Self-Instantiated Recurrent Units with Dynamic Soft Recursion (Self-IRU)

This repository contains the PyTorch implementation of the Self-IRU model in the paper Self-Instantiated Recurrent Units with Dynamic Soft Recursion at NeurIPS 2021.

Installation

One needs to install the following libraries

• PyTorch (e.g., v1.4.0)
• Sequence-to-Sequence Toolkit
• CuPy
• Python Bindings to NVRTC
• SciPy

Usage

The usage of this repository follows the TCN repository (e.g., for polyphonic music tasks). To run the Self-IRU model, set model = RNNModel(input_size, args.nhid, dropout=dropout, rnn_type='INFINITY', args=args) in the [TASK_NAME]_test.py file, where INFINITY is the alias of the Self-IRU in our implementation. If you encounter ModuleNotFoundError, try export PYTHONPATH="${PYTHONPATH}:.".

Citation

If you find this repository helpful, please cite our paper:

@article{zhang2021selfiru,
    title={Self-Instantiated Recurrent Units with Dynamic Soft Recursion
},
    author={Zhang, Aston and Tay, Yi and Shen, Yikang and Chan, Alvin and Zhang, Shuai},
    booktitle={Advances in neural information processing systems},
    year={2021}
}