Releases: LabeliaLabs/distributed-learning-contributivity
Releases · LabeliaLabs/distributed-learning-contributivity
v0.4.2
v0.4.1
Contributors
celinejacques
Assets 2
v0.4.0
Changes
Main changes vs. v0.3.1:
- #355 & #356: Upgrade to TF 2.5.1
- #354: Add DRFA/AFL learning approach
- #353: Plot data distribution
- #352: Add single partner learning to available methods
- #350: Add fashion mnist dataset
- #345: Federated averaging with local and global optimizations
- #342: Conduct first benchmarks on reference scenarios
- #338: Configure nightly builds
- #337: Change results saving behaviour with results saved by default
- #335: Clarify documentation
- #326: Update to Pytest 6.2.2
- #322: Improve log format
- #321: Define and describe reference scenarios for experiments and benchmarks
- #320: Create a fully flexible Splitter enabling more advanced and realistic Scenarios
- #318: Fix bug with aggregation_weighting schemes
- #315: Faster versions of FedAvg, FedGrads and Smodel
Contributors
Many thanks to the following contributors and participants:
Contributors
bowni, kat-leen, and 9 other contributors
Assets 2
v.0.3.1
v0.3.0
MPLC v0.3.0 Release notes
This release introduces several changes to the library which is now offering more modularity. It is also deployed on PyPI.
Features
- Multiple code refactorization: Refacto the multi-partner learning approaches in a more object oriented way. #255
- New
Experimentobject: An object which runs and repeats several scenarios and gathers results to simplify their analysis. #275 - New corruption methods and options: Refacto the corruption in a more object oriented way. Add
random,permutation,duplication,redundancyways to corrupt a dataset. #280 & #277 - Updated datasets integration: Change module architecture. The datasets are now a subclass of the
Datasetclass. #262 - Updated tutorial notebooks. #266
- Documentation: Update and format documentation. #264 & #294
- Add code coverage badge: A new badge on the README, to follow the extension of the test coverage. #268
- More tests added: There is now end-to-end tests for contributivity methods and multi-partner learning approaches. #304 & #300
- Add
FederatedGradient: New multi-partner learning method. The gradients are aggregated (instead of the weights). Then the optimizer updates the model with the aggregated gradient. #299 - Add S-model: New multi-partner learning method. It adds for each partner a NoiseAdaptative layer, which is supposed to adapt to potential label flip and thus improve the shared model performances, even with corrupted partners.
#281 & #301 - Add local/global option for test and validation datasets. The dataset is split (only once! and in a stratified way) between train/test/validation datasets. If
‘local’, these test/validation datasets are splitted between partners, in the exact same way than the training one. #288
Fixes
- Fix bugs in corruption by permutation: A transposition was missing. #284
- Update Cifar10 model optimizer: The learning rate was way too low. We also switched from RMSprop to Adam for the optimizer, as it showed better performances. #283
- Fix early stopping: Condition for early stopping was not reachable. #279
- Logs: Add epoch number when showing evaluation metrics. #274
- Dependencies updates: Removal of Keras (We now exclusively use the tensorflow's built-in Keras module), upgrade to Tensor Flow v2.4.0, Numpy v1.19.4, Scikit Learn v0.23.2, Pandas v1.1.5. #290
- Switch to Github Actions instead of Travis for Continuous Integration. #298
- Normalize contributivity scores in step-by-step methods. #295
Contributors
This release received successful contributions from: