In short, MPLC enables to:
- simulate collaborative multi-partner ML scenarios
- experiment and benchmark multi-partner learning approaches
- experiment and benchmark contributivity measurement approaches
This work focuses on collaborative data science projects where multiple partners want to train a model on multiple datasets, contributed by different data providing partners. Such scenarios are sometimes referenced as cross-silos federated learning (see for example the reference paper Advances and Open Problems in Federated Learning).
In the context of cross-silos federated learning scenarios, it addresses the following questions:
- how to train and test a predictive model, what federated learning strategies could be considered? Federated averaging seems to have become the default approach, but other could be explored.
- how to measure how much each dataset contributed to the performance of the collective model? This question can be of interest in some cases, for example as a basis to agree on how to share the reward of the ML challenge or the future revenues derived from the predictive model, or to detect possible corrupted datasets or partners not playing by the rules.
This library aims to help researchers and practitioners to explore these questions.
This work is being carried out in the context of collaborative research projects, open science and open source software. It is work in progress.
It depends in what capacity you are interested! For example:
- If you'd like to experiment right now by yourself multi-partner learning approaches and contributivity measurement methods, jump to section Run an experiment.
- If you'd like to get in touch with active members of the workgroup, jump to section Contacts, contributions, collaborations. If you are a student or a teacher, we are used to discuss student projects related to the
mplc
library. - If you are already familiar with this type of projects, you can either have a look at section Ongoing work and improvement plan or head towards issues and PRs to see what's going on these days. We use the
help wanted
tag to flag issues on which help is particularly wanted, but other open issues would also very much welcome contributions. There is also aCONTRIBUTING.md
with indications and best practices we recommend.
Should you have any question, don't hesitate reach out, we'll be happy to discuss how we could help.
In this repository, we benchmark different distributed learning and contributivity measurement approaches on public datasets artificially partitioned in a number of individual datasets, to mock a collaborative ML project (a cross-silos federated learning project).
The public datasets currently supported are: MNIST, CIFAR10, TITANIC, ESC50 and IMDB. They also act as templates for using the library on custom datasets.
The documentation is here
This library can be schematically broken down into 3 blocks:
A key capability is to easily define and simulate different multi-partner settings to be able to experiment on them. For that, the library enables to configure scenarios by specifying the number of partners, how the dataset is partitioned among them, etc. See the first tutorial, and the related documentation's section Definition of collaborative scenarios for all available parameters.
Once a given scenario is configured, it is then possible to configure the multi-partner learning approach. So far, 3 different approaches are implemented (federated averaging, sequential, sequential averaging). See the related documentation section Configuration of the collaborative and distributed learning for descriptive schemas and additional ML-related parameters.
Finally, with a given scenario and selected multi-partner learning approaches, it becomes possible to address contributivity measurement approaches. See the related documentation's sections Configuration of contributivity measurement methods to be tested and Contributivity measurement approaches studied and implemented.
pip install mplc
This installs the last packaged version available on PyPI.
If you want to install mplc from the repository, make sure that you got the latest version of pip
.
Then clone the repository, and trigger the installation from local sources.
git clone https://github.com/LabeliaLabs/distributed-learning-contributivity.git
cd distributed-learning-contributivity
pip install -e .
There are two ways to run an experiment of multi-partner learning approaches and/or contributivity methods.
You can first use the mplc library in a notebook, or regular python script, as it is demonstrated in the tutorials and in the below code snippet.
import mplc
from mplc.experiment import Experiment
from mplc.scenario import Scenario
# Let's configure different multi-partner scenarios
scenario1 = Scenario(partners_count=3,
amounts_per_partner=[0.2, 0.3, 0.5],
dataset_name='cifar10',
epoch_count=10,
minibatch_count=3,
contributivity_methods=["Shapley values", "S-Model"]
)
scenario2 = Scenario(4, [0.25]*4) # Here attributes (e.g. dataset, epochs...) will be default values
scenario3 = Scenario(4, [0.8, 0.1, 0.05, 0.05])
# Now let's instantiate an Experiment and add the above scenarios
my_exp = Experiment(experiment_name='my_first_experiment',
nb_repeats=10,
scenarios_list=[scenario1, scenario3],
)
my_exp.add_scenario(scenario2)
# Everything is now set to run the Experiment
my_exp.run()
Alternatively, you can also use the main.py
provided in the repository together with a .yml
config file.
- Define your scenario(s) in the
config.yml
file by changing the values of the suggested parameters of the example scenario (you can browse more available parameters in the documentation). For example:
experiment_name: my_custom_experiment
n_repeats: 5
scenario_params_list:
- dataset:
'mnist':
- 'random_initialization'
'cifar10':
- 'random_initialization'
dataset_proportion:
- 0.5
partners_count:
- 3
amounts_per_partner:
- [0.4, 0.3, 0.3]
samples_split_option:
- 'random'
- ['advanced', [[7, 'shared'], [6, 'shared'], [2, 'specific']]]
multi_partner_learning_approach:
- 'fedavg'
aggregation:
- 'data-volume'
- 'uniform'
contributivity_methods:
- ["Shapley values", "Independent scores", "TMCS"]
epoch_count:
- 20
minibatch_count:
- 20
gradient_updates_per_pass_count:
- 8
Under scenario_params_list
, enter a list of sets of scenario(s). Each set starts with - dataset:
and must have only one partners_count
value. The length of amount_per_partners
, corrupted_datasets
(and samples_split_option
when the advanced definition is used) must match the partner_counts
value. If for a given parameter multiple values are specified, e.g. like for aggregation
in the example scenario above, all possible combinations of parameters will be assembled as separate scenarios and run.
-
Then execute
main.py -f config.yml
. Add the-v
argument if you want a more verbose output. -
A
results.csv
file will be generated in a new folder for your experiment under/experiments/<your_experiment>
. You can read this rawresults.csv
file or use the notebooks in/notebooks
.
Note: example experiment(s) are stored in folder /saved_experiments
to illustrate the use of the library. The notebooks include graphs, like for example the following:
We defined 5 reference scenarios on which we propose to test and benchmark the different multi-partner learning approaches and contributivity measurement methods.
The 5 reference scenarios are described on the following schema (link to editable version):
In brief:
-
Scenarios 1 and 2 with 2 partners only enable simple baselines, with different data splits (each partner having samples of different classes in the first scenario, of all classes both in the second one), and introducing corrupted data;
-
Scenario 3 proposes a more realistic configuration, with partners having both samples of common classes and samples of different classes each;
-
Scenario 4 offers a case with 5 partners, and an identical distribution of data samples of all classes, but with 1 partner having its data entirely corrupted;
-
Scenario 5 is more complex, with 11 partners and a mix of identical distribution of data samples of several classes for a majority of partners, data samples of different classes for certain other partners, and corrupted data also.
Results of experiments and benchmarks of multi-partner learning approaches and contributivity methods on the reference scenarios will be summarised in this section in the upcoming months. Associated notebooks and full results will be shared on Open Science Framework.
The current work focuses on the following 4 priorities:
- Design and implement new multi-partner learning approaches
- Design and implement new contributivity measurement methods
- Perform experiments and gain experience about best-suited multi-partner learning approaches and contributivity measurement methods in different situations
- Make the library agnostic/compatible with other datasets and model architectures
There is also a transverse, continuous improvement effort on code quality, readability, optimization.
This work is collaborative, enthusiasts are welcome to comment open issues and PRs or open new ones.
Should you be interested in this open effort and would like to share any question, suggestion or input, you can use the following channels:
- This Github repository (issues or PRs)
- Labelia Labs' (ex- Substra Foundation) Slack workspace, channel
#workgroup-mpl-contributivity
- Email: contact@labelia.org