Model Loading

To use a model that has been trained on Substra, you need to first download the files with the download_algo_files function. These files are used to re-instantiate and load all variables to retrieve the Substra Algo in its wanted state by the load_algo function.

An example on how to download a model is available in the MNIST Substrafl FedAvg example.

substrafl.model_loading.download_algo_files(client: substra.sdk.client.Client, compute_plan_key: str, dest_folder: os.PathLike, round_idx: Optional[int] = None)

Download all the files needed to load the model:

  • hosted on the client organization

  • being part of the given compute plan

  • being the result of the associated strategy after round_idx steps

into memory.

Those files are:

  • the function used for this task

  • the output local state of the task

  • a metadata.json


This function supports only strategies with one composite traintask for a given organization and round.

  • client (substra.Client) – Substra client where to fetch the model from.

  • compute_plan_key (str) – Compute plan key to fetch the model from.

  • dest_folder (os.PathLike) – Folder where to download the files.

  • round_idx (Optional[int], None) – Round of the strategy to fetch the model from. If set to None, the last round will be used. (Defaults to None).

  • NotImplementedError – The given compute plan must have been submitted to Substra through the execute_experiment() function.

  • TrainTaskNotFoundError – If no composite matches the given requirements.

  • MultipleTrainTaskError – The experiment to get the model from can’t have multiple TrainDataNodes hosted on the same organization. In practice this means the presence of multiple composite train tasks with the same round number on the same rank.

  • UnfinishedTrainTaskError – The task from which the files are trying to be downloaded is not done.

substrafl.model_loading.load_algo(input_folder: os.PathLike) Any

Loads an algo from a specified folder. This folder should contains:

  • function.tar.gz

  • metadata.json

  • the file specified in metadata.local_state_file

This kind of folder can be generated with the download_algo_files() function.


input_folder (os.PathLike) – Path to folder containing the required files.


The serialized algo within the input_folder

Return type