TensorFlow 2.1.0 rc0 released, TensorFlow 2.1 will be the last to support Python version TF 2. Python2 support will be 1 January 2020 officially ended, TensorFlow also from that date to stop supporting Python 2, and will no longer be expected to release a new version in 2019.
Key features and improvements are as follows:
- tensorflow pip package is now included by default for Linux and Windows GPU support (and tensorflow-gpu same). It can run on machines with and without NVIDIA GPU's. tensorflow-gpu is still available for users concerned about the size of the package, the package can be downloaded only CPU on tensorflow-cpu.
tf.keras
Model.fit_generator
,Model.evaluate_generator
,Model.predict_generator
,Model.train_on_batch
,Model.test_on_batch
, AndModel.predict_on_batch
methods are now respected run_eagerly property, and the case will be used to run correctly tf.function default.Model.fit_generator
,Model.evaluate_generator
AndModel.predict_generator
it is deprecated endpoints. They contain Model.fit, Model.evaluate and Model.predict, they now supports the generation and sequence.- As long as in the range of building the model, you can be Keras .compile .fit .evaluate .predict and placed outside DistributionStrategy range.
- Keras model.load_weights now accept skip_mismatch as a parameter. It is available in external Keras, it has been copied to the tf.keras.
- Introduced TextVectorization layer that the original string as input, and for text normalization, labeled, n-gram index generation and vocabulary.
- Cloud TPU Pod provides, .fit, experimental support for Keras .compile .evaluate and .predict of.
- TPU is now cloud enabled automatic external compiler. Tf.summary This makes it easier to use with Cloud TPU.
- Cloud TPU support dynamic batch sizes and with DistributionStrategy of Keras.
- GPU and Cloud TPU provides experimental support for mixing accuracy.
- TensorFlow Model Garden offers many popular models Keras reference implementation.
tf.data
- Change tf.data dataset recataloged + distribution strategies to improve performance. Please note that the behavior is slightly different data sets, since relabeled data sets base will always be a multiple of the number of copies.
TensorRT
- Now supported and enabled by default TensorRT 6.0. It adds support for more TensorFlow operations, including Conv3D, Conv3DBackpropInputV2, AvgPool3D, MaxPool3D, ResizeBilinear and ResizeNearestNeighbor. In addition, TensorFlow-TensorRT python conversion API to export to tf.experimental.tensorrt.Converter.
For details, see Update:
https://github.com/tensorflow/tensorflow/releases/tag/v2.1.0-rc0