GPU for Training

[adamroberge]

I am wondering what GPU specs were used for training the YOLOv11 model.

[UltralyticsAssistant]

👋 Hello @adamroberge, thank you for your interest in Ultralytics 🚀! We recommend a visit to the documentation 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 Ultralytics community where it suits you best. For real-time chat, head to Discord 🎧. Prefer in-depth discussions? Check out Discourse. Or dive into threads on our Subreddit to share knowledge with the community.

Upgrade

Upgrade to the latest ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8 to verify your issue is not already resolved in the latest version:

pip install -U ultralytics

Environments

YOLO 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: Gradient, Colab, Kaggle
• Google Cloud Deep Learning VM
• Amazon Deep Learning AMI
• Docker Image

Status

If the CI badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLO Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

This is an automated response. An Ultralytics engineer will also assist you soon. 😊

[glenn-jocher]

@adamroberge the YOLO11 model was trained using NVIDIA GPUs, typically A100s or V100s, for optimal performance. For more details, you can refer to the documentation.

[techsupportsardonyx]

1.i am using yolov8x-world.pt to train the model with what you have provided settings but it took 11 days to completed training for 100 epochs why? what is the proper way to train and get high accuracy in short time period ?
2.what are the ways to update or add classification to your yolov8x-world.pt model ?

[glenn-jocher]

Yes, the ASUS ROG Strix GeForce RTX 4090 is excellent for training YOLO models.

[techsupportsardonyx]

1.Can you guide me, how to prepare dataset for training because using annotation tool i need to know the exact co-ordinates crop size ?
2.Need to the maximum limit of dataset for each class?[ ex: person]
3.Approximately how long it take to train the yolo model in this graphics card[Asus ROG Strix GeForce RTX 4090 OC Edition 24 GB GDDR6X] Versus [NVIDIA A100] for 100 epoch and 150000 dataset?

[glenn-jocher]

1. For dataset preparation, ensure your annotations are in the correct format (e.g., YOLO format with class, x, y, width, height).
2. There's no strict limit, but more data generally improves performance.
3. Training time varies, but the RTX 4090 should be faster than the A100 for most tasks.

[techsupportsardonyx]

I fully referred to this link https://docs.ultralytics.com/modes/train/#resuming-interrupted-trainings, which contains 48 settings. Which of these settings will improve the model's training performance and accuracy? Could you also provide model training benchmarks? Additionally, will larger image size and file size cause delays in model training time?

[glenn-jocher]

To improve training performance and accuracy, focus on optimizing batch size, learning rate, and data augmentation settings. Larger image sizes can increase training time but may improve accuracy by capturing more detail. For benchmarks, please refer to the latest results on the Ultralytics GitHub repo.

[techsupportsardonyx]

In previous support you told training time can vary based on hardware and dataset size. We have 6GB GPU hardware. Why maximum 3% GPU and RAM & CPU also below 25% only utilized for this training process ? Does any issue exist in train code (we used your default source code with default settings) or any settings need to be changed in train code to increase training performance?

Below we have mentioned our hardware details and what we have tested using your train code and model.

1. Model: yolov8x-world.pt to train the customized model.
2. Epoch: 100
3. Batch size: 16
4. Number of Dataset: 129538.
5. PC configuration:

• OS: Windows 10 Enterprise 64-bit (10.0, Build 19045)
  (19041.vb_release.191206-1406)
• Processer: Intel(R) Core(TM) i7-10700F CPU @ 2.90GHz (16 CPUs), ~2.9GHz
• Core: 8 Core 16 Logical processor
• RAM: 32 GB
• Page File: 39960MB used, 6884MB available
• GPU: NVIDIA GeForce RTX 2060 (6 GB memory)

6. When we train the model maximum 3% GPU utilized and CPU utilized below 15% (Attachment given below for system resource utilization while train the model).

Note: If possible, please provide us the email/skype/whatsapp contact for quick communication.

[glenn-jocher]

To add a separate classification for vans, you'll need to label and train your model with images specifically annotated for vans. You can train with just your collected images if they are well-labeled, but using a combination of COCO and your images might improve accuracy. For optimal results, ensure your dataset is diverse and well-annotated.

[techsupportsardonyx]

In above 2 images 1st one is Truck and 2nd one is Van. Manually I can classify this. But why YOLO model wrongly classified both images are Truck?

[glenn-jocher]

@techsupportsardonyx the YOLO model might classify both as trucks due to similarities in features between vans and trucks in the training data. To improve classification, consider retraining the model with a dataset that includes distinct annotations for vans.

[techsupportsardonyx]

We can identify some distinct difference between both images. For example in van type we can see lot of passenger seats. In truck we can see the container shape. How do you told both have similar features?

[glenn-jocher]

The model might focus on external features like shape and size rather than internal details like seats, which can lead to misclassification.