TensorFlow 2.2.0
Release 2.2.0
TensorFlow 2.2 discontinues support for Python 2, previously announced as following Python 2's EOL on January 1, 2020.
Coinciding with this change, new releases of TensorFlow's Docker images provide Python 3 exclusively. Because all images now use Python 3, Docker tags containing -py3
will no longer be provided and existing -py3
tags like latest-py3
will not be updated.
Major Features and Improvements
-
Replaced the scalar type for string tensors from
std::string
totensorflow::tstring
which is now ABI stable. -
A new Profiler for TF 2 for CPU/GPU/TPU. It offers both device and host performance analysis, including input pipeline and TF Ops. Optimization advisory is provided whenever possible. Please see this tutorial and guide for usage guidelines.
-
Export C++ functions to Python using
pybind11
as opposed toSWIG
as a part of our deprecation of swig efforts. -
tf.distribute
:- Support added for global sync
BatchNormalization
by using the newly addedtf.keras.layers.experimental.SyncBatchNormalization
layer. This layer will syncBatchNormalization
statistics every step across all replicas taking part in sync training. - Performance improvements for GPU multi-worker distributed training using
tf.distribute.experimental.MultiWorkerMirroredStrategy
- Update NVIDIA
NCCL
to2.5.7-1
for better performance and performance tuning. Please see nccl developer guide for more information on this. - Support gradient
allreduce
infloat16
. See this example usage. - Experimental support of all reduce gradient packing to allow overlapping gradient aggregation with backward path computation.
- Deprecated
experimental_run_v2
method for distribution strategies and renamed the methodrun
as it is no longer experimental. - Add CompositeTensor support for DistributedIterators. This should help prevent unnecessary function retracing and memory leaks.
- Update NVIDIA
- Support added for global sync
-
tf.keras
:Model.fit
major improvements:- You can now use custom training logic with
Model.fit
by overridingModel.train_step
. - Easily write state-of-the-art training loops without worrying about all of the features
Model.fit
handles for you (distribution strategies, callbacks, data formats, looping logic, etc) - See the default
Model.train_step
for an example of what this function should look like. Same applies for validation and inference viaModel.test_step
andModel.predict_step
. - SavedModel uses its own
Model._saved_model_inputs_spec
attr now instead of
relying onModel.inputs
andModel.input_names
, which are no longer set for subclass Models.
This attr is set in eager,tf.function
, and graph modes. This gets rid of the need for users to
manually callModel._set_inputs
when using Custom Training Loops(CTLs). - Dynamic shapes are supported for generators by calling the Model on the first batch we "peek" from the generator.
This used to happen implicitly inModel._standardize_user_data
. Long-term, a solution where the
DataAdapter
doesn't need to call the Model is probably preferable.
- You can now use custom training logic with
- The SavedModel format now supports all Keras built-in layers (including metrics, preprocessing layers, and stateful RNN layers)
- Update Keras batch normalization layer to use the running mean and average computation in the
fused_batch_norm
. You should see significant performance improvements when usingfused_batch_norm
in Eager mode.
-
tf.lite
:- Enable TFLite experimental new converter by default.
-
XLA
- XLA now builds and works on windows. All prebuilt packages come with XLA available.
- XLA can be enabled for a
tf.function
with “compile or throw exception” semantics on CPU and GPU.
Breaking Changes
tf.keras
:- In
tf.keras.applications
the name of the "top" layer has been standardized to "predictions". This is only a problem if your code relies on the exact name of the layer. - Huber loss function has been updated to be consistent with other Keras losses. It now computes mean over the last axis of per-sample losses before applying the reduction function.
- In
- AutoGraph no longer converts functions passed to
tf.py_function
,tf.py_func
andtf.numpy_function
. - Deprecating
XLA_CPU
andXLA_GPU
devices with this release. - Increasing the minimum bazel version to build TF to 2.0.0 to use Bazel's
cc_experimental_shared_library
. - Keras compile/fit behavior for functional and subclassed models have been unified. Model properties such as
metrics
,metrics_names
will now be available only after training/evaluating the model on actual data for functional models.metrics
will now include modelloss
and output losses.loss_functions
property has been removed from the model. This was an undocumented property that was accidentally public and has now been removed.
Known Caveats
- The current TensorFlow release now requires gast version 0.3.3.
Bug Fixes and Other Changes
tf.data
:- Removed
autotune_algorithm
from experimental optimization options.
- Removed
- TF Core:
tf.constant
always creates CPU tensors irrespective of the current device context.- Eager
TensorHandles
maintain a list of mirrors for any copies to local or remote devices. This avoids any redundant copies due to op execution. - For
tf.Tensor
&tf.Variable
,.experimental_ref()
is no longer experimental and is available as simply.ref()
. pfor/vectorized_map
: Added support for vectorizing 56 more ops. Vectorizingtf.cond
is also supported now.- Set as much partial shape as we can infer statically within the gradient impl of the gather op.
- Gradient of
tf.while_loop
emitsStatelessWhile
op ifcond
and body functions are stateless. This allows multiple gradients while ops to run in parallel under distribution strategy. - Speed up
GradientTape
in eager mode by auto-generating list of op inputs/outputs which are unused and hence not cached for gradient functions. - Support
back_prop=False
inwhile_v2
but mark it as deprecated. - Improve error message when attempting to use
None
in data-dependent control flow. - Add
RaggedTensor.numpy()
. - Update
RaggedTensor.__getitem__
to preserve uniform dimensions & allow indexing into uniform dimensions. - Update
tf.expand_dims
to always insert the new dimension as a non-ragged dimension. - Update
tf.embedding_lookup
to usepartition_strategy
andmax_norm
whenids
is ragged. - Allow
batch_dims==rank(indices)
intf.gather
. - Add support for bfloat16 in
tf.print
.
tf.distribute
:- Support
embedding_column
with variable-length input features forMultiWorkerMirroredStrategy
.
- Support
tf.keras
:- Added
experimental_aggregate_gradients
argument totf.keras.optimizer.Optimizer.apply_gradients
. This allows custom gradient aggregation and processing aggregated gradients in custom training loop. - Allow
pathlib.Path
paths for loading models via Keras API.
- Added
tf.function
/AutoGraph:- AutoGraph is now available in
ReplicaContext.merge_call
,Strategy.extended.update
andStrategy.extended.update_non_slot
. - Experimental support for shape invariants has been enabled in
tf.function
. See the API docs fortf.autograph.experimental.set_loop_options
for additonal info. - AutoGraph error messages now exclude frames corresponding to APIs internal to AutoGraph.
- Improve shape inference for
tf.function
input arguments to unlock more Grappler optimizations in TensorFlow 2.x. - Improve automatic control dependency management of resources by allowing resource reads to occur in parallel and synchronizing only on writes.
- Fix execution order of multiple stateful calls to
experimental_run_v2
intf.function
. - You can now iterate over
RaggedTensors
using a for loop insidetf.function
.
- AutoGraph is now available in
tf.lite
:- Migrated the
tf.lite
C inference API out of experimental into lite/c. - Add an option to disallow
NNAPI
CPU / partial acceleration on Android 10 - TFLite Android AARs now include the C headers and APIs are required to use TFLite from native code.
- Refactors the delegate and delegate kernel sources to allow usage in the linter.
- Limit delegated ops to actually supported ones if a device name is specified or
NNAPI
CPU Fallback is disabled. - TFLite now supports
tf.math.reciprocal1
op by lowering totf.div op
. - TFLite's unpack op now supports boolean tensor inputs.
- Microcontroller and embedded code moved from experimental to main TensorFlow Lite folder
- Check for large TFLite tensors.
- Fix GPU delegate crash with C++17.
- Add 5D support to TFLite
strided_slice
. - Fix error in delegation of
DEPTH_TO_SPACE
toNNAPI
causing op not to be accelerated. - Fix segmentation fault when running a model with LSTM nodes using
NNAPI
Delegate - Fix
NNAPI
delegate failure when an operand for Maximum/Minimum operation is a scalar. - Fix
NNAPI
delegate failure when Axis input for reduce operation is a scalar. - Expose option to limit the number of partitions that will be delegated to
NNAPI
. - If a target accelerator is specified, use its feature level to determine operations to delegate instead of SDK version.
- Migrated the
tf.random
:- Various random number generation improvements:
- Add a fast path for default
random_uniform
random_seed
documentation improvement.RandomBinomial
broadcasts and appends the sample shape to the left rather than the right.
- Add a fast path for default
- Added
tf.random.stateless_binomial
,tf.random.stateless_gamma
,tf.random.stateless_poisson
tf.random.stateless_uniform
now supports unbounded sampling ofint
types.
- Various random number generation improvements:
- Math and Linear Algebra:
- Add
tf.linalg.LinearOperatorTridiag
. - Add
LinearOperatorBlockLowerTriangular
- Add broadcasting support to tf.linalg.triangular_solve#26204, tf.math.invert_permutation.
- Add
tf.math.sobol_sample
op. - Add
tf.math.xlog1py
. - Add
tf.math.special.{dawsn,expi,fresnel_cos,fresnel_sin,spence}
. - Add a Modified Discrete Cosine Transform (MDCT) and its inverse to
tf.signal
.
- Add
- TPU Enhancements:
- Refactor
TpuClusterResolver
to move shared logic to a separate pip package. - Support configuring TPU software version from cloud tpu client.
- Allowed TPU embedding weight decay factor to be multiplied by learning rate.
- Refactor
- XLA Support:
- Add standalone XLA AOT runtime target + relevant .cc sources to pip package.
- Add check for memory alignment to MemoryAllocation::MemoryAllocation() on 32-bit ARM. This ensures a deterministic early exit instead of a hard to debug bus error later.
saved_model_cli aot_compile_cpu
allows you to compile saved models to XLA header+object files and include them in your C++ programs.- Enable
Igamma
,Igammac
for XLA.
- Deterministic Op Functionality:
- XLA reduction emitter is deterministic when the environment variable
TF_DETERMINISTIC_OPS
is set to "true" or "1". This extends deterministictf.nn.bias_add
back-prop functionality (and therefore also deterministic back-prop of bias-addition in Keras layers) to include when XLA JIT complilation is enabled. - Fix problem, when running on a CUDA GPU and when either environment variable
TF_DETERMINSTIC_OPS
or environment variableTF_CUDNN_DETERMINISTIC
is set to "true" or "1", in which some layer configurations led to an exception with the message "No algorithm worked!"
- XLA reduction emitter is deterministic when the environment variable
- Tracing and Debugging:
- Add source, destination name to
_send
traceme to allow easier debugging. - Add traceme event to
fastpathexecute
.
- Add source, destination name to
- Other:
- Fix an issue with AUC.reset_states for multi-label AUC #35852
- Fix the TF upgrade script to not delete files when there is a parsing error and the output mode is
in-place
. - Move
tensorflow/core:framework/*_pyclif
rules totensorflow/core/framework:*_pyclif
.
Thanks to our Contributors
This release contains contributions from many people at Google, as well as:
372046933, 8bitmp3, aaronhma, Abin Shahab, Aditya Patwardhan, Agoniii, Ahti Kitsik, Alan Yee, Albin Joy, Alex Hoffman, Alexander Grund, Alexandre E. Eichenberger, Amit Kumar Jaiswal, amoitra, Andrew Anderson, Angus-Luo, Anthony Barbier, Anton Kachatkou, Anuj Rawat, archis, Arpan-Dhatt, Arvind Sundararajan, Ashutosh Hathidara, autoih, Bairen Yi, Balint Cristian, Bas Aarts, BashirSbaiti, Basit Ayantunde, Ben Barsdell, Benjamin Gaillard, boron, Brett Koonce, Bryan Cutler, Christian Goll, Christian Sachs, Clayne Robison, comet, Daniel Falbel, Daria Zhuravleva, darsh8200, David Truby, Dayananda-V, deepakm, Denis Khalikov, Devansh Singh, Dheeraj R Reddy, Diederik Van Liere, Diego Caballero, Dominic Jack, dothinking, Douman, Drake Gens, Duncan Riach, Ehsan Toosi, ekuznetsov139, Elena Zhelezina, elzino, Ending2015a, Eric Schweitz, Erik Zettel, Ethan Saadia, Eugene Kuznetsov, Evgeniy Zheltonozhskiy, Ewout Ter Hoeven, exfalso, FAIJUL, Fangjun Kuang, Fei Hu, Frank Laub, Frederic Bastien, Fredrik Knutsson, frreiss, Frédéric Rechtenstein, fsx950223, Gaurav Singh, gbaned, George Grzegorz Pawelczak, George Sterpu, Gian Marco Iodice, Giorgio Arena, Hans Gaiser, Hans Pabst, Haoyu Wu, Harry Slatyer, hsahovic, Hugo, Hugo Sjöberg, IrinaM21, jacco, Jake Tae, Jean-Denis Lesage, Jean-Michel Gorius, Jeff Daily, Jens Elofsson, Jerry Shih, jerryyin, Jin Mingjian, Jinjing Zhou, JKIsaacLee, jojimonv, Jonathan Dekhtiar, Jose Ignacio Gomez, Joseph-Rance, Judd, Julian Gross, Kaixi Hou, Kaustubh Maske Patil, Keunwoo Choi, Kevin Hanselman, Khor Chean Wei, Kilaru Yasaswi Sri Chandra Gandhi, Koan-Sin Tan, Koki Ibukuro, Kristian Holsheimer, kurileo, Lakshay Tokas, Lee Netherton, leike666666, Leslie-Fang-Intel, Li, Guizi, LIUJIAN435, Lukas Geiger, Lyo Nguyen, madisetti, Maher Jendoubi, Mahmoud Abuzaina, Manuel Freiberger, Marcel Koester, Marco Jacopo Ferrarotti, Markus Franke, marload, Mbah-Javis, mbhuiyan, Meng Zhang, Michael Liao, MichaelKonobeev, Michal Tarnowski, Milan Straka, minoring, Mohamed Nour Abouelseoud, MoussaMM, Mrinal Jain, mrTsjolder, Måns Nilsson, Namrata Bhave, Nicholas Gao, Niels Ole Salscheider, nikochiko, Niranjan Hasabnis, Nishidha Panpaliya, nmostafa, Noah Trenaman, nuka137, Officium, Owen L - Sfe, Pallavi G, Paul Andrey, Peng Sun, Peng Wu, Phil Pearl, PhilipMay, pingsutw, Pooya Davoodi, PragmaTwice, pshiko, Qwerty71, R Gomathi, Rahul Huilgol, Richard Xiao, Rick Wierenga, Roberto Rosmaninho, ruchit2801, Rushabh Vasani, Sami, Sana Damani, Sarvesh Dubey, Sasan Jafarnejad, Sergii Khomenko, Shane Smiskol, Shaochen Shi, sharkdtu, Shawn Presser, ShengYang1, Shreyash Patodia, Shyam Sundar Dhanabalan, Siju Samuel, Somyajit Chakraborty Sam, Srihari Humbarwadi, srinivasan.narayanamoorthy, Srishti Yadav, Steph-En-M, Stephan Uphoff, Stephen Mugisha, SumanSudhir, Taehun Kim, Tamas Bela Feher, TengLu, Tetragramm, Thierry Herrmann, Tian Jin, tigertang, Tom Carchrae, Tom Forbes, Trent Lo, Victor Peng, vijayphoenix, Vincent Abriou, Vishal Bhola, Vishnuvardhan Janapati, vladbataev, VoVAllen, Wallyss Lima, Wen-Heng (Jack) Chung, wenxizhu, William D. Irons, William Zhang, Xiaoming (Jason) Cui, Xiaoquan Kong, Xinan Jiang, Yasir Modak, Yasuhiro Matsumoto, Yaxun (Sam) Liu, Yong Tang, Ytyt-Yt, yuan, Yuan Mingshuai, Yuan Tang, Yuki Ueda, Yusup, zhangshijin, zhuwenxi