Budgeted Super Networks

Original implementation of the Learning Time/Memory-Efficient Deep Architectures with Budgeted Super Networks

Installation

pip install -r requirements.txt

Running

python bsn_main.py

The available parameters can be seen using python bsn_main.py -h For exemple to run the Budgeted Super Networks on Cifar10 using the 8 layers/128 channels B-CNF architecture:

python bsn_main.py -arch CNF -layers 8 -channels 128 -dset CIFAR10

All plotting is done through Visdom. The server can be configured using the resources/visdom.json file.

CUDA usage can be enabled using the -cuda n flag, where n corresponds to the index of the GPU.

# To use the first GPU of the machine:
python bsn_main.py -arch CNF -layers 8 -channels 128 -dset CIFAR10 -cuda 0