yolo-v3-finetune

Overview

Finetuning YOLO V3 with the lvis dataset.

In order to extend the prediction vocabulary of the basic model pretrained with coco dataset. Based on the same set of images, the lvis dataset is based on contains 1203 annotated classes compared to just 80 classes in coco.

The underlying model and weights are from Torch Hub: https://pytorch.org/hub/ultralytics_yolov5/ uploaded by Ultralytics.

Apart from that, this repository contains my own implementation of a dataloader, loss function and evaluation/training code packaged with lightning.

Setup

Prequisites: install pyenv and install python version > 3.10

poetry install

TODO:

Train/evaluation

WIP

References

Here are various references I found helpful while working on the code :)

• YOLO papers 1, 2
• C4W3L09 YOLO Algorithm
• What’s new in YOLO v3?
• https://github.com/w86763777/pytorch-simple-yolov3