Deep learning for NeuroImaging in Python.

Note

This page is a reference documentation. It only explains the function signature, and not how to use it. Please refer to the gallery for the big picture.

nidl.experiment.fetch_experiment(expfile: str, selector: tuple[str] | None = None, cv: tuple[str] | None = None, logdir: str | None = None, verbose: int = 0)[source]¶

Fetch an experiement from an input configuration file.

Allowed keys are:

project: define here some usefull information about your experiment such as the ‘name’, ‘author’, ‘date’…
global: define here global variables that can be reused in other sections.
import: define here import that can be reused in other sections with the ‘auto’ mode (see desciption below).
scaler: dl interface
transform: dl interfaces
compose: dl interfaces
augmentation: dl interface
dataset: dl interface
dataloader: dl interface
model: dl interface
weights: dl interface
loss: dl interface
optimizer: dl interface
scheduler: dl interface
probe: dl interface
training: define here training settings.
environements: define here the interface to load in order to fullfil your needs and the constraint impose by the ‘interface_occurrences’ parameter (see desciption below).

Interface definition:

the ‘interface’ key contains a name that specifies what class to import in absolute terms.
the ‘interface_version’ key contains the expected version of the loaded interface. The ‘__version__’ module parameter is checked if available and a warning is displayed if is mismatched is detected or the version connot be checked.
other key are the interface parameters.
dynamic parameters can be defined by specifying where this parameter can be find a previously loaded interface, i.e., ‘auto|<name>.<param>’. Note that the order is important here.
cross validation can be defined by specifying a list of values, i.e. ‘cv|[1, 2, 3]’. This will automatically instanciate multiple interface, one for each input setting.

The codes works as follows:

multiple interfaces of the same type can be returned.
the different choices must be described in an ‘environments’ section.
the output name will be prefixed by the environment name.
use the selector input parameters to filter the available interfaces in the input configuration file.

How to define multiple building blocks:

the construction is hierarchic, i.e. child building blocks inherit the properties of the parent building block.
a child building block name contains the parent name as a prefix and use the ‘.’ separator.

The weights section special case:

model names specified in the form hf-hub:path/architecture_name@revision will be loaded from Hugging Face hub.
model names specifid with a path will be loaded from the local machine.

Parameters:

expfile : str

the experimental design file.

selector : tuple of str, default=None

if multiple interface of the same type are defined, this parameter allows you to select the appropriate environements.

cv : tuple of str, default=None

if a cross validation scheme is defined, this parameter allows you to select only some interfaces (i.e., the best set of hyperparapeters).

logdir : str, defautl=None

allows you to save a couple of information about the loaded interface: for the moment only the source code of each interface.

verbose : int, default=0

enable verbose output.

Returns:

data : Bunch

dictionaray-like object containing the experiment building blocks.