This is an implementation of DSOD in Pytorch. It is based on the code dsod.pytorch and torchcv

Origin implementation is here and here is the paper

I can train it on PASCAL VOC dataset and the loss also converges, but I am not sure it can achive the same scores as in the paper. Some more modifications need to be done.

python 2.7
pytorch 0.4
visdom

1. Download this repo

git clone git@github.com:qqadssp/DSOD-Pytorch  
cd DSOD-Pytorch  

2. Download Pascal VOC dataset and unzip it, its path should be {root_dir}/VOCdevkit

3. Modify opt.train_img_root in torchcv/utils/config.py with proper img_path

4. Start visdom server and begin to train

python -m visdom.server  
python train.py main  

1. After training some epochs, checkpoint will be saved with name 'dsod.pth' or '##.pth' like '39.pth'. Modify opt.load_path in config.py with 'checkpoint/dsod.pth' or 'checkpoint/39.pth'

2. Download Pascal VOC testset, modify opt.eval_img_root in config.py with proper path

3. Evaluate the model

python eval.py

Training dataset is used here. If you have trained some epochs and get checkpoint, modify opt.load_path and run

python demo.py