YOLO trainer

Implements scripts used to train and validate yolo models

Requirements

• pytorch <= 2.0.1
• ultralytics <= 8.0.157

Prepare dataset

To train a model properly, you need a dataset that represents your application Make sure the dataset is split into three sets:

• train
• val
• test These splits are preferably split equally, so each split contains the equal ratios of classes and scenarios Splitting the dataset in three sets randomly can be done by running split_data.py

Furthermore, to enable the ultralytics library to find the dataset, a .yaml file must be created This file contains five keywords:

• path
• train
• val
• test
• names

path contains the path to the dataset The train, val, and test keywords point to the .txt files that contain the split information names contains a list that contains all names of the classes (eg. 0=person, 1=car, ...)

Train model

The train.py script runs the training program. Make sure to set the dataset parameter to the just created .yaml file. The pretrained_weights parameter determines what model is loaded:

• nano (yolov8n.pt)
• small (yolov8s.pt)
• medium (yolov8m.pt)
• large (yolov8l.pt)
• extra large (yolov8x.pt)
• pretrained supermarket model (SKU-110K_pretrained_yolov6_medium.pt)