README.md

Model converter

Convert Pytorch model to onnx or tflite, and the converted model can be visualized by Netron

You can use this project to:

1. Pytorch -> onnx (float32)
2. Pytorch -> onnx -> tflite (float32)
3. Pytorch -> onnx -> tflite (int8)

Requirements

torch2onnx

pytorch
onnx
opencv-python

torch2tflite

tensorflow ~= 2.5
torch == 1.8.1
tensorflow-addons ~= 0.15
opencv-python ~= 4.5.4
onnx ~= 1.10
onnx-tf ~= 1.9
numpy >= 1.19

(opencv-python is optional)

Usage

torch2onnx (float32)

from converter import Torch2onnxConverter

converter = Torch2onnxConverter(model_path, target_shape=(3,224,224))
converter.convert()

torch2tflite (float32)

from converter import Torch2TFLiteConverter

converter = Torch2TFLiteConverter(tmp_path, tflite_model_save_path='model_float32.lite', target_shape=(224,224,3))
converter.convert()

torch2tflite (int8)

import torch
from converter import Torch2TFLiteConverter

dataset = torch.rand(16,3,224,224, dtype=torch.float32)
def representative_dataset():
    for data in dataset:
        data = data.unsqueeze(0) # (1,3,224,224)
        yield [data]

converter = Torch2TFLiteConverter(tmp_path, tflite_model_save_path='model_int8.lite', target_shape=(224,224,3),
                                    representative_dataset=representative_dataset)
converter.convert()

More details can be found in Torch2onnxConverter and Torch2TfliteConverter __init__ method.

Note that target_shape is different for Pytorch and Tensorflow.

Example

1. torch2onnx example
   

2. torch2tflite example
   

References

• PyTorch to TensorFlow Lite Converter