Script for training hexagonal convolutional neural networks on 6x6 SAM patch patterns

The easiest way to get all of the dependencies is to use the Anaconda Python Distribution (w/ Python 3.6.x)

You'll also need PyTorch and HexagDLy. HexagDLy provides support for convolutional and pooling operations for inputs arranged on hexagonal grids.

Make sure you're using the version of Pip installed with the Anaconda distribution

PyTorch:

with conda:

conda install pytorch torchvision -c pytorch

with pip:

pip3 install torch torchvision

HexagDLy:

pip install hexagdly

All of the training data (i.e, patch patterns and their free energies) are contained in sam_pattern_data.dat.npz.

You can train a model using model_fnot.py

For a list of training options, run:

python model_fnot.py -h

For instance, to train a CNN with 6 convolutional filters, 1 hidden layer, and 12 hidden nodes:

python model_fnot.py --augment-data --n-layers 1 --n-hidden 12 --do-conv --n-out-channels 6

Nick Rego ([email protected])