ONNX-TensorRT 8.6 EA Release

For more details, see the 8.6 EA release notes for new features added in TensorRT 8.6

Added

• Added support for GroupNormalization, LayerNormalization, IsInf operations
• Added support for INT32 input types for Argmin, Argmax, and TopK
• Added support for ReverseSequence operators with dynamic shapes
• Added support for TopK operators with dynamic K values
• Added OnnxParserFlag enum and setFlag interfaces to the ONNX parser to modify the default parsing behavior
• Added metadata tracking, now ONNX node metadata will be embedded into TensorRT layers

Changed

• All cast operations will now use the new CastLayer over the pervious IdentityLayer.

ONNX-TensorRT 8.5 GA Release

For more details, see the 8.5 GA release notes for new features added in TensorRT 8.5

Added

• Added the RandomNormal, RandomUniform, MeanVarianceNormalization, RoiAlign, Mod, Trilu, GridSample and NonZero operations
• Added native support for the NonMaxSuppression operator
• Added support for importing ONNX networks with UINT8 I/O types

Fixed

• Fixed an issue with output padding with 1D deconv
• Fixed an issue when flattening 1D tensors
• Fixed an issue when parsing String attributes from TRT plugins
• Fixed an issue when importing If subgraphs with shared initializer names
• Fixied an issue when importing Loop subgraphs with INT_MAX trip counts

Removed

• Removed onnx2trt binary. See the README.md for alternative binaries to run ONNX model with TensorRT.

ONNX-TensorRT 8.4 GA Release

For more details, see the 8.4 GA release notes for new features added in TensorRT 8.4

Added

• Added native FP16 support for importing and manipulating FP16 initializers
• Added support for Shrink
• Added support for Xor
• Added dynamic shape support for ArgMax and ArgMin
• Added dynamic shape support for Range for floating point types

Fixes

• Fixed an issue in tensor name scoping in ONNX models with nested subgraphs
• Fixed misc issues when dealing with empty tensors
• Fixed the operations in the Celu importer function
• Removed unnecessary reshapes in the GEMM importer function

ONNX-TensorRT 22.02 Release

Fixes

• Fixed naming issue in parsing If conditonal graphs

ONNX-TensorRT 8.2 GA Release

Added

See the 8.2 EA release notes for new features added in TensorRT 8.2.

Fixes

• Removed duplicate constant layer checks that caused some performance regressions
• Fixed expand dynamic shape calculations
• Added parser-side checks for Scatter layer support

ONNX-TensorRT 8.2 EA Release

Added

• Added support for the following ONNX operators:
  • Einsum
  • IsNan
  • GatherND
  • Scatter
  • ScatterElements
  • ScatterND
  • Sign
  • Round

Updated

• Updated Gather and GatherElements implementations to natively support negative indices
• Updated Pad layer to support ND padding, along with edge and reflect padding mode support
• Updated If layer with general performance improvements.

ONNX-TensorRT 8.0 Release

Added

• Rehauled resize operator, now fully supporting the following modes:
  • Coordinate Transformation modes: half_pixel, pytorch_half_pixel, tf_half_pixel_for_nn, asymmetric, and align_corners
  • Modes: nearest, linear
  • Nearest Modes: floor, ceil, round_prefer_floor, round_prefer_ceil
• QuantizeLinear/DequantizeLinear updates:
  • Added support for tensor scales
  • Added support for per-axis quantization
• Added support for multi-input ConvTranpose
• Added support for generic 2D padding
• Added experimental support for NonMaxSuppression

Updated

• Moved RefitMap API to core TensorRT.
• Added Datatype column to operators.md

For more information on the TensorRT 8.0 release, see the release notes.

ONNX-TensorRT 21.05 Release

Added

• Added support for InstanceNormalization on 5D tensors
• Added library only build target #659
• Added support for negative gather indices #681
• Added support for DOUBLE-typed inputs and weights through downcast to float #674
• Added support for optional plugin fields in FallbackPlugin path #676

Updated

• Updated license #657

Fixes

• Fixed index offset calculation in GatherElements #675
• Clarified dynamic shape support for ReverseSequence

ONNX-TensorRT 21.03 Release

This is the release of ONNX-TensorRT that is used in the 21.03 NVIDIA NGC containers.

Changes

Added

• Added opset13 support for SoftMax, LogSoftmax, Squeeze, and Unsqueeze
• Added support for the EyeLike operator
• Added support for the GatherElements operator

ONNX-TensorRT 21.02 Release

This is the release of ONNX-TensorRT that is used in the 21.02 NVIDIA NGC containers.

Changes

Added

• Added support for the ReverseSequence operator #590
• Updated parse() and supportsModel() API calls with an optional model_path parameter to support models with external weights #621
• Added support for the Celu operator
• Added support for the CumSum operator
• Added support for the LessOrEqual operator
• Added support for the LpNormalization operator
• Added support for the LpPool operator
• Added support for the GreaterOrEqual operator
• Added support for dynamic inputs in onnx_tensorrt python backend
• Added FAQ section for commonly asked questions

Fixes

• Fixed relative path imports for models with external weights #619
• Fixed importing loops operators with no loop-carried depedencies #619
• Worked around unsupported BOOL concats through casting #620
• Fixed compilation error with GCC9 #568

Removed

• Removed onnx_tensorrt/config.py as it is no longer needed