A distributed learning contributivity package
Project description
Simulate collaborative multi-partner ML projects & Experiment learning and contributivity measurement approaches
Introduction
In collaborative data science projects partners sometimes need to train a model on multiple datasets, contributed by different data providing partners. In such cases the partners might have to measure how much each dataset involved contributed to the performance of the model. This is useful 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. We explore this question and the opportunity to implement some mechanisms helping partners in such scenarios to measure each dataset's contributivity (as contribution to the performance of the model).
Context of this work
This work is being carried out in the context of collaborative research projects. It is work in progress. We would like to share it with various interested parties, research and business partners to get their feedback and potential contributions. This is why it is shared as open source content on Substra Foundation’s repositories.
How to interact with this?
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 the workgroup, jump to section Contacts, contributions, collaborations. If you are a student or a teacher, we love discussing student projects!
- If you are very familiar with this type of projects, well 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
Should you have any question, reach out and we'll be happy to discuss how we could help.
About this repository
In this repository, we benchmark different contributivity measurement approaches on a public dataset artificially partitioned in a number of individual datasets, to mock a collaborative ML project.
The objective is to compare the contributivity figures obtained with the different approaches, and try to see how potential differences could be interpreted.
The public datasets currently supported are: MNIST, CIFAR10, TITANIC, ESC50 and IMDB.
Structure of the library
This library can be broken down into 3 blocks:
- Scenarios
- Multi-partner learning approaches
- Contributivity measurement approaches
Scenarios
From the start of this work it seemed very useful to be able to 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, what are the relative volume of data they have, how the data are distributed among them, etc. (see our first Tutorial, and the related documentation's section Definition of collaborative scenarios for all available parameters).
Multi-partner learning approaches
Once a given scenario configured, it seemed useful to choose how the multi-partner learning would be done. So far, 3 different approaches are implemented (federated averaging, sequential learning, sequential averaging). See below section Configuration of the collaborative and distributed learning for descriptive schemas and additional ML-related parameters.
Contributivity measurement approaches
Finally, with given scenarios and multi-partner learning approaches, we can address contributivity measurement approaches. See below sections Configuration of contributivity measurement methods to be tested and Contributivity measurement approaches studied and implemented.
Run an experiment
There are two ways to run an experiment. You can first use the mplc library in a notebook, or regular python script, as it is demonstrated in the tutorials.
For instance :
Using this library
from subtest.scenario import Scenario
my_scenario = Scenario(partners_count=3,
amounts_per_partner=[0.2, 0.3, 0.5],
dataset_name='mnist',
epoch_count=10,
minibatch_count=3)
my_scenario.run()
# (...) Results' exploration
You can also use the main.py provided in the repository, with a yaml config file.
Using the main.py script
-
Define your mock scenario(s) in the
config.yml
file by changing the values of the suggested parameters of the 2 example scenarios (you can browse more available parameters in section Config file parameters below). For example:experiment_name: my_custom_experiment n_repeats: 5 scenario_params_list: - dataset_name: 'mnist': - 'random_initialization' 'cifar10': - 'random_initialization' partners_count: - 3 amounts_per_partner: - [0.4, 0.3, 0.3] samples_split_option: - ['advanced', [[7, 'shared'], [6, 'shared'], [2, 'specific']]] multi_partner_learning_approach: - 'fedavg' aggregation_weighting: - 'data_volume' - 'uniform' epoch_count: - 38 methods: - ["Shapley values", "Independent scores", "TMCS"] minibatch_count: - 20 gradient_updates_per_pass_count: - 8 dataset_proportion: - 1 - dataset_name: - 'mnist' partners_count: - 2 amounts_per_partner: - [0.5, 0.5] samples_split_option: - ['basic', 'random'] - ['basic', 'stratified'] multi_partner_learning_approach: - 'fedavg' aggregation_weighting: - 'data_volume' - 'uniform' epoch_count: - 38 methods: - ["Shapley values", "SMCS", "IS_lin_S", "IS_reg_S"] minibatch_count: - 20 gradient_updates_per_pass_count: - 8 dataset_proportion: - 1
Under
scenario_params_list
, enter a list of sets of scenario(s). Each set starts with- dataset_name:
and must have only onepartners_count
value. The length ofamount_per_partners
,corrupted_datasets
(andsamples_split_option
when the advanced definition is used) must match thepartner_counts
value. If for a given parameter multiple values are specified, e.g. like foragregation_weighting
in the first scenario set of the above example, 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:
Ongoing work and improvement plan
The current work focuses on the following 4 priorities:
- Improve the multi-partner learning approaches
- Continue developing 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.
Contacts, contributions, collaborations
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)
- Substra Foundation's Slack workspace, channel
#workgroup-mpl-contributivity
- Email: hello@substra.org
- Come meet with us at La Paillasse (Paris, France), Le Palace (Nantes, France) or Studio Iconosquare (Limoges, France)
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.