BlockCIrculantRNN

BlockCIrculantRNN (LSTM and GRU) using TensorFlow. This project implements the block circulant matrix based training of RNN models, The code is based on clean version of this project. For more details, please refer to the paper

@inproceedings{wang2018c,
  title={C-LSTM: Enabling Efficient LSTM using Structured Compression Techniques on FPGAs},
  author={Wang, Shuo and Li, Zhe and Ding, Caiwen and Yuan, Bo and Qiu, Qinru and Wang, Yanzhi and Liang, Yun},
  booktitle={Proceedings of the 2018 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays},
  pages={11--20},
  year={2018},
  organization={ACM}
}

Other references

@article{li2018efficient,
  title={Efficient Recurrent Neural Networks using Structured Matrices in FPGAs},
  author={Li, Zhe and Wang, Shuo and Ding, Caiwen and Qiu, Qinru and Wang, Yanzhi and Liang, Yun},
  journal={arXiv preprint arXiv:1803.07661},
  year={2018}
}

Make sure you convert the TIMIT dataset from NIST format wav to RIFF format wavform

# Install sox
sudo apt-get install sox
# Use provided tools to convert TIMIT wav files
cd preprocessing
bash nist2wav.sh /xxx/xxx/TIMIT

To preprocess the TIMIT dataset, run

cd preprocessing
# Training set
python timit_preprocess_main.py \
--input_path=/xxx/xxx/TIMIT \
--output_path=/xxx/xxx/timit_preproc \
--level=phn \
--split=TRAIN

# Test set
python timit_preprocess_main.py \
--input_path=/xxx/xxx/TIMIT \
--output_path=/xxx/xxx/timit_preproc \
--level=phn \
--split=TEST

To Train the model, run

# Training a baseline
python timit_train_eval.py \
--input_data_dir=/xxx/xxx/timit_preproc \
--exp_dir=/xxx/xxx/timit_exp \
--level=phn \
--cell=LSTM \
--is_training=True

# Training block circulant model
python timit_train_eval.py \
--input_data_dir=/xxx/xxx/timit_preproc \
--exp_dir=/xxx/xxx/timit_exp \
--level=phn \
--cell=LSTM \
--partition_size=8 \
--is_training=True

To test the model, run

# Test a baseline
python timit_train_eval.py \
--input_data_dir=/xxx/xxx/timit_preproc \
--exp_dir=/xxx/xxx/timit_exp \
--level=phn \
--cell=LSTM \
--is_training=False \
--restore=True

# Test block circulant model
python timit_train_eval.py \
--input_data_dir=/xxx/xxx/timit_preproc \
--exp_dir=/xxx/xxx/timit_exp \
--level=phn \
--cell=LSTM \
--partition_size=8 \
--is_training=False \
--restore=True

For more argument options, check the code.

HongxingGao/BlockCIrculantRNN

BlockCIrculantRNN