ESPnet-ONNX v0.2.0

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 onnxruntime==1.14.1 for espnet_onnx

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 onnxruntime for espnet_onnx

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 version 2 of onnxruntime=1.11.1-espnet_onnx

This version contains fusion logic and runtime for

• CrossAttention for CPU and GPU
• RelPosAttention for CPU

and contains an optimization script for Conformer.

Custom version of onnxruntime=1.11.1-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.