DEV Community

KinetiNode
KinetiNode

Posted on

Day 2 of making the EON-compiler open source alternative

I have decided to slow down my pace because i have to balanced WAY too many things while making this project.

For today i have decided to make a vision model.
totally from scratch no python libraries.

This of course is very hard especially for me whos still actively learning machine learning and how the EON compiler even works.

I won't be taking any courses for this i will study and learn concepts alongside using them in this project, this approach suits me the best and also helps me manage my time better.

Before i make this in pure C++ , i will first make it in python though because directly using C++ is like a baby trying to do calculus in his head. ( im scared of people who can)

why make a vision model in the first place? what does it have to do with my project? you may ask. heres how it fits.

I will do EVERYTHING by hand. so i understand the depth of what i'm making.

Heres roughly everything i will do:

  • write the training code (obviously , it's something i used python for ease since im beginner here)
  • Hand quantize the weights for ESP32. Int8 quantisation to be specific.
  • Write the C++ inference code by hand.
  • test it on real hardware.

unless i can do this i can't tell whether i will be able to make this project or not. And the reason im not using premade models for this is im still learning ML as i make this project, sounds insane right? Might sound dumb to do all this together but it just fits my style.

PROGRESS FOR DAY 2:

  • made the python training code.
  • trained the simple model with no actual data (using torch.rand)

may seem slow but im working on it.

Top comments (0)