layers that are referenced in alexnet.c? The inferences will be incorrect of course, but the performance numbers should be identical.

[HyunseokErikJung]

If you were to port a trained model parameters to work in RoCC/Gemmini, what steps would it take to do the necessary quantizations? Been looking around all available resources but couldn't get a solid answer for this.

[mbelda]

@k0nze Did you managed to generate the dummy parameters for alexnet? I am also interested in executing it.

[k0nze]

@mbelda No, I haven't I might have a look in the coming weeks.

[mik300]

@k0nze Did you generate the parameters for alexnet?

[k0nze]

@mbelda I ended up implementing Conv2D and FC AlexNet Layer from scratch

[mik300]

@k0nze Do you mind sharing the code please?

[k0nze]

I'm not able to share the code, but it was quite straightforward. As I was only interested in the runtime of AlexNet I filled the weights in the header with 0 and just set the correct dimensions taken from the torchvision implementation.

[mik300]

Oh, I see. I was asking because I needed trained parameters, but since you used all zeros, I can't use your code. Thank you anyway.