3.0.2

[3.0.01] - 2023-06-28

Changed

• Update documentation

3.0.0

[3.0.0] - 2022-01-24

Changed

• Combined verified release of 2.1.x-2.4.x improvements

[2.4.0] - 2021-12-14

Added

• Added ModelLoader.LoadAsync methods for coroutine-based model loading
• Added support for Model weights as half on disk and in memory for CPU backend.
• Vastly improved operator coverage for PixelShader worker.
• Added YoloV3Tiny demo to Starter Kit

Changed

• Added flipY as an optional parameter to TensorToRenderTexture
• Download from texture use pixel shader when compute shader are not available.
• Improved support for dynamic input shape and rank inference.

Fixed

• Fixed Resample2D when using input tensor as size parameter.
• Fixed OneHot when using rank 2 input tensor.
• Fix TextureToTensor for PixelShader backend.
• Fix Added importer support for ConvTranspose with 1D spatial inputs.

2.4.0-preview

[2.4.0] - 2021-12-14

Added

• Added ModelLoader.LoadAsync methods for coroutine-based model loading
• Added support for Model weights as half on disk and in memory for CPU backend.
• Vastly improved operator coverage for PixelShader worker.
• Added YoloV3Tiny demo to Starter Kit

Changed

• Added flipY as an optional parameter to TensorToRenderTexture
• Download from texture use pixel shader when compute shader are not available.
• Improved support for dynamic input shape and rank inference.

Fixed

• Fixed Resample2D when using input tensor as size parameter.
• Fixed OneHot when using rank 2 input tensor.
• Fix TextureToTensor for PixelShader backend.
• Fix Added importer support for ConvTranspose with 1D spatial inputs.

[2.3.1] - 2021-10-26

Fixed

• Fixed CSharpBurst backend failure on RandomNormal and RandomUniform ops

2.3.1-preview

[2.3.1] - 2021-10-26

Fixed

• Fixed CSharpBurst backend failure on RandomNormal and RandomUniform ops

[2.3.0] - 2021-10-15

Added

• PixelShader worker! Allowing to run on GPU without requiring support for compute shaders. Note: limited to 4D tensors for now.
• ONNX: Added support for transA and transB parameters for Gemm operator.
• ONNX: Added support for multi-axis Squeeze/Unsqueeze.
• Added support for RoiAlign.

Changed

• Improved Tensor storage handling for Burst backend, should reduce overall memory pressure when used.
• Improved Transpose performance when they are not changing memory layout.
• Improved automatic test coverage for consoles.
• Improved ConvTranspose2D performance for large spatial kernel.

Fixed

• Added ability to create tensor from RenderTexture when compute shaders or GPU are not available.

2.3.0-preview

[2.3.0] - 2021-10-15

Added

• PixelShader worker! Allowing to run on GPU without requiring support for compute shaders. Note: limited to 4D tensors for now.
• ONNX: Added support for transA and transB parameters for Gemm operator.
• ONNX: Added support for multi-axis Squeeze/Unsqueeze.
• Added support for RoiAlign.

Changed

• Improved Tensor storage handling for Burst backend, should reduce overall memory pressure when used.
• Improved Transpose performance when they are not changing memory layout.
• Improved automatic test coverage for consoles.
• Improved ConvTranspose2D performance for large spatial kernel.

Fixed

• Added ability to create tensor from RenderTexture when compute shaders or GPU are not available.

[2.2.1] - 2021-09-14

Fixed

• Fixed CSharpBurst Random layer failure when model is loaded from an old .nn file

2.2.1-preview

[2.2.1] - 2021-09-14

Fixed

• Fixed CSharpBurst Random layer failure when model is loaded from an old .nn file

[2.2.0] - 2021-09-01

Added

• Added Burst implementation for Random layers, improves scheduling of the whole network computation
• Added scale and bias options for TensorFromTexture
• Added Erf layer support
• Added LogSoftmax axis support

Changed

• Improved MatMul performance on CSharpBurst backend on Desktop platforms
• Improved temp Tensor allocation handling, should reduce GC pressure
• Improved Tensor storage handling for PrecompiledCompute backend, should reduce overall memory pressure when used.
• Exposed flag for GPU worker to take ownership of weights, reduces memory pressure on CPU side, but limits Model object re-use.
• Improved convolution performance on Mobiles
• Improved 1x1 convolution performance
• Minimal supported Unity version updated to 2019.4.29f1
• Improved Barracuda installation instructions

Fixed

• Fixed broadcasting between rank 4 and rank 2 operands
• Fixed GPU Flatten when channel order is NHWC

2.2.0-preview

[2.2.0] - 2021-09-01

Added

• Added Burst implementation for Random layers, improves scheduling of the whole network computation
• Added scale and bias options for TensorFromTexture
• Added Erf layer support
• Added LogSoftmax axis support

Changed

• Improved MatMul performance on CSharpBurst backend on Desktop platforms
• Improved temp Tensor allocation handling, should reduce GC pressure
• Improved Tensor storage handling for PrecompiledCompute backend, should reduce overall memory pressure when used.
• Exposed flag for GPU worker to take ownership of weights, reduces memory pressure on CPU side, but limits Model object re-use.
• Improved convolution performance on Mobiles
• Improved 1x1 convolution performance
• Minimal supported Unity version updated to 2019.4.29f1
• Improved Barracuda installation instructions

Fixed

• Fixed broadcasting between rank 4 and rank 2 operands
• Fixed GPU Flatten when channel order is NHWC

[2.1.0] - 2021-05-17

Added

• ONNX: Added support for the metadata_props field, which is exposed as a <string,string> dictionary via Model.Metadata

Changed

• Improved Slice range handling
• Empty tensor buffers are now filled with -1 instead of the random data, helps to reduce accidental errors
• Greatly reduced memory allocation when using GPU backends (no more CPU mem cache for intermediate tensors). Mostly helps mobiles with unified memory architecture
• Improved SoftMax performance
• Improved LSTM performance on CPU

Fixed

• Fixed default axis to channels in ModelBuilder.SoftMax
• Fixed Optimize Model checkbox in model importer UI
• Fixed crash when tensor is created on CPU and later is moved to GPU
• Fixed out of bounds write in NCHW path
• Fixed crash in Tensor finalizers when using CSharpBurst backend
• Fixed Gather behavior for rank3 tensors

Known issues

• There is small regression in ARMv7 performance when running CSharpBurst backend. It can be worked around on most of the phones by shipping ARMv8(ARM64) builds

2.0.0-pre.6

[2.0.0] - 2021-04-05

Changed

• Combined verified release of 1.1.x-1.4.x improvements
• Burst updated to 1.6.0-pre.4

Fixed

• Fixed Transpose removal pass for ML-Agents networks.

2.0.0-pre.5

[2.0.0] - 2021-04-05

Changed

• Combined verified release of 1.1.x-1.4.x improvements
• Burst updated to 1.6.0-pre.2

Fixed

• Fixed Transpose removal pass for ML-Agents networks.

2.0.0-pre.4

[2.0.0] - 2021-04-05

Changed

• Combined verified release of 1.1.x-1.4.x improvements

Fixed

• Fixed Transpose removal pass for ML-Agents networks.