TensorFlow Addons v0.15.0

Release Notes

• Built against TensorFlow 2.7
• CUDA kernels are compiled with CUDA11.2 and cuDNN 8.1.0
• API docs found on the website

Changelog

• Use multipython image for dev container (#2598)
• Add support for publishing macOS M1 ARM64 wheels for tfa-nightly (#2559)

Tutorials

• Update optimizers_cyclicallearningrate.ipynb (#2538)

tfa.activations

• Correct documentation for Snake activation to match literature and return statement (#2572) @fliptrail

tfa.iamge

• Fix euclidean distance transform float16 kernel (#2568)

tfa.layers

• Fix using NoisyNet with .fit() or .train_on_batch() (#2486)
• Fix spectral norm mixed precision (#2576)

tfa.optimizers

• Add AdaBelief optimizer (#2548)
• Make Rectified Adam faster (#2570)

tfa.text

• Add the CRF model wrapper (#2555)
• Add a codeowner for CRF (#2556)

Thanks to our Contributors

@MarkDaoust, @bhack, @eli-osherovich, @fliptrail, @fsx950223, @howl-anderson, @juntang-zhuang, @jvishnuvardhan, @lgeiger, @markub3327, @seanpmorgan, @szutenberg and @vtjeng

TensorFlow Addons v0.14.0

Release Notes

• Built against TensorFlow 2.6
• CUDA kernels are compiled with CUDA11.2 and cuDNN 8.1.0
• API docs found on the website

Changelog

• Remove compatibility code for TensorFlow < 2.4 (#2545)
• Modify configure.py to recognize 'aarch64' for 64-Bit Raspberry Pi OS (#2540)
• Apple silicon support (#2504)
• Build fix Raspberry Pi 4 Linux ARM64 (#2487)

tfa.layers

• Add EmbeddingBag gpu op and layer (#2352) (#2517)(#2505)
• Fix StochasticDepth layer error in training mixed_float16 (#2450)

tfa.optimizers

• Adding a tutorial on CyclicalLearningRate (#2463)

Thanks to our Contributors

@HeatfanJohn, @Rocketknight1, @RyanGoslingsBugle, @fsx950223, @kaoh, @leondgarse, @lgeiger, @maxhgerlach, @sayakpaul, @seanpmorgan, @singhsidhukuldeep and @tetsuyasu

TensorFlow Addons v0.13.0

Release Notes

• Built against TensorFlow 2.5
• CUDA kernels are compiled with CUDA11.2 and cuDNN 8.1.0
• API docs found on the website

Changelog

• Add python3.9 support (#2204)
• Fixed build on ppc (#2438)

tfa.activations

• Cleanup legacy codes for activations (#2394)

tfa.image

• Add python fallback for adjust_hsv_in_yiq (#2392)
• Remove ImageProjectiveTransform kernel (#2395)
• Fix EDT float16 and float64 kernels (#2412)
• Optimize EDT (#2402)
• Update cutout_ops.py (#2416)

tfa.metrics

• Initial commit of streaming Kendall's Tau algorithm. (#2423)
• Fix F1Score docs (#2462)
• Matthew Fix (#2406)
• Fix RSquare serialization (#2390)
• Make RSquare.reset_states to be able to run in tf.function (#2445)

tfa.optimizers

• Adding COntinuos COin Betting (COCOB) Backprop optimizer (#2063)
• Fix NovoGrad optimizer to work with float64 layers (#2467)
• Update cyclical_learning_rate.py (#2286)
• RectifiedAdam: Store 'total_steps' hyperparameter as float (#2369)

tfa.text

• fix wrong type hinting of crf_log_likelihood (#2471)

Thanks to our Contributors

@0x0badc0de, @bhack, @DragonPG2000, @Harsh188, @WindQAQ, @ashutosh1919, @fsx950223, @jeongukjae, @jonpsy, @juliangilbey, @lucasdavid, @lum4chi, @m-a-r-o-u, @nickswalker, @nleastaugh, @npanpaliya, @olesalscheider, @rehanguha, @seanpmorgan, @shubhanshu02, @sorensenjs, @whatwilliam and @xiedeping

TensorFlow Addons v0.12.1

Release Notes

• Built against TensorFlow 2.4.1
• CUDA kernels are compiled with CUDA 11
• API docs found on the website

Changelog

• Remove AVX2 Compilation by default (Aligns with https://github.com/tensorflow/tensorflow/releases/tag/v2.4.1)

tfa.image

• Fix image/fix sparse image warp unknown batch size (#2311)

TensorFlow Addons v0.12.0

Release Notes

• Built against TensorFlow 2.4
• CUDA kernels are compiled with CUDA 11
• API docs found on the website

Changelog

• Add AVX2 support (#2299)
• Drop TF2.2 compatibility (#2224)
• Drop python3.5 support (#2204)
• Expose tfa.types doc (#2162)
• Rename "Arguments:" to "Args:" (#2267)
• Add support for ARM architecture build from source (#2182)

tfa.activations

• Add tf.nn.gelu alias for TF >= 2.4 (#2265)
• Remove custom op activations (#2247)

tfa.image

• Speedup gaussian kernel generation (#2149)
• Support fill_mode for transform (#2153)
• Use ImageProjectiveTransformV3 for TF >= 2.4.0 (#2293)
• Support unknown rank image (#2300)
• Fix sparse_image_warp partially unknown shape (#2308)
• Make cutout compatible with keras layer (#2302)
• Remove unsupported data_format (#2296)
• Refactor sharpness (#2287)
• Rodert fix image random cutout 2276 (#2285)
• Remove tf.function decorator in tfa.image.equalize (#2264)
• Support empty batches in ResamplerOp (#2219)
• Make cutout op compatible with non eager mode (#2190)

tfa.layers

• Add stochastic depth layer (#2154)

• Add MaxUnpooling2D layer (#2272)

• Add noisy dense layers. (#2099)

• Add discriminative Layer Training (#969)

• Make MultiHeadAttention agnostic to dtype (float32 vs. float16) (#2253)

• Change CRF layer dtype (#2270)

• Change GroupNormalization default groups to 32. (#2241)

tfa.optimizers

• Standardized Testing Module (#2233)
• Fix LazyAdam resource variable ops performance issue (#2274)
• Add experimental_aggregate_gradients support (#2137)

tfa.rnn

• Bug fix for conflict variable name in layernorm cells. (#2284)

tfa.seq2seq

• Graduate _BaseAttentionMechanism to a public base class (#2209)
• Add a doctest example for BasicDecoder (#2214)
• Add a doctest example for AttentionWrapper (#2215)
• Improve sampler documentation, use doctest (#2213)
• Beam search decoding procedure added to seq2seq_nmt tutorial (#2140)

Thanks to our Contributors

@DanBmh, @DavidWAbrahams, @Harsh188, @JulianRodert, @LeonShams, @MHStadler, @MarkDaoust, @SamuelMarks, @WindQAQ, @aaronmondal, @abhishek-niranjan, @albertz, @bhack, @crccw, @edend10, @fsx950223, @gabrieldemarmiesse, @guillaumekln, @HMPH, @hp77-creator, @hwaxxer, @hyang0129, @kaixih, @lamberta, @marksandler2, @matwilso, @napsternxg, @nataliyah123, @perfinion, @qlzh727, @rmlarsen, @rushabh-v, @rybakov, @seanpmorgan, @stephengmatthews, @tgaddair and @thaink

TensorFlow Addons v0.11.2

Changelog

Improve API documentation

• Beautifier image doc (#2101)
• Beautifier callbacks doc (#2105)
• Beautifier losses doc (#2062)
• Fix broken link in NovoGrad docstring (#2096)
• Beautifier layers doc (#2072)

TensorFlow Addons v0.11.1

Release Notes

• Update TF compatibility warning to include all of 2.3.x as acceptable.

TensorFlow Addons v0.11.0

Release Notes

• Built against TensorFlow 2.3
• CUDA kernels are compiled with CUDA 10.1
• API docs found on the website

Changelog

• Support building against CUDA 11 and CUDNN 8 (#1950)

tfa.activations

• Add Snake layer and activation (#1967)
• Deprecate gelu (#2048)

tfa.image

• Set shape for dense image warp (#1993
• Drop data_format argument (#1980)
• Enable half and double for resampler GPU ops (#1852)

tfa.layers

• Add Spectral Normalization layer (#1244)
• Add CRF layer (#1999)
• Add Snake layer and activation (#1967)
• Add Spatial Pyramid Pooling layer (#1745)
• Add Echo State Network (ESN) layer (#1862)
• Incorporate low-rank techniques into DCN. (#1795)

tfa.metrics

• Add geometric mean (#2031)
• Fix R_Square shape issue in model.evaluate (#2034)

tfa.losses

• Change the default distance metric for tfa.losses.triplet_semihard_loss and tfa.losses.triplet_hard_loss from squared euclidean norm to euclidean norm. Users must change distance_metric to "squared-L2" in order to achieve the old behavior.

tfa.optimizers

• Add ProximalAdagrad optimizer (#1976)
• Add support for scheduled weight decays in RectifiedAdam. (#1974)
• Fixed lr/wd schedules for DecoupledWeightDecayExtension running on GPU (#2053) (#2029)
• Fixed sparse novograd (#1970)
• MovingAverage: add dynamic decay and swap weights (#1726)
• Remove RAdam optional float total steps (#1871)

tfa.rnn

• Move the tf.keras.layers.PeepholeLSTMCell to TFA (#1944)
• Added echo state network (ESN) recurrent cell (#1811)

tfa.seq2seq

• Improve support of global dtype policy in seq2seq layers (#1981)
• Add a Python alternative to seq2seq.gather_tree (#1925)
• Allow resetting embedding_fn when calling BeamSearchDecoder (#1917)
• Fixup returned cell state structure in BasicDecoder (#1905)
• Fixup returned cell state structure in BeamSearchDecoder (#1904)
• Fix AttentionWrapper type annotation for multiple attention mechanisms (#1872)
• Ensure cell state structure is unchanged on first AttentionWrapper call (#1861)
• Remove sequential_update from AverageWrapper (#1807)

Thanks to our Contributors

@AakashKumarNain, @AntPeixe, @JakeTheWise, @MHStadler, @PRUBHTEJ, @Smankusors, @Squadrick, @Susmit-A, @WindQAQ, @autoih, @bhack, @brunodoamaral, @cgarciae, @charlielito, @csachs, @failure-to-thrive, @feyn-aman, @fsx950223, @gabrieldemarmiesse, @gugarosa, @guillaumekln, @jaeyoo, @jaspersjsun, @jlsneto, @ksachdeva, @lc0, @leandro-gracia-gil, @marload, @nluehr, @pedrolarben, @qlzh727, @seanpmorgan, @tanzhenyu, @tf-marissaw and @xvr-hlt

TensorFlow Addons v0.10.0

Release Notes

• Built against TensorFlow 2.2
• CUDA kernels are compiled with CUDA 10.1
• API docs found on the website

Changelog

• Enable ppc64le build (#1672)

tfa.activations

• Added the DepreciationWarning for the custom op version of activations functions (#1791)

tfa.image

• Fix condition tracing in scale_channel (#1830)
• Expose sharpness and equalize image op (#1827)
• Clarify flow definition for dense_image_warp (#1817)
• Added gaussian_blur_op (#1450)

tfa.layers

• Added Adaptive MaxPooling layers (#1727)
• Added AdaptiveAveragePooling2D layer (#1383)

tfa.metrics

• Add sample_weight support to FScore metrics (#1816)

tfa.losses

• Added angular distance option to triplet loss (#1730)
• Enable npairs loss on windows (#1742)
• Added float16 and bfloat16 support for TripletSemiHardLoss, TripletHardLoss and LiftedStructLoss (#1683)
• Add Soft Weighted Kappa Loss (#762)

tfa.optimizers

• Fixed serializability bug in yogi (#1728)

Thanks to our Contributors

@Dagamies, @HauserA, @MarkDaoust, @Squadrick, @Susmit-A, @WindQAQ, @ageron, @amascia, @ashutosh1919, @autoih, @ben-arnao, @bhack, @fsx950223, @gabrieldemarmiesse, @ghosalsattam, @guillaumekln, @henry-eigen, @jharmsen, @olesalscheider, @seanpmorgan, @shun-lin, @terrytangyuan and @wenmin-wu

TensorFlow Addons v0.9.1

Release Notes

• Include CUDA kernels missing from 0.9.0
• Fix serialization for cyclical learning rate (#1623)