by Yuesong Shen
This is the source code for experiments in paper "Deriving Neural Network Design and Learning from the Probabilistic Framework of Chain Graphs" by Yuesong Shen and Daniel Cremers.
This source code is released under the GPL v3 license. A part of the code is derived from the source code from https://github.com/akamaster/pytorch_resnet_cifar10 which is released under the BSD license.
If you find our implementation useful for your research, please consider citing our paper (Arxiv page):
@misc{shen2020deriving,
title={Deriving Neural Network Design and Learning from the Probabilistic Framework of Chain Graphs},
author={Yuesong Shen and Daniel Cremers},
year={2020},
eprint={2006.16856},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
Enjoy ;)
Package structure:
This code package has two independent folders:
-
mnist_simple/: This corresponds to the source code for the "Simple dense network" experiments. -
cifar_resnet/: This corresponds to the source code for the "Convolutional residual network" experiments. It is based on the source code from https://github.com/akamaster/pytorch_resnet_cifar10.
Dependencies:
- python (tested on Python 3.7)
- pytorch (tested on PyTorch 1.4.0)
- torchvision (tested on torchvision 0.5.0)
Usage of code in mnist_simple/:
Run python run_mnist.py --help to get the list of command line arguments and their descriptions.
Example usages:
-
Train a baseline with Tanh activation on MNIST, no dropout, no PCFF:
python run_mnist.py -d MNIST -a tanh -
Train a dropout baseline with ReLU activation on FashionMNIST with drop rate 0.5, no PCFF, use GPU (cuda):
python run_mnist.py -d FashionMNIST -a relu -p 0.5 -c -
Train a PCFF model with ReLU activation on MNIST with sample rate 1.0, no dropout, use GPU (cuda):
python run_mnist.py -d MNIST -a relu -s 1.0 -c -
Evaluate a pretrained model stored at "models/mymodel.pickle" with MNIST dataset, Tanh activation and PCFF sample rate 0.4:
python run_mnist.py -d MNIST -a tanh -s 0.4 -r "models/mymodel.pickle"
Usage of code in cifar_resnet/:
To run the full set of experiments, simply run bash run2.sh (might take a while to run all trainings one by one).
In case your system doesn't support bash script or you want to customize the training, you can have a look at the run2.sh file for example usages, as well as running python trainer2.py --help to get the list of command line arguments and their descriptions.