Flower 1.5.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

achiverram28, Adam Narozniak, Anass Anhari, Charles Beauville, Dana-Farber, Daniel J. Beutel, Daniel Nata Nugraha, Edoardo Gabrielli, eunchung, Gustavo Bertoli, Heng Pan, Javier, Mahdi, Ruth Galindo, Steven Hé (Sīchàng), Taner Topal

What's new?

• Introduce new simulation engine (#1969, #2221, #2248)

  The new simulation engine has been rewritten from the ground up, yet it remains fully backwards compatible. It offers much improved stability and memory handling, especially when working with GPUs. Simulations transparently adapt to different settings to scale simulation in CPU-only, CPU+GPU, multi-GPU, or multi-node multi-GPU environments.

  Comprehensive documentation includes a new how-to run simulations guide, new simulation-pytorch and simulation-tensorflow notebooks, and a new YouTube tutorial series.

• Restructure Flower Docs (#1824, #1865, #1884, #1887, #1919, #1922, #1920, #1923, #1924, #1962, #2006, #2133, #2203, #2215, #2122, #2223, #2219, #2232, #2233, #2234, #2235, #2237, #2238, #2242, #2231, #2243, #2227)

  Much effort went into a completely restructured Flower docs experience. The documentation on flower.dev/docs is now divided into Flower Framework, Flower Baselines, Flower Android SDK, Flower iOS SDK, and code example projects.

• Introduce Flower Swift SDK (#1858, #1897)

  This is the first preview release of the Flower Swift SDK. Flower support on iOS is improving, and alongside the Swift SDK and code example, there is now also an iOS quickstart tutorial.

• Introduce Flower Android SDK (#2131)

  This is the first preview release of the Flower Kotlin SDK. Flower support on Android is improving, and alongside the Kotlin SDK and code example, there is now also an Android quickstart tutorial.

• Introduce new end-to-end testing infrastructure (#1842, #2071, #2072, #2068, #2067, #2069, #2073, #2070, #2074, #2082, #2084, #2093, #2109, #2095, #2140, #2137, #2165)

  A new testing infrastructure ensures that new changes stay compatible with existing framework integrations or strategies.

• Deprecate Python 3.7

  Since Python 3.7 reached its end of life (EOL) on 2023-06-27, support for Python 3.7 is now deprecated and will be removed in an upcoming release.

• Add new FedTrimmedAvg strategy (#1769, #1853)

  The new FedTrimmedAvg strategy implements Trimmed Mean by Dong Yin, 2018.

• Introduce start_driver (#1697)

  In addition to start_server and using the raw Driver API, there is a new start_driver function that allows for running start_server scripts as a Flower driver with only a single-line code change. Check out the mt-pytorch code example to see a working example using start_driver.

• Add parameter aggregation to mt-pytorch code example (#1785)

  The mt-pytorch example shows how to aggregate parameters when writing a driver script. The included driver.py and server.py have been aligned to demonstrate both the low-level way and the high-level way of building server-side logic.

• Migrate experimental REST API to Starlette (2171)

  The (experimental) REST API used to be implemented in FastAPI, but it has now been migrated to use Starlette directly.

  Please note: The REST request-response API is still experimental and will likely change significantly over time.

• Introduce experimental gRPC request-response API (#1867, #1901)

  In addition to the existing gRPC API (based on bidirectional streaming) and the experimental REST API, there is now a new gRPC API that uses a request-response model to communicate with client nodes.

  Please note: The gRPC request-response API is still experimental and will likely change significantly over time.

• Replace the experimental start_client(rest=True) with the new start_client(transport="rest") (#1880)

  The (experimental) start_client argument rest was deprecated in favour of a new argument transport. start_client(transport="rest") will yield the same behaviour as start_client(rest=True) did before. All code should migrate to the new argument transport. The deprecated argument rest will be removed in a future release.

• Add a new gRPC option (#2197)

  We now start a gRPC server with the grpc.keepalive_permit_without_calls option set to 0 by default. This prevents the clients from sending keepalive pings when there is no outstanding stream.

• Improve example notebooks (#2005)

  There's a new 30min Federated Learning PyTorch tutorial!

• Example updates (#1772, #1873, #1981, #1988, #1984, #1982, #2112, #2144, #2174, #2225, #2183)

  Many examples have received significant updates, including simplified advanced-tensorflow and advanced-pytorch examples, improved macOS compatibility of TensorFlow examples, and code examples for simulation. A major upgrade is that all code examples now have a requirements.txt (in addition to pyproject.toml).

• General improvements (#1872, #1866, #1884, #1837, #1477, #2171)

  Flower received many improvements under the hood, too many to list here.

Incompatible changes

None

Flower 1.4.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Adam Narozniak, Alexander Viala Bellander, Charles Beauville, Chenyang Ma (Danny), Daniel J. Beutel, Edoardo, Gautam Jajoo, Iacob-Alexandru-Andrei, JDRanpariya, Jean Charle Yaacoub, Kunal Sarkhel, L. Jiang, Lennart Behme, Max Kapsecker, Michał, Nic Lane, Nikolaos Episkopos, Ragy, Saurav Maheshkar, Semo Yang, Steve Laskaridis, Steven Hé (Sīchàng), Taner Topal

What's new?

• Introduce support for XGBoost (FedXgbNnAvg strategy and example) (#1694, #1709, #1715, #1717, #1763, #1795)

  XGBoost is a tree-based ensemble machine learning algorithm that uses gradient boosting to improve model accuracy. We added a new FedXgbNnAvg strategy, and a code example that demonstrates the usage of this new strategy in an XGBoost project.

• Introduce iOS SDK (preview) (#1621, #1764)

  This is a major update for anyone wanting to implement Federated Learning on iOS mobile devices. We now have a swift iOS SDK present under src/swift/flwr that will facilitate greatly the app creating process. To showcase its use, the iOS example has also been updated!

• Introduce new "What is Federated Learning?" tutorial (#1657, #1721)

  A new entry-level tutorial in our documentation explains the basics of Fedetated Learning. It enables anyone who's unfamiliar with Federated Learning to start their journey with Flower. Forward it to anyone who's interested in Federated Learning!

• Introduce new Flower Baseline: FedProx MNIST (#1513, #1680, #1681, #1679)

  This new baseline replicates the MNIST+CNN task from the paper Federated Optimization in Heterogeneous Networks (Li et al., 2018). It uses the FedProx strategy, which aims at making convergence more robust in heterogenous settings.

• Introduce new Flower Baseline: FedAvg FEMNIST (#1655)

  This new baseline replicates an experiment evaluating the performance of the FedAvg algorithm on the FEMNIST dataset from the paper LEAF: A Benchmark for Federated Settings (Caldas et al., 2018).

• Introduce (experimental) REST API (#1594, #1690, #1695, #1712, #1802, #1770, #1733)

  A new REST API has been introduced as an alternative to the gRPC-based communication stack. In this initial version, the REST API only supports anonymous clients.

  Please note: The REST API is still experimental and will likely change significantly over time.

• Improve the (experimental) Driver API (#1663, #1666, #1667, #1664, #1675, #1676, #1693, #1662, #1794)

  The Driver API is still an experimental feature, but this release introduces some major upgrades. One of the main improvements is the introduction of an SQLite database to store server state on disk (instead of in-memory). Another improvement is that tasks (instructions or results) that have been delivered will now be deleted. This greatly improves the memory efficiency of a long-running Flower server.

• Fix spilling issues related to Ray during simulations (#1698)

  While running long simulationa, ray was sometimes spilling huge amounts of data that would make the training unable to continue. This is now fixed! 🎉

• Add new example using TabNet and Flower (#1725)

  TabNet is a powerful and flexible framework for training machine learning models on tabular data. We now have a federated example using Flower: https://github.com/adap/flower/tree/main/examples/tabnet.

• Add new how-to guide for monitoring simulations (#1649)

  We now have a documentation guide to help users monitor their performance during simulations.

• Add training metrics to History object during simulations (#1696)

  The fit_metrics_aggregation_fn can be used to aggregate training metrics, but previous releases did not save the results in the History object. This is now the case!

• General improvements (#1659, #1646, #1647, #1471, #1648, #1651, #1652, #1653, #1659, #1665, #1670, #1672, #1677, #1684, #1683, #1686, #1682, #1685, #1692, #1705, #1708, #1711, #1713, #1714, #1718, #1716, #1723, #1735, #1678, #1750, #1753, #1736, #1766, #1760, #1775, #1776, #1777, #1779, #1784, #1773, #1755, #1789, #1788, #1798, #1799, #1739, #1800, #1804, #1805)

  Flower received many improvements under the hood, too many to list here.

Incompatible changes

None

Flower 1.3.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Adam Narozniak, Alexander Viala Bellander, Charles Beauville, Daniel J. Beutel, JDRanpariya, Lennart Behme, Taner Topal

What's new?

• Add support for workload_id and group_id in Driver API (#1595)

  The (experimental) Driver API now supports a workload_id that can be used to identify which workload a task belongs to. It also supports a new group_id that can be used, for example, to indicate the current training round. Both the workload_id and group_id enable client nodes to decide whether they want to handle a task or not.

• Make Driver API and Fleet API address configurable (#1637)

  The (experimental) long-running Flower server (Driver API and Fleet API) can now configure the server address of both Driver API (via --driver-api-address) and Fleet API (via --fleet-api-address) when starting:

  flower-server --driver-api-address "0.0.0.0:8081" --fleet-api-address "0.0.0.0:8086"

  Both IPv4 and IPv6 addresses are supported.

• Add new example of Federated Learning using fastai and Flower (#1598)

  A new code example (quickstart_fastai) demonstrates federated learning with fastai and Flower. You can find it here: quickstart_fastai.

• Make Android example compatible with flwr >= 1.0.0 and the latest versions of Android (#1603)

  The Android code example has received a substantial update: the project is compatible with Flower 1.0 and later, the UI received a full refresh, and the project is updated to be compatible with newer Android tooling.

• Add new FedProx strategy (#1619)

  This strategy is almost identical to FedAvg, but helps users replicate what is described in this paper. It essentially adds a parameter called proximal_mu to regularize the local models with respect to the global models.

• Add new metrics to telemetry events (#1640)

  An updated event structure allows, for example, the clustering of events within the same workload.

• Add new custom strategy tutorial section #1623

  The Flower tutorial now has a new section that covers implementing a custom strategy from scratch: Open in Colab

• Add new custom serialization tutorial section (#1622)

  The Flower tutorial now has a new section that covers custom serialization: Open in Colab

• General improvements (#1638, #1634, #1636, #1635, #1633, #1632, #1631, #1630, #1627, #1593, #1616, #1615, #1607, #1609, #1608, #1603, #1590, #1580, #1599, #1600, #1601, #1597, #1595, #1591, #1588, #1589, #1587, #1573, #1581, #1578, #1574, #1572, #1586)

  Flower received many improvements under the hood, too many to list here.

• Updated documentation (#1629, #1628, #1620, #1618, #1617, #1613, #1614)

  As usual, the documentation has improved quite a bit. It is another step in our effort to make the Flower documentation the best documentation of any project. Stay tuned and as always, feel free to provide feedback!

Incompatible changes

None

Flower 1.2.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Adam Narozniak, Charles Beauville, Daniel J. Beutel, Edoardo, L. Jiang, Ragy, Taner Topal, dannymcy

What's new?

• Introduce new Flower Baseline: FedAvg MNIST (#1497, #1552)

  Over the coming weeks, we will be releasing a number of new reference implementations useful especially to FL newcomers. They will typically revisit well known papers from the literature, and be suitable for integration in your own application or for experimentation, in order to deepen your knowledge of FL in general. Today's release is the first in this series. Read more.

• Improve GPU support in simulations (#1555)

  The Ray-based Virtual Client Engine (start_simulation) has been updated to improve GPU support. The update includes some of the hard-earned lessons from scaling simulations in GPU cluster environments. New defaults make running GPU-based simulations substantially more robust.

• Improve GPU support in Jupyter Notebook tutorials (#1527, #1558)

  Some users reported that Jupyter Notebooks have not always been easy to use on GPU instances. We listened and made improvements to all of our Jupyter notebooks! Check out the updated notebooks here:

  • An Introduction to Federated Learning
  • Strategies in Federated Learning
  • Building a Strategy
  • Client and NumPyClient

• Introduce optional telemetry (#1533, #1544, #1584)

  After a request for feedback from the community, the Flower open-source project introduces optional collection of anonymous usage metrics to make well-informed decisions to improve Flower. Doing this enables the Flower team to understand how Flower is used and what challenges users might face.

  Flower is a friendly framework for collaborative AI and data science. Staying true to this statement, Flower makes it easy to disable telemetry for users that do not want to share anonymous usage metrics. Read more..

• Introduce (experimental) Driver API (#1520, #1525, #1545, #1546, #1550, #1551, #1567)

  Flower now has a new (experimental) Driver API which will enable fully programmable, async, and multi-tenant Federated Learning and Federated Analytics applications. Phew, that's a lot! Going forward, the Driver API will be the abstraction that many upcoming features will be built on - and you can start building those things now, too.

  The Driver API also enables a new execution mode in which the server runs indefinitely. Multiple individual workloads can run concurrently and start and stop their execution independent of the server. This is especially useful for users who want to deploy Flower in production.

  To learn more, check out the mt-pytorch code example. We look forward to you feedback!

  Please note: The Driver API is still experimental and will likely change significantly over time.

• Add new Federated Analytics with Pandas example (#1469, #1535)

  A new code example (quickstart_pandas) demonstrates federated analytics with Pandas and Flower. You can find it here: quickstart_pandas.

• Add new strategies: Krum and MultiKrum (#1481)

  Edoardo, a computer science student at the Sapienza University of Rome, contributed a new Krum strategy that enables users to easily use Krum and MultiKrum in their workloads.

• Update C++ example to be compatible with Flower v1.2.0 (#1495)

  The C++ code example has received a substantial update to make it compatible with the latest version of Flower.

• General improvements (#1491, #1504, #1506, #1514, #1522, #1523, #1526, #1528, #1547, #1549, #1560, #1564, #1566)

Flower received many improvements under the hood, too many to list here.

• Updated documentation (#1494, #1496, #1500, #1503, #1505, #1524, #1518, #1519, #1515)

  As usual, the documentation has improved quite a bit. It is another step in our effort to make the Flower documentation the best documentation of any project. Stay tuned and as always, feel free to provide feedback!

  One highlight is the new first time contributor guide: if you've never contributed on GitHub before, this is the perfect place to start!

Incompatible changes

None

Flower 1.1.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Akis Linardos, Christopher S, Daniel J. Beutel, George, Jan Schlicht, Mohammad Fares, Pedro Porto Buarque de Gusmão, Philipp Wiesner, Rob Luke, Taner Topal, VasundharaAgarwal, danielnugraha, edogab33

What's new?

• Introduce Differential Privacy wrappers (preview) (#1357, #1460)

  The first (experimental) preview of pluggable Differential Privacy wrappers enables easy configuration and usage of differential privacy (DP). The pluggable DP wrappers enable framework-agnostic and strategy-agnostic usage of both client-side DP and server-side DP. Head over to the Flower docs, a new explainer goes into more detail.

• New iOS CoreML code example (#1289)

  Flower goes iOS! A massive new code example shows how Flower clients can be built for iOS. The code example contains both Flower iOS SDK components that can be used for many tasks, and one task example running on CoreML.

• New FedMedian strategy (#1461)

  The new FedMedian strategy implements Federated Median (FedMedian) by Yin et al., 2018.

• Log Client exceptions in Virtual Client Engine (#1493)

  All Client exceptions happening in the VCE are now logged by default and not just exposed to the configured Strategy (via the failures argument).

• Improve Virtual Client Engine internals (#1401, #1453)

  Some internals of the Virtual Client Engine have been revamped. The VCE now uses Ray 2.0 under the hood, the value type of the client_resources dictionary changed to float to allow fractions of resources to be allocated.

• Support optional Client/NumPyClient methods in Virtual Client Engine

  The Virtual Client Engine now has full support for optional Client (and NumPyClient) methods.

• Provide type information to packages using flwr (#1377)

  The package flwr is now bundled with a py.typed file indicating that the package is typed. This enables typing support for projects or packages that use flwr by enabling them to improve their code using static type checkers like mypy.

• Updated code example (#1344, #1347)

  The code examples covering scikit-learn and PyTorch Lightning have been updated to work with the latest version of Flower.

• Updated documentation (#1355, #1558, #1379, #1380, #1381, #1332, #1391, #1403, #1364, #1409, #1419, #1444, #1448, #1417, #1449, #1465, #1467)

  There have been so many documentation updates that it doesn't even make sense to list them individually.

• Restructured documentation (#1387)

  The documentation has been restructured to make it easier to navigate. This is just the first step in a larger effort to make the Flower documentation the best documentation of any project ever. Stay tuned!

• Open in Colab button (#1389)

  The four parts of the Flower Federated Learning Tutorial now come with a new Open in Colab button. No need to install anything on your local machine, you can now use and learn about Flower in your browser, it's only a single click away.

• Improved tutorial (#1468, #1470, #1472, #1473, #1474, #1475)

  The Flower Federated Learning Tutorial has two brand-new parts covering custom strategies (still WIP) and the distinction between Client and NumPyClient. The existing parts one and two have also been improved (many small changes and fixes).

Incompatible changes

None

Flower 1.0.0

Highlights

• Stable Virtual Client Engine (accessible via start_simulation)
• All Client/NumPyClient methods are now optional
• Configurable get_parameters
• Tons of small API cleanups resulting in a more coherent developer experience

Thanks to our contributors

We would like to give our special thanks to all the contributors who made Flower 1.0 possible (in reverse GitHub Contributors order):

@rtaiello, @g-pichler, @rob-luke, @andreea-zaharia, @kinshukdua, @nfnt, @tatiana-s, @TParcollet, @vballoli, @negedng, @RISHIKESHAVAN, @hei411, @SebastianSpeitel, @AmitChaulwar, @Rubiel1, @FANTOME-PAN, @Rono-BC, @lbhm, @sishtiaq, @remde, @Jueun-Park, @architjen, @PratikGarai, @mrinaald, @zliel, @MeiruiJiang, @sandracl72, @gubertoli, @Vingt100, @MakGulati, @cozek, @jafermarq, @sisco0, @akhilmathurs, @CanTuerk, @mariaboerner1987, @pedropgusmao, @tanertopal, @danieljanes.

Incompatible changes

• All arguments must be passed as keyword arguments (#1338)

  Pass all arguments as keyword arguments, positional arguments are not longer supported. Code that uses positional arguments (e.g., start_client("127.0.0.1:8080", FlowerClient())) must add the keyword for each positional argument (e.g., start_client(server_address="127.0.0.1:8080", client=FlowerClient())).

• Introduce configuration object ServerConfig in start_server and start_simulation (#1317)

  Instead of a config dictionary {"num_rounds": 3, "round_timeout": 600.0}, start_server and start_simulation now expect a configuration object of type flwr.server.ServerConfig. ServerConfig takes the same arguments that as the previous config dict, but it makes writing type-safe code easier and the default parameters values more transparent.

• Rename built-in strategy parameters for clarity (#1334)

  The following built-in strategy parameters were renamed to improve readability and consistency with other API's:

  • fraction_eval --> fraction_evaluate
  • min_eval_clients --> min_evaluate_clients
  • eval_fn --> evaluate_fn

• Update default arguments of built-in strategies (#1278)

  All built-in strategies now use fraction_fit=1.0 and fraction_evaluate=1.0, which means they select all currently available clients for training and evaluation. Projects that relied on the previous default values can get the previous behaviour by initializing the strategy in the following way:

  strategy = FedAvg(fraction_fit=0.1, fraction_evaluate=0.1)

• Add server_round to Strategy.evaluate (#1334)

  The Strategy method evaluate now receives the current round of federated learning/evaluation as the first parameter.

• Add server_round and config parameters to evaluate_fn (#1334)

  The evaluate_fn passed to built-in strategies like FedAvg now takes three parameters: (1) The current round of federated learning/evaluation (server_round), (2) the model parameters to evaluate (parameters), and (3) a config dictionary (config).

• Rename rnd to server_round (#1321)

  Several Flower methods and functions (evaluate_fn, configure_fit, aggregate_fit, configure_evaluate, aggregate_evaluate) receive the current round of federated learning/evaluation as their first parameter. To improve reaability and avoid confusion with random, this parameter has been renamed from rnd to server_round.

• Move flwr.dataset to flwr_baselines (#1273)

  The experimental package flwr.dataset was migrated to Flower Baselines.

• Remove experimental strategies (#1280)

  Remove unmaintained experimental strategies (FastAndSlow, FedFSv0, FedFSv1).

• Rename Weights to NDArrays (#1258, #1259)

  flwr.common.Weights was renamed to flwr.common.NDArrays to better capture what this type is all about.

• Remove antiquated force_final_distributed_eval from start_server (#1258, #1259)

  The start_server parameter force_final_distributed_eval has long been a historic artefact, in this release it is finally gone for good.

• Make get_parameters configurable (#1242)

  The get_parameters method now accepts a configuration dictionary, just like get_properties, fit, and evaluate.

• Replace num_rounds in start_simulation with new config parameter (#1281)

  The start_simulation function now accepts a configuration dictionary config instead of the num_rounds integer. This improves the consistency between start_simulation and start_server and makes transitioning between the two easier.

New features

• Support Python 3.10 (#1320)

  The previous Flower release introduced experimental support for Python 3.10, this release declares Python 3.10 support as stable.

• Make all Client and NumPyClient methods optional (#1260, #1277)

  The Client/NumPyClient methods get_properties, get_parameters, fit, and evaluate are all optional. This enables writing clients that implement, for example, only fit, but no other method. No need to implement evaluate when using centralized evaluation!

• Enable passing a Server instance to start_simulation (#1281)

  Similar to start_server, start_simulation now accepts a full Server instance. This enables users to heavily customize the execution of eperiments and opens the door to running, for example, async FL using the Virtual Client Engine.

• Update code examples (#1291, #1286, #1282)

  Many code examples received small or even large maintenance updates, among them are

  • scikit-learn
  • simulation_pytorch
  • quickstart_pytorch
  • quickstart_simulation
  • quickstart_tensorflow
  • advanced_tensorflow

• Remove the obsolete simulation example (#1328)

  Removes the obsolete simulation example and renames quickstart_simulation to simulation_tensorflow so it fits withs the naming of simulation_pytorch

• Update documentation (#1223, #1209, #1251, #1257, #1267, #1268, #1300, #1304, #1305, #1307)

  One substantial documentation update fixes multiple smaller rendering issues, makes titles more succinct to improve navigation, removes a deprecated library, updates documentation dependencies, includes the flwr.common module in the API reference, includes support for markdown-based documentation, migrates the changelog from .rst to .md, and fixes a number of smaller details!

• Minor updates

  • Add round number to fit and evaluate log messages (#1266)
  • Add secure gRPC connection to the advanced_tensorflow code example (#847)
  • Update developer tooling (#1231, #1276, #1301, #1310)
  • Rename ProtoBuf messages to improve consistency (#1214, #1258, [#1259](https://github.com/a...

0.19.0

What's new:

• Flower Baselines (preview): FedOpt, FedBN, FedAvgM (919, 1127, 914)

  The first preview release of Flower Baselines has arrived! We're kickstarting Flower Baselines with implementations of FedOpt (FedYogi, FedAdam, FedAdagrad), FedBN, and FedAvgM. Check the documentation on how to use Flower Baselines. With this first preview release we're also inviting the community to contribute their own baselines.

• C++ client SDK (preview) and code example (1111)

  Preview support for Flower clients written in C++. The C++ preview includes a Flower client SDK and a quickstart code example that demonstrates a simple C++ client using the SDK.

• Add experimental support for Python 3.10 and Python 3.11 (1135)

  Python 3.10 is the latest stable release of Python and Python 3.11 is due to be released in October. This Flower release adds experimental support for both Python versions.

• Aggregate custom metrics through user-provided functions (1144)

  Custom metrics (e.g., accuracy) can now be aggregated without having to customize the strategy. Built-in strategies support two new arguments, fit_metrics_aggregation_fn and evaluate_metrics_aggregation_fn, that allow passing custom metric aggregation functions.

• User-configurable round timeout (1162)

  A new configuration value allows the round timeout to be set for start_server and start_simulation. If the config dictionary contains a round_timeout key (with a float value in seconds), the server will wait at least round_timeout seconds before it closes the connection.

• Enable both federated evaluation and centralized evaluation to be used at the same time in all built-in strategies (1091)

  Built-in strategies can now perform both federated evaluation (i.e., client-side) and centralized evaluation (i.e., server-side) in the same round. Federated evaluation can be disabled by setting fraction_eval to 0.0.

• Two new Jupyter Notebook tutorials (1141)

  Two Jupyter Notebook tutorials (compatible with Google Colab) explain basic and intermediate Flower features:

  An Introduction to Federated Learning: Open in Colab

  Using Strategies in Federated Learning: Open in Colab

• New FedAvgM strategy (Federated Averaging with Server Momentum) (1076)

  The new FedAvgM strategy implements Federated Averaging with Server Momentum [Hsu et al., 2019].

• New advanced PyTorch code example (1007)

  A new code example (advanced_pytorch) demonstrates advanced Flower concepts with PyTorch.

• New JAX code example (906, 1143)

  A new code example (jax_from_centralized_to_federated) shows federated learning with JAX and Flower.

• Minor updates

  • New option to keep Ray running if Ray was already initialized in start_simulation (1177)
  • Add support for custom ClientManager as a start_simulation parameter (1171)
  • New documentation for implementing strategies (1097, 1175)
  • New mobile-friendly documentation theme (1174)
  • Limit version range for (optional) ray dependency to include only compatible releases (>=1.9.2,<1.12.0) (1205)

Incompatible changes:

• Remove deprecated support for Python 3.6 (871)
• Remove deprecated KerasClient (857)
• Remove deprecated no-op extra installs (973)
• Remove deprecated proto fields from FitRes and EvaluateRes (869)
• Remove deprecated QffedAvg strategy (replaced by QFedAvg) (1107)
• Remove deprecated DefaultStrategy strategy (1142)
• Remove deprecated support for eval_fn accuracy return value (1142)
• Remove deprecated support for passing initial parameters as NumPy ndarrays (1142)

0.18.0

What's new?

• Improved Virtual Client Engine compatibility with Jupyter Notebook / Google Colab (866, 872, 833, 1036)

  Simulations (using the Virtual Client Engine through start_simulation) now work more smoothly on Jupyter Notebooks (incl. Google Colab) after installing Flower with the simulation extra (pip install flwr[simulation]).

• New Jupyter Notebook code example (833)

  A new code example (quickstart_simulation) demonstrates Flower simulations using the Virtual Client Engine through Jupyter Notebook (incl. Google Colab).

• Client properties (feature preview) (795)

  Clients can implement a new method get_properties to enable server-side strategies to query client properties.

• Experimental Android support with TFLite (865)

  Android support has finally arrived in main! Flower is both client-agnostic and framework-agnostic by design. One can integrate arbitrary client platforms and with this release, using Flower on Android has become a lot easier.

  The example uses TFLite on the client side, along with a new FedAvgAndroid strategy. The Android client and FedAvgAndroid are still experimental, but they are a first step towards a fully-fledged Android SDK and a unified FedAvg implementation that integrated the new functionality from FedAvgAndroid.

• Make gRPC keepalive time user-configurable and decrease default keepalive time (1069)

  The default gRPC keepalive time has been reduced to increase the compatibility of Flower with more cloud environments (for example, Microsoft Azure). Users can configure the keepalive time to customize the gRPC stack based on specific requirements.

• New differential privacy example using Opacus and PyTorch (805)

  A new code example (opacus) demonstrates differentially-private federated learning with Opacus, PyTorch, and Flower.

• New Hugging Face Transformers code example (863)

  A new code example (quickstart_huggingface) demonstrates usage of Hugging Face Transformers with Flower.

• New MLCube code example (779, 1034, 1065, 1090)

  A new code example (quickstart_mlcube) demonstrates usage of MLCube with Flower.

• SSL-enabled server and client (842, 844, 845, 847, 993, 994)

  SSL enables secure encrypted connections between clients and servers. This release open-sources the Flower secure gRPC implementation to make encrypted communication channels accessible to all Flower users.

• Updated FedAdam and FedYogi strategies (885, 895)

  FedAdam and FedAdam match the latest version of the Adaptive Federated Optimization paper.

• Initialize start_simulation with a list of client IDs (860)

  start_simulation can now be called with a list of client IDs (clients_ids, type: List[str]). Those IDs will be passed to the client_fn whenever a client needs to be initialized, which can make it easier to load data partitions that are not accessible through int identifiers.

• Minor updates

  • Update num_examples calculation in PyTorch code examples in (909)
  • Expose Flower version through flwr.__version__ (952)
  • start_server in app.py now returns a History object containing metrics from training (974)
  • Make max_workers (used by ThreadPoolExecutor) configurable (978)
  • Increase sleep time after server start to three seconds in all code examples (1086)
  • Added a new FAQ section to the documentation (948)
  • And many more under-the-hood changes, library updates, documentation changes, and tooling improvements!

Incompatible changes:

• Removed flwr_example and flwr_experimental from release build (869)

  The packages flwr_example and flwr_experimental have been deprecated since Flower 0.12.0 and they are not longer included in Flower release builds. The associated extras (baseline, examples-pytorch, examples-tensorflow, http-logger, ops) are now no-op and will be removed in an upcoming release.

0.17.0

What's new?

• Experimental virtual client engine (781 790 791)

  One of Flower's goals is to enable research at scale. This release enables a first (experimental) peek at a major new feature, codenamed the virtual client engine. Virtual clients enable simulations that scale to a (very) large number of clients on a single machine or compute cluster. The easiest way to test the new functionality is to look at the two new code examples called quickstart_simulation and simulation_pytorch.

  The feature is still experimental, so there's no stability guarantee for the API. It's also not quite ready for prime time and comes with a few known caveats. However, those who are curious are encouraged to try it out and share their thoughts.

• New built-in strategies (828 822)

  • FedYogi - Federated learning strategy using Yogi on server-side. Implementation based on https://arxiv.org/abs/2003.00295
  • FedAdam - Federated learning strategy using Adam on server-side. Implementation based on https://arxiv.org/abs/2003.00295

• New PyTorch Lightning code example (617)

• New Variational Auto-Encoder code example (752)

• New scikit-learn code example (748)

• New experimental TensorBoard strategy (789)

• Minor updates

  • Improved advanced TensorFlow code example (769)
  • Warning when min_available_clients is misconfigured (830)
  • Improved gRPC server docs (841)
  • Improved error message in NumPyClient (851)
  • Improved PyTorch quickstart code example (852)

Incompatible changes:

• Disabled final distributed evaluation (800)

  Prior behaviour was to perform a final round of distributed evaluation on all connected clients, which is often not required (e.g., when using server-side evaluation). The prior behaviour can be enabled by passing force_final_distributed_eval=True to start_server.

• Renamed q-FedAvg strategy (802)

  The strategy named QffedAvg was renamed to QFedAvg to better reflect the notation given in the original paper (q-FFL is the optimization objective, q-FedAvg is the proposed solver). Note the the original (now deprecated) QffedAvg class is still available for compatibility reasons (it will be removed in a future release).

• Deprecated and renamed code example simulation_pytorch to simulation_pytorch_legacy (791)

  This example has been replaced by a new example. The new example is based on the experimental virtual client engine, which will become the new default way of doing most types of large-scale simulations in Flower. The existing example was kept for reference purposes, but it might be removed in the future.

0.16.0

What's new?

• New built-in strategies (#549)

  • (abstract) FedOpt
  • FedAdagrad

• Custom metrics for server and strategies (#717)

  The Flower server is now fully task-agnostic, all remaining instances of task-specific metrics (such as :code:accuracy) have been replaced by custom metrics dictionaries. Flower 0.15 introduced the capability to pass a dictionary containing custom metrics from client to server. As of this release, custom metrics replace task-specific metrics on the server.

  Custom metric dictionaries are now used in two user-facing APIs: they are returned from Strategy methods :code:aggregate_fit/:code:aggregate_evaluate and they enable evaluation functions passed to build-in strategies (via :code:eval_fn) to return more than two evaluation metrics. Strategies can even return aggregated metrics dictionaries for the server to keep track of.

  Stratey implementations should migrate their :code:aggregate_fit and :code:aggregate_evaluate methods to the new return type (e.g., by simply returning an empty :code:{}), server-side evaluation functions should migrate from :code:return loss, accuracy to :code:return loss, {"accuracy": accuracy}.

  Flower 0.15-style return types are deprecated (but still supported), compatibility will be removed in a future release.

• Migration warnings for deprecated functionality (#690)

  Earlier versions of Flower were often migrated to new APIs, while maintaining compatibility with legacy APIs. This release introduces detailed warning messages if usage of deprecated APIs is detected. The new warning messages often provide details on how to migrate to more recent APIs, thus easing the transition from one release to another.

• Improved docs and docstrings (#691, #692, #713)

• MXNet example and documentation

• FedBN implementation in example PyTorch: From Centralized To Federated (#696, #702, #705)

Incompatible changes:

• Serialization-agnostic server (#721)

  The Flower server is now fully serialization-agnostic. Prior usage of class :code:Weights (which represents parameters as deserialized NumPy ndarrays) was replaced by class :code:Parameters (e.g., in :code:Strategy). :code:Parameters objects are fully serialization-agnostic and represents parameters as byte arrays, the :code:tensor_type attributes indicates how these byte arrays should be interpreted (e.g., for serialization/deserialization).

  Built-in strategies implement this approach by handling serialization and deserialization to/from :code:Weights internally. Custom/3rd-party Strategy implementations should update to the slighly changed Strategy method definitions. Strategy authors can consult PR #721 to see how strategies can easily migrate to the new format.

• Deprecated :code:flwr.server.Server.evaluate, use :code:flwr.server.Server.evaluate_round instead (#717)