Export models

Environments

Install modelscope and funasr

The installation is the same as funasr
shell pip3 install torch torchaudio pip install -U modelscope funasr # For the users in China, you could install with the command: # pip install -U modelscope funasr -i https://mirror.sjtu.edu.cn/pypi/web/simple

Install the quantization tools

pip install torch-quant # Optional, for torchscript quantization
pip install onnxruntime # Optional, for onnx quantization

Export model

Tips: torch>=1.11.0

shell python -m funasr.export.export_model \ --model-name [model_name] \ --export-dir [export_dir] \ --type [onnx, torch] \ --quantize [true, false] \ --fallback-num [fallback_num]
model-name: the model is to export. It could be the models from modelscope, or local finetuned model(named: model.pb).

export-dir: the dir where the onnx is export.

type: onnx or torch, export onnx format model or torchscript format model.

quantize: true, export quantized model at the same time; false, export fp32 model only.

fallback-num: specify the number of fallback layers to perform automatic mixed precision quantization.

Performance Benchmark of Runtime

Paraformer on CPU

onnx runtime

libtorch runtime

Paraformer on GPU

nv-triton

For example

Export onnx format model

Export model from modelscope
shell python -m funasr.export.export_model --model-name damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch --export-dir ./export --type onnx
Export model from local path, the model'name must be model.pb.
shell python -m funasr.export.export_model --model-name /mnt/workspace/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch --export-dir ./export --type onnx

Export torchscripts format model

Export model from modelscope
shell python -m funasr.export.export_model --model-name damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch --export-dir ./export --type torch

Export model from local path, the model'name must be model.pb.
shell python -m funasr.export.export_model --model-name /mnt/workspace/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch --export-dir ./export --type torch

Acknowledge

Torch model quantization is supported by BladeDISC, an end-to-end DynamIc Shape Compiler project for machine learning workloads. BladeDISC provides general, transparent, and ease of use performance optimization for TensorFlow/PyTorch workloads on GPGPU and CPU backends. If you are interested, please contact us.