XNOR-Net-PyTorch

This is a PyTorch implementation of XNOR-Net, adapted to event-based datasets like N-MNIST.

---

📦 Environment

Tested with:

• Python 3.10.16
• PyTorch 2.7.0 + CUDA 11.8
• TorchVision 0.22.0
• OpenCV 4.11.0
• NumPy 1.24.4
• Pandas 2.2.3
• Matplotlib 3.10.3
• Scikit-learn 1.6.1
• tqdm 4.67.1

---

🧪 Setup

🔹 Option 1: Conda (recommended)

conda create -n xnor_net python=3.10 -y
conda activate xnor_net
pip install -r requirements.txt

If you're using CUDA 11.8 and need the official PyTorch binaries:

pip install torch==2.7.0+cu118 torchvision==0.22.0+cu118 --index-url https://download.pytorch.org/whl/cu118

---

🧠 N-MNIST: LeNet-5 Training

For local testing:

cd N-MNIST/
python main.py --batch-size 2 --test-batch-size 2 --epochs 10

For dgx-1 training

 python main.py --batch-size 32 --test-batch-size 32 --epochs 10 --input-dir ../input/nmnist_events_frame/ --output-dir ../output/ --num-workers 4

🧠 CIFAR10DVS: VGG-16 Training

For dgx-1 training

python main_vgg16.py --batch-size 32 --test-batch-size 32 --epochs 10 --input-dir ../input/cifar10dvs_event_frames_ts/ --output-dir ../output/cifar10dvs_event_frames_ts --num-workers 4

---

📁 Outputs

• Trained model: output/LeNet_5.best.pth.tar
• Confusion matrix: output/best_confusion_matrix.jpg
• Prediction log: output/test_preds.csv
• Training accuracy curve: output/train_accuracy_plot.pdf

---

📄 License

MIT. Based on the original XNOR-Net.

C++ Compile

g++ -std=c++11 test_binactive.cpp -I /usr/include/eigen3 -o test_binactive
./test_binactive

g++ fullbwresnet.cpp test_binactive.cpp     -I /usr/include/eigen3     -O3 -march=native     -DEIGEN_MAX_THREADS=4     -pthread     -o test_binactive
./test_binactive

g++ fullbwresnet.cpp test_binactive.cpp \
    -std=c++17 -O3 -ffast-math -funroll-loops \
    -march=native -mtune=native -mfma \
    -DNDEBUG -DEIGEN_NO_DEBUG \
    -I/usr/include/eigen3 -pthread -fopenmp \
    -o test_binactive

g++ fullbwresnet.cpp test_binactive.cpp \
    -std=c++17 -O3 -ffast-math -funroll-loops \
    -march=native -mtune=native -mfma \
    -flto \
    -fopenmp -pthread \
    -DEIGEN_NO_DEBUG -DEIGEN_DONT_PARALLELIZE -DNDEBUG \
    -I/usr/include/eigen3 \
    -o test_binactive

cOMP_PROC_BIND=true OMP_NUM_THREADS=4 ./test_binactive

XNOR ACTIVE

g++ xnorresnet.cpp test_xnoractive.cpp \
  -std=c++17 -O3 -ffast-math -funroll-loops \
  -march=native -mtune=native -mavx2 -mfma -mbmi2 \
  -flto -fopenmp -fopenmp-simd -pthread \
  -DEIGEN_NO_DEBUG -DNDEBUG \
  -I/usr/include/eigen3 -I/usr/local/include \
  -o test_xnoractive


cOMP_PROC_BIND=true OMP_NUM_THREADS=4 ./test_xnoractive

perf stat ./test_xnoractive perf record ./test_xnoractive perf report --stdio > perf_report.txt

git clone https://github.com/xtensor-stack/xsimd.git

cd xsimd mkdir build cd build cmake -DCMAKE_INSTALL_PREFIX=/usr/local .. make -j$(nproc) sudo make install