imgz [w,h] and GPU memory requirement

[apiszcz]

Specifying a single value for imgz 640 suggests using 640x640 for the image size and a certain memory allocation is noted on the GPU.
Specifying imgz: [640,160] continues to use the same memory on the GPU and the single 640 setting.
Should I expect a lower GPU memory utilization with [640, 160] vs 640?
imgsz: [640, 160] # (int | list) input images size as int for train and val modes, or list[w,h] for predict and export modes

[UltralyticsAssistant]

👋 Hello @apiszcz, thank you for your interest in Ultralytics 🚀! This is an automated response to help guide you while an Ultralytics engineer will also assist soon.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us understand and debug the issue. This will allow us to see the problem in action and diagnose it more effectively.

If this is a custom training ❓ Question, please include as much information as possible, such as dataset image examples and training logs, and ensure you are following our guidelines for best training results.

Additionally, you can engage with the Ultralytics community where it suits you best. For real-time conversations, consider joining our Discord. For detailed questions and discussions, check out our Discourse. You can also share your insights and learn from others on our Subreddit.

Upgrade

To ensure the issue isn't resolved, make sure to upgrade to the latest ultralytics package in a suitable Python environment. This step often resolves many common issues:

pip install -U ultralytics

Environments

YOLO can be run effectively on various pre-configured environments, ensuring all dependencies are met, including CUDA, CuDNN, Python, and PyTorch.

Status

Our CI testing ensures the stable operation of all YOLO Modes and Tasks on macOS, Windows, and Ubuntu. Keep an eye on the status badge to confirm all system tests are passing.

Don't hesitate to ask further questions or provide additional details needed to assist you better! 😊

[glenn-jocher]

@apiszcz using imgsz: [640, 160] should reduce GPU memory usage compared to 640, as it specifies a smaller image size for one dimension. If you're not seeing a difference, ensure the model is configured correctly for prediction or export modes.