FastGRNN: A Fast, Accurate, Stable and Tiny Kilobyte Sized Gated Recurrent Neural Network

Adapted from https://github.com/microsoft/EdgeML/

Paper: https://arxiv.org/pdf/1901.02358

Instructions

Download the google speech dataset

wget http://download.tensorflow.org/data/speech_commands_v0.02.tar.gz
mkdir google_speech
tar -xvzf speech_commands_v0.02.tar.gz -C google_speech
rm speech_commands_v0.02.tar.gz

This implementation of the FastGRNN model has trouble handling datasets as large as the Google Speech Command Dataset. So we create a subset of the dataset with 12 classes of word utterances rather than 35. The classes are: 'yes', 'no', 'up', 'down', 'left', 'right', 'on', 'off', 'stop', 'go', 'zero', 'one'.

python subdatasetCreate.py

Install requirements.txt (and ensure dependency compatibility(#todo))

pip install -r requirements.txt

Install CUDA kernel
There is an unresolved issue with the CUDA kernel currently. Working on a fix.

cd cuda
python setup.py install

Run Model

python trainClassifier.py --dataset ./google_12 --epochs 30 --batch_size 64 --outdir OUTPUT_DIRECTORY --use_batchnorm

Adjust parameters in the TrainingConfig.py file to control various aspects of training