TensorFlow Addons v0.9.0

Release Notes

• Built against TensorFlow 2.1
• CUDA kernels are compiled with CUDA 10.1
• API docs found on the website

Changelog

• Add strict TF2.1 version match warning (#1281, #1492, #1482)
• Move to pytest as default test runner

tfa.activations

• Support for python composite ops instead of custom-ops where possible
  • export TF_ADDONS_PY_OPS=1
  • tfa.options.TF_ADDONS_PY_OPS=True

tfa.image

• Add blend image op (#1275)
• Add cutout image op (#1338)
• Add sharpness op (#1452)

tfa.layers

• Add MultiHeadAttention layer (#1062)
• Add NetVLAD layer (#1237)
• Add Filter response normalization layer (#765)
• Add Thresholded linear unit layer (#857)

tfa.metrics

• Add Multivariate r2 metric (#1310)

tfa.optimizers

• Fix serialization bug for rectified ADAM (#1375)
• Fix typing check for LAMB optimizer (#1532)

Thanks to our Contributors

Aakash Kumar Nain, Cristian Garcia, Dheeraj R Reddy, Gabriel de Marmiesse, George Sterpu, Guillaume Klein, Harald Husum, Jared T Nielsen, Jeremy Oldfather, Jhuo IH, Joel Shor, Lucas Frey, Mark Sandler, Michael Reneer, NoobMaster, Qianli Scott Zhu, Rushabh Vasani, Sean Morgan, Sergii Khomenko, Xiaoquan Kong, failure-to-thrive, manzilz, pkan2, us

TensorFlow Addons v0.8.5

Release Notes

• Botched release due to GitHub action disk space limitation.
• This was never published to pypi

TensorFlow Addons v0.8.4

Release Notes

• Botched release due to GitHub action disk space limitation.
• This was never published to pypi

TensorFlow Addons v0.8.3

Release Notes

• Updated compatibility for macos to >= 10.13 (#1119)
• Fixed typing error in cyclical learning rates (#1160)

TensorFlow Addons v0.8.2

Release Notes

• Cherrypick cohen kappa fix (#1047)
• Cherrypick typing for normalization layers (#1069)

TensorFlow Addons v0.8.1

Release Notes

• CUDA kernels are compiled with CUDA 10.1
• API docs found on the website

Changelog

• Dropped Python2 support
• PyPi package has no longer has an installation dependency on TF.
  • Enables TFA for tensorflow / tensorflow-gpu / tensorflow-cpu (#974)
• Much of the repository is typed and has run-time type checking using typeguard
  • Adds a new pip dependency
• Lazy loading of custom ops to prevent ABI crashes when unnecessary (#869)

tfa.activations

• Renamed GeLU to GELU (#843)

tfa.callbacks

• Add AverageModelCheckpoint
• Bugfix for TimeStopping callback (#1010)

tfa.layers

• Add PolynomialCrossing (#1018)
• Add Thresholded Linear Unit (#857)

tfa.losses

• Add TripletHardLoss with soft margin (#959)

tfa.optimizers

• Add Novograd Optimizer (#836)

tfa.rnn

• Add LayernormSimpleRNN (#841)

Tutorials

• Updated NMT Seq2seq tutorial (#1003)
• Fixed KerasTQDM tutorial (#936)

Thanks to our Contributors

Aakash Kumar Nain, Calvin, Dheeraj R Reddy, Gabriel de Marmiesse, Guillaume Klein, HauserA, Jhuo IH, Luke Miner, Mark Daoust, NoobMaster, Romain Brault, Samuel Marks, Tzu-Wei Sung

TensorFlow Addons v0.7.1

Release Notes

• Correct import structure to enable Yogi optimizer (#922)

TensorFlow Addons v0.7.0

Release Notes

• Built against: TensorFlow 2.1.0
• Available for Linux/MacOS/Windows
• Last release to support python2
• CUDA kernels are compiled with CUDA 10.1
• API docs found on the website
• Keras subclasses now use the public registration decorator as package Addons
  • The prefix of registration name is Addons>. For example, users should change the serialization from tf.keras.activations.get('gelu') to tf.keras.activations.get('Addons>gelu')

Changelog

tfa.activations

• Add mish activation (#569)
• Add rrelu activation (#573)
• Add softshrink activation (#570)

tfa.callbacks

• Add Keras TQDM callback (#610)
• Add TimeStopping callback (#757)

tfa.image

• Add resampler kernel (#662)
• Various bugfixes

tfa.layers

• Support for WeightNorm in RNNs (#769)
• Fix WeightNorm support for layers without bias (#696)
• Support for WeightNorm serialization in keras hd5 format (#691)
• Fix bug in GroupNormalization axis (#611)

tfa.losses

• Fix keras model compile for lifted_struct_loss (#700)
• Speed up CPU CorrelationCost layer (#689)

tfa.metrics

• Fix bug to enable CohenKappa with distributed strategy scope (#721)

tfa.optimizers

• Add Yogi Optimizer (#745)
• Add LAMB Optimizer (#491)
• Add Cyclical learning rate schedulers (#644)
• Add Stochastic Weight Averaging Optimizer (#620)
• Allow MovingAverage to support dynamic number of updates (#758)

Tutorials

• Add Keras TQDM progress bar tutorial
• Add TimeStopping callback tutorial

Various bug fixes and improvements

Thanks to our Contributors

Aakash Kumar Nain, Amirhossein Kazemnejad, Brian Lee, Brian Patton, Calvin, Daniel, David Honzátko, Dero Gharibian, Dheeraj R Reddy, Guillaume Klein, Ishant Mrinal Haloi, John Arevalo, Mark Daoust, Matt Elsey, Philip May, Rahul Nair, Raphael Meudec, Saishruthi, Sean Morgan, Sergii Khomenko, Shreyash Patodia, Shun Lin, Tzu-Wei Sung, Vladimir Vlasov, Yan Facai (颜发才), junjiek, manzilz, omalleyt12, pkan2, prcvih, tzachar, who who who

TensorFlow Addons v0.6.0

Release Notes

• Built against: tensorflow-gpu==2.0.0 (Linux); tensorflow==2.0.0 (Mac)
• API docs found here in repository
• Ops are now namespaced Addons>OpName per RFC #126

Changelog

tfa.activations

• Add tanhshrink (#493)
• Add hardshrink (#500)
• Add lisht (#529)

tfa.losses

• Add npairs_multilabel loss (#466)

tfa.optimizers

• Add RAdam (#506)
• Add LookAhead wrapper (#506)
• Add conditional_gradient optimizer (#469)

tfa.seq2seq

• Minor bugfixes and improvements
• Make SequenceLoss Keras compatible (#503)
• Allow manual memory reset in AttentionMechanism (#547)

tfa.text

• Add ParseTime Op (#530)

Tutorials

• Examples moved to docs/tutorials

Thanks to our Contributors

Aakash Kumar Nain, Amirhossein Kazemnejad, Chen Yang, Dheeraj R Reddy, George Sterpu, Guillaume Klein, Helin Wang, Mark Daoust, Philip May, PyExtreme, Qianli Scott Zhu, Saishruthi, Sean Morgan, Tomer Kaftan, Tzu-Wei Sung, Yan Facai (颜发才)

TensorFlow Addons v0.5.2

Release Notes

• Built against: tensorflow-gpu==2.0.0 (Linux); tensorflow==2.0.0 (Mac)
• Bugfix: #532 Import crashed on a non-GPU Linux system. (#539 #548)