v0.2.7

PySyft 0.2.7 Release Notes

New functionality:

• Implement msgpack serde for RoleAssignments (#3595 by @Jasopaum)
• Create AbstractSendable class to provide comms methods generically (#3611 by @karlhigley)
• Support serialization of lists in Arg (#3617 by @vvmnnnkv and @karlhigley)
• Add FV homomorphic encryption scheme (#3276 @IamRavikantSingh)
• Proto messaging support (#3638 by @tudorcebere)
• BaseDataset proto support (#3641 by @tudorcebere & @karlhigley)
• Add iterable behavior for PointerTensor (#3659 by @gmuraru)
• Add connection speed test for workers (#3609 by @vkkhare)
• Add mean and var methods for AST (#3682 by @LaRiffle)
• Add new gradient functions to Autograd (#3681 by @LaRiffle)
• Implemented Subtraction operation of FV Scheme (#3775 by @IamRavikantSingh)
• Add a mechanism for arbitrary remote exec on Python workers (#3689 by @LaRiffle)
• Add Multiplication operation of FV scheme. (#3711 by @IamRavikantSingh)

Bug fixes

• Fix federated_avg side effect (#3614 @joaolcaas)
• serde, torch: fix typo in numpy tensor serializer. (#3644 by @shubham3121)
• Fixed precision tensor support (#3650 by @tudorcebere & @karlhigley)
• Fix: Copy issue with ciphertext in FV scheme. (#3672 by @IamRavikantSingh)
• Fix dtype issues in AST (#3679 by @LaRiffle)
• Fix decorator issues for FSS in AST (#3680 by @LaRiffle)
• Fix bug in set_model_params() in Create Plan (#3781 @sachin-101)
• Fix grid model hosting for large models (#3719 by @IonesioJunior)
• Fix Plan trace for inplace ops, add Role actions pruning (#3798 by @vvmnnnkv and @LaRiffle)
• Fix issues in handle_func_command for syft tensors (#3807 by @LaRiffle)

Refactoring:

• Use serde for tags serialization in PointerTensor (#3566 by @gmuraru)
• Move torch arg hooking into the relevant classes to reduce circular imports (#3612 by @karlhigley)
• Remove protocol parameter in decrypt (#3599 by @youben11)
• Add utilities for hooking and chaining methods calls along tensor chain (#3636 by @karlhigley)
• Add an __init__.py file to syft.generic.abstract (#3642 by @karlhigley)
• Move permissions from TorchTensor.send() to PrivateTensor._before_send hook (#3643 by @karlhigley)
• Remove owner attribute from FrameworkWrapper (#3652 by @Jasopaum)
• Move tests for generic functionality out of test/torch/tensors dir (#3645 by @karlhigley)
• Clean up AbstractSendable to remove extraneous methods (#3660 by @karlhigley)
• Improve @hookable decorator to allow modifying return values after (#3655 by @karlhigley)
• Adding support in msgpack for external libs. (#3653 by @tudorcebere)
• Extract MessageHandler classes from BaseWorker (#3661 by @karlhigley)
• Change whitelist to allow_list (#3675 by @karlhigley)
• Update Operations: For non same size polynomials (#3684 @IamRavikantSingh )
• Removed wrap() on non-tensor pointers in respond_to_search() (#3685 by @AlanAboudib)
• feat: add plan translation to tfjs (#3371 @Nolski)
• Update hook_args list (#3677 by @LaRiffle)
• Customizable timeout in workers (#3760 by @youben11)
• Serialize Plan's roles and base_framework fields (#3801 by @vvmnnnkv and @gmuraru)
• Grid module architecture refactoring (#3818 by @hericlesme)

Docs & tutorials:

• Update CODEOWNERS file to reflect new team names (#3620 by @karlhigley)
• Update Part 11 - Secure Deep Learning Classification.ipynb (#3444 by @znreza)
• Simplify encrypted aggregation in the tutorials (#3666 by @bobsonlin26 )

Tests & CI

• Delete pull_request_template.md by @cereallarceny
• Register the translation mark with PyTest to suppress warnings (#3618 by @karlhigley)
• Update requirements.txt with tornado compatible notebook version (#3606 by @rajathpatel23)
• Update the label used to mark stale issues/PRs (#3630 by @karlhigley)
• Added functionality to create a cluster using autoscale API (#3637 @rimijoker)

v0.2.6

PySyft 0.2.6 Release Notes

Breaking changes since v0.2.5:

• Message.contents field has now been fully removed after previous deprecation (#3477 by @karlhigley)
• TrainConfig has been fully removed (#3532 by @karlhigley)

New functionality:

• Add initial support for auto-scaling PyGrid workers and notebooks on Google Cloud Platform (#3437 and #3483 by @rimijoker)
• Add asynchronous send_command method to WebsocketClient (#3454 by @LaRiffle and @Syzygianinfern0)
• Build Protocol Roles from @func2protocol decorator (#3488 by @Jasopaum)
• Trace CommunicationActions into Protocol Roles (#3448 by @Prtfw)
• Add a reset() method on Role (#3495 by @Jasopaum)
• Add Plan argument type-checking along with support for lists and dicts as Plan arguments (#3386 by @tudorcebere)
• Support additional ML/DL frameworks during Plan tracing (#3461 by @gmuraru)
• Provide a way to trace module-level methods like torch.rand() into Plan/Protocol Roles (#3522 by @karlhigley)
• Allow Protocol Roles to have State for initial values (#3499 by @Jasopaum)
• Add ability to seed the Syft IdProvider and safely toggle verbose mode (#3457 by @Syzygianinfern0)
• Add support for WebRTC connections between peer Workers (#3548 and several more by @IonesioJunior)
• Add Protobuf serialization support for torch.dtype, syft.generic.String, and PointerDataset (#3584, #3594, #3600 by @tudorcebere)
• Print useful information on syft.grid.register() (#3589 by @IonesioJunior)
• Add RoleAssignments to Protocols (#3544 by @Jasopaum)
• Add ways to fetch Grid network peers to Network and WebRTCManager #3582 by @joaolcaas)
• Host private datasets locally with Grid (#3596 by @IonesioJunior)
• Add a method to return the supported ML/DL frameworks in BaseWorker (#3535 by @shubham3121)

Bug fixes:

• Fix the type of the allowed users parameter to PrivateTensors (#3446 by @gmuraru)
• Fix PointerTensor.move() when the source and destination workers are the same (#3442 by @Syzygianinfern0)
• Remove stray objects from FSS Worker/Plan initialization (#3525 by @karlhigley)
• Fix Batchnorm for Federated Learning (#3472 by @LaRiffle)
• Add scaffolding for Federated Learning client worker API (#3426 by @vvmnnnkv)
• Avoid registering Grid peer Workers multiple times (#3591 by @jefersonf)
• Remove test_spinup_time which has been flaky (#3603 by @robert-wagner)

Refactoring:

• Clean up Syft implementation of RNNs (#3441 by @dcato98)
• Store logged message in BaseWorker in an unserialized format for readability (#3473 by @karlhigley)
• Update Role.register_action() to allow registering CommunicationActions (#3465 by @karlhigley)
• Refactor BaseWorker.send_command() to accept explicit arguments (#3487 by @shubham3121)
• Add a SyftSerializable base class for Syft classes that need to be serialized to msgpack or Protobuf (#3386 by @tudorcebere)
• Convert ObjectStore to a stand-alone class instead of a base class for Workers (#3484 by @karlhigley)
• Extract @abstractmethods from BaseWorker to AbstractWorker (#3519 by @karlhigley)
• Create a VirtualWorker for each Plan/Protocol Role and use it for tracing (#3521 by @karlhigley)
• Improve the serde interface with lazy data loaders (#3530 by @tudorcebere)
• Split FrameworkHook into multiple parts (#3561 by @karlhigley)
• Clean up the code style of PySyft (#3535 by @gmuraru)

Examples and Tutorials:

• Add an experimental notebook for how to create a Plan that trains a NN on MNIST using Syft autograd (#3567 by @vvmnnnkv)

Translations:

• Updates to Bengali translation of Parts 4-5, 12-13 (#3415 by @jabertuhin, #3445, #3449, #3466, and #3500 by @znreza)

Documentation:

• Update the Pull Request templates (#3506 by @tkrsh, #3520 by @karlhigley, additional work by @cereallarceny)

Tests and Builds:

• Add a check for docstring coverage to the Github builds (#3462 and #3505 by @karlhigley)
• Remove verbose flag from code format check to highlight necessary changes (#3475 by @karlhigley)

v0.2.5

PySyft v0.2.5 Release Notes

Breaking changes since v0.2.4:

• CRTPrecisionTensor and LargePrecisionTensor have been removed (#2982 by @Yugandhartripathi)
• Autograd for sqrt has been disabled due to garbage collection issues (#3387 by @Syzygianinfern0)
• Support for custom dtype in SecureNN has been dropped (#3392 by @Syzygianinfern0)
• API and attributes of Protocol have changed significantly during on-going rework (#3409 by @Jasopaum)

New functionality:

• Add multi-worker computation graph support to allow gradients to be calculated on tensors moved across workers (#3083 by @LaRiffle)
• Add dropout support for SMPC (#3195 by @sukhadj)
• Add a Chebyshev approximation for sigmoid (#3205 by @gmuraru)
• Add host_federated_training_method to GridClient (#3253 by @hericlesme)
• Add a protocol kwarg to encrypt() to select between MPC and Paillier HE (#3234 by @Syzygianinfern0)
• Add device option to async_fit to enable optional GPU training on workers (#3299 by @wmlba)
• Add Plan translation to Torchscript (#3263 and #3360 by @vvmnnnkv)
• Make it possible to send and get datasets (#3266 by @abogaziah)
• Add requires_grad and no_wrap kwargs to encrypt() for MPC secret sharing (#3368 @Syzygianinfern0)
• Add buffers iterator for tensor elements of torch.nn.module (#3273 by @ratmcu)
• Increase field size to 2**64 in AdditiveSharingTensor and FixedPrecisionTensor (#2982 by @Yugandhartripathi)
• Add Syft version checking for Grid nodes (#3380 by @IonesioJunior)
• Make Placeholders forward unknown methods to child tensors (#3377 by @gmuraru)
• Add average and max-pooling with SMPC (#3399 by @abogaziah)
• Add an initial implementation of Function Secret Sharing using Plans (#3057 by @LaRiffle)
• Make Plan tracing work with backward() and autograd (#3383 by @vvmnnnkv)
• Make SecureNN work with more than 3 parties (#3148 by @knexator)
• Rework Protocols to trace multiple Roles from a single function (#3409 by @Jasopaum)

Bug fixes:

• Fix the Placeholder serialization (#3256 by @rotalex)
• Remove implicit in-place fix_precision for PointerTensor ( #3271 by @sukhadj)
• Fix overloading for torch.nn and torch.nn.functional modules (#3269 by @Syzygianinfern0)
• Fix indexing Syft tensors with lists (#3285 by @rotalex)
• Fix circular import and overloading in RNNs (#3286 by @Syzygianinfern0)
• Don’t assume self is in known_workers in BaseWorker (#3370 by @karlhigley)
• Fix memory leak in AdditiveSharingTensor operations (#3343 by @Syzygianinfern0)
• Fix garbage collection for AutogradTensor (#3387 by @Syzygianinfern0)
• Call native Torch functions with their original names if not overloaded (#3323 by @youben11)
• Fix a bug in the AdditiveSharingTensor.dtype serde (#3396 by @Jasopaum)
• Auto-register the Torchscript Plan translator (#3405 by @vvmnnnkv)
• Fix garbage collection in BaseWorker.execute_computation_action() and Torch in-place method identification (#3403 by @Syzygianinfern0)
• Fix a double-tracing bug in Plans (#3428 by @Jasopaum)

Refactoring:

• Decouple Plan methods from each other by removing temporary attributes (#3203 by @karlhigley)
• Consolidate run_websocket_server.py files (#3225 and 3265 by @vineetjai)
• Use PlaceholderIds in Actions and Plans instead of full Placeholders (#3223 by @Jasopaum)
• Clean up FixedPrecisionTensor and related classes (#3253 by @sachin-101)
• Extract a Role class from Plans (#3201 by @Jasopaum)
• Generalize Role to any type of Action, including computation or communication (#3295 by @karlhigley)
• Remove a redundant is_built check from Plans (#3292 by @gmuraru)
• Remove eval from Role execution with framework package store (#3295 by @karlhigley)
• Disambiguate method args and kwargs from arguments and local variables args_ and kwargs_ (#3314 by @karlhigley)
• Handle Plan state tensors as arguments to Torchscript (#3361 by @vvmnnnkv)
• Rework PlanTranslatorTorchscript so it doesn’t strip Parameters from tensor chains (#3376 by @karlhigley and @Syzygianinfern0)
• Make Plan tracing happen from the Placeholders instead of the Torch hooks (#3303 by @Jasopaum)
• Add a name field to CommunicationAction (#3352 by @Prtfw)
• Remove init_plan attribute from BaseWorker to decouple from Plans (#3404 by @karlhigley)
• Remove a value() method left over from Promises (#3418 by @Jasopaum)
• Remove deprecated properties and methods from Plan (#3408 by @karlhigley)
• Fix a warning about getargspec deprecation in Plans (#3421 by @gmuraru)

Examples and Tutorials:

• Update FL training plan notebooks to report diff to PyGrid (#3224 by @vvmnnnkv)
• Add a tutorial on model inversion attacks (#3258 by @TTitcombe)
• Fix typo in Grid Part 3 (#3351 by @sfmajors373)
• Explicitly specify encryption protocol in PaillierTensor notebook (#3353 by @Syzygianinfern0)
• Fix smart_city_aggregation in Grid Part 3 notebook (#3342 by @IonesioJunior)
• Clean up cells in Tutorial Part 13a (#3290 by @ivynasantino)

Translations:

• Add Pidgin translations of Parts 1-7, 9 (#3149 by @TemitopeOladokun)
• Add Korean translations of Part 2 (#3051 by @wonderit), Part 4 (#3270 by @seungjaeryanlee)
• Add Japanese translations of Parts 1-12 (#3182 by @kouohhashi), Part 13 (#3242 by @MADONOKOUKI)
• Add German translations of Parts 2-8,10 (#3294 and others by @JMBehnken)
• Add Marathi translations of Parts 1-7 (#3173 by @Noob-can-Compile)

Documentation:

• Update the docstrings and argument types for federated_avg method (#3251 by @DanyEle)
• Add docstrings to ObjectStorage tests (#3177 by @vineeth14)
• Fix installation command in the README (#3280 by @stephenjfox)
• Update PointerTensor docstrings (#3379 by @jaintj95)
• Update the CONTRIBUTING.md doc (#3414 by @Prince326)
• Add a link the PySyft ReadTheDocs page to the README (#3412 by @lucasshenv)

Tests and Builds:

• Start remote worker in test suite on 0.0.0.0 instead of localhost (#3218 by @rotalex)
• Add codecov integration (#3230 and others by @imskr)
• Add additional code owners for automated PR review requests (#3239, #3240 by @karlhigley)
• Run the tests in random orders to shake out flaky tests (#3127 by @karlhigley)
• Skip the main test suite for translation-only PRs (#3328 by @karlhigley)
• Add sklearn dependency for notebooks (#3339 and #3347 by @ucalyptus)
• Fix missing device param in test_federated_client.py::test_fit (#3346 by @karlhigley)
• Add dtype specification to test_pate.py::test_torch_ref_match (#3348 by @gmuraru)
• Extract State and Translator tests from Plan tests (#3378 by @karlhigley)
• Add tolerance to gradient clipping test (#3327 by @gmuraru)
• Add a PR template (#3268 by @sachin-101)
• Create a manifest file to include pip-dep/ directory for Conda package (#3420 by @systemshift)
• Increase tolerance for sigmoid approximation tests (#3427 by @gmuraru)

v0.2.4

Changes since 0.2.3

New functionality:

• Added the option to compute tanh using sigmoid or Chebyshev approximation (#3113 by @gmuraru)
• Added model parameters as a default optimizer parameter in FederatedClient (#3117 by @gconst02)
• Added an Optims class to track remote optimizers in federated learning (#3179 by @rimijoker)
• Added the ability to use a Plan input as an output (#3199 by @tudorcebere)
• Added the ability to send and get Datasets (#2960 by @abogaziah)

Refactoring:

• Extensive improvements Plan, Placeholder, and the Torch hook (#3080 by @LaRiffle)
• Split Operation class into separate message and action classes (#3132 and #3168 by @Jasopaum and @karlhigley)
• Added a new message for executing non-tensor operations on remote workers (#3048 by @midokura-silvia)
• Reworked BaseWorker's message router to use message classes in handler functions (#3175 by @Jasopaum and @karlhigley)
• Standardized the names of command message to TensorCommandMessage, PlanCommandMessage, and WorkerCommandMessage (#3198 by @karlhigley)
• Moved the Placeholder class to the execution package (#3211 by @karlhigley)
• Converted existing formatted strings to f-strings for consistency and performance (#3186 by @TTitcombe
• Replaced zstd compression library with zlib (#3150 and #3164 by @refactormyself)

Bug fixes:

• Fixed VirtualWorker message sending test (#3139 by @gmuraru)
• Fixed event loop error when starting a WebsocketServerWorker in a notebook (#3196 by @imskr and #3204 by @gmuraru)
• Stopped creating grad tensors for frozen layers (#3200 by @tudorcebere)

Documentation:

• Added homomorphic encryption to the Introduction section (#3135 by @LaRiffle)
• Updated supported PyTorch version to 1.4 (#3140 by @codeboy5)
• Added a note about support Python versions (#3154 by @J-Yash)
• Added a note about using Python 3.7 with conda (#3162 by @teddykoker)
• Added docstrings to callable pointer tests (#3130 by @steph-en-m)
• Added instructions for installing TF Encrypted (#3197 by @rimijoker)

Tutorials:

• Fixed FL training plan tutorials and added a PyGrid FL hosting example (#3185 by @vvmnnnkv)
• Cleaned up the tutorials directory structure (#3159 by @bicycleman15)
• Removed a stray cell from Part 5 - Welcome to the Sandbox (#3155 by @bicycleman15)
• Removed unnecessary requires_grad in Part 2 (#3216 by @bicycleman15)

Translations:

• Bengali:
  • Part 3 (#3133 by @adventuroussrv), Parts 4 and 5 (#3121 by @jabertuhin)
• German:
  • Part 1 (#3178 by @vineetjai)

Builds:

• Set up code owners to automatically request reviews from the relevant OpenMined teams for various parts of the PySyft codebase (#3192 and #3215 by @karlhigley, @LaRiffle, and @AlanAboudib)

v0.2.3

Changes since 0.2.2

New functionality:

• Migrates PySyft to PyTorch 1.4 (#2930 by @gmuraru)
• Implements tanh for FixedPrecisionTensors using Chebyshev approximation (#3004 by @gmuraru)
• Adds the ability to simulate latency with VirtualWorkers (#3070 by @jefersonf)
• Adds Protobuf serialization for Placeholders, Plans, and States (#2972 by @karlhigley)

Refactoring:

• Reworks Plans for smoother serialization to multiple formats (#2910 by @LaRiffle and @vvmnnnkv)
• Moves Plans, Protocols, and States from the messaging package to the execution package (#3078 by @karlhigley)
• Renames Operation class to OperationMessage (#3090 by @karlhigley)

Bug fixes:

• Fixes retrieval of the fit() result in WebsocketClientWorker (#2948 by @brandonhee)
• Fixes numeric issues in handcrafted Conv and Pool implementations (#2945 and #2964 by @arshjot)
• Removes an insecure eval in native tensor interpreter (#2951 by @karlhigley)
• Fixes PyTorch JIT tracing compatibility for Plans (#2988 by @karlhigley)
• Removes workarounds for previous versions of PyTorch (#2999 by @gmuraru)
• Pins requests dependency specification to version 2.22.0 (#2970 by @ADMoreau)
• Fixes interoperability of AutogradTensors with other tensors vis a vis requires_grad (#2998 by @gmuraru)
• Improves logging, typing, and documentation of PATE implementation (#3033 by @TTitcombe)
• Fixes a potential security issue with unsafe YAML loading (#3037 by @systemshift)
• Raises an error when attempting to additively share FloatTensors (#3094 by @pierrepocreau)
• Improves testing for Syft's RNN implementation (#3092 by @jimboH)
• Changes dependency specifications to require compatible versions (#3119 by @karlhigley)
• Fixes compatibility with msgpack 1.0 serialization library (#3067 by @IonesioJunior and #3073 by @hdodenhof)

Documentation:

• Adds Sphinx documentation (#3017 by @Benardi)
• Fixes notebook test badge (#3028 by @jefersonf)
• Adds a link to the Udacity Secure And Private AI course (#3016 by @AVJdataminer)
• Improves instructions for developing protocol changes with syft-proto (#2818 by @refactormyself)

Tutorials:

• Adds model evaluation to SplitNN tutorial (#2983 by @midokura-silvia)
• Adds a note to Part 10 FL with Secure Aggregation tutorial about hooking Numpy (#3022 by @fdroessler)

Translations:

• Bengali:
  • Parts 1, 2, 6, and 7 (#2938, #2942, #3052, and #3053 by @ucalyptus)
  • Parts 8, 9, 11, 12, and 12bis (#3096, #3050, #3041, #3039, and #3038 by @adventuroussrv)
• French:
  • Part 1 (#3107 by @r0cketr1kky)
• Hindi:
  • Parts 5, 6, 7, 13b, and 13c (#2909 and #3055 by @raheja)
  • Part 13a (#2958 by @Yugandhartripathi)
• Italian:
  • Part 1 (#3056 by @DanyEle)
• Portuguese:
  • Part 1 (#3035 by @MarcioPorto)
  • Parts 7, 8, and 8bis (#2977 by @joaolcaas)
  • Parts 9, 10 and 11 (#2980 by @jefersonf)
  • Parts 12 and 13a (#3015 by @marcusvlc)
  • Parts 12bis and 13b (#3020 by @hericlesme)
  • Part 13c (#3023 by @Izabellaaaq)
• Spanish:
  • Parts 6 (#2941 by @ricardopretelt)
  • Parts 7, 8, 9, 10, 11, and 13abc (#2944, #2962, #3031 and #3032 by @arturomf94)
  • Part 12 (#2822 by @socd06)
• Ukrainian:
  • Part 1 (#2854 by @KyloRen1)

Builds:

• Moves automated testing of PRs from Travis to Github Actions (#2936, #3012, and #3013 by @karlhigley and @systemshift)
• Adds a security scan for every PR (#3036 by @systemshift)
• Runs automated translation tests only on the notebooks that changed to speed up the builds (#3064 by @arturomf94)
• Automatically updates the pysyft-notebook Docker image when changes are merged to master (#3030 by @linamnt)
• Caches dependencies in Github Actions (#3124 by @imskr)

v0.2.3.a3

Third release of v0.2.3 to incorporate Syft protocol schema updates.

v0.2.3.a2

Second release of v0.2.3 to incorporate msgpack 1.0 compatibility and Syft protocol schema updates.

v0.2.3.a1

New functionality:

• Migrates PySyft to PyTorch 1.4 (#2930 by @gmuraru)
• Reworks Plans for smoother serialization to multiple formats (#2910 by @LaRiffle and @vvmnnnkv)

Bug fixes:

• Fixes numeric issues in handcrafted Conv and Pool implementations (#2945 and #2964 by @arshjot)
• Removes an insecure eval in native tensor interpreter (#2951 by @karlhigley)
• Fixes parameters to ObjectRequestMessage in websocket_client.py (#2948 by @brandonhee)

Tutorial updates:

• Bengali: Parts 1 and 2 (#2938 and #2942 by @ucalyptus)
• Hindi: Part 13a (#2958 by @Yugandhartripathi)
• Spanish: Parts 6, 7, and 8 (#2941, #2944, and #2962 by @ricardopretelt and @arturomf94)

v0.2.2.a1

Includes some noteworthy new functionality:

• CUDA processing enabled in PySyft (#2735 and #2772 by @midokura-silvia)
• A PrivateTensor type (#2709 by @IonesioJunior)
• Promise and PromiseTensor types (#2610 by @Jasopaum)
• A complete test suite for msgpack serialization and standardization of the serialization format (#2762 and #2812 by @vvmnnnkv)
• An implementation of Distributed Association Scan Hammer (DASH) algorithm (#2658 by @andrelmfarias)
• String and StringPointer types added to support NLP applications (#2684 by @AlanAboudib)
• The ability to nest Plans within other Plans (#2791 by @gmuraru)
• Serializability for the grad_fn in AutogradTensors (#2871 by @sukhadj)
• PyGrid module moved to PySyft (#2760 by @IonesioJunior)
• A NumpyTensor type (#2913 by @iamtrask)
• Python implementations of torch.nn.Conv2d and torch.nn.AvgPool2d (#2896 by @iamtrask)
• Approximate exp log inverse and sigmoid for SMPC (#2659 by @LaRiffle)

Also includes updates to the tutorials:

• A tutorial on Promises and PromiseTensors (#2786 by @Jasopaum)
• A split neural network tutorial (#2808 by @H4LL)
• Tutorial notebooks translated into Chinese, Spanish, Hindi, Indonesian, Korean, Portuguese, and Romanian (many PRs by @dljgs1, @Bingyy, @socd06, @ricardopretelt, @arturomf94, @darkmatter18, @Yugandhartripathi, @nggih, @wonderit, @jefersonf, and @gmuraru.)

And finally, includes:

• Many bug fixes, which are too numerous to list but nonetheless much appreciated!