Model Functionality Coverage

Model initialization is handled as-is: construct YOLO(model, task, verbose) directly, then pass that instance as the Model input to any operator. The operators wrap individual model functions; they do not replace the native model façade.

Upstream reference: Ultralytics Model.

Ultralytics Model functionality	ml-pipes equivalent
model(source, **kwargs) / predict(source, predictor=None, **kwargs)	yolo.Predict(model, **predict_options)*
embed(source, **kwargs)	yolo.Embed(model, **embed_options)*
track(source, persist=..., **kwargs)	yolo.Track(model, persist=..., **track_options)*
val(validator=None, **kwargs)	None
train(trainer=None, **kwargs)	None
benchmark(**kwargs)	None
export(**kwargs)	None
tune(use_ray=False, iterations=10, *args, **kwargs)	None
reset_weights()	AS-IS
load(weights)	AS-IS
save(filename)	AS-IS
info(detailed=False, verbose=True)	AS-IS
fuse()	AS-IS
add_callback(event, func)	AS-IS
clear_callback(event)	AS-IS
reset_callbacks()	AS-IS
names	AS-IS
device	AS-IS
transforms	AS-IS
task_map	AS-IS
is_hub_model(model)	AS-IS
is_triton_model(model)	AS-IS

[!IMPORTANT] * stream=True is deliberately unsupported at this boundary. For videos or large datasets, use a pipeline specifically designed to decode, batch, and schedule frames before invoking these operators.

[!TIP] AS-IS: To use native model functionality, create or retain an ultralytics.engine.model.Model instance (for example, YOLO), pass that same instance as the model constructor parameter for any yolo.* operator, and call state, introspection, and callback methods directly on the native model—not on the operator:

from ultralytics import YOLO
from ml_pipes.ultralytics import yolo

# Initialize model directly
model = YOLO("yolo26n.pt")
operator = yolo.Predict(model, conf=0.25)

# or simply access it through operator
operator = yolo.Predict("yolo26n.pt", conf=0.25)
model = operator.model

# call your desired function
summary = model.info(verbose=False)

Result Functionality Coverage

Results fields—orig_img, orig_shape, boxes, masks, probs, keypoints, obb, semantic_mask, depth, speed, names, path, and save_dir—are pure native data. They remain accessible in a pipeline through any operator that accepts Results, such as Map; no dedicated adapter is needed.

Upstream reference: Ultralytics Results.

Ultralytics Results functionality	ml-pipes operator / access
cpu()	Direct call: Results.cpu
numpy()	Direct call: Results.numpy
cuda()	Direct call: Results.cuda
new()	Direct call: Results.new
verbose()	Direct call: Results.verbose
update(...)	None
to(...)	results.To(...)
plot(...)	results.Plot(...)
show(...)	results.Show(...)
to_df(...)	results.ToDataFrame(...)
to_csv(...)	results.ToCSV(...)
to_json(...)	results.ToJSON(...)
save_txt(...)	results.SaveTXT(...)
save_crop(...)	results.SaveCrop(...)
save(...)	results.Save(...)
summary(...)	results.Summary(...)

results.Show, results.SaveTXT, results.SaveCrop, and results.Save are pipeline side effects: after performing the native action, they pass the same Results object to the next operator. Transformations retain their native return values. update(...) remains unimplemented while its mutation semantics are decided.

[!TIP] Methods marked Direct call have no required arguments beyond the native Results instance, so their unbound method can be placed directly in a pipeline. For example, use Results.verbose for a Results → str step.