The TensorFlow team released the first release candidate of TensorFlow 1.13.0-rc0 yesterday. TensorFlow 1.13.0-rc0 explores major bug fixes, improvements and other changes.
Let’s have a look at the major highlights in TensorFlow 1.13.0-rc0.
Major improvements
- In TensorFlow 1.13.0-rc0, TensorFlow Lite has been moved from contrib to core. What this means is that Python modules are now under tf.lite and the source code is now under tensorflow/lite instead of tensorflow/contrib/lite.
- TensorFlow GPU binaries have now been built against CUDA 10.
- NCCL has been moved to core in TensorFlow 1.13.0-rc0.
Behavioural and other changes
- Conversion of python floating types to uint32/64 (i.e. matching behaviour of other integer types) in tf.constant has been disallowed in TensorFlow 1.13.0-rc0.
- Doc consisting of details about the rounding mode used in quantize_and_dequantize_v2 has been updated.
- The performance of GPU cumsum/cumprod has been increased by up to 300x.
- Support has been added for weight decay in most TPU embedding optimizers such as AdamW and MomentumW.
- An experimental Java API has been added for injecting TensorFlow Lite delegates.
- New support is added for strings in TensorFlow Lite Java API.
- tf.spectral has been merged into tf.signal for TensorFlow 2.0.
Bug fixes
- tensorflow::port::InitMain() now gets called before using the TensorFlow library. Programs that fail to do this are not portable to all platforms.
- saved_model.loader.load has been deprecated and is replaced by saved_model.load.
- Saved_model.main_op has also been deprecated and is replaced by saved_model.main_op in V2.
- tf.QUANTIZED_DTYPES has been deprecated and is changed to tf.dtypes.QUANTIZED_DTYPES.
- sklearn imports has been updated for deprecated packages.
- confusion_matrix op is now exported as tf.math.confusion_matrix instead of tf.train.confusion_matrix.
- An ignore_unknown argument is added in TensorFlow 1.13.0-rc0 to parse_values that suppresses ValueError for unknown hyperparameter types. Such * Add tf.linalg.matvec convenience function.
- tf.data.Dataset.make_one_shot_iterator() has been deprecated in V1 and added tf.compat.v1.data.make_one_shot_iterator()`.
- tf.data.Dataset.make_initializable_iterator() is deprecated in V1, removed it from V2, and added another tf.compat.v1.data.make_initializable_iterator().
- The XRTCompile op is can now return the ProgramShape resulted from the XLA compilation as a second return argument.
- XLA HLO graphs are rendered as SVG/HTML in TensorFlow 1.13.0-rc0.
For more information, check out the complete TensorFlow 1.13.0-rc0 release notes.
Read Next
Building your own Snapchat-like AR filter on Android using TensorFlow Lite [ Tutorial ]