0.37.0rc1

0.37.0rc1 - 2023-06-12

Added

• ComputePlanBuilder base class to define which method are needed to implement a custom strategy in SubstraFL.
  These methods are build_compute_plan, load_local_states and save_local_states. #120
• Check and test on string used as metric name in test data nodes (#122).
• Add default exclusion patterns when copying file to avoid creating large Docker images (#118)
• Add the possibility to force the Dependency editable_mode through the environment variable SUBSTRA_FORCE_EDITABLE_MODE (#131)
• Check on the Python version used before generating the Dockerfile ([#123])(#123)).

Changed

• BREAKING: depreciate the usage of model_loading.download_algo_files and model_loading.load_algo functions. New utils functions are now available. (#125)
  model_loading.download_algo_state to download a SubstraFL algo of a given round or rank.
  model_loading.download_shared_state to download a SubstraFL shared object of a given round or rank.
  model_loading.download_aggregated_state to download a SubstraFL aggregated of a given round or rank.
  The API change goes from:

  algo_files_folder = str(pathlib.Path.cwd() / "tmp" / "algo_files")
  download_algo_files(
    client=client_to_download_from,
    compute_plan_key=compute_plan.key,
    round_idx=round_idx,
    dest_folder=algo_files_folder,
  )
  model = load_algo(input_folder=algo_files_folder).model

  to

  algo = download_algo_state(
    client=client_to_download_from  ,
    compute_plan_key=compute_plan.key,
    round_idx=round_idx,
  )
  model = algo.model

• BREAKING: rename build_graph to build_compute_plan. (#120)

• BREAKING: move schema.py into the strategy module. (#120)

  from substrafl.schemas import FedAvgSharedState
  # Become
  from substrafl.strategies.schemas import FedAvgSharedState

• Way to copy function files (#118)

• download_train_task_models_by_rank uses new function list_task_output_assets instead of using value that has been removed (#129)

• Python dependencies are resolved using pip compile during function registration ([#123])(#123)).

• BREAKING: local_dependencies is renamed local_installable_dependencies([#123])(#123)).

• BREAKING: local_installable_dependencies are now limited to local modules or Python wheels (no support for bdist, sdist...)([#123])(#123)).

Fixed

• New dependencies copy method in Docker mode.(#130)

0.36.0

Fixed

• Close issue #114. Large batch size are set to the number of samples in predict for NR and FedPCA. (#115)

Changed

• BREAKING: Metrics are now given as metric_functions and not as metric_key. The functions given as metric functions to test data nodes are automatically registered in a new Substra function by SubstraFL. (#117).
  The new argument of the TestDataNode class metric_functions replaces the metric_keys one and accepts a dictionary (using the key as the identifier of the function given as value), a list of functions or directly a function if there is only one metric to compute (function.__name__ is then used as identifier).
  Installed dependencies are the algo_dependencies passed to execute_experiment, and permissions are the same as the predict function.

  From a user point of view, the metric registration changes from:

  def accuracy(datasamples, predictions_path):
    y_true = datasamples["labels"]
    y_pred = np.load(predictions_path)
  
    return accuracy_score(y_true, np.argmax(y_pred, axis=1))
  
  metric_deps = Dependency(pypi_dependencies=["numpy==1.23.1", "scikit-learn==1.1.1"])
  
  permissions_metric = Permissions(public=False, authorized_ids=DATA_PROVIDER_ORGS_ID)
  
  metric_key = add_metric(
      client=client,
      metric_function=accuracy,
      permissions=permissions_metric,
      dependencies=metric_deps,
  )
  
  test_data_nodes = [
      TestDataNode(
          organization_id=org_id,
          data_manager_key=dataset_keys[org_id],
          test_data_sample_keys=[test_datasample_keys[org_id]],
          metric_keys=[metric_key],
      )
      for org_id in DATA_PROVIDER_ORGS_ID
  ]

  to:

  def accuracy(datasamples, predictions_path):
    y_true = datasamples["labels"]
    y_pred = np.load(predictions_path)
  
    return accuracy_score(y_true, np.argmax(y_pred, axis=1))
  
  test_data_nodes = [
      TestDataNode(
          organization_id=org_id,
          data_manager_key=dataset_keys[org_id],
          test_data_sample_keys=[test_datasample_keys[org_id]],
          metric_functions={"Accuracy": accuracy},
      )
      for org_id in DATA_PROVIDER_ORGS_ID
  ]

• Enforce kwargs for user facing function with more than 3 parameters (#109)

• Remove references to composite. Replace by train_task. (#108)

Added

• Add the Federated Principal Component Analysis strategy (#97)

0.36.0rc2

Fixed

• Close issue #114. Large batch size are set to the number of samples in predict for NR and FedPCA. (#115)

0.36.0rc1

Fixed

• Close issue #114. Large batch size are set to the number of samples in predict for NR and FedPCA. (#115)

0.35.1

This is a promotion of 0.35.1rc1.

Changed

• Change order of layers in the Dockerfile: files are copied as needed before the installation layers, and the final copy

0.35.1rc1

Changed

• Change order of layers in the Dockerfile: files are copied as needed before the installation layers, and the final copy

0.35.0

This is a promotion of 0.35.0rc1

Added

• Initialization task to each strategy in SubstraFL. (#89)

This allows to load the Algo and all its attributes to the platform before any training? Once on the platform, we can perform a testing task before any training.
This init task consists in submitting an empty function, coded in the BaseAlgoclass.

@remote
def initialize(self, shared_states):
     return

The init task return a local output that will be passed as input to a test task, and to the first train task.

The graph pass from:

flowchart LR
    TrainTask1_round0--Local-->TestTask1_r0
    TrainTask1_round0--Shared-->TestTask1_r0

 TrainTask2_round0--Shared-->AggregateTask
    TrainTask2_round0--Local-->TestTask2_r0
    TrainTask2_round0--Shared-->TestTask2_r0

    AggregateTask--Shared-->TrainTask1_r1
    TrainTask1_round0--Local-->TrainTask1_r1

    AggregateTask--Shared-->TrainTask2_r1
    TrainTask2_round0--Local-->TrainTask2_r1
    TrainTask1_round0--Shared-->AggregateTask

    TrainTask1_r1--Local-->TestTask1_r1
    TrainTask1_r1--Shared-->TestTask1_r1

    TrainTask2_r1--Local-->TestTask2_r1
    TrainTask2_r1--Shared-->TestTask2_r1

to:

flowchart LR
    InitTask1_round0--Local-->TestTask1_r0
    InitTask2_round0--Local-->TestTask2_r0

    InitTask1_round0--Local-->TrainTask1_r1
    InitTask2_round0--Local-->TrainTask2_r1

    TrainTask2_r1--Shared-->AggregateTask
    TrainTask1_r1--Shared-->AggregateTask
    TrainTask1_r1--Local-->TestTask1_r1
    TrainTask2_r1--Local-->TestTask2_r1

    TrainTask1_r1--Local-->TrainTask1_r2
    TrainTask2_r1--Local-->TrainTask2_r2
    AggregateTask--Shared-->TrainTask1_r2
    AggregateTask--Shared-->TrainTask2_r2

    TrainTask1_r2--Local-->TestTask1_r2
    TrainTask2_r2--Local-->TestTask2_r2

Changed

• BREAKING: algo are now passed as parameter to the strategy and not to execute_experiement anymore (#98)
• BREAKING A strategy need to implement a new method build_graph to build the graph of tasks to be execute in execute_experiment (#98)
• BREAKING: predict method of strategy has been renamed to perform_predict (#98)
• Test tasks don't take a shared as input anymore (#89)
• BREAKING: change eval_frequency default value to None to avoid confusion with hidden default value (#91)
• BREAKING: rename Algo to Function (#82)
• BREAKING: clarify EvaluationStrategy arguments: change rounds to eval_frequency and eval_rounds (#85)
• replace schemas.xxx by substra.schemas.xxx (#105)

Fixed

• BREAKING: Given local code dependencies are now copied to the level of the running script systematically (#99)
• Docker images are pruned in main check of Github Action to free disk space while test run (#102)
• Pass aggregation_lr to the parent class for Scaffold. Fix issue 103 (#104)

Removed

• from substra import schemas in aggregation_node.py, test_data_node.py and train_data_node.py (#105)

chore: release 0.35.0rc1

Added

• Initialization task to each strategy in SubstraFL. (#89)

This allows to load the Algo and all its attributes to the platform before any training? Once on the platform, we can perform a testing task before any training.
This init task consists in submitting an empty function, coded in the BaseAlgoclass.

@remote
def initialize(self, shared_states):
     return

The init task return a local output that will be passed as input to a test task, and to the first train task.

The graph pass from:

flowchart LR
    TrainTask1_round0--Local-->TestTask1_r0
    TrainTask1_round0--Shared-->TestTask1_r0

 TrainTask2_round0--Shared-->AggregateTask
    TrainTask2_round0--Local-->TestTask2_r0
    TrainTask2_round0--Shared-->TestTask2_r0

    AggregateTask--Shared-->TrainTask1_r1
    TrainTask1_round0--Local-->TrainTask1_r1

    AggregateTask--Shared-->TrainTask2_r1
    TrainTask2_round0--Local-->TrainTask2_r1
    TrainTask1_round0--Shared-->AggregateTask

    TrainTask1_r1--Local-->TestTask1_r1
    TrainTask1_r1--Shared-->TestTask1_r1

    TrainTask2_r1--Local-->TestTask2_r1
    TrainTask2_r1--Shared-->TestTask2_r1

to:

flowchart LR
    InitTask1_round0--Local-->TestTask1_r0
    InitTask2_round0--Local-->TestTask2_r0

    InitTask1_round0--Local-->TrainTask1_r1
    InitTask2_round0--Local-->TrainTask2_r1

    TrainTask2_r1--Shared-->AggregateTask
    TrainTask1_r1--Shared-->AggregateTask
    TrainTask1_r1--Local-->TestTask1_r1
    TrainTask2_r1--Local-->TestTask2_r1

    TrainTask1_r1--Local-->TrainTask1_r2
    TrainTask2_r1--Local-->TrainTask2_r2
    AggregateTask--Shared-->TrainTask1_r2
    AggregateTask--Shared-->TrainTask2_r2

    TrainTask1_r2--Local-->TestTask1_r2
    TrainTask2_r2--Local-->TestTask2_r2

Changed

• BREAKING: algo are now passed as parameter to the strategy and not to execute_experiement anymore (#98)
• BREAKING A strategy need to implement a new method build_graph to build the graph of tasks to be execute in execute_experiment (#98)
• BREAKING: predict method of strategy has been renamed to perform_predict (#98)
• Test tasks don't take a shared as input anymore (#89)
• BREAKING: change eval_frequency default value to None to avoid confusion with hidden default value (#91)
• BREAKING: rename Algo to Function (#82)
• BREAKING: clarify EvaluationStrategy arguments: change rounds to eval_frequency and eval_rounds (#85)
• replace schemas.xxx by substra.schemas.xxx (#105)

Fixed

• BREAKING: Given local code dependencies are now copied to the level of the running script systematically (#99)
• Docker images are pruned in main check of Github Action to free disk space while test run (#102)
• Pass aggregation_lr to the parent class for Scaffold. Fix issue 103 (#104)

Removed

• from substra import schemas in aggregation_node.py, test_data_node.py and train_data_node.py (#105)

0.34.0

0.34.0 - 2023-02-20

Added

• Possibility to test on an organization where no training has been performed (#74)
• Add contributing, contributors & code of conduct files (#68)

Removed

• Test only field for datasamples (#67)

Changed

• Remove RemoteDataMethod and change RemoteMethod class to be fully flexible regarding function name.
  The substra-tools methods is now generic, and load the inputs depending on the inputs dictionary content (#59)
• BREAKING: rename tuple to task (#79)

0.34.0rc1

0.34.0 - 2023-02-06

Added

• Possibility to test on an organization where no training have been performed (#74)
• Add contributing, contributors & code of conduct files (#68)

Removed

• chore: Test only field for datasamples (#67)

Changed

• Remove RemoteDataMethod and change RemoteMethod class to be fully flexible regarding function name.
  The substra-tools methods is now generic, and load the inputs depending on the inputs dictionary content (#59)
• BREAKING: rename tuple to task (#79)