[Quantization] onnx quantization

wanchen.swc

2023-03-10 525f5d77564f016acdd03ff71197f7a4a9177840

 funasr/export/export_model.py

@@ -153,11 +153,19 @@

        if self.quant:
            from onnxruntime.quantization import QuantType, quantize_dynamic
            import onnx
            quant_model_path = os.path.join(path, f'{model.model_name}_quant.onnx')
            onnx_model = onnx.load(model_path)
            nodes = [n.name for n in onnx_model.graph.node]
            nodes_to_exclude = [m for m in nodes if 'output' in m]
            quantize_dynamic(
                model_input=model_path,
                model_output=quant_model_path,
                op_types_to_quantize=['MatMul'],
                per_channel=True,
                reduce_range=False,
                weight_type=QuantType.QUInt8,
                nodes_to_exclude=nodes_to_exclude,
            )