Quickrun mode

Summary

1. Build a quickrun mode, with a simple example to start using the package. Two use cases:

• A way to discover the codebase in a guided way
• A config system for repeating experiments

2. Make the quickstart section of the doc the sole entry point for a user to discover the lib, and improve the doc to be easier to use.

3.Introduce a data splitting utility

Done (detailed) :

v0

• Run
  • Create main run file using multiprocessing
  • Make multiprocessing a main util and revamp for mac compatibility
• TOML
  • Write the TOML file for MNIST
  • Add options in from_toml method of Config classes
• Create a model file and MyModel in a clean way
• Data loading : generalize split_data
  • Update the data loading function to follow our naming
  • remove csv export
  • Split the data between train and test
  • Generalize data loading function
  • Take out with to csv and data format of fedbiomed
  • Review the argparser
  • Check as_array accepts Sparse inputs > it does not
• Add run from package using the main .toml project.scripts (source)

Prep Magnet au Lac

• Finish Debug:
  • pytorch version of the model (data input shape issue)
• Refine TOML and argparsers:
  • Provide a TOML instead of a full folder as input to quickrun
  • Add custom names for files in TOML (see if can pickup name from file name)
  • Add parsing of Client and Server kwargs in TOML
    • Checkpointer (incl single expe destination in toml)
    • metric
  • Genenerate client network from server TOML and deal with client name more elegantly
  • Remove argparser in _split_data
• Dataset
  • Add shuffling in label split function
• Test all options in config toml
  • If there is an existing data split, and you request a new one, check for the new one to decide if splitting is needed
  • Parse metrics
• Simplify MNIST default model
• Allow for empty fields in TOML
• Improve doc (including toml comments)

Integrate feedback from Magnet au Lac:

• Separate the data split functionality
  • use python fire parser (not solved for now)
  • See if I can do add toml and fire at the same time : yes I can (https://dev.to/bowmanjd/build-a-command-line-interface-with-poetry-and-python-fire-c8b_)
• Properly document quickrun
  • Split quickrun and example doc
  • Rewrite quickstart section
  • Write a 'how to use TOML' entry point (strat in quickstrat, move it if needed)
  • Do collab
  • Modify doc for full follow up (Should I use the original objects or the config ?) > No need
• Clarify absolute vs relative path
• Lint
• Rebase -i

To do

P1

• Drop full list of client names and infer ?
• Update heart UCI to MNIST
• Add global seed > opening a separate branch and MR
• Update collab for with a more scalable solution and losing the reference to the experimental branch

P2

• Checkpointer saves a lot of things by default - could trim it for a quickrun
• Use "dataclass_from_func" in building my config for clarity of doc
• Literal["iid", "labels", "biased"] not parsed in DataSplitConfig
• Add torch and sklearn examples
• Reintroduce DataSplitConfig (find a way to use Doc of parent class)
• Remaining 'bug': Slow pytorch runtime