Original (messy) source code for Principal Component Networks: Parameter Reduction Early in Training.
Code for the paper can be found in the pcn folder. The beginnings of a
refactor meant to clean up the code can be found in pcns_v2.
Forward (input) transformation for CIFAR complete. ImageNet code & backward (output) transformation not started. Additional adversarial training functionality has been added.
The imports are with respect to the base directory. This may cause errors depending on what directory the code is run from. To avoid this simply:
export PYTHONPATH=$PYTHONPATH:{path to base directory}
or in your Dockerfile:
ENV PYTHONPATH "${PYTHONPATH}:{path to base directory}"
python src/pcns_v2/experiment_scripts/vgg16A_on_cifar.py
python src/pcns_v2/experiment_scripts/resnet_on_cifar.py
python src/pcns_v2/adversarial/adv_resnet_on_cifar.py
python src/pcns/vgg16.py
python src/pcns/resnet/cifar10.py