Update outputs

[Laughing-q]

@AyushExel @glenn-jocher

• make the Predictor a python generator, then there won't be a memory problem.
• return segments coordinate instead of masks.

python interface usage:
segmentation:

model = YOLO("weights/yolov8n-seg.pt")
for output in model.predict(source="ultralytics/assets"):
    if output:
        print(output['det'].shape)   # ndarray, Nx6, include box, cls score and cls
        print(len(output['segment']))  # List[ndarray], each element is the corresponding coordinate of segment

detection:

model = YOLO("weights/yolov8n.pt")
for output in model.predict(source="ultralytics/assets"):
    if output:
        print(output['det'].shape)   # ndarray, Nx6, include box, cls score and cls

cls:

model = YOLO("weights/yolov8n-cls.pt")
for output in model.predict(source="ultralytics/assets"):
    if output:
        print(output['prob'].shape)   # ndarray, shape: (num_class, )

set a arg return_outputs to control whether to return outputs.
return_outputs=True by default in python interface.
return_outputs=False by default in cli mode.

If don't want the outputs in python interface, just set it to False.

model = YOLO("weights/yolov8n-cls.pt")
model.predict(source="ultralytics/assets", return_outputs=False):

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Refactoring prediction functionality to support iterative result generation and optional result return in Ultralytics YOLO models.

📊 Key Changes

• Modified tests to iterate through predictions instead of relying on a direct list length check.
• Added an optional return_outputs flag to the predict() method to control the output behavior.
• Updated predictors to act as generators yielding results upon each iteration if return_outputs is set to True.
• Altered client-facing scripts to reflect generator behavior and handle the option of not returning outputs.
• Removed unused accumulation of all_outputs in favor of yielding or returning individual predictions.

🎯 Purpose & Impact

• Purpose: To enhance the prediction methods by providing more flexible output handling and allowing incremental processing of results.
• Impact for Developers: More control over how prediction results are processed and presented, making integration into larger applications smoother.
• Impact for Users: Potential for improved performance and memory usage when handling large numbers of predictions, as not all results need to be kept in memory at once. 🚀

[github-actions]

CLA Assistant Lite bot All Contributors have signed the CLA. ✅

[AyushExel]

@Laughing-q okay perfect! let's discuss this later today. I'll add some docs so we have them at the same time as release

[glenn-jocher]

recheck

[AyushExel]

@Laughing-q do we have a function in place to convert segment coordinates to masks?

[Laughing-q]

@AyushExel here, but it serves dataset. never tested independently.

ultralytics/ultralytics/yolo/data/utils.py

Lines 133 to 151 in 55b313b

	def polygon2mask(imgsz, polygons, color=1, downsample_ratio=1):
	"""
	Args:
	imgsz (tuple): The image size.
	polygons (np.ndarray): [N, M], N is the number of polygons, M is the number of points(Be divided by 2).
	color (int): color
	downsample_ratio (int): downsample ratio
	"""
	mask = np.zeros(imgsz, dtype=np.uint8)
	polygons = np.asarray(polygons)
	polygons = polygons.astype(np.int32)
	shape = polygons.shape
	polygons = polygons.reshape(shape[0], -1, 2)
	cv2.fillPoly(mask, polygons, color=color)
	nh, nw = (imgsz[0] // downsample_ratio, imgsz[1] // downsample_ratio)
	# NOTE: fillPoly firstly then resize is trying the keep the same way
	# of loss calculation when mask-ratio=1.
	mask = cv2.resize(mask, (nw, nh))
	return mask

[AyushExel]

okay got it. We can make something similar later

[glenn-jocher]

@Laughing-q good to merge?

[AyushExel]

@glenn-jocher yes..But we need to update docs before putting out stable release.. You can merge this.. I'll start another PR for docs

[glenn-jocher]

Merged into 8.0.3 fixes