[q-implant] PyTorch Quantization Parameter exporter

[ssafyWooJeong]

What

Let's develop PyTorch Quantization Parameter exporter, a tool to export quantization parameters generated by PyTorch, to be used by q-implant.

Why

• ONE has its own quantization tool, but it has limitations in applying latest quantization techniques implemented in PyTorch.
• There is q-implant module which can import quantization parameters into circle, in development.
• We will implement PyTorch Quantization Parameter exporter, which can export quantization parameters for q-implant

• Expected benefits

  • The latest quantization techniques implemented in PyTorch can be applied to the circle model
  • Circle model will have improved accuracy and performance.

Tasks

• Define exporter's input arguments.
• Define export format (quantization parameters + weights of the model).
• Implement PyTorch Quantization Parameter exporter.
• Mapping names between PyTorch model and Circle model.

Usage

TorchQParamExporter.export(original_model, quantized_model, sample_input, json_path, tflite2circle_path)

Calling TorchQParamExporter export method with PyTorch original model(not quantized), PyTorch Quantized model, sample input of network, json(qparam.json) save path and tflite2cirlce module's path

The above method will generate quantization parameters by qparam.json and numpy files. The generated files will be used by q-implant

Notes

It uses tflite2circle and pics to make name mapping between PyTorch and Circle.

[Judgement9882]

What에다가 관련된 이슈의 # 번호 붙이면 좋을것같아요!