ML / Jetson Nano / Google CoLab / Training ML in the cloud ...

[Jon Russell]

Hi,

I got a Jetson nano for Xmas, and I've been training it to recognise birds on my garden bird feeder. I have a basic model working, trained locally on the Jetson, with about 500 images.

Ideally I want to train the model in the "cloud" and then download it to the Jetson. I could do with a few pointers, from someone who's done it before.

I think I can use Google CoLabs, or AWS SageMaker or Jupyter Notebooks, which I think all do roughly the same thing ?

I am currently training it with pyTorch locally on the Jetson.

I have started porting the code to CoLabs, just because its the first one I tried. I am using a MobileNet v2 SSD model and TensorRT on the Jetson, with a bunch of python libraries provided by nVidia.

I think once I train it on CoLabs I need to export it to Onnx.

I'm very new to this. I've been playing for about 4 weeks in total! ;-)

Has anyone done this before. Fancy a chat to give me some pointers. I have lots of questions like where the best place to store the source images, do I need to download them all locally in to CoLabs. What is the best / cheapest platform to train a MobileNet SSD. What other pitfall will I face transferring the trained model from the cloud, back to the Jetson?

Thanks !

Jon.

[Erica Calogero]

Hi Jon,

Sounds interesting. I haven't played with any hardware, but have done a totally web-based project with Google Teachable machine. It was super easy. Ans once you've trained the model, you can download it and use it locally or point to it from the web. Here's an article on how to use it. I understand that Amazon also has a product called DeepLens with it's own hardware, but I'm not sure what the real uses are for the hobbyist hacker when you have Teachable Machine, which is free and works in the browser with any webcam.

https://medium.com/analytics-vidhya/a-fun-project-pose-detector-with-googles-teachable-machine-6c7c8d650be1

Thank you for sharing the other approaches, it's great to know more about the ecosystem.

Best,

Erica.

[Ant -]

I did see a documentary once where someone trained a smart fridge to recognise what food was being put into and removed from it. Unfortunately the training process was quite involved so they only trained it to recognise a hot-dog or not a hot-dog. It did seem a bit of a failure until someone realised they could sell it to a social media platform to help them identify unsolicited "artistic" male photos...

Thinking about it that may not have been a documentary.

Ant

--
You received this message because you are subscribed to the Google Groups "London Hackspace" group.
To unsubscribe from this group and stop receiving emails from it, send an email to london-hack-sp...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/london-hack-space/7c23944c-548e-4b9f-bf74-2f926396d89bn%40googlegroups.com.

[Dragos M.]

Hi Jon,

nice purchase of the jetson nano!

I haven't been on this bandwagon for a couple of years, but in 2018 Jensen was actively pushing for nvidia to create their own cloud solutions for these tasks.

They've created the Nvidia GPU Cloud, which by the looks of it now is a container registry you can pull in an AWS EC2 / Google Cloud compute instance and train your models using their software. You will likely need a compute instance with a nvidia gpu attached so the software can run CUDA. https://developer.nvidia.com/blog/jump-start-ai-training-with-ngc-pretrained-models-on-premises-and-in-the-cloud/ [The demo uses a pre-trained model but I assume you can upload your own model instead]

Another resource being developed was DIGITS https://developer.nvidia.com/digits . 

I'm sure AWS/GCP make it much easier to work with models as a beginner, however if you've invested in a nano - might as well look at some of their proprietary software.

Frankly, after I left the company I have no clue what they are up to with their ML/AI stuff, aside from selling DGX-2 and Xavier systems :)

Hope any of this would be somewhat useful.

Thanks,

Dragos

To view this discussion on the web, visit https://groups.google.com/d/msgid/london-hack-space/CAHZn-VMmJEF2e%2BZrF1tAAcMd0vV7eVsm52-VDkwotHJ1zAUetA%40mail.gmail.com.

[Jon Russell]

Hi Erica.

Thanks. I'll have a look at Teachable Machines.

I also have a DeepLens someone has given me too. I just need to work out how to get it to boot ! :-)

It seems tightly integrated with AWS Services.

I like the Jetson better, as it allows me to understand exactly what's going on, rather than it be a black box.

Regards,

Jon.

[Jon Russell]

Hi Dragos,

Thanks ! That's really helpful. That was exactly what I was hoping for. My Google Fu is clearly off today.

As I am familiar with the nVidia libraries, its much easier for me to run the same code in the cloud, if possible, until I really understand how it all work.

Having a container in the cloud, that's the same as the container on the Jetson is perfect !

Thank you ... That's my weekend sorted :-)

Regards,

Jon.

[Jon Russell]

For those interested ... here's my first "Hello ML World" video.

I trained it to recognise cuddly toys, instead of real birds, so I don't have to stand in the garden while I learn :-)

https://www.dropbox.com/s/xbqvwzfl9uhbepu/recognised%20x3%20with%20feeder.mp4?dl=0

Regards,

Jon.