Convolutional kernel network with Pytorch

Re-implementation of Convolutional Kernel Network (CKN) from Mairal (2016) in Python based on the Pytorch framework. The package is available under the GPL-v3 license.

Author: Dexiong Chen

Credits: Ghislain Durif, Mathilde Caron, Alberto Bietti, Julien Mairal

The code is based on

Mairal, Julien. End-to-end kernel learning with supervised convolutional kernel networks. NIPS 2016.

If you have any issues, please contact dexiong.chen@inria.fr.

Installation

We strongly recommend users to use anaconda to install the following packages

numpy
scipy
scikit-learn
pytorch=1.2.0
miso_svm

The Python package miso_svm can be installed with (original repository)

cd third-party/miso_svm-1.0
python setup.py install

Results

Reproduction of the results from Mairal (2016) with this package. The results from the original paper (Mairal, 2016) were achieved using cudnn-based Matlab code available here. To run the following experiments, please first download the data, put into the folder ./data/cifar-10 and then do

export PYTHONPATH=$PWD:$PYTHONPATH
cd experiments

Unsupervised CKN

Here is a summary of the results of unsupervised CKN on CIFAR10 image classification dataset with pre-whitening and without data augmentation or model ensembling.

# Code examples
python cifar10_unsup.py --filters 64 256 --subsamplings 2 6 --kernel-sizes 3 3

#layers	#filters	filter size	subsampling	sigma	Accuracy
2	64, 256	3, 3	2, 6	0.6	77.5
2	256, 1024	3, 3	2, 6	0.6	82.0
2	512, 8192	3, 2	2, 6	0.6	84.0

Supervised CKN

Here is a summary of the results of supervised CKN on CIFAR10 image classification dataset with pre-whitening and without data augmentation or model ensembling.

# Code examples
python cifar10_sup.py --epochs 105 --lr 0.1 --alpha 0.001 --loss hinge --alternating --model ckn5
python cifar10_sup.py --epochs 105 --lr 0.1 --alpha 0.1 --loss hinge --alternating --model ckn14

Architecture	Accuracy	training time (GTX1080_ti)
CKN-5	86.1	~60 min
CKN-14	90.2	~260 min