/PeepholeLSTM-Apple-Issue

Primary LanguagePython

RNN conversion

The python file convert_models.py details the architecture and the conversion method used for both TorchScript and CoreML RNN models.

TorchScript vs CoreML inference time benchmark

A simple Swift project can be found in the TorchScriptVSCoreML directory to assess the inference time both TorchScript and CoreML RNN models. To reproduce this assessment follow those steps:

  • open a terminal and run
cd TorchScriptVSCoreML
pod install
  • open TorchScriptVSCoreML.xcworkspace
  • move the slider and set to a sequence length above 2000 (20 sec of audio in our use-case) as displayed down below
  • see in the terminal how the different models react