Hub

Bases: Hub

new(name, task=None, model_type=None, model_size=None, categories=None, root_dir=None, *args, **kwargs) classmethod

Create Ultralytics Hub.

Parameters:

Name	Type	Description	Default
name	str	Hub name	required
task	str	Task Name. See UltralyticsHub.TASKS. Defaults to None.	None
model_type	str	Model Type. See UltralyticsHub.MODEL_TYPES. Defaults to None.	None
model_size	str	Model Size. See UltralyticsHub.MODEL_SIZES. Defaults to None.	None
categories	Union[list[dict], list]	class dictionary or list. [{"supercategory": "name"}, ] or ["name",].	None
root_dir	str	Root directory of hub repository. Defaults to None.	None

artifact_dir: Path property cached

Artifact Directory. This is raw output of each backend.

backend: str property writable

Backend name

best_ckpt_file: Path property cached

Best Checkpoint File

draw_dir: Path property cached

Draw Results Directory

evaluate_file: Path property cached

Evaluate Json File

hub_dir: Path property cached

Hub(Model) Directory

inference_dir: Path property cached

Inference Results Directory

inference_file: Path property cached

Inference Results File

last_ckpt_file: Path property cached

Last Checkpoint File

metric_file: Path property cached

Metric Csv File

model_config_file: Path property cached

Model Config yaml File

model_size: str property writable

Model Size

model_type: str property writable

Model Type

name: str property writable

Hub name

onnx_file: Path property cached

Best Checkpoint File

root_dir: Path property writable

Root Directory

task: str property writable

Task Name

train_config_file: Path property cached

Train Config yaml File

version: str property writable

Version

benchmark(image_size=None, batch_size=16, device='0', half=False, trial=100)

Benchmark Model

Parameters:

Name	Type	Description	Default
image_size	Union[int, list[int]]	inference image size. None for same with train_config (recommended).	None
batch_size	int	dynamic batch size. Defaults to 16.	16
device	str	device. "cpu" or "gpu_id". Defaults to "0".	'0'
half	bool	half. Defaults to False.	False
trial	int	number of trials. Defaults to 100.	100

Example

hub.benchmark( image_size=640, batch_size=16, device="0", half=False, trial=100, ) { "inference_time": 0.123, "fps": 123.123, "image_size": [640, 640], "batch_size": 16, "device": "0", "cpu_name": "Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz", "gpu_name": "GeForce GTX 1080 Ti", }

Returns:

Name	Type	Description
dict	dict	benchmark result

check_train_sanity()

Check if all essential files are exist.

Returns:

Name	Type	Description
bool	bool	True if all files are exist else False

delete_artifact()

Delete Artifact Directory. It can be trained again.

delete_hub()

Delete all artifacts of Hub. Hub name can be used again.

evaluate(dataset, dataset_root_dir=None, set_name='test', batch_size=4, image_size=None, letter_box=None, confidence_threshold=0.25, iou_threshold=0.5, half=False, workers=2, device='0', draw=False, hold=True)

Start Evaluate

Parameters:

Name	Type	Description	Default
dataset	Union[Dataset, str]	Waffle Dataset object or path or name.	required
dataset_root_dir	str	Waffle Dataset root directory. Defaults to None.	None
batch_size	int	batch size. Defaults to 4.	4
image_size	Union[int, list[int]]	image size. Defaults to None.	None
letter_box	bool	letter box. Defaults to None.	None
confidence_threshold	float	confidence threshold. Defaults to 0.25.	0.25
iou_threshold	float	iou threshold. Defaults to 0.5.	0.5
half	bool	half. Defaults to False.	False
workers	int	workers. Defaults to 2.	2
device	str	device. Defaults to "0".	'0'
draw	bool	draw. Defaults to False.	False
hold	bool	hold. Defaults to True.	True

Raises:

Type	Description
FileNotFoundError	if can not detect appropriate dataset.
e	something gone wrong with ultralytics

Examples:

>>> evaluate_result = hub.evaluate(
        dataset=detection_dataset,
        batch_size=4,
        image_size=640,
        letterbox=False,
        confidence_threshold=0.25,
        iou_threshold=0.5,
        workers=4,
        device="0",
    )
# or you can use train option by passing None
>>> evaluate_result = hub.evaluate(
        ...
        image_size=None,  # use train option
        letterbox=None,  # use train option
        ...
    )
>>> evaluate_result.metrics
[{"tag": "mAP", "value": 0.1}, ...]

Returns:

Name	Type	Description
EvaluateResult	EvaluateResult	evaluate result

export(image_size=None, batch_size=16, opset_version=11, half=False, device='0', hold=True)

Export Model

Parameters:

Name	Type	Description	Default
image_size	Union[int, list[int]]	inference image size. None for same with train_config (recommended).	None
batch_size	int	dynamic batch size. Defaults to 16.	16
opset_version	int	onnx opset version. Defaults to 11.	11
half	bool	half. Defaults to False.	False
device	str	device. "cpu" or "gpu_id". Defaults to "0".	'0'
hold	bool	hold or not. If True then it holds until task finished. If False then return Inferece Callback and run in background. Defaults to True.	True

Example

export_result = hub.export( image_size=640, batch_size=16, opset_version=11, )

or simply use train option by passing None

export_result = hub.export( ..., image_size=None, # use train option ... ) export_result.export_file hubs/my_hub/weights/model.onnx

Returns:

Name	Type	Description
ExportResult	ExportResult	export result

from_model_config(name, model_config_file, root_dir=None) classmethod

Create new Hub with model config.

Parameters:

Name	Type	Description	Default
name	str	hub name.	required
model_config_file	str	model config yaml file.	required
root_dir	str	hub root directory. Defaults to None.	None

Returns:

Name	Type	Description
Hub	Hub	New Hub instance

get_available_backends() classmethod

Get available backends

Returns:

Type	Description
list[str]	list[str]: Available backends

get_available_model_sizes(backend=None, task=None, model_type=None) classmethod

Get available model sizes

Parameters:

Name	Type	Description	Default
backend	str	Backend name	None
task	str	Task name	None
model_type	str	Model type	None

Raises:

Type	Description
ModuleNotFoundError	If backend is not supported

Returns:

Type	Description
list[str]	list[str]: Available model sizes

get_available_model_types(backend=None, task=None) classmethod

Get available model types

Parameters:

Name	Type	Description	Default
backend	str	Backend name	None
task	str	Task name	None

Raises:

Type	Description
ModuleNotFoundError	If backend is not supported

Returns:

Type	Description
list[str]	list[str]: Available model types

get_available_tasks(backend=None) classmethod

Get available tasks

Parameters:

Name	Type	Description	Default
backend	str	Backend name	None

Raises:

Type	Description
ModuleNotFoundError	If backend is not supported

Returns:

Type	Description
list[str]	list[str]: Available tasks

get_default_train_params(backend=None, task=None, model_type=None, model_size=None) classmethod

Get default train params

Parameters:

Name	Type	Description	Default
backend	str	Backend name	None
task	str	Task name	None
model_type	str	Model type	None
model_size	str	Model size	None

Raises:

Type	Description
ModuleNotFoundError	If backend is not supported

Returns:

Name	Type	Description
dict	dict	Default train params

get_evaluate_result()

Get evaluate result from evaluate file.

Example

hub.get_evaluate_result() [ { "tag": "mAP", "value": 0.5, }, ]

Returns:

Name	Type	Description
dict	list[dict]	evaluate result

get_hub_class(backend=None) classmethod

Get hub class

Parameters:

Name	Type	Description	Default
backend	str	Backend name	None

Raises:

Type	Description
ModuleNotFoundError	If backend is not supported

Returns:

Name	Type	Description
Hub	Hub	Backend hub Class

get_hub_list(root_dir=None) classmethod

Get hub name list in root_dir.

Parameters:

Name	Type	Description	Default
root_dir	str	hub root directory. Defaults to None.	None

Returns:

Type	Description
list[str]	list[str]: hub name list

get_image_loader()

Get image loader function.

Returns:

Type	Description
tuple[torch.Tensor, ImageInfo]	tuple[torch.Tensor, ImageInfo]: input transform function

Example

transform = hub.get_image_loader() image, image_info = transform("path/to/image.jpg") model = hub.get_model() output = model(image.unsqueeze(0))

get_inference_result()

Get inference result from inference file.

Example

hub.get_inference_result() [ { "id": "00000001", "category": "person", "bbox": [0.1, 0.2, 0.3, 0.4], "score": 0.9, }, ]

Returns:

Type	Description
list[dict]	list[dict]: inference result

get_metrics()

Get metrics per epoch from metric file.

Example

hub.get_metrics() [ [ { "tag": "epoch", "value": "1", }, { "tag": "train_loss", "value": "0.0012", } ], ]

Returns:

Type	Description
list[list[dict]]	list[dict]: metrics per epoch

get_model_config()

Get model config from model config file.

Returns:

Name	Type	Description
ModelConfig	ModelConfig	model config

get_train_config()

Get train config from train config file.

Returns:

Name	Type	Description
TrainConfig	TrainConfig	train config

inference(source, recursive=True, image_size=None, letter_box=None, batch_size=4, confidence_threshold=0.25, iou_threshold=0.5, half=False, workers=2, device='0', draw=False, hold=True)

Start Inference

Parameters:

Name	Type	Description	Default
source	str	source directory	required
recursive	bool	recursive. Defaults to True.	True
image_size	Union[int, list[int]]	image size. None for using training config. Defaults to None.	None
letter_box	bool	letter box. None for using training config. Defaults to None.	None
batch_size	int	batch size. Defaults to 4.	4
confidence_threshold	float	confidence threshold. Defaults to 0.25.	0.25
iou_threshold	float	iou threshold. Defaults to 0.5.	0.5
half	bool	half. Defaults to False.	False
workers	int	workers. Defaults to 2.	2
device	str	device. "cpu" or "gpu_id". Defaults to "0".	'0'
draw	bool	draw. Defaults to False.	False
hold	bool	hold. Defaults to True.	True

Raises:

Type	Description
FileNotFoundError	if can not detect appropriate dataset.
e	something gone wrong with ultralytics

Example

inference_result = hub.inference( source="path/to/images", batch_size=4, image_size=640, letterbox=False, confidence_threshold=0.25, iou_threshold=0.5, workers=4, device="0", draw=True, )

or simply use train option by passing None

inference_result = hub.inference( ... image_size=None, # use train option letterbox=None, # use train option ... ) inference_result.predictions [{"relative/path/to/image/file": [{"category": "1", "bbox": [0, 0, 100, 100], "score": 0.9}, ...]}, ...]

Returns:

Name	Type	Description
InferenceResult	InferenceResult	inference result

load(name, root_dir=None) classmethod

Load Hub by name.

Parameters:

Name	Type	Description	Default
name	str	hub name.	required
root_dir	str	hub root directory. Defaults to None.	None

Raises:

Type	Description
FileNotFoundError	if hub is not exist in root_dir

Returns:

Name	Type	Description
Hub	Hub	Hub instance

new(name, backend=None, task=None, model_type=None, model_size=None, categories=None, root_dir=None, *args, **kwargs) classmethod

Create Hub.

Parameters:

Name	Type	Description	Default
name	str	Hub name	required
backend	str	Backend name. See Hub.BACKENDS. Defaults to None.	None
task	str	Task Name. See Hub.TASKS. Defaults to None.	None
model_type	str	Model Type. See Hub.MODEL_TYPES. Defaults to None.	None
model_size	str	Model Size. See Hub.MODEL_SIZES. Defaults to None.	None
categories	Union[list[dict], list]	class dictionary or list. [{"supercategory": "name"}, ] or ["name",].	None
root_dir	str	Root directory of hub repository. Defaults to None.	None

Returns:

Name	Type	Description
Hub	Hub	Hub instance

train(dataset, dataset_root_dir=None, epochs=None, batch_size=None, image_size=None, learning_rate=None, letter_box=None, pretrained_model=None, device='0', workers=2, seed=0, verbose=True, hold=True)

Start Train

Parameters:

Name	Type	Description	Default
dataset	Union[Dataset, str]	Waffle Dataset object or path or name.	required
dataset_root_dir	str	Waffle Dataset root directory. Defaults to None.	None
epochs	int	number of epochs. None to use default. Defaults to None.	None
batch_size	int	batch size. None to use default. Defaults to None.	None
image_size	Union[int, list[int]]	image size. None to use default. Defaults to None.	None
learning_rate	float	learning rate. None to use default. Defaults to None.	None
letter_box	bool	letter box. None to use default. Defaults to None.	None
pretrained_model	str	pretrained model. None to use default. Defaults to None.	None
device	str	"cpu" or "gpu_id" or comma seperated "gpu_ids". Defaults to "0".	'0'
workers	int	number of workers. Defaults to 2.	2
seed	int	random seed. Defaults to 0.	0
verbose	bool	verbose. Defaults to True.	True
hold	bool	hold process. Defaults to True.	True

Raises:

Type	Description
FileExistsError	if trained artifact exists.
FileNotFoundError	if can not detect appropriate dataset.
ValueError	if can not detect appropriate dataset.
e	something gone wrong with ultralytics

Example

train_result = hub.train( dataset=dataset, epochs=100, batch_size=16, image_size=640, learning_rate=0.001, letterbox=False, device="0", # use gpu 0 # device="0,1,2,3", # use gpu 0,1,2,3 # device="cpu", # use cpu workers=2, seed=123 ) train_result.best_ckpt_file hubs/my_hub/weights/best_ckpt.pt train_result.metrics [[{"tag": "epoch", "value": 1}, {"tag": "train/loss", "value": 0.1}, ...], ...]

Returns:

Name	Type	Description
TrainResult	TrainResult	train result