What's new in Sentis 2.5

This is a summary of the changes from Sentis 2.4 to Sentis 2.5.

Added

  • PyTorch model import to directly import PyTorch files (.pt2) to Sentis without using ONNX.
  • LRN (LocalResponseNormalization) operator is now implemented on all backends.
  • 3D MaxPool and AveragePool operators are now implemented on all backends.
  • Sentis Importer now allows users to specify dynamic dimensions as static on Sentis model import, same as we do for ONNX.
  • Tokenizer now parses Hugging Face models.
  • Wider coverage of all the components of the Tokenizer.

Updated

  • Model Visualizer now supports background loading of models.
  • Resize operator on CPU no longer uses main (mono) thread path.
  • All model converters use switch-case instead of if-else cascade for improved performance
  • Mono APIs are migrated to CoreCLR-compatible APIs

Fixed

  • Editor crash when quitting in play mode.
  • Memory leak in FuseConstantPass was fixed.
  • Clip operator no longer need CPU fallback for min/max parameters.
  • Mod operator fix on some platform with float operands.
  • Faulty optimization pass was corrected.
  • Fix in existing burst code for 2D pooling vectorization calculations.
  • TopK issue on GPUCompute when dimension is specified.
  • Many fixes to the Tokenizer.

What's new in Sentis 2.4

Sentis is the new name for this package.

This is a summary of the changes from Inference Engine 2.3 to Sentis 2.4.

Added

  • Tokenizer API for tokenization and detokenization of strings with language models.
  • LiteRT model import to directly import .tflite files to Sentis without using ONNX.
  • Spectral operators to enable audio models.
  • Many new operators corresponding to LiteRT and torch operators with functional API and optimization passes.

Updated

  • Import of ONNX models has been greatly sped up and optimized to match the ONNX specification.

Fixed

  • Many small import, inference and documentation issues.

What's new in Inference Engine 2.3

This is a summary of the changes from Inference Engine 2.2 to Inference Engine 2.3.

Added

  • Model Visualizer for inspecting models as node-based graphs inside the Unity Editor.
  • GatherND and Pow operators now support Tensor<int> inputs more widely.
  • ConvTranspose and Constant operators now support more input arguments.

What's new in Inference Engine 2.2

Inference Engine is the new name for the Sentis package.

This is a summary of the changes from Sentis 2.1 to Inference Engine 2.2.

For information on how to upgrade, refer to the Upgrade Guide.

Added

  • Dynamic input shape dimensions support at import time for better model optimization.
  • Custom input and output names for models created with the functional API.
  • The model stores the shapes and data types of intermediate and output tensors and displays them in the Model Asset Inspector.
  • New Mish operator.
  • Improved shape inference for model optimization.

Updated

  • ScatterElements and ScatterND operators now support min and max reduction modes.
  • DepthToSpace and SpaceToDepth now support integer tensors.
  • TopK supports integer tensors.
  • Functional.OneHot now allows negative indices.
  • RoiAlign now supports the coordinate_transformation_mode parameter.
  • Reduction operators return correct results when reducing a tensor along an axis of length 0.
  • Reshape operator can now infer unknown dimensions even when reshaping a length 0 tensor like in PyTorch.
  • Improved documentation for Model Asset Inspector.

Removed

  • Obsolete Unity Editor menu items.
  • Slow CPU support for 4-dimensional and higher Convolution layers.

Fixed

  • Out-of-bounds errors for certain operators on GPUCompute backend.
  • The TextureConverter methods now correctly performs sRGB to RGB conversions.
  • Incorrect graph optimizations for certain models.
  • Issues with negative padding values in pooling and convolutions.
  • Accurate handling of large and small integer values in the GPUPixel backend.
  • Proper destruction of allocated render textures in the GPUPixel backend.
  • LeakyRelu now supports alpha greater than 1 on all platforms.
  • Fixed Async behaviour for CPU tensor data.