How to solve the YOLO original model mAP50-95 results the more you run the better

[lroy615]

Search before asking

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

Question

Training the original YOLOv5s model results ran about three or four times, each time the results are better and better, precison/recall/mAP50 values are almost the same each time, so the comparison is to look at the value of mAP50-95, but the original model the value of this value is better and better, the first 100 rounds of training of the mAP50-95 results are 0.86 or so, the latest The first 100 rounds of training the mAP50-95 result is 0.86 or so, the latest run is 0.89 or so, but I want to reproduce the first result, because my improved model although the result is trained, precison/recall/mAP50 value is almost the same as the original, it is already 0.98 or so, but the mAP50-95 run is 0.87 or so, I would like to ask how to solve this problem, can you help I reproduce the first result.

Additional

No response

[glenn-jocher]

👋 Hello @lroy615, thank you for reaching out and for your interest in YOLOv5 🚀!

It seems you have a detailed question regarding training reproducibility and mAP50-95 results across multiple runs. To assist you better and debug this behavior effectively, we kindly ask that you provide the following:

• A minimum reproducible example (MRE). This includes your training command, hyperparameter modifications, and any changes you’ve made to the codebase, if applicable. Please also include details on the dataset you are using (even just general information, such as dataset size and structure).
• Logs or metrics from your training runs, if available, that highlight the observed behavior of mAP50-95 improving with each run.
• Your current random seed setting (if any), as randomization can impact training reproducibility.

If you're aiming to reproduce a specific past result, you may want to revisit the conditions of that run, including software versions, random seed settings (if manually set), and training environment configuration.

To get started with YOLOv5 or ensure you have the required setup, please follow these steps:

Requirements

Python>=3.8.0 with all requirements.txt installed, including PyTorch>=1.8. To set up YOLOv5:

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

Environments

YOLOv5 can be run in any of the following environments with 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

Tutorials

You may also find the following resources helpful:

• Custom Training Guide for ensuring optimal hyperparameters and setups for your use case.
• Model Training Tips to achieve the best training results.

Status

If this badge is green, all YOLOv5 GitHub Actions Continuous Integration (CI) tests are currently passing. CI tests verify the proper operation of YOLOv5 training, validation, inference, export, and benchmarks across macOS, Windows, and Ubuntu daily.

🌟 This is an automated response. Rest assured, an Ultralytics engineer will follow up to assist you further. We look forward to your updates and additional details!