v0.8.1

Release v0.8.1

IMPORTANT: fluke is now in beta!

New features

• The federation orchestrator is now the federated algorithm and not the server
• Decentralized FL: base class for implementing a fully decentralized FL algorithm is now available (DecentralizedFL). Standard Gossip Learning (GossipDFL) has also been added.
• Added TimedMessage class
• Added Topology class
• Added documentation on evaluation configuration
• Added receive_all and has_message to Channel

Fixes

• Fixed documentation issues
• Fixed issue with the personalized model of personalized algorithms
• Fixed GEAR hyper-paramaters setting
• Fixed _last_round on Client which was off by 1

Deprecations

• The fit and finalize methods of Server are now deprecated because the orchestration is now performed by the algorithm's class

v0.7.9

Release v0.7.9

This release fixes many minor issues and introduced multi-gpu parallel training for clients (experimental features).

New features

• [Experimental] It is now possible to perform parallel clients training using multiple GPU (via torch.distributed)
• Now the logging and the tracking of the evaluation performances are done using two different classes
• The evaluation log now includes more information, like the support (number of clients involved in the computation of the mean/sum) and/or the round

Minor changes

• Revised code formatting
• Collection[Client] -> Sequence[Client]

Fixes

• Fixed issue in FedProto related to the generator of client models
• Fixed different sections in the documentation
• Fixed issue with the logging of the evaluation performance

v0.7.8

Release notes 0.7.8

This release fixes many minor issues and introduced multi-gpu training for clients (experimental features).

New features and improvements

• [Exprimental] It is now possible to perform the training client-side using multiple GPU (via torch.nn.DataParallel)
• The channel now notify also when messages are sent/broadcasted
• It is now possible to set a global logger in FlukeENV
• All classes/methods related to the configuration is now in the submodule fluke.config

Major changes

• Iterable -> Collection
• No more server reference on the client
• Receiver and sender of messages are no longer references to client/server but are identifiers (int/str). The server is identified with the string "server" and the client by its ID
• _ObjectRef -> ObjectRef
• Client/Server: no more _notify_xyz, but a single _notify with argument event

Fixes

• Fixed issue on the Server that caused to select 0 clients
• Fixed issue with clients-only when logging on wandb
• Fixed code style and documentation
• Fixed requirements.txt
• Added a clean close to ClearMLLog

v0.7.4

Release notes 0.7.4

This release adds some improvements to and fix some minor issues

Fixed issues

• Removed useless _load_from_cache from some clients
• Other minor fixes including docs and tests

Improvements

• Improved ClassificationEval that is now extensible with custom metrics
• The fluke command accepts now overriding arguments (via hydra)
• Added configuration validation with cerberus

New features

• It is now possible to install fluke via Docker
• Added notification for message sent, broadcasted and for the early stopping of the federation
• Added DebugLog to have a more verbose logger
• Added ConfigurationError to handle these types of errors

v0.7.2

Fixed issues

• Fixed issue related to torch.device witth centralized and clients-only command
• Fixed minor issues with the API documentation

Improvements

• Improved start up speed
• Improved some printing

New features

• Added fluke ASCII art banner
• Added possibility to early stop the federation
• Added possibility to track custom value with the logger

v0.7.1

Fixed issues

• Fixed issue with inmemory=false in the experiment configuration

New features

• Added FlukeCache
• Added FL algorithm FedLD
• Added FL algorithm FAT
• Added FL algorithm GEAR
• Added client-side gradient clipping

Important modifications

• ⚠️ From this version fluke does not support python <= 3.9
• GeneralSettings is now FlukeENV
• The arguments of the fluke commands are changed

v0.5.1

Fixed issues

• Fixed issue with FEMNIST (#classes 64 -> 62) and an inefficiency on DataSplitter.assign
• Fixed issue related to batch normalization layers

New features

• Added option to unload models on disk to save memory (inmemory)
• Added fine_tuning_epochs to clients so the fine-tuning is now possible with a custom number of local epochs
• Added output_size argument to some nets
• Added the possibility to pass arguments to build the model through the configuration (net_args option)
• Added Saving options
• Added FL algorithm DPFedAVG

v0.4.0

Fixed issues

• Fixed FedNH, FedHP and FedProto
• Fixed issues on Dirichlet data distribution when balanced=True
• Removed the required output_size from the networks

New features

• Added get_activation_size in utils
• Added dataset FCUBE from https://arxiv.org/pdf/2102.02079

v0.3.4

Fixed issues

• Fixed problem with the cuda:N setting. Now it works properly and selects the correct device.
• Fixed warning related to the redundancy of wandb.require("core") in the last versions of the wand library

Known issues with torch

If you get the error AttributeError: module 'torch.serialization' has no attribute 'add_safe_globals' please upgrade your torch installation to the latest version.

Important note

Release 0.3.3 has been published and deleted right after due to an error!

v0.3.2

Fixed issues

• Fixed problem with the validation of the eval configuration key
• Fixed minor issue with AllLayerOutputModel class