yiweifengyan / Pytorch-XNOR-Net

XNOR-Net, with binary gemm and binary conv2d kernels, support both CPU and GPU.

Pytorch-XNOR-Net

Build

cd <Repository Root>/csrc/binop
make

MNIST

Usage

Train:

cd <Repository Root>/MNIST/
python3 main.py --arch Bin_LeNet
python3 main.py --arch LeNet

Evaluate:

cd <Repository Root>/MNIST/
python3 main.py --arch Bin_LeNet --evaluate --pretrained ./models/Bin_LeNet.best.pth # --no_cuda (Use CPU)
python3 main.py --arch LeNet --evaluate --pretrained ./models/LeNet.best.pth # --no_cuda (Use CPU)

Result

Network	Accuracy	Size
LeNet	99.50%	1.7 MB
Bin_LeNet	99.45%	102 KB

Cifar10

Usage

Train:

cd <Repository Root>/Cifar10/
python3 main.py --arch Bin_VGG16 #(11, 13, 16, 19)
python3 main.py --arch VGG16 #(11, 13, 16, 19)

Evaluate:

cd <Repository Root>/Cifar10/
python3 main.py --arch Bin_VGG16 --evaluate --pretrained ./models/Bin_VGG16.best.pth # --no_cuda (Use CPU)
python3 main.py --arch VGG16 --evaluate --pretrained ./models/VGG16.best.pth # --no_cuda (Use CPU)

Result

Network	Accuracy	Size
VGG13	92.40%	37.7 MB
Bin_VGG13	88.74%	1.3 MB
VGG16	92.29%	59.0 MB
Bin_VGG16	87.78%	2.0 MB

Pre-trained models

Google Drive

Environment

Software

• Ubuntu 16.04
• Python 3.5
• Pytorch 0.3.1
• CUDA 8.0
• gcc 5.4

Hardware

• NVIDIA GTX 1080
• Intel i5-6500 CPU @ 3.20GHz × 4

Reference

• Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1
• XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks
• https://github.com/jiecaoyu/XNOR-Net-PyTorch
• cpu-gemm
• cpu-conv2d
• gpu-gemm and gpu-conv2d
• popcount