Show HN: Try running deep learning inference on Raspberry Pi | Hacker News Reader

Show HN: Try running deep learning inference on Raspberry Pi

actcast.io

53 pointsnineties11 comments

We don't need more proprietary machine learning devices in our homes. I'd appreciate this so much more, if it was open source, so I can reshape it to whatever use case I have.

There are plenty of viable business models, that give you your well earned money and us the option to customize and understand our devices.

> We don't need more proprietary machine learning devices in our homes. I'd appreciate this so much more, if it was open source

It doesn't seem to do anything special? You can probably run your favorite machine learning framework on the Raspberry Pi, and it will work - albeit using the ARM cores and NEON only. Now, machine learning and inference _using the Raspberry Pi's GPU part_ (which is broadly documented, unlike most GPU hardware) would be a gamechanger, if only for educational scenarios.

There is PlaidML[1] - a cross-platform deep learning framework that works, among others, on Raspberry Pi's GPU.

[1] - https://github.com/plaidml/plaidml

Edit: they don't support RPI GPU yet - https://github.com/plaidml/plaidml/issues/141

I agree, so I looked at running image detection offline on a raspberry pi and wrote a post about it. It can't do anywhere near real time object detection on the larger YOLO models, but real time detection is often unnecessary. Also there are smaller models (e.g. yolo-mini) that can likely give you an acceptable frame rate.

https://medium.com/ml-everything/offline-object-detection-an...

You can run TensorFlow Lite on a Pi with no problems at all. You can even train and run basic gesture recognition with full TensorFlow on a Pi.

Source: wrote a tutorial doing this for Arm (https://github.com/ARM-software/ML-examples/blob/master/mult...)

There are gesture recognition sensors available that don't even need a general purpose CPU, eg. Broadcom APDS-9960. https://cdn.sparkfun.com/datasheets/Sensors/Proximity/apds99...

Security major issue for edge computing. Also I guess as it is isolated you assume it will be hacked and then handle the risk as well.

Non-open source? would it make it harder to deal with security and deal with as we do not know it is hacked or not.

Blog article about this release: https://blog.idein.jp/post/181016515935/alphareleaseen

Why would I use this rather than Snips? Edge computing is great, but in a closed source, hard to audit form it will be a tough sell to get 3rd party developers onboard.

Since this is about computing at the edge, I was wondering if anyone had an opinion on the intel neural compute stick 2?

I got one recently and I am not convinced, but I suppose someone on HN might have legitimate use cases.

I think the NCS2 has potential.. it's still early though. One thing on their roadmap is to support ARM, which I view as a must-have for low powered edge applications. Right now it only works with OpenVINO and certain x86 systems. So I view the real use cases coming when it can be used in settings with limited power (and limited or variable network connectivity).

Reply on news.ycombinator.com