Is there a problem with the way I fine-tuned the YOLOv5?

[CYH040306]

Search before asking

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

Question

Hello team,

Initially, I trained using YOLOv5s YAML on a large dataset I created, comprising approximately 15,500 images with a validation set of 2000 images. There were only two classes in the dataset, with a class ratio of approximately 1:3. After training, the precision (P) and recall (R) were around 90% and 87%, respectively.

Later, I switched to a new dataset, which had fewer images compared to the previous one but included different scenes. My intention was to fine-tune the model based on the previously trained weights, freezing the backbone and using the Adam optimizer with a lower learning rate. However, the results of subsequent training were not satisfactory, with P and R reaching only around 90% and 83%, respectively. My goal is to achieve around 95% precision without significantly increasing the number of parameters. How can I improve the performance under these constraints?

Thank you, team!

Additional

No response

[github-actions]

👋 Hello @CYH040306, 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

[CYH040306]

I made modifications to the network structure of YOLOv5s, such as introducing auxiliary head loss based on the v7 structure and replacing C3 with the ELAN structure. Currently, the parameter count is 14,145,982, and the GFLOPs is 32.0.

[glenn-jocher]

Hello!

Your modifications to the YOLOv5 architecture sound quite advanced! Modifying the C3 module with ELAN and incorporating auxiliary head loss is an interesting approach. When dealing with changes like these, it’s important to keep an eye on how these affect the overall balance between model complexity and inference speed.

If you're running into specific issues or not seeing the expected performance improvements, consider:

• Reviewing how the new architectural changes integrate with the existing YOLOv5 pipeline.
• Ensuring the additional complexity (auxiliary loss, ELAN structure) is effectively contributing to learning.
• Analyzing if the new components are properly trained, maybe further tweaking training parameters or adding more data could help.

Feel free to share more details if you encounter specific issues or have more results to share! 😊