Tinh chỉnh siêu tham số hiệu quả với Ray Tune và YOLO26

Yes — you can load best_hyperparameters.yaml and unpack it directly into model.train() instead of copying values by hand; see the Ray Tune integration guide and feel free to keep follow-ups here in Discussions so the Ultralytics team can see them.

from pathlib import Path
import yaml
from ultralytics import YOLO

model = YOLO("yolo26n.pt")

hyp_path = Path("runs/detect/tune/best_hyperparameters.yaml")
with hyp_path.open() as f:
    hyp = yaml.safe_load(f)

model.train(
    data=str(dataset_yaml),
    epochs=epochs,
    imgsz=imgsz,
    batch=batch,
    project=project,
    name=f"fold_{k}",
    workers=0,
    patience=100,
    exist_ok=True,
    cache=False,
    device=0,
    **hyp,
)

Disregard...completely misunderstood what tuning was actually doing.

Glad it’s sorted—quick tip: to tune your existing segmentation model, load your .pt first (model = YOLO("path/to/your.pt")) and then run model.tune(..., use_ray=True); Ray trials will reuse that same model/task—verify in the logs that task=segment.

@Fatima-khanjar Thank you for your detailed questions about Ray Tune integration with YOLO11!

1. Naming Trials: Currently, the high-level model.tune() API doesn't provide direct control over individual trial naming. Ray Tune automatically generates these names. You can use the name parameter to name the overall experiment (as you've done with "tuning"), but individual trial names remain auto-generated.

2. Retrieving Best Hyperparameters: When using model.tune() with Ray Tune, the best hyperparameters are automatically saved to runs/[task]/tune/best_hyperparameters.yaml. This file contains the optimal configuration and can be loaded for future use. You can also examine tune_results.csv for all trial results.

3. Using tune.Tuner(): While possible, creating a custom trainable function for YOLO with the low-level Ray Tune API requires more work. You'd need to wrap the YOLO training process in a function that configures and trains the model while reporting metrics to Ray Tune. For most users, the high-level model.tune() API is recommended as it handles this integration automatically.

For more advanced customization, you might consider extending the Ray Tune integration in Ultralytics by referring to the source code in the ultralytics/utils/tuner.py file.

Hello! Thank you for your clarification. I will continue using model.tune() and use the following line to retrieve the best hyperparameters, since the best_hyperparameters.yaml file wasn’t automatically saved in the output directory of the Ray tuning: best_config = analysis.get_best_config(metric="metrics/mAP50(B)", mode="max")`

I have a question: I read that Optuna uses the TPE (Tree-structured Parzen Estimator) search algorithm, which is more powerful than the default search algorithms in Ray Tune. Since Ray Tune provides scalability and distributed tuning, is it possible to combine both—using Optuna’s TPE algorithm with Ray Tune while still calling model.tune()? If so, how can I integrate them?

I also have another question that’s unrelated. I noticed that YOLOv11 comes in different versions, from n to xl. I understand that the choice depends on the dataset size and the task complexity. But what are the actual criteria used to determine the complexity of a task?

best regrads,

@Fatima-khanjar

Regarding using TPE with Ray Tune - yes, it's possible! Ray Tune already supports TPE through the HyperOpt integration. You can use it by installing the HyperOpt package and specifying the search algorithm:

pip install hyperopt

from ray.tune.search.hyperopt import HyperOptSearch
from ray.tune.search import ConcurrencyLimiter

hyperopt_search = HyperOptSearch(metric="metrics/mAP50(B)", mode="max")
limited_search = ConcurrencyLimiter(hyperopt_search, max_concurrent=4)

# Then in your model.tune() call
model.tune(
    data="data.yaml",
    search_alg=limited_search,
    use_ray=True,
    # other parameters
)

For your question about YOLO11 model size selection, the criteria typically include:

• Computational resources: Larger models (l, x, xl) require more GPU memory and processing power
• Dataset complexity: More complex datasets with small/numerous objects benefit from larger models
• Inference speed requirements: Smaller models (n, s) offer faster inference for real-time applications
• Accuracy needs: Larger models generally provide higher accuracy at the cost of speed

You can check the performance trade-offs in the Ultralytics documentation to make an informed decision based on your specific constraints.

Hello, I hope this message finds you well.

I followed your instructions to integrate Ray Tune with the TPE algorithm using HyperOpt:

hyperopt_search = HyperOptSearch(metric="metrics/mAP50(B)", mode="max")
limited_search = ConcurrencyLimiter(hyperopt_search, max_concurrent=1)

Then, I used the following code to tune hyperparameters with Ray Tune:

tuning_results = model.tune(
    data="data.yaml",
    epochs=50,                # Number of epochs for each tuning iteration
    iterations=100,           # Total number of tuning iterations
    batch=16,
    device=0,
    workers=0,
    grace_period=30,
    search_alg=limited_search,
    use_ray=True,
    gpu_per_trial=1,
    project=tune_dir
)

However, I encountered a syntax error: SyntaxError: 'search_alg' is not a valid YOLO argument. Could you please advise me on how to fix this?

I have another question regarding the output directory of Ray Tune. It shows that inside the runs folder, there are subfolders named train (equal to the number of iterations) and a tune folder that contains subdirectories also equal to the number of iterations.

My question is: should I focus on the train directories? Why are they generated, and what's their purpose?

Also, if I want to monitor the overall progress of the fine-tuning process — for example, in genetic tuning, we usually look at a graph of the fitness score vs. iteration — where should I look for that kind of visualization or information?

best regards,

@Fatima-khanjar

Regarding your issues with HyperOptSearch and Ray Tune integration:

1. The search_alg argument is passed through model.tune(use_ray=True). It defaults to "optuna" (TPE) and accepts either a string (such as "optuna", "hyperopt", "bohb" or "random") or a Ray Tune searcher instance like your HyperOptSearch, so you do not need to modify ultralytics/utils/tuner.py.

2. About the directory structure:

   • The train directories (train1, train2, etc.) contain individual training runs for each trial with different hyperparameter sets
   • The tune directory contains the aggregated results of all trials and should be your primary focus

3. For monitoring progress:

   • Check tune/tune_results.csv for all trial results
   • Look for tune/best_fitness.png which shows fitness score progression over iterations
   • tune/best_hyperparameters.yaml contains the optimal configuration found

The plots in the tune directory (especially tune_scatter_plots.png if available) provide valuable visualizations of the hyperparameter search space and how different parameters affected model performance.

Hello Community!!

Just following up on my own post — I managed to solve the issue on Windows 11.

Here’s what worked for me:

1. Opened gpedit.msc as Administrator
2. Navigated to: Computer Configuration > Administrative Templates > System > Filesystem
3. Enabled the setting “Enable Win32 long paths"

After doing this, the problem was resolved.

Posting this in case it helps someone else!

@Fatima-khanjar Thank you for sharing your solution to the Windows path length issue! This is helpful information that could benefit other users experiencing similar problems.

Regarding your question about Ray Tune versus Genetic Algorithms: They're different approaches to hyperparameter optimization. Ray Tune offers several advantages including:

1. Multiple search strategies beyond genetic algorithms (Bayesian optimization, grid search, random search)
2. Built-in parallel execution that efficiently utilizes available resources
3. Early stopping via ASHA scheduler to terminate underperforming trials
4. Distributed computing capabilities for scaling across multiple machines

Ultralytics actually supports both approaches - the internal tuner uses genetic algorithms focusing on mutation, while the Ray Tune integration provides more advanced search strategies and better parallelization. Ray Tune is generally recommended for more complex search spaces and when you have multiple GPUs available.

tambien necesito:

model.add_callback("on_train_start", self.on_train_start) model.add_callback("on_train_end", self.on_train_end)

pero dichas funciones no se disparan

agradezco mucho sus valiosas aportaciones, de verdad que es una funcionalidad increible y magnifica

@wisrovi to integrate MLFlow with Ultralytics YOLO hyperparameter tuning, you'll need to create a custom Ray Tune reporter that logs metrics to MLFlow. While we don't have native MLFlow integration like Weights & Biases, you can implement it by adding MLFlow logging callbacks to your tuning script. The Ray Tune documentation provides guidance on MLFlow integration that you can adapt for YOLO workflows.

@NegroAmigo yes, YOLOv8 fully supports Ray Tune integration for hyperparameter tuning. You can use the same model.tune() method shown in our Ray Tune integration guide by simply loading your YOLOv8 model (YOLO('yolov8n.pt')). The tuning process works identically across supported YOLO versions. For implementation details, see the Ultralytics Ray Tune documentation.

@alifazeli00 the documentation currently focuses on the default hyperparameters and their tuning ranges. If you're looking for specific details on beta2 for AdamW, it might not be explicitly covered. For more information on hyperparameter tuning, you can explore the Ray Tune documentation linked in the guide.

thank you, It has been mentioned that momentum is essentially beta1. Is this the same for beta2?

Beta2 is not explicitly mentioned in the documentation as it typically defaults to 0.999 in AdamW, which is a common standard setting. For specific adjustments, you may need to modify the optimizer settings directly in your training script.

@sirtris currently, there are no plans to integrate the .tune() function with MLflow, but we appreciate your feedback and will consider it for future updates.

Could you point me to some resources so I could maybe look into it doing it myself?

For resources on integrating MLflow with Ray Tune, I recommend checking the Ray Tune documentation for any updates or community contributions. This might provide insights into logging metrics and managing runs with MLflow.

@sachmatkris if you're seeing identical results for different learning rates, it might be worth checking the hyperparameter search space and ensuring the dataset configuration is correct. Adjusting these settings could help resolve the issue. For further guidance, please refer to the Ray Tune documentation.

@sachmatkris it seems like you've identified the issue with the device setting and path specification. If you need further assistance, feel free to ask!

Another issue I have is using the following call all of my experiments with ray_tune return the same metrics:

search_space = { 'lr0': tune.loguniform(1e-5, 1e-1), 'momentum': tune.uniform(0.6, 0.95), }

result_grid = model.tune(data = 'C:/Users/Kristijonas/Desktop/Mokslo šaknys karčios/AI/Projects/masinu_numeriai/formated_dataset_new/data.yaml', imgsz = 1280, space = search_space, batch = 16, epochs = 50, patience = 5, use_ray = True, iterations = 20, device='cuda' )

e.g.

@sachmatkris it seems like the issue might be with the search space configuration. Ensure that the hyperparameters have enough variability and that the dataset path is correctly set. If the problem persists, consider adjusting the search space or reviewing the dataset configuration. For more details, you can refer to the Ray Tune documentation.