Rationale: the model must be an instance, while "MyModel" has the
naming convention for a class.

BIGAUD Nathan authored 1 year ago and ANDREY Paul committed 1 year ago

* Separating `split_data` and `quickrun`
* Improving doc and leveraging better existing doc
* Using fire for parsing

- Move functional data loading/saving utils to it.
- Properly deprecate `InMemoryDataset.(load|save)_data_array`.
- Improve docstrings formatting and exhaustiveness.

BIGAUD Nathan authored 1 year ago and ANDREY Paul committed 1 year ago

* Creating "ModelConfig","DataSplitConfig", and "ExperimentConfig"
* Relocating and updating parser to accept toml file as input
* Update split data to use config input
* Modularization and updating of run.py
* Small changes to MNIST toml example

* Add exceptions and outputs catching to multiprocessed routines.
* Change the output signature of `run_as_processes` to return a bool
  flag indicating success, and a list of routine-wise output value or
  RuntimeError (that may either result from an actual failure or from
  the process having been interrupted due to another one's failure).
* Add `auto_stop` parameter to enable disabling the default automated
  interruption of processes once any of them has failed.

* Add optional parameters to the Python functions and the argparser
  so as to enable running the UCI-Heart-Dataset examples over an
  actual network.
* Make the documentation more verbose about these options, and the
  difference between localhost-based simulations and real-network
  use.
* Rationale: having a minimal example to run is a valuable feature
  when deploying declearn to a new machines~network setting, so as
  to quickly assess (and/or debug) the framework's usability.

BIGAUD Nathan authored 2 years ago and ANDREY Paul committed 2 years ago

Details :
* In `/optimizer`, `/metrics`, and `/aggregator` :`name` and `aux_name`
* In `/communications` : `protocol`
* In `/dataset` : `_type_key`
* In `/data_info` : `field`, `type`, and `doc`

Excluding all classes decorated with `@dataclasses.dataclass`,
see https://stackoverflow.com/a/52099457

* Rewrite the `declearn.main.utils.Checkpointer` util.
  - Implement Model, Optimizer and Metric-output results saving and loading methods.
  - Add a `max_history` parameter to garbage-collect past files.
  - Revise the labeling of output files to ease collection and re-use.
* Revise the use of `Checkpointer` in `FederatedClient` and `FederatedServer`.
  - Have both entities checkpoint at the end of evaluation rounds.
  - For now, clients save the global model (not local one), local optimizer and metrics.
  - For now, server saves the global model (not local ones), server optimizer and all
    metrics: aggregated ones and client-wise ones (in separate files).
  - Move lowest-loss-yielding weights tracking and restoring to the private backend
    of `FederatedServer`, and ensure they are still saved and sent to clients at the
    end of the federated learning process.
* Add unit tests for the `Checkpointer` class.

* Add `declearn.main.config` submodule:
  - Move dataclasses (`TrainingConfig`, `EvaluateConfig`
    and `RegisterConfig`) to this new submodule.
  - Implement `FLRunConfig`, that binds together all arguments and
    hyper-parameters used by the `FederatedServer.run` method, and
    provides with the possibility to write them up as a TOML config
    file (by inheriting `declearn.utils.TomlConfig`).
* Deploy `FLRunConfig` to `FederatedServer`:
  - Replace the argument to `run` and `async_run` by `FLRunConfig`,
    adding the possibility to either provide (legacy-style) kwargs
    or the path to a TOML file to the `run` one.
  - Replace the argument to `initialize` by the full `FLRunConfig`,
    to facilitate overriding by subclasses that may need to send or
    use information on the process in addition to conducting client
    registration.
  - Preserve the use of targetted fields of hyper-parameters in the
    `training_round` and `evaluation_round` methods.

BIGAUD Nathan authored 2 years ago and ANDREY Paul committed 2 years ago

Co-authored-by: Paul Andrey <paul.andrey@inria.fr>