Training ConvNet models using Google Commands Dataset, implemented in PyTorch.

• Training and testing ConvNets.
• Arrange Google Commands Dataset, in an Train, Test, Valid folders for easy loading.
• Dataset loader.

Several libraries are needed to be installed in order to extract spectrograms and train the models.

• Install PyTorch if you haven't already :)
• Install LibRosa

Download and extract the Google Commands Dataset.

To make the arrange the data run the following command:

python make_dataset.py <google-command-folder> --out_path <path to save the data the new format>

You can also use the data loader and training scripts for your own custom dataset. In order to do so the dataset should be arrange in the following way:

root/up/kazabobo.wav
root/up/asdkojv.wav
root/up/lasdsa.wav
root/right/blabla.wav
root/right/nsdf3.wav
root/right/asd932.wav

Use python run.py --help for more parameters and options.

python run.py --train_path <train_data_path> --valid_path <valid_data_path> --test_path <test_data_path>

Accuracy results for the train, validation and test sets using two ConvNet models (LeNet5 and VGG11).

In order to reproduce the below results just exec the run.py file with default parameters. Results may be improved using deeper models (VGG13, VGG19), or better hyper-parameters optimization.