This page discusses how to save and load configuration objects:

  1. Saving configuration
  2. How to specify files/directories to be serialized
  3. HuggingFace integration

Saving/Loading objects with configurations

Configuration objects can be loaded and saved. You can even embed them within any standard Python structure (i.e. dictionary, list, tuple).

from experimaestro import load, save

# Saves the object
save([obj1, obj2, {key: obj3, key2: obj4}], "/my/folder")

# Load the object
[obj1, obj2, obj_dict] = load("/my/folder")

load(path, as_instance=False)

Load data from disk

  • path (Union[str, Path, SerializedPathLoader]) –

    A directory or a function that transforms relative file path into absolute ones

  • as_instance (bool, default: False ) –

    returns instances instead of configuration objects

save(obj, save_directory)

Saves an object into a disk file

  • save_directory (Path) –

    The directory in which the object and its data will be saved (by default, the object is saved in "definition.json")

A task configuration/instance can be loaded with

from_task_dir(path, as_instance=False)

Loads a task object

The serialization context is controlled by a specific object named SerializationContext:


Context when serializing experimaestro configurations

__init__(*, save_directory=None)

Creates a new serialization context

  • save_directory (Optional[Path], default: None ) –

    Defines the directory where SerializedPath are stored

If you need more control over saved data, you can use state_dict and from_state_dict that respectively returns Python data structures and loads from them.

state_dict(context, obj)

Returns a state dictionary of the object

  • context (SerializationContext) –

    The serialization context

  • obj (Any) –

    the object to serialize

from_state_dict(state, path=None, *, as_instance=False)

Load an object from a state dictionary

  • state (Dict[str, Any]) –

    The state

  • path (Union[None, str, Path, SerializedPathLoader], default: None ) –

    A directory or a function that transforms relative file path into absolute ones

  • as_instance (bool, default: False ) –

    returns instances instead of configuration objects

Saving/Loading from running experiment

To ease saving/loading configuration from experiments, one can use methods from the experiment objects as follows:

from experimaestro import experiment, Param, Config

class MyConfig(Config):
    a: Param[int]

if __name__ == "__main__":
    # Saving configurations
    with experiment("/tmp/load_save", "xp1", port=-1) as xp:
        cfg = MyConfig(a=1)[cfg])

    # Loading configurations
    with experiment("/tmp/load_save", "xp2", port=-1) as xp:
        # Loads MyConfig(a=1)
        cfg, = xp.load("xp1")

Specifying paths to be serialized

Configurations can be serialized with the data necessary to restore their state. This can be useful to share a model (e.g. with HuggingFace hub).

from experimaestro import DataPath

class MyConfig(Config):
    to_serialize: DataPath
    """This path will be serialized on the hub"""

HuggingFace integration

# ExperimaestroHFHub implements the interface from ModelHubMixin
from experimaestro.huggingface import ExperimaestroHFHub

# True if the object should be an instance (and not a configuration)
as_instance = False

# Save and load a configuration
ExperimaestroHFHub().from_pretrained(hf_id_or_folder, as_instance=as_instance)

# Save and load a configuration (with a variant)
ExperimaestroHFHub(config).push_to_hub(hf_id, variant)
ExperimaestroHFHub().from_pretrained(hf_id_or_folder, variant=variant, as_instance=as_instance)


Bases: ModelHubMixin

Defines models that can be uploaded/downloaded from the Hub