This repository implements SDCRKL-GP (Scalable Deep Convolutional Random Kernel Learning in Gaussian Process for Image Recognition).We directly present a demo of the algorithm on MNIST dataset.Training on other data sets, such as CIFAR10, caltech4,CIFAR100,tiny-ImageNet-200, etc., requires users to modify the network architecture design in model.py by themselves.The user should modify all absolute paths in the .py file as well as the path of the model
python 3.7.4
pytorch 1.7.0
git clone https://github.com/w-tingting/deep-rff-pytorch.git
cd experiments
python run.py
cd experiments
python test.py
| datasets | error rate (%) | nlpp | Parameters | FLOPs | MAC |
|---|---|---|---|---|---|
| MNIST | 0.60 | 0.020 | 19.088k | 0.984M | 1.057M |
| FMNIST | 7.22 | 0.229 | 12.176k | 0.336M | 0.892M |
| CIFAR10 | 27.28 | 0.811 | 22.448k | 2.454M | 1.582M |
| CALTECH4 | 2.81 | 0.086 | 47.024k | 15.434M | 19.021M |
| CIFAR100 | 60.68 | 2.468 | - | - | - |
| tiny-ImageNet-200 | 73.52 | 3.746 | - | - | - |