Properties

• name Name of this Experiment

• status Status of this Experiment

Methods

update

update(
   name: Optional[str] = None, description: Optional[str] = None,
   base_experiment: Optional['Experiment'] = None,
   base_model_version: Optional['ModelVersion'] = None,
   status: Union[ExperimentStatus, str, None] = None
)

Description

Update this experiment with a given name, description, a base experiment, a base model version or a status.

Examples

my_experiment.update(description="First try Yolov5")

Arguments

• name (str, optional) : Name of the experiment. Defaults to None.

• description (str, optional) : Description of the experiment. Defaults to None.

• base_experiment (Experiment, optional) : Base experiment of the experiment. Defaults to None.

• base_model_version (ModelVersion, optional) : Base model version of the experiment. Defaults to None.

• status (Union[ExperimentStatus, str, None], optional) : Status of the experiment. Defaults to None.

delete

delete()

Description

Delete this experiment.

Examples

my_experiment.delete()

list_artifacts

list_artifacts(
   limit: Optional[int] = None, offset: Optional[int] = None,
   order_by: Optional[list[str]] = None
)

Description

List artifacts stored in the experiment.

Examples

artifacts = my_experiment.list_artifacts()

Arguments

• limit (int, optional) : limit of artifacts to retrieve. Defaults to None.

• offset (int, optional) : offset to start retrieving artifacts. Defaults to None.

• order_by (list[str], optional) : fields to order by. Defaults to None.

Returns

A list of (Artifact) objects that you can manipulate

delete_all_artifacts

delete_all_artifacts()

Description

Delete all stored artifacts for experiment

⚠️ DANGER ZONE: This will definitely remove the artifacts from our servers

Examples

experiment.delete_all_artifacts()

create_artifact

create_artifact(
   name: str, filename: str, object_name: str, large: bool = False
)

Description

Create an artifact for this experiment.

Examples

self.create_artifact(name="a_file", filename="file.png", object_name="some_file_in_s3.png", large=False)

Arguments

• name (str) : name of the artifact.

• filename (str) : filename.

• object_name (str) : s3 object name.

• large (bool, optional) : >5Mb or not. Defaults to False.

Returns

An (Artifact) object

store

store(
   name: str, path: Union[str, Path, None] = None, do_zip: bool = False
)

Description

Store an artifact and attach it to the experiment.

Examples

# Zip and store a folder as an artifact for the experiment
# you can choose an arbitrary name or refer to our 'namespace'
# for certain artifacts to have a custom behavior

trained_model_path = "my_experiment/saved_model"
experiment.store("model-latest", trained_model_path, do_zip=True)

Arguments

• name (str) : name for the artifact. Defaults to "".

• path (str or Path) : path to the file or folder. Defaults to None.

• do_zip (bool, optional) : Whether to compress the file to a zip file. Defaults to False.

Raises

• FileNotFoundException : No file found at the given path

Returns

An (Artifact) object

store_local_artifact

store_local_artifact(
   name: str
)

Description

Store an artifact in platform that is locally stored

This artifact shall have the name: config, checkpoint-data-latest, checkpoint-index-latest or model-latest

It will look for special file into current directory.

Examples

my_experiment.store_local_artifact("model-latest")

Arguments

• name (str) : Name of the artifact to store

Returns

An (Artifact) object

get_base_model_version

get_base_model_version()

Description

Retrieve the base model version of this experiment.

Examples

model_version = experiment.get_base_model_version()

Returns

A ModelVersion object representing the base model.

get_base_experiment

get_base_experiment()

Description

Retrieve the base experiment of this experiment.

Examples

previous = experiment.get_base_experiment()

Returns

An Experiment object representing the base experiment.

get_artifact

get_artifact(
   name: str
)

Description

Retrieve an artifact from its name in this experiment.

Examples

model_artifact = experiment.get_artifact("model-latest")
assert model_artifact.name == "model-latest"
assert model_artifact.object_name == "d67924a0-7757-48ed-bf7a-322b745e917e/saved_model.zip"

Arguments

• name (str) : Name of the artifact to retrieve

Returns

An (Artifact) object

list_logs

list_logs(
   limit: Optional[int] = None, offset: Optional[int] = None,
   order_by: Optional[list[str]] = None
)

Description

List everything that has been logged.

List everything that has been logged to an experiment.

Examples

logs = experiment.list_logs()
assert logs[0].type == LogType.Table
assert logs[0].data == {"batch_size":4, "epochs":1000}

Arguments

• limit (int, optional) : limit of logs to retrieve. Defaults to None.

• offset (int, optional) : offset to start retrieving logs. Defaults to None.

• order_by (list[str], optional) : fields to order by. Defaults to None.

Returns

A list of Log objects

delete_all_logs

delete_all_logs()

Description

Delete everything that has been logged.

Delete everything that has been logged into this experiment.

Examples

experiment.delete_all_logs()

get_log

get_log(
   name: str
)

Description

Retrieve a log from its name in this experiment.

Examples

parameters = experiment.get_log("parameters")
assert log.parameters == { "batch_size":4, "epochs":1000}

Arguments

• name (str) : name of the log to retrieve

Returns

A Log object

log

log(
   name: str, data: LogDataType, type: Union[LogType, str], replace: bool = True
)

Description

Record some data in an experiment.

It will create a Log object, that you can manipulate in SDK.
All logs are displayed in experiment view on Picsellia.

If a Log with this name already exists, it will be updated unless parameter replace is set to False.

If it is a LogType LINE and replace is True, then it will append data at the end of stored data.
If you want to replace a line, delete this log and create another one.

Examples

parameters = {
    "batch_size":4,
    "epochs":1000
}
exp.log("parameters", parameters, type=LogType.TABLE)

Arguments

• name (str) : Name of the log.

• data (Any) : Data to be saved.

• type (LogType, optional) : Type of the data to log.
  This will condition how it is displayed in the experiment dashboard. Defaults to None.

• replace (bool, optional) : If true and log already exists and it is not a line, it will replace log data.
  Defaults to True.

Raises

• Exception : Impossible to upload the file when logging an image.

Returns

A Log object

log_parameters

log_parameters(
   parameters: dict
)

Description

Record parameters of an experiment into Picsellia.

If parameters were already setup, it will be replaced.

Examples

parameters = {
    "batch_size":4,
    "epochs":1000
}
exp.log_parameters(parameters)

Arguments

• parameters (Any) : Parameters to be saved.

Returns

A Log object

store_logging_file

store_logging_file(
   path: Union[str, Path]
)

Description

Store a logging file for this experiment.

Examples

experiment.store_logging_file("logs.txt")

Arguments

• path (str or Path) : path to the file or folder.

Raises

• FileNotFoundException : No file found at the given path

Returns

A (LoggingFile) object

get_logging_file

get_logging_file()

Description

Retrieve logging file of this experiment.

Examples

logging_file = experiment.get_logging_file()
logging_file.download()

Returns

A (LoggingFile) object

send_logging

send_logging(
   log: Union[str, list], part: str, final: bool = False, special: Union[str, bool,
   list] = False
)

Description

Send a logging experiment to the experiment.

Examples

experiment.send_logging("Hello World", "Hello", final=True)

Arguments

• log (str) : Log content

• part (str) : Logging Part

• final (bool, optional) : True if Final line. Defaults to False.

• special (bool, optional) : True if special log. Defaults to False.

start_logging_chapter

start_logging_chapter(
   name: str
)

Description

Print a log entry to the log.

Examples

experiment.start_logging_chapter("Training")

Arguments

• name (str) : Chapter name

start_logging_buffer

start_logging_buffer(
   length: int = 1
)

Description

Start logging buffer.

Examples

experiment.start_logging_buffer()

Arguments

• length (int, optional) : Buffer length. Defaults to 1.

end_logging_buffer

end_logging_buffer()

Description

End the logging buffer.

Examples

experiment.end_logging_buffer()

update_job_status

update_job_status(
   status: Union[JobStatus, str]
)

Description

Update the job status of this experiment.

Examples

experiment.update_job_status(JobStatus.FAILED)

Arguments

• status (JobStatus) : Status to send

publish

publish(
   name: str
)

Description

export_as_model

export_as_model(
   name: str
)

Description

Publish an Experiment as a ModelVersion to your registry.
A Model with given name will be created, and its first version will be the exported experiment

Examples

model_version = experiment.export_as_model("awesome-model")

Arguments

• name (str) : Target Name for the model in the registry.

Returns

A ModelVersion just created from the experiment

export_in_existing_model

export_in_existing_model(
   existing_model: Model
)

Description

Publish an Experiment as a ModelVersion of given already existing Model

Examples

my_model = client.get_model("foo_model")
model_version = experiment.export_in_existing_model(my_model)

Arguments

• existing_model (Model) : Model in the registry were this experiment should be exported.

Returns

A ModelVersion just created from the experiment

launch

launch(
   gpus: int = 1
)

Description

Launch a job on a remote environment with this experiment.

ℹ️ The remote environment has to be setup prior launching the experiment.
It defaults to our remote training engine.

Examples

experiment.launch()

Arguments

• gpus (int, optional) : Number of GPU to use for the training. Defaults to 1.

download_artifacts

download_artifacts(
   with_tree: bool
)

Description

Download all artifacts from the experiment to the local directory.

Examples

experiment.download_artifacts(with_tree=True)

Arguments

• with_tree : If True, the artifacts will be downloaded in a tree structure.

attach_model_version

attach_model_version(
   model_version: ModelVersion, do_attach_base_parameters: bool = True
)

Description

Attach model version to this experiment.
There is only one model version attached to an experiment

Examples

foo_model = client.get_model("foo").get_version(3)
my_experiment.attach_model_version(foo_model)

Arguments

• model_version ModelVersion : A model version to attach to the experiment.

• do_attach_base_parameters (bool) : Attach base parameters of model version to experiment. Defaults to True.

attach_dataset

attach_dataset(
   name: str, dataset_version: DatasetVersion
)

Description

Attach a dataset version to this experiment.

Retrieve or create a dataset version and attach it to this experiment.

Examples

foo_dataset = client.get_dataset("foo").get_version("first")
my_experiment.attach_dataset("training", foo_dataset)

Arguments

• name (str) : Name to label this attached dataset. Use it like a descriptor of the attachment.

• dataset_version DatasetVersion : A dataset version to attach to the experiment.

detach_dataset

detach_dataset(
   dataset_version: DatasetVersion
)

Description

Detach a dataset version from this experiment.

Examples

foo_dataset = client.get_dataset("foo").get_version("first")
my_experiment.attach_dataset(foo_dataset)
my_experiment.detach_dataset(foo_dataset)

Arguments

• dataset_version DatasetVersion : A dataset version to attach to the experiment.

list_attached_dataset_versions

list_attached_dataset_versions()

Description

Retrieve all dataset versions attached to this experiment

Examples

datasets = my_experiment.list_attached_dataset_versions()

Returns

A list of DatasetVersion object attached to this experiment

get_dataset

get_dataset(
   name: str
)

Description

Retrieve the dataset version attached to this experiment with given name

Examples

pics = dataset.list_assets()
annotations = dataset.list_annotations()

Arguments

• name (str) : Name of the dataset version in the experiment

Returns

A DatasetVersion object attached to this experiment

run_train_test_split_on_dataset

run_train_test_split_on_dataset(
   name: str, prop: float = 0.8, random_seed: Optional[Any] = None
)

Description

Run a train test split on a dataset attached to this experiment.

Examples

pics = dataset.list_assets()
annotations = dataset.list_annotations()

Arguments

• name (str) : Name of the dataset version in the experiment

• prop (float) : Proportion of the dataset to use for training

• random_seed (int) : Random seed to use for the split

Returns

• train_assets MultiAsset : assets to use for training
  • eval_assets MultiAsset: assets to use for evaluation
  • train_label_count (dict): number of assets per label in train set
  • eval_label_count (dict): number of assets per label in eval set
  • labels (list[Label]): list of labels in the dataset

A tuple containing:

add_evaluation

add_evaluation(
   asset: Asset, add_type: Union[str,
   AddEvaluationType] = AddEvaluationType.REPLACE,
   rectangles: Optional[list[tuple[int, int, int, int, Label, float]]] = None,
   polygons: Optional[list[tuple[list[list[int]], Label, float]]] = None,
   classifications: Optional[list[tuple[Label, float]]] = None
)

Description

Add an evaluation of the asset by this experiment.

By default, if given asset had already been evaluated, evaluation will be replaced.
You can add different shapes but will only be able to compute evaluation metrics on one kind of inference type.

Examples

asset = dataset_version.find_asset(filename="asset-1.png")
experiment.add_evaluation(asset, rectangles=[(10, 20, 30, 40, label_cat, 0.8), (50, 60, 20, 30, label_dog, 0.7)])
job = experiment.compute_evaluations_metrics(InferenceType.OBJECT_DETECTION)
job.wait_for_done()

Arguments

• asset Asset : asset to add evaluation on

• add_type (str ou AddEvaluationType) : replace or keep old evaluation, defaults to

• rectangles (optional) : list of tuples representing rectangles with scores

• polygons (optional) : list of tuples representing polygons with scores

• classifications (optional) : list of tuples representing classifications with scores

list_evaluations

list_evaluations(
   limit: Optional[int] = None, offset: Optional[int] = None,
   page_size: Optional[int] = None, order_by: Optional[list[str]] = None,
   q: Optional[str] = None
)

Description

List evaluations of this experiment.
It will retrieve all evaluations made by this experiment.
You will then be able to manipulate them.

Arguments

• limit (int, optional) : if given, will limit the number of evaluations returned

• offset (int, optional) : if given, will return evaluations that would have been returned
  after this offset in given order

• page_size (int, optional) : page size when returning evaluations paginated, can change performance

• order_by (list[str], optional) : if not empty, will order evaluations by fields given in this parameter

• q (str, optional) : if given, will try to apply query to filter evaluations

Returns

A (MultiEvaluation)

compute_evaluations_metrics

compute_evaluations_metrics(
   inference_type: InferenceType, evaluations: Union[list[Union[str, UUID]],
   list[Evaluation], MultiEvaluation, None] = None, worker: Optional[Worker] = None,
   status: Optional[AnnotationStatus] = None
)

Description

Compute evaluation metrics across evaluations added to this experiment.
Picsellia will compute coco metrics on each evaluation and compare to existing annotations.

Examples

experiment.add_evaluation(rectangles=[(10, 20, 30, 40, label_cat, 0.8), (50, 60, 20, 30, label_dog, 0.7)])
experiment.compute_evaluations_metrics(InferenceType.OBJECT_DETECTION)

Arguments
status (AnnotationStatus, optional): Existing annotations will be filtered to only retrieve those that have this status.

Returns

A Job that you can wait for done.