yester31/torch_to_tf-lite

Convert Pytorch model to Tensorflow lite model

pytorch_to_tensorflow

Enviroments

• Windows 10 laptop
• CPU 11th Gen Intel(R) Core(TM) i7-11375H @ 3.30GHz (cpu)
• tensorflow 2.7.0

0. Pytorch Resnet18

• Standard pytorch resnet18 Model
• Performace evaluation(Execution time of 100 iteration for one 224x224x3 image)
• PyTorch (cpu) -> 3235 [ms]
• PyTorch (gpu) -> 363 [ms]

1. Pytorch to Tensorflow by ONNX

• Conversion pytorch to tensorflow by onnx
  • Tensorflow (cpu) -> 3748 [ms]
  • Tensorflow (gpu) -> 832 [ms]

2. Pytorch to Tensorflow by functional API

• Conversion pytorch to tensorflow by using functional API
  • Tensorflow (cpu) -> 4804 [ms]
  • Tensorflow (gpu) -> 3227 [ms]

3. Tensorflow lite on CPU

• Conversion pytorch to tensorflow by functional API
  • Tensorflow lite f32 -> 7781 [ms], 44.5 [MB]
    • max index : 388 , prob : 13.55378, class name : giant panda panda panda bear coon ...
  • Tensorflow lite f16 -> 5447 [ms], 22.3 [MB]
    • max index : 388 , prob : 13.54807, class name : giant panda panda panda bear coon ...
  • Tensorflow lite int8 -> 977569 [ms], 11.2 [MB]
    • max index : 388 , prob : 13.71834, class name : giant panda panda panda bear coon ...
• Conversion pytorch to tensorflow by onnx
  • Tensorflow lite f32 -> 6133 [ms], 44.5 [MB]
    • max index : 388 , prob : 13.80411, class name : giant panda panda panda bear coon ...
  • Tensorflow lite f16 -> 6297 [ms], 22.3 [MB]
    • max index : 388 , prob : 13.79882, class name : giant panda panda panda bear coon ...
  • Tensorflow lite int8 -> 1072768 [ms], 11.2 [MB]
    • max index : 388 , prob : 14.05152, class name : giant panda panda panda bear coon ...
• Calculated on 2021-12-09

4. Tensorflow lite on raspberry pi 4

Reference :

• https://www.tensorflow.org/lite/convert?hl=ko
• https://dmolony3.github.io/Pytorch-to-Tensorflow.html