TOSA 0.80.0 Release Notes
We have now tagged and released TOSA 0.80.0 on https://git.mlplatform.org/
The rendered specification is available at https://mlplatform.org/tosa/tosa_spec.html
A git diff of the specification between the 0.70 and 0.80 releases can be found here: https://git.mlplatform.org/tosa/specification.git/diff?id=v0.70.0&id2=v0.80.0
Specification
- Change specification to refer to signless types consistently.
– RESCALE takes new arguments to determine whether inputs and outputs are unsigned - Add ERF activation function
- Improve conformance definitions for floating-point operations
– New flag for convolution operators to determine accuracy requirements - Add stateful operators - VARIABLE / VARIABLE_READ / VARIABLE_WRITE
- Add DIM operator to allow for improved shape inference calculations
- Allow bias tensors to be broadcast
- Various other cleanup, see the git revision history for complete details
Reference Model
- Conformance test repository is now deprecated, conformance test generation scripts are in the reference model repository
- Deprecated TensorFlow Lite based conformance tests
– Framework tests still generate TFL tests - Add support for new operators
– DIM operator for shape inference
– New unsigned attributes for RESCALE input and output - Improve TOSA flatbuffer compatibility checks
- Fixes for test and reference model cleanup
- Enable reading/writing native data types for numpy files
- Various other cleanup, see the git revision history for complete details
Serialization Library
- Add support for new operators
– DIM operator for shape inference
– New attributes for RESCALE operator
– ERF activation function support - Improved NumPy file handling
- Various other cleanup, see the git revision history for complete details
TOSA MLIR Translator
- Add support for new operators
– DIM operator for shape inference
– ERF activation function
– New attributes for RESCALE function - Add precommit to verify clang-format has been run
- Data type cleanup to align with TOSA specification
