soumith / spectral-lib

To run experiments, cd to test. All parameters are listed in params.lua. To modify the GFT matrix, change the path at the beginning of model.lua to load the proper thing.

The -model parameter specfies model architecture. fc2 refers to fully connected with 2 layers, etc. gconv2 refers to 2 layers of graph convolution followed by a fully connected layer. The -nhidden parameter refers to number of hidden units for fully connected models and number of feature maps for graph convolutions. The -k parameter refers to the number of subsampled weights in the graph convolution layers.

To train on TIMIT with a 3-layer fully connected net with 2000 hidden units:

th -i train.lua -dataset timit -model fc3 -nhidden 2000

To train on TIMIT with a net consisting of 2 layers of graph conv followed by a fully connected, with 20 feature maps and 20 subsampled weights at each layer:

th -i train.lua -dataset timit -model gconv2 -nhidden 20 -k 20