yolov5/tutorials/model_ensembling/

[giscus[bot]]

yolov5/tutorials/model_ensembling/

Learn how to ensemble YOLOv5 models for improved mAP and Recall! Clone the repo, install requirements, and start testing and inference.

https://docs.ultralytics.com/yolov5/tutorials/model_ensembling/

[TanWeeKang]

Hi, is there a way to ensemble models using python instead of CLI? I am trying to test if the ensembled models would perform better or the current ones of individual will do.

[TanWeeKang]

Hi, when i run it says list has no attribute boxes on the combined_boxes line

[PhucPhamSy]

Dear author, is there a way to ensemble 2 models yolo9e-seg and yolo9c-seg, then visualize only ensemble's result?

[glenn-jocher]

To ensemble YOLO9e-seg and YOLO9c-seg models using a Python script, you can load both models and aggregate their predictions. Here's a concise example:

from ultralytics import YOLO

# Load the models
model1 = YOLO("yolo9e-seg.pt")
model2 = YOLO("yolo9c-seg.pt")

# Run inference on an image
results1 = model1("path/to/image.jpg")
results2 = model2("path/to/image.jpg")

# Combine results (e.g., averaging bounding boxes, masks)
ensemble_results = (results1 + results2) / 2

# Visualize the ensemble results
ensemble_results.show()

This script loads both models, runs inference, combines the results, and visualizes the ensemble output. For more details, refer to the Model Ensembling guide.

[PhucPhamSy]

I got the error like this, it seems like can't combination because not same data type.
ensemble_results = (result1 + result2) / 2
TypeError: unsupported operand type(s) for /: 'list' and 'int'

[glenn-jocher]

@PhucPhamSy the error you're encountering is due to the fact that results1 and results2 are lists, and you cannot directly perform arithmetic operations on lists. To ensemble the models correctly, you need to aggregate their predictions in a way that combines the results appropriately. You can achieve this by processing the results from each model and then combining them. Ensure you are using the latest versions of the models and the Ultralytics package to avoid any compatibility issues. For detailed guidance on model ensembling, please refer to the Model Ensembling guide.

[PhucPhamSy]

thank you! I try to use detect.py in the Model Ensembling guide. And got this error like this
Traceback (most recent call last): File "/raid/data/phuc/segmentation/SDI-yolo/yolov5/detect.py", line 425, in <module> main(opt) File "/raid/data/phuc/segmentation/SDI-yolo/yolov5/detect.py", line 420, in main run(**vars(opt)) File "/home/ubuntu/anaconda3/envs/yolo5/lib/python3.9/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context return func(*args, **kwargs) File "/raid/data/phuc/segmentation/SDI-yolo/yolov5/detect.py", line 250, in run s += f"{n} {names[int(c)]}{'s' * (n > 1)}, " # add to string KeyError: 230

[glenn-jocher]

It looks like you're encountering a KeyError related to the class index during the ensembling process. This issue often arises when the class indices in the models do not match. Please ensure that both models are trained on the same dataset with identical class indices. Additionally, verify that you are using the latest versions of the Ultralytics package and the models. If the issue persists, try running the script with a single model to isolate the problem. For further details, you can refer to the Model Ensembling guide.

[SarvananFreeStone]

Hi, in python how do we ensemble a yolo object detection with a following yolo image classification, any module created with optimized/parallel computing

[glenn-jocher]

@SarvananFreeStone to ensemble a YOLO object detection model with a YOLO image classification model, you can run them sequentially in your Python script. First, perform object detection to identify regions of interest, then pass these regions to the classification model. While there's no specific module for optimized parallel computing in this context, you can use Python's multiprocessing or threading for efficiency. For more on model ensembling, you might find the YOLOv5 documentation helpful.