After the TensorFlow 1.13.0-rc0 release last month, the TensorFlow team is out with another update 1.13.0-rc2, unveiling major features and updates. The new release explores minor bug fixes, improvements, and other changes.
Let’s have a look at the noteworthy features in TensorFlow 1.13.0-rc2.
Major Improvements
- TensorFlow Lite has moved from contrib to core.
- TensorFlow GPU binaries are built against CUDA 10 and TensorRT 5.0.
- There’s newly added support for Python3.7 on all operating systems.
- NCCL has been moved to core.
Behavioral and other changes
- Conversion of python floating types to uint32/64 in tf.constant is not allowed.
- The gain argument of convolutional orthogonal initializers has consistent behavior with the tf.initializers.orthogonal initializer.
- Subclassed Keras models can be saved via tf.contrib.saved_model.save_keras_model.
- LinearOperator.matmul now returns a new LinearOperator.
- Performance of GPU cumsum/cumprod has improved by up to 300x.
- Support has been added for weight decay in most TPU embedding optimizers, including AdamW and MomentumW.
- Tensorflow/contrib/lite has been moved to tensorflow/lite.
- An experimental Java API is added to inject TensorFlow Lite delegates.
- Support has been added for strings in TensorFlow Lite Java API.
- All the occurences of tf.contrib.estimator.DNNLinearCombinedEstimator has been replaced with tf.estimator.DNNLinearCombinedEstimator.
- Regression_head has been updated to the new Head API for Canned Estimator V2.
- XLA HLO graphs can be rendered as SVG/HTML.
Bug Fixes
- Documentation has been updated with the details regarding the rounding mode used in quantize_and_dequantize_v2.
- OpenSSL compatibility has been fixed by avoiding EVP_MD_CTX_destroy.
- CUDA dependency has been upgraded to 10.0.
- All occurences of tf.contrib.estimator.InMemoryEvaluatorHook and tf.contrib.estimator.make_stop_at_checkpoint_step_hook have been replaced with tf.estimator.experimental.InMemoryEvaluatorHook and tf.estimator.experimental.make_stop_at_checkpoint_step_hook.
- tf.data.Dataset.make_one_shot_iterator() has been deprecated in V1, removed from V2, and tf.compat.v1.data.make_one_shot_iterator() has instead been added.
- keep_prob is deprecated and Dropout now takes rate argument.
- NUMA-aware MapAndBatch dataset has been added.
- Apache Ignite Filesystem plugin has been added to support accessing Apache IGFS.
For more information, check out the official TensorFlow 1.13.0-rc2 release notes
Read Next
Building your own Snapchat-like AR filter on Android using TensorFlow Lite [ Tutorial ]