number of Bbox support limit in yolov8!

[MehrnazFani]

Search before asking

• I have searched the YOLOv8 issues and found no similar bug report.

YOLOv8 Component

No response

Bug

With training yolov8s_obb, I get cuda out of memory, although my batch_size=1. It happens because I have training images with more than 100 overlapping bboxes in them. If I remove those images and their corresponding bboxs from my dataset, the issue will be resolved. Why this is happning?

Environment

No response

Minimal Reproducible Example

No response

Additional

No response

Are you willing to submit a PR?

• Yes I'd like to help by submitting a PR!

[github-actions]

👋 Hello @MehrnazFani, thank you for your interest in Ultralytics YOLOv8 🚀! We recommend a visit to the Docs for new users where you can find many Python and CLI usage examples and where many of the most common questions may already be answered.

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.

Join the vibrant Ultralytics Discord 🎧 community for real-time conversations and collaborations. This platform offers a perfect space to inquire, showcase your work, and connect with fellow Ultralytics users.

Install

Pip install the ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8.

pip install ultralytics

Environments

YOLOv8 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 Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLOv8 Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

[glenn-jocher]

@MehrnazFani hi there! It looks like you're encountering a CUDA out of memory error due to handling many overlapping bounding boxes in your images. Here’s a quick rundown of why this might be happening:

1. High GPU Memory Usage: Each bounding box computation requires a certain amount of GPU memory. Having a lot of overlapping bounding boxes can increase the memory requirement significantly, even if your batch size is set to 1.

2. Complexity of Bboxes: Overlapping bounding boxes can lead to complex loss calculations and more memory being allocated to manage the overlaps during training.

To resolve this:

• Reduce the complexity or number of bounding boxes if possible, as you've already noticed improvements by doing so.
• Increase your GPU memory, if upgrading hardware is an option, to accommodate larger datasets with more complex annotations.
• Optimize memory usage: Try using half precision (float16) during training to reduce memory consumption, which can be done by setting half=True in your training command if the YOLOv8 implementation supports it.

Here's how you might use half precision:

model.train(data='data.yaml', imgsz=640, batch=1, epochs=100, half=True)

Good luck with your further training, and thanks for offering to help with a PR! 😊

[MehrnazFani]

Hi Glenn, Thank you very much for your thorough explanation. Best, Mehrnaz

…

On Thu, May 16, 2024 at 4:01 AM Glenn Jocher ***@***.***> wrote: @MehrnazFani <https://github.com/MehrnazFani> hi there! It looks like you're encountering a CUDA out of memory error due to handling many overlapping bounding boxes in your images. Here’s a quick rundown of why this might be happening: 1. *High GPU Memory Usage*: Each bounding box computation requires a certain amount of GPU memory. Having a lot of overlapping bounding boxes can increase the memory requirement significantly, even if your batch size is set to 1. 2. *Complexity of Bboxes*: Overlapping bounding boxes can lead to complex loss calculations and more memory being allocated to manage the overlaps during training. To resolve this: - *Reduce the complexity or number of bounding boxes* if possible, as you've already noticed improvements by doing so. - *Increase your GPU memory*, if upgrading hardware is an option, to accommodate larger datasets with more complex annotations. - *Optimize memory usage*: Try using half precision (float16) during training to reduce memory consumption, which can be done by setting half=True in your training command if the YOLOv8 implementation supports it. Here's how you might use half precision: model.train(data='data.yaml', imgsz=640, batch=1, epochs=100, half=True) Good luck with your further training, and thanks for offering to help with a PR! 😊 — Reply to this email directly, view it on GitHub <#12711 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AF75TQBNHFZMKHLNIJVOLKLZCRRWNAVCNFSM6AAAAABHY345XOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMJUGM2DQOJXGQ> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[MehrnazFani]

I have access to more powerful hardware. That solved my problem for now. half=True didnot help!

[glenn-jocher]

@MehrnazFani hi Mehrnaz,

Great to hear that upgrading your hardware resolved the issue! It's useful to know that half=True didn't make a difference in your case. This feedback helps us understand the performance under different setups. If you encounter any more questions or need further assistance, feel free to reach out.