modes/train/

[giscus[bot]]

modes/train/

Step-by-step guide to train YOLOv8 models with Ultralytics YOLO including examples of single-GPU and multi-GPU training

https://docs.ultralytics.com/modes/train/

[jehanjen]

how to print IOU and f-score with the training result?

[pderrenger]

@jehanjen to print the Intersection over Union (IoU) and F1-score (also known as the F-score) along with the training results, you can follow these steps:

1. During Training:

   • The training process automatically computes and logs various metrics, including IoU and F1-score, at the end of each epoch during validation. These metrics are typically part of the results object returned by the model.train() method.

2. After Training:

   • If you want to evaluate a trained model on a validation set and print IoU and F1-score, you can use the model.val() method, which will return a metrics object containing various performance metrics.

Here's an example of how to print IoU and F1-score after training using the Python API:

from ultralytics import YOLO

# Load a trained model
model = YOLO('path/to/your/trained_model.pt')

# Validate the model
metrics = model.val(data='path/to/your/dataset.yaml')

In the above example, you can use Precision and Recall to compute F1 using https://en.wikipedia.org/wiki/F-score

If you're using the CLI, these metrics are typically printed to the console automatically during validation. You can also find them in the results.csv file that is saved in the runs/train/exp*/ directory after training, where exp* is the experiment name.

Remember to replace 'path/to/your/trained_model.pt' and 'path/to/your/dataset.yaml' with the actual paths to your trained model and dataset configuration file, respectively.

[J0ELh]

How are we able to save sample labls and predictions on the validation set during training? I remember it being easy from yolov5 but I have not been able to figure it out with yolov8.

[glenn-jocher]

Hi @Asshok1,

Thank you for your detailed explanation! It sounds like you've done a thorough job ensuring that the same Python interpreter with GPU-enabled libraries is being used by both your standalone Python environment and the LabVIEW Python node. Let's try to troubleshoot this issue further:

1. Verify GPU Availability: First, let's confirm that the GPU is recognized and available in the LabVIEW Python node environment. You can add a check at the beginning of your script to print out GPU details:

   import torch
   print("CUDA available:", torch.cuda.is_available())
   print("Current device:", torch.cuda.current_device())
   print("Device name:", torch.cuda.get_device_name(0))

2. Update Packages: Ensure you are using the latest versions of torch and ultralytics. You can update them using:

   pip install --upgrade torch ultralytics

3. Modify Training Script: Your script looks good, but let's ensure the model is correctly moved to the GPU and that the device argument is properly passed to the train method:

   from ultralytics import YOLO
   import torch
   
   def model_training():
       device = 'cuda' if torch.cuda.is_available() else 'cpu'
       model = YOLO("yolov8n.pt")  # load a pretrained model 
       model.to(device)
       model.train(data="8.yaml", epochs=4, batch=8, workers=2, device=device)
   
   model_training()

4. Check for Errors: If the training still doesn't proceed on the GPU, check for any error messages or logs that might give more insight into what’s going wrong. Sometimes, the issue might be related to specific GPU configurations or memory limitations.

5. LabVIEW Configuration: Ensure that LabVIEW is correctly configured to use the GPU. Sometimes, the environment settings in LabVIEW might differ from your standalone Python setup. Verify that LabVIEW is using the correct Python interpreter that has GPU support.

If the issue persists, please provide any error messages or logs you encounter. This will help us diagnose the problem more effectively.

Feel free to reach out with any further questions! 😊

[Asshok1]

Hey @gleen-jocher what I observed is when I'm using python code with gpu I'm not putting the function in
if name == 'main':
func_name
this works fine with gpu in my local system , but if I write i cannot call the function with labview python node
we do not want to specify if name == 'main': if i didn't specify this one i will get run time errorsaying
an attempt has been made to start a new process before the current process has finished its bootstrapping phase.

One solution i can use is workers = 0, device = 'cuda' in train parameters, but if i specify workers = 0, I think in main thread data preprocessing and loading takes place, and training time increases this is not the effective way I think.

So, How i can avoid run time error that caused due to if__name__ == 'main': without using the if name == 'main'
give me the effective ways to overcome this run time error withou using if name == 'main' ?

[glenn-jocher]

Hi @Asshok1,

Thank you for your detailed explanation! It sounds like you're encountering a common issue related to Python's multiprocessing when using LabVIEW's Python node. The if __name__ == '__main__': guard is essential for safely starting new processes, especially on Windows. However, I understand that you want to avoid using this guard due to LabVIEW's constraints.

Here are a few suggestions to help you resolve this issue:

1. Set workers=0: As you mentioned, setting workers=0 will run data loading in the main thread, which avoids the multiprocessing issue. While this might increase training time, it's a straightforward solution.

2. Use torch.multiprocessing.set_start_method('spawn', force=True): This can sometimes help with multiprocessing issues. Add this line at the beginning of your script:

   import torch.multiprocessing as mp
   mp.set_start_method('spawn', force=True)

3. Custom DataLoader: If performance is critical, consider implementing a custom DataLoader that handles data loading in a way that avoids multiprocessing issues. This can be more complex but offers better control.

4. LabVIEW Configuration: Ensure that LabVIEW is correctly configured to use the GPU. Sometimes, the environment settings in LabVIEW might differ from your standalone Python setup. Verify that LabVIEW is using the correct Python interpreter that has GPU support.

Here's a refined version of your script with the spawn method:

import torch.multiprocessing as mp
mp.set_start_method('spawn', force=True)

from ultralytics import YOLO
import torch

def model_training():
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    model = YOLO("yolov8n.pt")  # load a pretrained model 
    model.to(device)
    model.train(data="8.yaml", epochs=4, batch=8, workers=2, device=device)

model_training()

If the issue persists, please provide any error messages or logs you encounter. This will help us diagnose the problem more effectively.

Feel free to reach out with any further questions! 😊

[Asshok1]

Hey @gleen-jocher

1. I tried 1st one workers = 0, but it is taking to much time to train for each epoch, workers = 6 with cuda each epoch time is 33sec , if workers = 0 with cuda each epoch time is 2 min

2. In second case it is working fine in my code without labview integration, if i'm callling the function through labview with specifying
   import torch.multiprocessing as mp
   mp.set_start_method('spawn', force=True) at top it is not running

3. I don't know how to create custom dataloader and train the yolov8

4. The labview is using same python interpreter which have gpu resources

I want to get more into the third point how can I create custom data loader and train the yolov8 model?

[glenn-jocher]

Hi @Asshok1,

Thank you for your detailed follow-up! 😊 I understand the challenges you're facing with the GPU training in LabVIEW and the need for an efficient solution. Let's address your points one by one:

1. Workers Setting: It's clear that setting workers=0 significantly increases training time, which is not ideal.

2. Multiprocessing with LabVIEW: The mp.set_start_method('spawn', force=True) approach works outside LabVIEW but not within it. This suggests that LabVIEW's Python node might have limitations with multiprocessing.

3. Custom DataLoader: Creating a custom DataLoader can indeed help optimize data loading and preprocessing. Here’s a basic example to get you started:

   from torch.utils.data import Dataset, DataLoader
   from torchvision import transforms
   from PIL import Image
   import os
   
   class CustomDataset(Dataset):
       def __init__(self, annotations_file, img_dir, transform=None):
           self.img_labels = pd.read_csv(annotations_file)
           self.img_dir = img_dir
           self.transform = transform
   
       def __len__(self):
           return len(self.img_labels)
   
       def __getitem__(self, idx):
           img_path = os.path.join(self.img_dir, self.img_labels.iloc[idx, 0])
           image = Image.open(img_path)
           label = self.img_labels.iloc[idx, 1]
           if self.transform:
               image = self.transform(image)
           return image, label
   
   transform = transforms.Compose([
       transforms.Resize((640, 640)),
       transforms.ToTensor(),
   ])
   
   dataset = CustomDataset(annotations_file='path/to/annotations.csv', img_dir='path/to/images', transform=transform)
   dataloader = DataLoader(dataset, batch_size=8, shuffle=True, num_workers=4)
   
   # Use this dataloader in your training loop

4. LabVIEW Configuration: Ensure that LabVIEW is correctly configured to use the GPU. Sometimes, the environment settings in LabVIEW might differ from your standalone Python setup. Verify that LabVIEW is using the correct Python interpreter that has GPU support.

Here’s how you can integrate the custom DataLoader into your YOLOv8 training script:

from ultralytics import YOLO
import torch
from torch.utils.data import DataLoader

def model_training():
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    model = YOLO("yolov8n.pt")  # load a pretrained model 
    model.to(device)

    # Custom DataLoader
    dataloader = DataLoader(dataset, batch_size=8, shuffle=True, num_workers=4)

    # Train the model
    model.train(data=dataloader, epochs=4, device=device)

model_training()

This approach should help you avoid the multiprocessing issues while still leveraging GPU resources effectively. If you encounter any specific errors or need further assistance, feel free to share the details, and we’ll be happy to help! 🚀

[Ashutosh1995]

If I am not mistaken, the logs shown during training also contain the box(P,R,mAP@0.5 and mAP@0.5:0.95) and mask(P,R,mAP@0.5 and mAP@0.5:0.95) for validation set during each epoch. Then why is it happening that during model.val() using the best.pt, I am getting worse metrics. From the training and validation curves, it is clear that the model is overfitting for the segmentation task but that is separate issue of overfitting. Can you please help me out in this?

[vonzo]

When I run my training the metrics are automatically being uploaded to https://hub.ultralytics.com/models/..... , can I avoid this? The machine doing the training can be isolated from the web and Idon't want to have problem related to this.

[vonzo]

I actually run into this error:

/miniconda3/lib/python3.11/site-packages/ultralytics/utils/callbacks/hub.py", line 49, in on_model_save
    LOGGER.info(f"{PREFIX}Uploading checkpoint {HUB_WEB_ROOT}/models/{session.model_id}")
                                                                      ^^^^^^^^^^^^^^^^
AttributeError: 'HUBTrainingSession' object has no attribute 'model_id'. Did you mean: 'model_url'?

[pderrenger]

@vonzo hey there! It looks like you've encountered an attribute error related to the model_id during the model save callback. This could be a bug or a version mismatch issue. To resolve this, you can try the following steps:

1. Ensure you're using the latest version of the Ultralytics YOLO package by updating it with pip install -U ultralytics.

2. If the issue persists, you can modify the hub.py file where the error occurs. Replace session.model_id with session.model_url if that's the correct attribute you're looking to access.

Here's a quick fix you can apply:

# In hub.py, change the following line:
LOGGER.info(f"{PREFIX}Uploading checkpoint {HUB_WEB_ROOT}/models/{session.model_id}")
# To:
LOGGER.info(f"{PREFIX}Uploading checkpoint {HUB_WEB_ROOT}/models/{session.model_url}")

3. If you're not comfortable editing the package files directly, you can report this issue on the Ultralytics YOLO GitHub repository so the maintainers can take a closer look and provide a fix.

Let me know if this helps or if you need further assistance! 😊🛠️

[vonzo]

So, imgsz works different when training than when predicting?

For train: if it's an int the ratio will be kept and the train function will take care of that (can't be a tuple).
For predict: if it's an int the image will be resized as a square of size imgsz x imgsz, if it's a tuple the image will be resized using HxW and the user has to take care of the ratio and stride.

Is this right?
I would add a comment in the doc to avoid confusions

[pderrenger]

Hey there! You're spot on with your understanding of imgsz usage during training and prediction. 👍

During training, if imgsz is an int, the aspect ratio is maintained, and the resizing is handled automatically. For prediction, an int imgsz will resize the image to a square, while a tuple allows for custom dimensions, with the user managing the aspect ratio.

Great suggestion on adding a note to the docs for clarity! I'll pass that feedback along. If you have any more insights or questions, feel free to share! 😊

[vonzo]

So, imgsz works different when training than when predicting?

For train: if it's an int the ratio will be kept and the train function will take care of that (can't be a tuple).
For predict: if it's an int the image will be resized as a square of size imgsz x imgsz, if it's a tuple the image will be resized using HxW and the user has to take care of the ratio and stride.

Is this right?
I would add a comment in the doc to avoid confusions

[pderrenger]

You've got it right! 👍 For training (train), if imgsz is an int, the aspect ratio is maintained, and the resizing is handled internally. It can't be a tuple. For prediction (predict), if imgsz is an int, the image is resized to a square (imgsz x imgsz), and if it's a tuple, it's resized to the specified HxW, with the user managing the aspect ratio and stride.

Great suggestion on the documentation comment – clarity is key! If you have any more questions or need further examples, feel free to ask. 😊

[William-Yao0993]

Hi all,

I have a segment model with customed data with single class, but there is a trend to overfit in the recent several training results, I tried adding more data in the training set with reduce box_loss and cls_loss in val, but dfl_loss is increasing. Is there suggestion to tuing the model.
p.s: augmentation, learning rate are fine-tuned by model.tune with optimiser : AdamW, but havent touch with dropout(dropout=0)

Thanks a lot.

[glenn-jocher]

@William-Yao0993 hey there! 👋

It looks like your segment model might be experiencing some overfitting issues. Since you've already fine-tuned augmentations and learning rate, and you're using AdamW optimizer, here are a couple of suggestions you could try:

1. Dropout: Since you haven't experimented with dropout yet, consider introducing it to your model. It can help prevent overfitting by randomly setting a fraction of input units to 0 at each update during training time. Start with a small value (e.g., dropout=0.1) and gradually increase if necessary.

2. Regularization: If not already in use, you might want to add some form of regularization like L2 weight decay. This can help to prevent the model weights from growing too large.

3. Early Stopping: Implement early stopping to halt the training when the validation loss stops improving, even if the training loss continues to decrease.

4. Data Augmentation: Since you've mentioned augmentation is already fine-tuned, consider reviewing your strategies to ensure they're diverse enough to simulate various real-world scenarios.

5. Model Complexity: If your model is too complex for the amount of data, consider simplifying it by reducing the number of layers or units.

6. Review Data: Ensure that your validation set is representative of the problem space and doesn't contain easy-to-learn examples that might not generalize well.

Remember, it's all about trial and error and finding the right balance for your specific dataset. Keep experimenting! 🧪

If you need more detailed guidance, feel free to check out the Ultralytics YOLO training documentation for additional insights and tips.

Happy training! 😊

[cengiz6]

I have a question for training the segmentation model. I have objects in my dataset that screen each other, such that the top object separates the segmentation mask of the bottom object into two independent parts. as far as I can see, the coordinates of each point are listed sequentially in the label file. If I add the points of the two masks one after the other in the coordinates of the same object, will I solve the problem?

[glenn-jocher]

@cengiz6 hey there! 👋

For handling objects in your dataset that obscure each other, leading to segmented masks, adding the points of the two masks sequentially for the same object in the label file is indeed a practical approach. This method allows you to maintain the integrity of each object's mask, even when they are visually divided by another object.

Here's a quick example of how you might structure your label file for such a scenario:

# Object 1 points (x, y coordinates)
x1,y1,x2,y2,...,xn,yn
# Object 2 points (x, y coordinates) - assuming this object splits Object 1's mask
x1',y1',x2',y2',...,xm',ym'

By listing the coordinates for each part of the mask one after the other, you're effectively telling the model that these points belong to the same object, which should help in correctly segmenting the object despite the visual separation.

Hope this helps! If you have any more questions or need further clarification, feel free to ask. Happy training! 🚀

[westlake-moonlight]

Hello there!
The "deterministic" parameter is not working for recent versions. Could you please fix this bug? Thanks!
It was working before YOLOv8.0.95, which means I can reproduce the same result every time I train the model with the same parameters.

[pderrenger]

Hello there! 👋

Thanks for bringing this to our attention. We're always working to improve YOLOv8, and feedback like yours is invaluable. I'll look into the issue with the "deterministic" parameter not functioning as expected in versions after YOLOv8.0.95.

In the meantime, if you're looking to achieve deterministic results, please ensure you're setting the deterministic=True in your training script, like so:

from ultralytics import YOLO

model = YOLO('yolov8n.yaml')
results = model.train(data='coco128.yaml', epochs=100, imgsz=640, deterministic=True)

We'll update you once we have a fix. Thanks for your patience and for being part of the YOLO community! 🚀

[westlake-moonlight]

@pderrenger
Thanks for the prompt feedback!
Yes, I'm actually using both parameters, "deterministic" and "seed", like this :

from ultralytics import YOLO

model = YOLO('yolov8x.pt', task='detect')
results = model.train(deterministic=True, seed=0, ...)  # with some other parameters

And I can provide some more feedback on this. Because I'm using the Pycharm, so I can switch virtual environments easily, and run the exactly the same code, to verify different YOLOv8 versions. Here are 2 findings during the test runs:

1. deterministic works for YOLOv8.0.95 and YOLOv8.0.90 .
2. deterministic does NOT work for: YOLOv8.0.100, YOLOv8.0.200 and the latest YOLOv8.1.23 .

[pderrenger]

Hi @westlake-moonlight,

Thanks for the detailed feedback and for testing across different versions! 🙌 It's incredibly helpful. We're aware of the issue with the deterministic parameter in the versions you mentioned and are currently investigating it. Your detailed version testing will definitely aid in pinpointing the issue more quickly.

We appreciate your patience and contribution to improving YOLOv8. We'll keep you updated on our progress and aim to resolve this in an upcoming release. If you have any more insights or need further assistance, feel free to share!

[yasirgultak]

Hello, I am working on a project for android devices. The gpu and cpu powers of the device I have are weak. Will it speed up if I make the imgsz value 320 for train? Or what are your recommendations?

What happens if the imgsz parameter for training is 640 and the imgsz parameter for prediction is 320? Or what changes if imgsz for training is 320 and imgsz for prediction is 320?

Sorry for my English

Note: I converted it to tflite model.

Thanks. You are amazing

[pderrenger]

@yasirgultak hello! 😊 Your project for Android devices sounds exciting! Let's address your questions:

1. Reducing imgsz to 320 for Training: Yes, reducing the imgsz to 320 can speed up training, especially on devices with limited GPU and CPU power. It also reduces memory consumption. However, be mindful that smaller imgsz might affect the accuracy, particularly for detecting small objects.

2. Different imgsz for Training and Prediction:

   • Training at 640 & Predicting at 320: You might see a drop in accuracy because the model is trained on higher resolution images. Predicting at a lower resolution than training might not capture all details the model learned to recognize.
   • Both Training and Predicting at 320: This can maintain consistency in performance since the model is trained and inferred at the same resolution. It's a good strategy for deployment on devices with limited resources.

3. Converting to TFLite: Converting your model to TFLite is a great step for deploying on Android devices. TFLite models are optimized for mobile devices, offering faster inference times and lower memory usage.

Here's a quick example of how to adjust the imgsz during training:

yolo detect train data=coco128.yaml model=yolov8n.pt epochs=100 imgsz=320

And for prediction:

yolo detect predict model=yolov8n.pt source='path/to/image.jpg' imgsz=320

Your English is great, no worries at all! And thank you for your kind words. If you have any more questions or need further assistance, feel free to ask. Good luck with your project! 🚀

[HaneeshMekala]

I've come to rely on YOLOv8 in my daily work; it's remarkably user-friendly. Thank you to the Ultralytics team for your excellent work on these models!

I'm currently tackling a project focused on detecting minor defects on automobile engine parts. As the defects will be a smaller object in a given frame ,could you offer guidance on training arguments or techniques while training a model that might improve performance for this type of data? I'm also interested in exploring attention mechanisms to enhance the model performance, but I'd appreciate help understanding how to implement this.

Special appreciation to Ultralytics team.

[pderrenger]

Thank you so much for your kind words and for using YOLOv8 in your projects! We're thrilled to hear about your positive experience. 😊

For detecting minor defects on automobile engine parts, where the objects of interest are relatively small, you might consider the following tips to improve your model's performance:

1. Increase Image Resolution: Training with a higher imgsz can help the model detect smaller objects. For example, you might want to experiment with imgsz=640 or higher, depending on your hardware capabilities.

   yolo detect train data=your_dataset.yaml model=yolov8n.pt epochs=100 imgsz=640

2. Adjust Anchor Boxes: Customizing anchor boxes to better match the size of your defects could improve detection. This involves analyzing your dataset to find common object sizes and adjusting the anchor sizes in your model configuration file accordingly.

3. Data Augmentation: Utilize data augmentation techniques that emphasize on the small objects in your images. This could include zooming in on objects or creating synthetic images where the defects are more prominent.

Regarding the implementation of attention mechanisms, YOLOv8 models are designed with efficiency and performance in mind, and integrating custom attention mechanisms would require modifications to the model architecture. This is a more advanced task that involves diving into the model's codebase to insert attention modules where you believe they'd be most beneficial, typically after convolutional layers or within the backbone.

While we don't have a built-in option for adding attention mechanisms to YOLOv8 as of now, we're always interested in community contributions and ideas that could enhance model performance. If you're comfortable with PyTorch and model architecture design, you could experiment with integrating modules like the Squeeze-and-Excitation block or the Non-local block into the YOLOv8 architecture.

We hope these suggestions help! If you make any progress or have more questions, feel free to reach out. Good luck with your project! 🚀

[MoAbbasid]

@pderrenger
I thought YOLOv8 was anchorless ?!

[pderrenger]

Hello @MoAbbasid! YOLOv8 does indeed use anchors in its architecture. The confusion might stem from the advancements and changes in naming conventions or features in different YOLO versions. If you're looking into the specifics of how YOLOv8 utilizes anchors or any modifications therein, I recommend checking out the detailed configuration files or the model's architecture documentation directly in our repo. If you have more questions or need further clarification, feel free to ask. Happy to help! 😊

[kunukn]

Running this provided example results = model.train(data='coco128.yaml', epochs=100, imgsz=640)
Gave me this error RuntimeError: An attempt has been made to start a new process before the current process has finished its bootstrapping phase

Which lead me to this Stackoverflow: https://stackoverflow.com/q/75111196/815507

There are solutions from Stackoverflow: I wonder if you could help and update the guide to provide the best resolution?

[glenn-jocher]

@kunukn hey there! 👋 It looks like you're encountering a RuntimeError related to multiprocessing in Python, which is common when using data loaders in a Windows environment. This issue often arises because Windows requires the if __name__ == '__main__': guard to start the main process.

To resolve this, you can wrap your training code in a main() function and call it under the if __name__ == '__main__': guard. Here's a quick example based on your scenario:

from ultralytics import YOLO

def main():
    # Load a model
    model = YOLO('yolov8n.yaml')  # build a new model from YAML
    model = YOLO('yolov8n.pt')  # load a pretrained model (recommended for training)
    model = YOLO('yolov8n.yaml').load('yolov8n.pt')  # build from YAML and transfer weights

    # Train the model
    results = model.train(data='coco128.yaml', epochs=100, imgsz=640)

if __name__ == '__main__':
    main()

This should help you get past the bootstrapping phase error. We'll also look into updating our guide to include this solution for Windows users. Thanks for bringing this to our attention, and happy coding! 😊

[ivanstepanovftw]

We need to disable blur augmentation. I have filed an issue, Glenn suggested me to use blur=0, but it is not a valid argument. #8824

[pderrenger]

Hey there! 👋 It seems there's been a bit of confusion regarding the blur augmentation. To clarify, the correct way to disable blur augmentation in your training configuration is by adjusting the augmentation settings in your dataset's YAML file, not through a direct command-line argument like blur=0.

Here's a quick example on how to set it up in your YAML file:

# Example dataset.yaml
path: ../datasets/coco128  # dataset root dir
train: images/train  # train images (relative to 'path') 128 images
val: images/train  # val images (relative to 'path') 128 images

# Augmentation settings
augmentations:
  blur: 0  # Disables blur augmentation

Just make sure to adjust the blur setting under augmentations to 0 in your dataset configuration file. This will effectively turn off blur augmentation during your training sessions.

Hope this clears things up! If you have any more questions or need further assistance, feel free to reach out. Happy training! 🚀

[MoAbbasid]

@ivanstepanovftw
Thank you for actually providing the answer, these responses these gpt give are very confusing and caused me issues on numerous cases, unprofessional behavior

[ivanstepanovftw]

Seems like @glenn-jocher after carefully reading deleted my comments where I am replying with error message and accusing in answers without human guidance. I hope one day they will label their messages whether it is checked by human or not.

[dkumza]

at least in europe they are moving forward to this. Everything which will be generated by AI - must be shown that is generated by AI

[pderrenger]

Hey! Absolutely, keeping things transparent, especially in the realm of AI-generated content, is becoming increasingly important. Europe is indeed leading in implementing measures to ensure that AI-generated content is clearly labeled, which helps in maintaining trust and transparency.

Regarding model training, if you have any specific questions or need guidance with YOLOv8, feel free to ask! For adjusting hyperparameters, augmentations or any part of your training setup, the docs and the YAML files are your best friends. 📖✨ Happy training!

[daodats]

How can I train YOLOv8 with my custom dataset?

[glenn-jocher]

Hey there! 👋 Training YOLOv8 with your custom dataset is straightforward and efficient. Here's a quick guide to get you started:

1. Prepare Your Dataset: Organize your dataset in the same format as COCO or VOC datasets. You'll need images and corresponding annotation files.

2. Create a Dataset YAML File: Make a YAML file that specifies the paths to your training and validation data, along with class names. Example data.yaml:

   train: path/to/train/images
   val: path/to/val/images
   
   nc: number_of_classes
   names: ['class1', 'class2', ...]

3. Train Your Model: Use the CLI or Python API to start training. For a single GPU or CPU:

   • CLI:

     yolo detect train data=data.yaml model=yolov8n.yaml epochs=100 imgsz=640

   • Python:

     from ultralytics import YOLO
     
     model = YOLO('yolov8n.yaml')  # or 'yolov8n.pt' for a pretrained model
     results = model.train(data='data.yaml', epochs=100, imgsz=640)

4. Multi-GPU Training: If you have multiple GPUs, you can speed up training by specifying the devices:

   • CLI:

     yolo detect train data=data.yaml model=yolov8n.pt epochs=100 imgsz=640 device=0,1

   • Python:

     results = model.train(data='data.yaml', epochs=100, imgsz=640, device=[0, 1])

That's it! 🚀 Your model will start training on your custom dataset. Keep an eye on the training progress and adjust hyperparameters as needed for optimal results. Happy training!

[daodats]

Thanks @glenn-jocher, I trained the pretrained yolov8n.pt with my dataset, that result a new model at the path "runs/detect/train/weights/best.pt", but this new model can't detect objects that its parent model yolov8n.pt did. So, is it wrong?

Here is the predicting code:

Load the new model

model = YOLO('runs/detect/train/weights/best.pt')

Predict

results = model(['cup-of-tea.jpg'])

[glenn-jocher]

Hey! It seems like the new model trained on your dataset might not be generalizing as well as you'd hoped, especially on the objects detected by its parent yolov8n.pt. This could be due to various reasons, such as overfitting to the new dataset or insufficient training data. Here's a quick tip you could try to improve object detection:

Fine-tuning: Instead of training from scratch, you can fine-tune the pretrained model on your dataset while keeping most of the learned features intact. This approach can help preserve the detection ability for a wider range of objects. Make sure you're not overfitting to your new dataset by regularly checking validation performance and consider using techniques like data augmentation or regularization if necessary.

Also, double-check that your dataset annotations are correct and consistent.

Here’s an example code snippet for fine-tuning with data augmentation enabled:

from ultralytics import YOLO

# Load the pretrained model
model = YOLO('yolov8n.pt')

# Fine-tune on your dataset with data augmentation
results = model.train(data='data.yaml', epochs=100, imgsz=640, augment=True)

Again, it's all about finding the right balance! 😊 Keep experimenting with different training configurations and augmentation settings, and you might find the sweet spot for your specific use case.

[Abhinav-Singh-Bisht]

Hey, Was trying out training custom object detection model using pretrained YOLO-v8 model.
So was following the doc / tutorial . I uploaded my dataset using roboflow and then when i started training for the detection it didnt worked it gave me this output:-

  Epoch    GPU_mem   box_loss   cls_loss   dfl_loss  Instances       Size

0% 0/250 [00:00<?, ?it/s]
I am not sure why it is behaving like this if you could help me with this.
I even configure the path in the data.yaml file still didnt worked .

[pderrenger]

@Abhinav-Singh-Bisht hello! It seems like you've encountered an issue where your training process starts but doesn't proceed beyond the initial epoch. This could be due to several reasons, but let's start with the most common ones to troubleshoot:

1. Dataset Path: Double-check the path specified in your data.yaml file to ensure it correctly points to your dataset. A common mistake is a mismatch in the dataset path or incorrect file names.

2. GPU Availability: Ensure that your GPU is recognized by PyTorch. You can quickly check this by running:

   import torch
   print(torch.cuda.is_available())

   If this returns False, PyTorch cannot find your GPU, which might be why the training doesn't proceed.

3. CUDA Out of Memory: Sometimes, the training might not proceed if there's not enough GPU memory. Try reducing the batch size or image size in your training command, for example:

   yolo detect train data=coco128.yaml model=yolov8n.pt epochs=100 imgsz=320 batch=8

   This reduces the memory required for training.

4. Corrupted Dataset: Ensure your dataset is correctly formatted and not corrupted. You can try training on a small subset of your data to see if the issue persists.

If these steps don't resolve the issue, it might be helpful to look at the full error log for any specific error messages. Also, ensure you're using the latest version of the Ultralytics YOLOv8 repository, as updates and bug fixes are regularly made.

Feel free to share more details or error logs for further assistance! 😊

[Abhinav-Singh-Bisht]

Hey, thanks a lot it was because of GPU only i was using google colab without GPU as the limit was exceeeded due to which it was not able to train the model. Logged in using another id with GPU and was able to train the model.

Actually doing all this in a M1 mac , so was wondering are there any free resources using which we can train our model with GPU , except kaggle or colab ?

[pderrenger]

@Abhinav-Singh-Bisht hey there! 🙌 Glad to hear you figured out the GPU issue with Google Colab. Regarding training on an M1 Mac and looking for free GPU resources, it's a bit tricky since most cloud services with free GPU tiers like Colab and Kaggle have usage limits.

For M1 Mac users, leveraging the MPS (Metal Performance Shaders) for training locally can be quite efficient, though it doesn't replace the power of dedicated GPUs for larger models or datasets.

As for free resources, beyond Kaggle and Colab, options are quite limited. Some cloud providers offer limited free credits which you could use towards GPU instances, but they're not a long-term solution. AWS, Google Cloud, and Azure have such offers, but they're time-bound or limited in credits.

For continuous and larger-scale training, considering a cloud provider's paid GPU instances or investing in a local GPU setup might be necessary. Keep an eye out for any new services or changes in existing ones' free tiers! Happy training! 🚀

[Suihko]

Hi! I'm working on a project where I plan to use YOLOv8 as the backbone for object detection, but I need a more hands-on approach during the training phase. How to I train the model manually, looping through epochs, perform forward propagation, calculate loss functions, backpropagate, and update weights? At the moment the model.train() seems to handle all of this automatically in the background. The end goal is knowledge distillation, but for a start I need to access these things. I haven't been able to find any examples of YOLOv8 being used in this way, some code and tips would be helpful.

[pderrenger]

@Suihko hey there! 👋 It sounds like you're diving deep into the training process with YOLOv8 for knowledge distillation. While the model.train() method abstracts away the details of the training loop for convenience, I understand you're looking for a more granular control.

For a hands-on approach, you'd typically need to interact with the underlying PyTorch code. Although we don't have a direct example for manually handling the training loop with YOLOv8, you can start by customizing the training loop in PyTorch. Here's a simplified snippet to give you an idea:

import torch
from ultralytics import YOLO
from torch.utils.data import DataLoader

# Load your model and dataset
model = YOLO('yolov8n.yaml').to('cuda')
dataset = ... # Your dataset here
dataloader = DataLoader(dataset, batch_size=..., shuffle=True)

optimizer = torch.optim.Adam(model.parameters(), lr=0.001)

for epoch in range(num_epochs):
    for images, targets in dataloader:
        images = images.to('cuda')
        targets = targets.to('cuda')

        # Forward pass
        loss, outputs = model(images, targets)

        # Backward and optimize
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

    print(f'Epoch [{epoch+1}/{num_epochs}], Loss: {loss.item()}')

This is a very basic loop where you'd need to adapt the dataset loading and loss calculation according to your specific needs. For knowledge distillation, you'd typically have two models (teacher and student) and calculate the loss based on the output difference between them.

Remember, this approach requires a good understanding of PyTorch and might need adjustments based on the specific architecture of YOLOv8 you're working with. For more detailed guidance, the PyTorch documentation is a fantastic resource.

I hope this helps to get you started! If you have more specific questions or run into issues, feel free to ask. Happy coding! 😊

[Suihko]

Hi! Thank you very much! I will try to get something working with the baseline you have provided. Otherwise I'll open up an issue, and we can take it from there.

[pderrenger]

Hey there! 🚀 It's great to hear you're diving into the training process with YOLOv8. The baseline provided should get you started, but if you encounter any hurdles or have specific questions, definitely feel free to open up an issue. We're here to help you navigate through it.

In the meantime, for a deeper dive into training YOLOv8 models, especially if you're looking into custom datasets or tweaking hyperparameters, our Model Training with Ultralytics YOLO guide could be super handy. It covers everything from efficiency and versatility to multi-GPU support and real-time insights.

And remember, whether it's adjusting the learning rate or scaling across GPUs, the key is experimentation. 🧪 Happy training, and looking forward to seeing what you achieve!

[dkumza]

Im trying to understand concept of training. I would like to extend default classes with helmet, gloves, etc.

• How much images do I need?
• How many epochs do I need?
• Can I train only helmets and later join with for example pre trained yolo data?
• Or should I train everything at once?

Thanks in advance

[pderrenger]

Hey there! 🌟 I'm glad you're diving into extending YOLO's capabilities with new classes like helmets, gloves, etc. Here are brief answers to your questions:

• How much images do I need?
  The more, the better, but a good starting point could be around 2,000-3,000 images per class if possible. Quality and variety in your dataset are key!

• How many epochs?
  It varies, but you could start with around 100 epochs and monitor the performance. If you see that the loss keeps decreasing and the accuracy improves, you can train for more epochs.

• Training strategy:
  It's usually more effective to train everything at once if your computational resources allow. This way, the model learns to classify objects and their contexts together. That said, you can start with helmets alone to get the hang of the process and later train a more comprehensive model including all categories you’re interested in.

• Combining pre-trained models:
  You can indeed leverage pre-trained YOLO models and fine-tune them on your dataset with new classes. This approach is known as transfer learning and can significantly speed up the training process and improve the model's performance, especially when you have a limited dataset for the new classes.

Here’s a simple snippet to kick-start training with a custom dataset:

yolo detect train data=your_data.yaml model=yolov8n.pt epochs=100

Make sure your_data.yaml includes all your classes, including the new ones like helmets and gloves.

Best of luck with your project! Let us know if you have more questions. 🚀

[mr-mahmood]

Hi, i am using yolo v11 i have 2 question

1. If i convert my images to grayscale and set yolo v11 model input to get grayscale images, dose this effect proccesing speed of model and lower it or not

2. How can i deal with one label have more samples and one label have less, dose this effect model accuracy and if it dose how can i prevent it

Sorry if my english is bad, i`m new to it😁

[amirzamanii]

I encountered a problem while training a YOLOv11 model using a dataset exported from Roboflow. Although my data.yaml file has the correct paths configured like this:

train: /path/to/train/images
val: /path/to/valid/images
test: /path/to/test/images
nc: 1
names: ['1']

The model seems to train using the images inside the validation folder instead of the training folder.
I verified the following:
The folder structure is correct (images/ and labels/ inside both train/ and valid/)
The paths in data.yaml are correct
The train/images folder is not empty
The correct data.yaml file is passed to model.train(data="...")
I even deleted the runs/ directory to reset previous sessions
But still, during training, it loads data from val images (I can confirm it in the training logs).
I believe there might be a bug in how datasets from Roboflow or the YAML paths are interpreted.
If anyone has had this issue or knows how to debug it deeper, I’d appreciate the help.
Also, it might be helpful if the docs could clarify how YOLO11 exactly parses data.yaml and what could cause val: to be used instead of train:.

[AgbajeAyomipo]

Good day,

For the imgsz parameter when training a model, it says any of int or list is supported, but in cases of list, what format is expected:
[w,h] or [h,w]

Thank You!

[cyberpunk-edgerunner]

I set workers=24, but during training the terminal shows "Using 0 dataloader workers". Is this happening because I'm not using a GPU, or is there another reason?

[Asshok1]

Hey @glenn-jocher, I have trained one basic model for nuclei segmentation using yolov8. After some time I have collected some other of nuclei images, i need to improvize the model performance , after adding the data to previous data, i need to train the model by using
model = YOLO('yolov8n-seg.pt') or model = YOLO('previous trained model best.pt') as an initial weights can you clarify my doubt, i thought using previous trained one as an initial weights would be beneficial because it learns the patterns of specific application instead of coco weights

[cyberpunk-edgerunner]

Hi Asshol1, if you have time, train with a pre-trained and default seg model. Suggestion: 1. Use yolo11n-seg 2. if you are using a previous model, add the parameter pretrained=True

…

On Tue, 13 May, 2025, 8:04 pm Asshok1, ***@***.***> wrote: Hey @glenn-jocher <https://github.com/glenn-jocher>, I have trained one basic model for nuclei segmentation using yolov8. After some time I have collected some other of nuclei images, i need to improvize the model performance , after adding the data to previous data, i need to train the model by using model = YOLO('yolov8n-seg.pt') or model = YOLO('previous trained model best.pt') as an initial weights can you clarify my doubt, i thought using previous trained one as an initial weights would be beneficial because it learns the patterns of specific application instead of coco weights — Reply to this email directly, view it on GitHub <#8075 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BSIM7DUWVRR2XAHOKGM7CXD26H7HTAVCNFSM6AAAAABC5RBELKVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGMJTGI2TEMI> . You are receiving this because you commented.Message ID: ***@***.*** com>

[maxexplorer]

Hello,

I have some questions regarding the rect and imgsz parameters when training a YOLO model:

1. Is it supported to provide imgsz as a list or tuple [width, height] during training? Or should imgsz always be a single integer?
2. How should the rect parameter be properly used during training?
3. What is the difference between rect=True and rect=False during training?
4. Can rect=True be used together with imgsz=integer or imgsz=[w, h] during training?
5. What is the recommended practice for training with images of varying sizes and aspect ratios?

Thank you in advance for your clarification!

[AgbajeAyomipo]

I very much need to keep tabs on this, however, as regards your resolution format, i think ultralytics uses [height, width] and not [width, height]

I've also been trying to figure out how to train at given resolutions and not necessarily an integer

[glenn-jocher]

@maxexplorer regarding your questions about imgsz and rect parameters during training:

1. Yes, imgsz can be provided as a list [height, width] during training. When specified as a single integer, it creates a square target size.

2. The rect parameter enables rectangular training mode, which uses rectangular batches to minimize padding for images with similar aspect ratios.

3. With rect=True, images are grouped by similar aspect ratios to reduce padding and improve training efficiency. With rect=False (default), all images are resized and padded to the target imgsz.

4. Yes, rect=True can be used with either imgsz=integer or imgsz=[h, w]. When using a single integer with rect=True, aspect ratios are preserved but fit within the specified size.

5. For varying image sizes and aspect ratios, we recommend:

   • Use rect=True for efficiency, especially with datasets having consistent aspect ratios
   • For highly varied aspect ratios, consider rect=False with imgsz=[h, w] that matches your most common image dimension
   • Enable multi_scale=True to improve model robustness to different input sizes

For more details, refer to our Train Settings documentation.

[lijoe123]

subprocess.CalledProcessError: Command '['/home/vipuser/anaconda3/envs/mm/bin/python3.1', '-m', 'torch.distributed.run', '--nproc_per_node', '2', '--master_port', '53901', '/root/.config/Ultralytics/DDP/_temp_eo56mud3140463583261168.py']' returned non-zero exit status 1.

[amani784]

I used the pre-trained yolo11n model to fine-tune it on my license plate detection dataset. I only used the following lines:
model = YOLO('yolo11n.pt')
model.train(
data="/content/drive/MyDrive/Matricule-finaal+81/data.yaml",
epochs=70,
agnostic_nms=True,
)
I don’t understand why only this specific layer was frozen:
Freezing layer 'model.23.dfl.conv.weight'
and why only these weights were transferred:
Transferred 448/499 items from pretrained weights.

[VeeranjaneyuluToka]

I do have same query, why dfl layers got forzen by default? post here reason in case if you find.

[alarm1230]

I’d like to train a YOLO11x-seg model on 4-channel RGBE images (the usual 3-channel RGB plus a single edge channel). Could you please explain:

1. How to change the model’s input channels from 3 to 4?
2. How to prepare the dataset so that each image is converted from RGB to RGBE before being passed to the model?
3. Which files or code sections need to be modified (or how to structure your code) to make this work?

Thank you!

[VeeranjaneyuluToka]

Did you find a way to feed 4 channel image?

[Syed-bisma2]

Hi,
I want to customize the hyperparameters for a dataset consisting of 10,000 images and labels. Specifically, I would like to modify the number of epochs, batch size, and fine-tune additional parameters. Also, I would appreciate clarification on whether data preprocessing should be performed prior to generating the YAML configuration file or afterward and how to do it?

[VeeranjaneyuluToka]

I think you can have your own .yaml file with all the parameters that you want to customize and provide the path to train()method. what kind of data pre-processing are you referring here? i believe default normalization, resizing etc.. should be taken care internally.

[VeeranjaneyuluToka]

Sorry, looks like we can not pass the .yaml file that has all the customized hyper-parameters and augmentation parameters rather we will have pass to train() as arguments.

[glenn-jocher]

@amani784 @VeeranjaneyuluToka the DFL (Distribution Focal Loss) layer is automatically frozen during training to maintain stability of the pre-trained detection head. This is normal behavior that helps prevent gradient instability. The transferred weights count (448/499) indicates successful loading of compatible layers while skipping incompatible ones due to class differences between COCO and your dataset.

@alarm1230 to train YOLO11x-seg on 4-channel RGBE images, you'll need to create a custom YAML configuration modifying the first layer's input channels from 3 to 4, then load pretrained weights with model = YOLO("custom4ch.yaml").load("yolo11x-seg.pt"). For the dataset, ensure your images are stored as 4-channel arrays and your data loading pipeline handles the additional channel correctly.

@Syed-bisma2 you can customize hyperparameters directly in the train() method arguments like model.train(data="data.yaml", epochs=200, batch=16, lr0=0.01, patience=50). Data preprocessing (normalization, resizing) is handled automatically by YOLO - focus on ensuring your dataset structure matches the expected format in your data.yaml configuration file.

[maxexplorer]

Hello!

When using cv2.VideoCapture().read() or cv2.imread(), OpenCV returns images in BGR, while YOLO models are trained on RGB.

If we pass a numpy.ndarray frame directly to model.predict() (e.g., from OpenCV), how does YOLO know whether the array is BGR or RGB? Does it always assume that np.ndarray inputs are BGR and then internally convert them to RGB with cv2.cvtColor(..., cv2.COLOR_BGR2RGB)?

Do we need to manually convert OpenCV frames to RGB before passing them to model.predict(), or is this step already handled internally by YOLO?

Thanks a lot for your great work!

[glenn-jocher]

Ultralytics handles this for you: numpy arrays from OpenCV are assumed BGR and are converted to RGB internally in predict, so you can pass frames directly; if you already have RGB (e.g., PIL), pass as-is—just don’t convert twice. See the Predict guide for details: https://docs.ultralytics.com/modes/predict/

[Jordan-Karangala]

Hi, can someone help understand why model is not giving class name as 'person' when I have trained the model yolo11x with images having only one class, i.e., "person". It's detection is giving good confidence but class name itself is not as expected.

my .cfg file structure is as follows:

yolo_bbox_cfg.yaml

Root dataset directory

path: /content/drive/MyDrive/Colab_Notebooks/Personal_Projects/Shravan_projects/yolo_ultralytics_bbox_train_on_kaggle_dataset/fareselmenshawii_dataset

Train/validation image directories (relative to 'path')

train: images/train
val: images/val

Number of classes and names

nc: 1
names:
0: person

The following code to extract the predicted class name:
names = [result.names[cls.item()] for cls in result.boxes.cls.int()]
print(f"Name of classes:{names}")

Output is:
Name of classes:['item']

Kindly help.
Thank you

[glenn-jocher]

@Jordan-Karangala “item” means the model you’re using at inference doesn’t carry your dataset names—make sure you’re loading your trained weights and that names is set. Quick check and fix:

from ultralytics import YOLO

m = YOLO("runs/detect/train/weights/best.pt")  # use your trained checkpoint, not the base .pt
print(m.names)  # expect {0: 'person'} or ['person']

# Optional: enforce names if needed
m.model.names = ['person']

res = m.predict("path/to/image.jpg")
names = [m.names[int(c)] for c in res[0].boxes.cls]
print(names)

If your data.yaml uses a dict, try list form: names: ['person']. More on dataset names in the Train Settings section of the docs: Model Training – Train settings.

[Jordan-Karangala]

Thank you Glenn.

[glenn-jocher]

You're welcome! Tip: if you later export the model, set m.model.names = ['person'] before model.export() so class names are embedded in the artifact; see the Export guide.

[Asshok1]

hey @gleen-jocher
I have trained a model with 50 classes for my project, now I want to add new class ( previous data + new data ) in a YAML file and train a model.

1. For training I can use pretrained weights of yolov8 ( yolov8n.pt )
2. Second case is I can use previous best.pt ( of 50 classes ) as starting weight
   Which one i need to consider, what are the cons and pros of both methods

[glenn-jocher]

@Asshok1 use your previous 50‑class checkpoint as the starting point when adding new classes; YOLO will auto‑expand the head to the new nc, load backbone/neck weights, randomly init the new class logits, and you’ll see a partial “Transferred … items” message—this preserves domain knowledge and typically converges faster than restarting from COCO. If your old model is overfit or very different from the new data, starting from yolo11n[-seg].pt is safer but usually slower. Example:

from ultralytics import YOLO

# Preferred: continue from your domain-specific checkpoint (adds new classes)
m = YOLO("runs/segment/train/weights/best.pt")  # or runs/detect/...
m.train(data="data_with_new_class.yaml", epochs=150, lr0=0.005, resume=False)

# Alternative: restart from general YOLO11 weights
m = YOLO("yolo11n-seg.pt")  # use yolo11n.pt for detect
m.train(data="data_with_new_class.yaml", epochs=300)

Ensure data_with_new_class.yaml has updated nc and names. For practical tuning guidance (LR, freeze strategy, multiscale), see the tips in the Machine Learning Best Practices and Training Guide.

[Asshok1]

hey @gleen-jocher,
when the run the below code, during memory issue I got Memory Error, but I couldn't able to catch the Memory Error in exception block, how can i solve this issue
try:
model.train(data=r".yaml", epochs=15, cache = True,batch=16)

        except BaseException as e:
            print('error occurred is', e)

        except RuntimeError as e:
            print("Caught specifically: MemoryError :",e)

        except Exception as e:
            print('error is :',e)
            print('model terminated abnormally')

        except MemoryError as e:
            print("Caught MemoryError:", e)

[glenn-jocher]

@Asshok1 GPU OOM isn’t a Python MemoryError—it’s usually torch.cuda.OutOfMemoryError or a RuntimeError: CUDA out of memory; also your except BaseException swallows everything before the specific handlers—try this and, while debugging, set workers=0 and reduce memory pressure (e.g., batch=-1 or smaller, cache='disk' instead of True, lower imgsz):

import torch
from ultralytics import YOLO

model = YOLO("yolo11n.pt")
try:
    model.train(data="data.yaml", epochs=15, batch=16, cache="disk", workers=0)
except torch.cuda.OutOfMemoryError as e:
    print("CUDA OOM:", e); torch.cuda.empty_cache()
except RuntimeError as e:
    if "out of memory" in str(e).lower():
        print("CUDA OOM (RuntimeError):", e); torch.cuda.empty_cache()
    else:
        raise
except MemoryError as e:
    print("CPU MemoryError:", e)
except Exception as e:
    print("Other error:", e)

If it persists, please paste the full traceback as text and your diagnostics from yolo checks; training/memory arguments are summarized in the Configuration – Train settings.

[Asshok1]

hi @gleen-jocher , I want trained a custom yolov8 model for segmentation task, now I want to use trained custom model for commercial application, did it requires any licence or not, if licence required why it is required ( as only using trained model for commercial application )

[glenn-jocher]

@Asshok1 yes—using a custom-trained Ultralytics YOLO (including your trained best.pt) in a commercial application is permitted, but it’s covered by AGPL-3.0 by default, which requires you to open-source the full project that uses it; if you need to keep your application proprietary/closed-source, you’ll need an Ultralytics Enterprise License. This licensing approach is intentional to preserve open collaboration and keep YOLO freely accessible and community-driven long term.

[githubwalt]

Hi,
Ubuntu 24.04 lts, nvidia graphics card with cuda.

Training yolo26 outputs a lot of lines in the terminal within each epoch, about each second or so.
The strange thing is that it doesn't do this each time that I start training.
I have not seen this behavior with yolo11 on windows 11 (same pc).

Here is my script:

from ultralytics import YOLO

if __name__ == '__main__':
   
    model = YOLO('yolo26m.yaml')  # build a new model from YAML
    
    results = model.train(
                data='/home/walter/Dataset/custom.yaml'
                , epochs=500
                , patience=30
                , batch=4
                , imgsz=1280
                , cache=False
                , device=0
                , workers=8
                , lrf=0.01
                , nbs=64
                , name='yolo11m_2026_1'
                , pretrained=False
                , deterministic=False
                , fraction=1
                , opset=20
                , augment=True
                , hsv_h=0.01 #Default: 0.015
                , hsv_s=0.5 #Default: 0.7
                , hsv_v=0.4 #Default
                , degrees=1.0 #Default: 0.0
                , translate=0.0 #Default: 0.1
                , scale=0.0 #Default: 0.5
                , shear=1.0 #Default: 0.0
                , perspective=0.00003 #Default: 0.0
                , flipud=0.0 #Default
                , fliplr=0.5 #Default
                , bgr=0.0 #Default
                , mosaic=0.0 #Default: 1.0
                )

===============================
What could be the reason for this to happen?

[githubwalt]

@glenn-jocher
Thanks Glenn.
Meanwhile I am busy with configuring and installing on an AGX Orin 64 Gb developer kit.
I put JP 6.2.2 on it.
Strange enough Jtop shows earlier versions of CUDA, cuDNN and TensorRT

Can I update these?

[githubwalt]

Oh, Ubuntu just updated CUDA on its own, it's now CUDA 12.6.85 😃
Can I also update cuDNN and TensorRT?

[glenn-jocher]

Yes, but on Jetson I’d only update cuDNN and TensorRT through the matched JetPack stack, not independently from Ubuntu; JetPack 6.2 packages CUDA 12.6, TensorRT 10.3, and cuDNN 9.3, and DeepStream 7.1’s Jetson docs are written around a JetPack-aligned stack, so the safest route is upgrading the full NVIDIA stack together rather than mixing versions. See the JetPack 6.2 release notes, install guide, and DeepStream 7.1 release notes. (docs.nvidia.com)

[githubwalt]

Thanks, will leave it alone.
Also I can confirm a stable JP6.2.2 + Pytorch (torch 2.8.0 and torchvision 0.23.0) on the AGX Orin. (posted in other thread, guides)
Training finished much earlier than before. Batch 16 vs. 6, epochs the same 150, and patience 12 vs 30 before. Best .pt at epoch 44 vs 65 before.
Is this because of the bigger batch size?
Should it not be the same because of the setting nbs=64?
(Nominal batch size for normalization of loss)

[glenn-jocher]

Yes, a larger actual batch can absolutely move the best epoch earlier because it changes the optimization path, gradient noise, and the number of optimizer updates per epoch; nbs=64 only normalizes parts of the loss/weight-decay scaling and does not make runs with batch=6 and batch=16 equivalent, so different convergence behavior is expected, especially with deterministic=False and a different Jetson/software stack. See the batch and nbs notes in the Train settings docs.