This repository provides all the data and code required to reproduce our paper.

We provide a notebook analysis.ipynb to reproduce the results of our paper.

All the data can be found under the ./data folder. The folder structure is organized in three layers (framework -> model -> language) as follows:

data
├── pytorch
│   ├── gru
│   │   ├── dotnet
│   │   ├── py
│   │   ├── rs
│   │   └── ts
│   ├── lenet1
│   │   └── ...
│   ├── lenet5
│   │   └── ...
│   ├── lstm
│   │   └── ...
│   └── vgg16
│   │   └── ...
└── tensorflow
    └── ...

each language (e.g., py) folder contains data for:

• Training accuracy in each epoch (files named training_errors_*)
• Test accuracy in each epoch (files named testing_errors_*)
• Training time cost (files named time_cost_*)
• Inference time and cross-binding test accuracy are stored in files name deploy_eval_*

All the code for building up models are shared under tensorflow_bindings and torch_bindings.

Conda environment for reference: ./torch_bindings/py/environment.yml. Also, check ./torch_bindings/py/README.md for setting up the environment.

For each binding, please refers to the README files for more information.

Conda environment for reference: ./tensorflow_bindings/py/cv/environment.yml. Also, check ./tensorflow_bindings/py/cv/README.md for setting up the environment.