Guide on how to utilize 'Weighted Loss' method in yolov5 custom training

[dilwolf]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

I am trying to detect football players, referees and ball on videos of mounted camera in football matches. From the nature of football analysis, I have an imbalanced dataset, overrepresented for player class and under-representation for class referee and ball, specifically for ball class, which is highly underrepresented. I am sure that my labels are accurate.

Below is the method I tried but did not help:

• manipulated ball and referee class instances in big part of dataset by excluding player class (because I have enough player class instances). However, this decreased model performance (usually not recommended by experts. I found it out later).

Now, I am going to try the Weighted Loss method but I am not sure how to do it. I also found this for yolov8 but it was not clear. So, it would be great, if you give me a detailed guide on how to modify code. I am new.

FIY: My dataset class instance distribution 459:80:14 in data.yaml file order

Thank you so much !!!

Additional

No response

[github-actions]

👋 Hello @dilwolf, thank you for your interest in YOLOv5 🚀! Please visit our ⭐️ Tutorials to get started, where you can find quickstart guides for simple tasks like Custom Data Training all the way to advanced concepts like Hyperparameter Evolution.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug it.

If this is a custom training ❓ Question, please provide as much information as possible, including dataset image examples and training logs, and verify you are following our Tips for Best Training Results.

Requirements

Python>=3.8.0 with all requirements.txt installed including PyTorch>=1.8. To get started:

git clone https://github.com/ultralytics/yolov5  # clone
cd yolov5
pip install -r requirements.txt  # install

Environments

YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all YOLOv5 GitHub Actions Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training, validation, inference, export and benchmarks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

Introducing YOLOv8 🚀

We're excited to announce the launch of our latest state-of-the-art (SOTA) object detection model for 2023 - YOLOv8 🚀!

Designed to be fast, accurate, and easy to use, YOLOv8 is an ideal choice for a wide range of object detection, image segmentation and image classification tasks. With YOLOv8, you'll be able to quickly and accurately detect objects in real-time, streamline your workflows, and achieve new levels of accuracy in your projects.

Check out our YOLOv8 Docs for details and get started with:

pip install ultralytics

[glenn-jocher]

Hello! 😊 Great choice evolving your approach to tackle class imbalance using the Weighted Loss method in YOLOv5.

To implement weighted loss during your custom training with YOLOv5, you'll need to adjust your training configuration to account for the variance in class representation.

Here's a concise way to do it:

1. Modify your data.yaml file by adding a class_weights field. You want to give higher weights to the underrepresented classes. Given your class instance distribution 459:80:14, you might consider a weighting like [0.5, 2, 10] reflecting the inverse proportion of the instances.

   # data.yaml
   train: path/to/train/dataset
   val: path/to/val/dataset
   
   nc: 3
   names: ['player', 'referee', 'ball']
   class_weights: [0.5, 2, 10]

2. Adjust the loss computation in your training script. Unfortunately, out-of-the-box YOLOv5 doesn't directly support weighted loss modification via the configuration file, so you'll need to manually tweak the loss computation within your training code, specifically when calculating the classification loss.

Understanding and making these changes can be somewhat involved, especially if you're new to the codebase. If you need a deeper dive into implementing these changes, I highly recommend checking our detailed methodological guides and model training strategies in the YOLOv5 documentation here: https://docs.ultralytics.com/yolov5/

Good luck with your model training, and feel free to reach out with more questions as you progress! 🌟