A pytorch Implementation of VoVNet Backbone Networks

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.

What does this repo provide?

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.

ImageNet results

Notes:

• 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

Pretrained models

• VoVNet-39

• VoVNet-57

Training & Inferecne

Installation

0. python  3.6

1. Install pytorch > 0.4

2. git clone https://github.com/stigma0617/VoVNet.pytorch.git

3. install depenecies : pip install -r requirements.txt

Imagenet data preparation

You can follow the instruction of pytorch/examples/imagenet

Train

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

Related projects

VoVNet-DetectronVoVNet-DeepLabV3

TO DO

• VoVNet-27-slim

• VoVNet-27-slim-depthwise

• VoVNet-99