yolov5/tutorials/hyperparameter_evolution/

[giscus[bot]]

yolov5/tutorials/hyperparameter_evolution/

Learn how to optimize YOLOv5 hyperparameters using genetic algorithms for improved training performance. Step-by-step instructions included.

https://docs.ultralytics.com/yolov5/tutorials/hyperparameter_evolution/

[alexchans]

The multi-gpu command does not work on my side, as the code I shown below in the slurm command file:

#!/bin/bash
#SBATCH -J yolov5_alex # job name
#SBATCH -A jiazhang_next_generation_ai_0001
#SBATCH -N 1 # request 1 node
#SBATCH -c 2 # request 2 CPUs
#SBATCH -G 4 # request 4 GPU
#SBATCH --time=48:00:00 # time limit
#SBATCH --mem=128G # request 128GB of memory
#SBATCH --mail-user=[email protected] # email for notifications
#SBATCH --mail-type=end # notify when job completes

module load conda
module load spack
conda activate yolov5

for i in 0 1 2 3; do
echo 'Starting GPU '$i'...' &&
nohup python train.py --epochs 50 --data ../crack_dataset/data.yaml --weights yolov5x.pt --cache --device $i --evolve > evolve_gpu_$i.log &
done

The output is just Starting GPU 1, 2, 3, 4 and then the job finishes within seconds, with no tuning being done, while the single GPU training command work perfectly, can someone point out what is wrong and what need to be fixed?

[alexchans]

I tested the --resume__evolve flag as follow: python train.py --epochs 50 --data ../crack_dataset/data.yaml --weights yolov5x.pt --cache --evolve --batch 64 --resume_evolve runs/evolve/exp4
But it give me an error: AssertionError: evolve population path is wrong!
So I change the command to python train.py --epochs 50 --data ../crack_dataset/data.yaml --weights yolov5x.pt --cache --evolve --batch 64 --resume_evolve runs/evolve/exp4/evolve_population.yaml, with evolve_population.yaml added in the end.
Now it is running, but the result is saved to exp5 instead of exp4, how should I fix this?

[glenn-jocher]

@alexchans the --resume_evolve option resumes evolution but saves results to a new experiment folder by design. To continue saving to the same folder (e.g., runs/evolve/exp4), use --resume_evolve runs/evolve/exp4 and ensure that folder exists and is writable. If issues persist, verify your folder structure and paths are correct. Let me know if you need further clarification!

[alexchans]

The file path is correct as shown in terminal:
alexchen@slogin-02:~/work/yolov5$ ls -l runs/evolve/exp4
total 3708
-rw-r--r-- 1 alexchen hpcusers 61992 Dec 10 01:10 evolve.csv
-rw-r--r-- 1 alexchen hpcusers 16599 Dec 9 06:31 evolve_population.yaml
-rw-r--r-- 1 alexchen hpcusers 827 Dec 10 01:10 hyp_evolve.yaml
-rw-r--r-- 1 alexchen hpcusers 219141 Dec 10 01:11 labels_correlogram.jpg
-rw-r--r-- 1 alexchen hpcusers 257477 Dec 10 01:11 labels.jpg
-rw-r--r-- 1 alexchen hpcusers 1191288 Dec 10 01:11 results.csv
-rw-r--r-- 1 alexchen hpcusers 432806 Dec 10 01:10 results.png
-rw-r--r-- 1 alexchen hpcusers 501247 Dec 10 01:11 train_batch0.jpg
-rw-r--r-- 1 alexchen hpcusers 576662 Dec 10 01:11 train_batch1.jpg
-rw-r--r-- 1 alexchen hpcusers 509394 Dec 10 01:11 train_batch2.jpg

Which I do not know why it is giving me the error: AssertionError: evolve population path is wrong!

[alexchans]

Another question would be on the next command, since now the newest results are saved in runs/evolve/exp5, should I use the evolve_population.yaml from exp4 (original) or exp5 (newest) to continue the experiment?

[glenn-jocher]

@alexchans you should use the evolve_population.yaml file from exp5 (the newest folder) to ensure continuity of the evolution process with the most updated population. This file includes the latest hyperparameter configurations based on the completed trials. Let me know if you encounter further challenges!

[UltralyticsAssistant]

👋 Hello, thank you for your interest in Ultralytics 🚀! The hyperparameter evolution tutorial is an excellent resource for exploring how to optimize YOLO models using genetic algorithms. We recommend checking the direct link you shared and also visiting the Docs for other tutorials and resources. New to YOLO or custom training? We suggest reviewing the Tips for Best Training Results to maximize your outcomes.

If this post is related to a 🐛 bug or issue, please provide a minimum reproducible example to help us investigate further. For any ❓ questions or customizations, please include detailed information such as your command, training logs, or configuration.

To stay connected with the community:

• Join live conversations on Discord 🎧
• Share insights and ask detailed questions on Discourse
• Engage with peers on our Subreddit

Upgrade

Ensure you are using the latest ultralytics version and all dependencies in a Python>=3.8 environment by running:

pip install -U ultralytics

Environments

YOLO can be tested and run in several verified environments, complete with pre-installed dependencies:

• 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 Ultralytics CI tests are currently passing. CI tests ensure all YOLO Modes and Tasks behave as expected across macOS, Windows, and Ubuntu.

This is an automated response to help guide you, but no worries—a member of the Ultralytics team will follow up to assist you further soon! 😊