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

Clone the repository.

git clone https://github.com/adithom/KWS.git

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

pip install -r requirements.txt

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' and 'noise'.

python subdatasetCreate.py

Created augmented dataset from the 12 classes. Applies audio augmentation techniques to the dataset, this approach with multiple scripts is a bit redundant for dataset manipulation so this is a temporary solution.

python audioAugmentation.py

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 

Adjust parameters in the TrainingConfig.py file to control various aspects of training. There are currently issues with inference for batch norm and training on GPU using custom CUDA acceleration.

Inference

Inference is currently segmented for models trained with and without batchnorm. So there are two inference scripts.