I’m learning Embedded ML, and experimenting with data collection and training.
Currently working on the project to recognize handwritten data using MNIST dataset and pre-trained model.
My project ID: https://studio.edgeimpulse.com/studio/319646
Unfortunately, I can’t make my model reliable due to the limitations in EI Community version:
For the custom model, I can apply only “Number of training cycles = 2”, for transfer learning only 1 cycle… So my model can’t work fine on real data and I have no way to improve it using EI platform.
I’ve tried to get the price (for instance for one month ) to remove the limitation, but it’s impossible as I have to provide the company @-mail, and the form doesn’t accept my private @-mail.
Is there any way to get prices for Hobbyists to use for non-commercial projects and studying?
We currently don’t provide a middle-tier pricing for Hobbyist and Students.
This is something we are evaluating but nothing has been defined yet.
In the meantime, I invite you to read that documentation page to try to stay within the community limitations: Lower compute time - Edge Impulse Documentation
My first suggestion if you use the MNIST dataset is to:
- Reduce your number of data samples for the training
- Reduce your image size in the create impulse tab.
Hi @louis ,
Thanks for the information.
I’ve tried increasing batch size and decreasing epoch, but it didn’t help me.
last chance to use less data.
Can I train my model locally, using export like in the picture?
Yes this is definitely possible
although I remember seeing an issue lately with ARM-based devices ([Fixed] "Edit block locally broken" on Arm (e.g. Macbooks, aarch64 Linux)) This has been fixed!
You can then import your model back on Edge Impulse using the BYOM feature (either through the Python SDK or manually).
Hi @louis ,
thanks for the great news.
I’ll try it.