v0.3.2

Changes

• Networks with outputs marked as shape tensors (INetworkDefinition.mark_output_for_shapes) now emit tensoracle.ShapeTensorOutputWarning when the engine is built and again when an execution unit is created, before TensorRT shape inference. The warning names the affected tensors: create_execution() cannot handle these engines, and calling it crashes the process inside TensorRT.
• To stop safely instead, turn the warning into an exception: warnings.filterwarnings("error", category=tensoracle.ShapeTensorOutputWarning). Building then stops without leaving cache files, and creating an execution unit stops before shape inference.
• Saved engines record shape output names in the file header outside the cache conditions. Existing saved engines remain valid and are reused without rebuilding.

Networks without shape tensor outputs behave as before.

Install

pip install tensoracle==0.3.2

Documentation: https://sync-dev-org.github.io/tensoracle/