Train ResNet18/34/50 on Encrypted CIFAR10/100 with Pytorch

This is a modified variant of ResNet trained on encrypted CIFAR10/100 dataset using pytorch. Trying to apply different tricks on modified ResNet to get better performance with encrypted dataset.

Features:

• Modified ResNet architecture to fit CIFAR dataset.
• Added visualization tool to get feature maps of layers.

Results

Train Result

Feature Map

Layer Output

Get Started

Environment

Make sure you have Nvidia GPU and installed CUDA and CUDNN.

cd ResNetToy
pip install -r requirements.txt
mkdir checkpoint dataset runs

Structure of dataset should be like:

└─dataset
    └─10
        │  test_label.txt
        │  train_label.txt
        │
        ├─test
        └─train

Train

Change to your paths in:

train.py
utils/readData.py

Then run:

python train.py

Test

Change to your paths in:

test.py
utils/readData.py

Then run:

python test.py

Visualization

Change to your paths in:

vis_tool.py

Then run:

python vis_tool.py