Note
ESPnet now features a new ONNX-compatible implementation of the `make_pad_mask` function, and ESPnet-ONNX fully supports this updated version. Developers can easily convert their ESPnet models without requiring any special handling, unless the model includes unsupported operations. In the event that you encounter an exportation error, please refer to the `doc/DevelopersGuide.md` for guidance.
**Please note that model conversion with ESPnet-ONNX v0.2.0 is compatible with espnet>=202308.**
custom_ort_v1.14.1.espnet
This release supports model optimization and inference of the optimized model.
The supported models are the same as `onnxruntime==1.11.1.espnet`
`1.14.1.espnet` supports the onnx model exported with `torch==2.0.1`.
Note that developers can export with `opset_version>13`, but this custom version is required for optimized inference. This is because the `axes` attribute in the `ReduceSum` node is moved to input from `opset_version=13`.
custom_ort_v1.11.1.espnet.3
This release supports model optimization and inference of optimized model.
The supported models are:
- Transformer encoder / decoder
- Conformer encoder
Both CPU and GPU operation is supported for optimize and inference.
Currently python3.8 version
custom_ort_v1.11.1-espnet_onnx.2
This version contains fusion logic and runtime for
- CrossAttention for CPU and GPU
- RelPosAttention for CPU
and contains an optimization script for Conformer.
custom_ort_v1.11.1-espnet_onnx
**This is an experimental release**
This is a custom version of onnxruntime. You need to install this version of onnxruntime if you want to fully optimize your ESPnet model.
I added some ops to the onnxruntime=1.11.1, so basically, almost all of the source code is built with v1.11.1
The source for this file is stored [here](https://github.com/Masao-Someki/onnxruntime).