This is a pytorch implementation of VoVNet backbone networks as described in the paper An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection. This is implemented through pytorch/vision style.
This repo provides VoVNet-39/57 models trained on ImageNet classification dataset with same training protocols as pytorch/examples/imagenet (e.g., 128 batch size, 90 epoch, data augmentations, lr_policy, etc) and compares to ResNet and DenseNet.
- For fair comparison, same training protocols used as pytorch/examples/imagenet
- 90 epoch
- step learning rate schedule (learning rate decayed at every 30 epoch)
- 256 batch size
- default augmentations (e.g., crop, flip, same mean/std normalization)
- @224x224 training/validation
- Inference time is measured on TITAN X PASCAL GPU
- CUDA v9.2, cuDNN v7.3, pytorch 1.0
Model | Top-1 | Top-5 | Inference time |
---|---|---|---|
ResNet-50 | 23.85% | 7.13% | 12 ms |
DenseNet-201 | 22.80% | 6.43% | 39 ms |
VoVNet-39 | 23.23% | 6.57% | 10 ms |
ResNet-101 | 22.63% | 6.44% | 20 ms |
DenseNet-161 | 22.35% | 6.20% | 27 ms |
VoVNet-57 | 22.27% | 6.28% | 13 ms |
- python 3.6
- Install pytorch > 0.4
git clone https://github.com/stigma0617/VoVNet.pytorch.git
- install depenecies :
pip install -r requirements.txt
You can follow the instruction of pytorch/examples/imagenet
run main.py
specifying data path, the desired model name--arch
, and save-directory--savedir
.
python main.py [imagenet-folder with train and val folders] --arch vovnet39 --savedir VoVNet39
VoVNet-Detectron VoVNet-DeepLabV3
- VoVNet-27-slim
- VoVNet-27-slim-depthwise
- VoVNet-99