alvises

Train a license-plate detector from scratch with YOLOX object detection models, on a public ~7k image dataset. In this video I start from a clean machine, install everything, configure YOLOX, fine-tune on plates, debug the usual setup hiccups, export to ONNX, evaluate real metrics (mAP/AP50-95) and finally run the model on dashcam footage.

That’s great! I’m going to support it then :)

It’s not built around ultralytics model and it doesn’t use its code, just del model and weights as a demo. The library is meant to be extensible and to run custom yolo models soon!

Yeah, v11 has been released recently.

Totally agree about ONNX, without it I wouldn't probably even started making this library. However, something I want to try is to code the network with Nx/Axon, without using onnx. The main advantage I see moving from onnx to Axon is to take advantage of all the new cool stuff like Nx sharding: https://dockyard.com/blog/2024/11/06/2024/nx-sharding-update...

The main difference is that this library is built for Elixir, which means you can take advantage of the Elixir real-time and distributed capabilities to easily integrate YOLO into scalable, fault-tolerant systems.

Thanks! About erlport/export I just didn’t want to use a library for something that is relatively simple.

In this step-by-step tutorial we see in depth how to build a Phoenix app from scratch, letting users upload their files using a multipart form. We learn how to easily run PostgreSQL with Docker, to receive uploads using Plug, and how to use Ecto to store upload details into the database.

In this article we see what a hash function is and what it can be useful for. We see how to calculate the hash of strings and files using the Erlang :crypto module and Elixir streams.

This is a second part of the article of last week: https://www.poeticoding.com/elixir-streams-to-process-large-...

This time we are going to process on the fly a 30 million lines remote file, summing the numbers of each line. We will see how to implement the functions we need to convert a stream of chunks (the one built in the part 1) to a stream of lines. We'll then run a benchmark of two different implementations.

We will then see how easy and quick is to process just the first 30 lines of the same 125mb remote file.