Classification of NMNIS datasets on the SpiNNaker platform

48 views
Skip to first unread message

Yuyang Liu

unread,
Aug 11, 2022, 3:44:28 AM8/11/22
to SpiNNaker Users Group
Hi,

I'm modifying a training code for NMNIST that I want to use for digit classification on the test set. I put all the numbers 0-9 in the training set into training, and I labeled 100 neurons after training. Almost all neurons have a maximum spike count of 1, so that all neurons are classified as 1. If the test set is classified by the number of spikes and the corresponding neuron labels, all the test set classification results will be 1. If you just put the three numbers 6, 7, and 9 from the training set into training, the neurons marked by the result will have 6, 7, and 9, not just 1. I don't know what is causing this result, and attached is my code file.


Many thanks,

Yuyang
timer.py
eventvision.py
testx.ipynb

Muhammad Aitsam

unread,
Sep 28, 2022, 10:42:03 AM9/28/22
to SpiNNaker Users Group
Hi,

I am also having the similar problem. I tried changing the hyperparameters but resulting in 0.0% accuracy. I am attaching the code.
Thank you in advance.
newCodeNMNIST.ipynb

Andrew Rowley

unread,
Sep 28, 2022, 10:53:04 AM9/28/22
to SpiNNaker Users Group
Hi,

I don't know where this code was from, but I don't recognise it so I have no idea what the network is supposed to do beyond learning MNIST digits i.e. how it is supposed to operate.  Do you know if it was run on SpiNNaker before and now doesn't work, or is this something that has been run on another platform and you are now trying it on SpiNNaker?

Thanks,

Andrew :)

Muhammad Aitsam

unread,
Sep 28, 2022, 11:08:06 AM9/28/22
to SpiNNaker Users Group
I modified this code.

I am running it on SpiNNaker 48chip board.

My major goal is to do hand gesture recognition but before that I was trying to implement SNN for NMNIST dataset. If you have better example code for NMNIST and/or for hand gesture dataset then kindly let me know.

Andrew Rowley

unread,
Sep 28, 2022, 11:15:53 AM9/28/22
to Muhammad Aitsam, SpiNNaker Users Group

Hi,

 

Sorry, I’m not sure if that ever worked unfortunately!  I don’t specifically have any better example code.  From a quick look at the repository, it seems that the person who wrote it was trying to implement something from a paper:

https://www.frontiersin.org/articles/10.3389/fncom.2015.00099/full

 

It may be a good idea to go back to that and see if there is any information about what each of the parts should do and then see if you can work out what might or might not be working in this example.

 

Andrew :)

 

--
You received this message because you are subscribed to the Google Groups "SpiNNaker Users Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to spinnakeruser...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/spinnakerusers/477a0ae4-76a0-45a0-b579-d2479bf3a526n%40googlegroups.com.

Muhammad Aitsam

unread,
Sep 28, 2022, 11:30:36 AM9/28/22
to SpiNNaker Users Group
Thank you so much for reply.

One last question. Can you suggest article or code to read spiking data and train it with SpiNNaker.

Andrew Gait

unread,
Sep 28, 2022, 12:21:41 PM9/28/22
to Muhammad Aitsam, SpiNNaker Users Group
Hi,

The notebook / repository that you shared originally is a good starting point for reading the relevant data.

Andy

----------------------------------------------------
Dr. Andrew Gait (he/him),
Research Software Engineer,
APT Group,
Department of Computer Science,
The University of Manchester,
Oxford Road, Manchester M13 9PL

email: andre...@manchester.ac.uk
web: http://personalpages.manchester.ac.uk/staff/Andrew.Gait/

I do not expect a reply to this email outside of normal working hours

From: spinnak...@googlegroups.com <spinnak...@googlegroups.com> on behalf of Muhammad Aitsam <aitsam...@gmail.com>
Sent: 28 September 2022 16:30

To: SpiNNaker Users Group <spinnak...@googlegroups.com>
Subject: Re: [SpiNNaker Mailing List] Re: Classification of NMNIS datasets on the SpiNNaker platform
 
Reply all
Reply to author
Forward
0 new messages