TensorFlow is an end-to-end open source platform for machine learning. It has a comprehensive and flexible ecosystem of tools, libraries, and community resources that make it easy for developers to build and deploy ML-powered applications.

TensorFlow was originally developed for machine learning and deep neural network research. But the system is general enough to apply to a wide range of other fields as well.

important changes

• Build, compile and package
  • Removed redundant packages tensorflow-gpu and tf-nightly-gpu.These packages were removed, and instead, users were directed to switch to tensorflow or tf-nightly package. Since TensorFlow 2.1, the only difference between the two sets of packages is their names, so there is no loss of functionality or GPU support.
• tf.function:
  • tf.function Now use the Python inspect library directly to parse the signature of the Python function it decorates. This change may break code that has wrong function signatures but was previously ignored, for example:
    • on a function with a different signature using functools.wraps
    • in void tf.function use under input functools.partial
  • tf.function It is now mandatory that input parameter names must be valid Python identifiers. Incompatible names are handled automatically, similar to the existing SavedModel signature behavior.
  • parameterless tf.function is assumed to have an empty input_signaturerather than undefined, even though input_signature is unspecified.
  • tf.types.experimental.TraceType It is now necessary to define an additional placeholder_value method.
  • tf.function Traces are now done using placeholder values ​​generated by TraceType instead of the values ​​themselves.
• experimental API tf.config.experimental.enable_mlir_graph_optimization and tf.config.experimental.disable_mlir_graph_optimization was removed.

Key Features and Improvements

• Added support for Python 3.11
• Removed support for Python 3.7
• tf.lite:
  • for built-in operations fill Added support for 16-bit floating point types
  • Transpose now supports 6D tensors
  • Float LSTM now supports diagonal recursive tensors
• tf.experimental.dtensor:
  • The coordination service can now communicate with dtensor.initialize_accelerator_system work together and are enabled by default.
  • Add to tf.experimental.dtensor.is_dtensor to check if a tensor is a DTensor instance.
• tf.test:
  • increased tf.test.experimental.sync_deviceswhich is useful for accurately measuring the performance of the benchmark
• tf.experimental.dtensor:
  • Added experimental support for ReduceScatter fuse on GPU
• …

More details can be viewed at: https://github.com/tensorflow/tensorflow/releases/tag/v2.12.0