com.unity.sentis 1.4.0-pre.1

com.unity.sentis

Source docs
Notes: processedReleased April 6, 2024

Browse versions

PreviousNext

Unity Compatibility

Minimum Unity: 2023.2

Package Dependencies

Features

  • Fast path for Scatter and Gather ops.
  • Quantization API for quantizing model weights to Float16 or Uint8.
  • TensorByte and TensorShort tensor types for quantization.
  • Pad operator to support 'axes' input, 'TensorInt' input and 'wrap' mode.
  • GeluFast op for tiny stories optimization.
  • Tensor 'Reshape' method for changing shape without backend.
  • Functional API for compiling models with torch-like syntax.
  • Analytics reporting on model import.

📈 Improvements

  • Reworked async API with awaitable methods on tensor readback.
  • Reworked tensor allocation scheme.
  • Renamed tensor fields and methods.
  • Reworked model serialization to use flatbuffers.
  • Random layers accept integer seed rather than float.
  • Improved NonMaxSuppression inference drastically and added GPUCompute backend.

🔧 Bug Fixes

  • Dispatch limit issues on Split operator.
  • Optimization pass where Dense has transposed weights.
  • Import settings for Resize op on opset 10.
  • Links to external sources in docs.
  • Edge cases in ScatterElements and GatherElements infer correctly.
  • Conv operator going out of bounds on GPUCompute and GPUCommandBuffer.

API Changes

  • Broken optimization pass for Dense > ScaleBias.
  • ArrayTensorData class and methods.
  • 'CreateWorker' method from model extensions.
  • Mapping of param symbolic tensor dims to original names.
  • Model API for adding inputs, constants, layers to model (in favour of Functional API).
  • Ops for tensor operations (in favour of Functional API).