This is a Pytorch implementation of DoreFa with CIFAR-100 dataset.
Use git clone to run this project.
git clone https://github.com/phuocphn/dorefa-cifar100Train full precision baseline model.
mkdir logs
python main.py --bit 32 \
--log_name resnet18_w32a32-ceres-01 \
--arch resnet18 --wd 5e-4 \
--max_epochs 250 >> logs/resnet18_w32a32-ceres-01.txtTrain W2A2 model.
mkdir logs
python main.py --bit 2 \
--log_name resnet18_w2a2-ceres-01 \
--arch resnet18 --wd 5e-4 \
--max_epochs 250 >> logs/resnet18_w2a2-ceres-01.txtTrain W4A4 model.
mkdir logs
python main.py --bit 4 \
--log_name resnet18_w4a4-ceres-01 \
--arch resnet18 --wd 5e-4 \
--max_epochs 250 >> logs/resnet18_w4a4-ceres-01.txt(Quantized models are trained from scratch.)
| Model | Top@1 Accuracy |
|---|---|
| FP32 | 76.00 |
| W4A4 | 76.20 |
| W2A2 | 74.02 |
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.
Please make sure to update tests as appropriate.