August 11, 2019

868 words 5 mins read



High quality, fast, modular reference implementation of SSD in PyTorch

repo name lufficc/SSD
repo link
language Python
size (curr.) 773 kB
stars (curr.) 710
created 2018-12-06
license MIT License

High quality, fast, modular reference implementation of SSD in PyTorch 1.0

This repository implements SSD (Single Shot MultiBox Detector). The implementation is heavily influenced by the projects ssd.pytorch, pytorch-ssd and maskrcnn-benchmark. This repository aims to be the code base for researches based on SSD.

Losses Learning rate Metrics
losses lr metric


  • PyTorch 1.0: Support PyTorch 1.0 or higher.
  • Multi-GPU training and inference: We use DistributedDataParallel, you can train or test with arbitrary GPU(s), the training schema will change accordingly.
  • Modular: Add your own modules without pain. We abstract backbone,Detector, BoxHead, BoxPredictor, etc. You can replace every component with your own code without change the code base. For example, You can add EfficientNet as backbone, just add (ALREADY ADDED) and register it, specific it in the config file, It’s done!
  • CPU support for inference: runs on CPU in inference time.
  • Smooth and enjoyable training procedure: we save the state of model, optimizer, scheduler, training iter, you can stop your training and resume training exactly from the save point without change your training CMD.
  • Batched inference: can perform inference using multiple images per batch per GPU.
  • Evaluating during training: eval you model every eval_step to check performance improving or not.
  • Metrics Visualization: visualize metrics details in tensorboard, like AP, APl, APm and APs for COCO dataset or mAP and 20 categories' AP for VOC dataset.
  • Auto download: load pre-trained weights from URL and cache it.



  1. Python3
  2. PyTorch 1.0 or higher
  3. yacs
  4. Vizer
  5. GCC >= 4.9
  6. OpenCV

Step-by-step installation

git clone
cd SSD
#Required packages
pip install torch torchvision yacs tqdm opencv-python vizer

# Optional packages
# If you want visualize loss curve. Default is enabled. Disable by using --use_tensorboard 0 when training.
pip install tensorboardX

# If you train coco dataset, must install cocoapi.
cd ~/github
git clone
cd cocoapi/PythonAPI
python build_ext install


If your torchvision >= 0.3.0, nms build is not needed! We also provide a python-like nms, but is very slower than build-version.

# For faster inference you need to build nms, this is needed when evaluating. Only training doesn't need this.
cd ext
python build_ext develop


Setting Up Datasets

Pascal VOC

For Pascal VOC dataset, make the folder structure like this:

|__ VOC2007
    |_ JPEGImages
    |_ Annotations
    |_ ImageSets
    |_ SegmentationClass
|__ VOC2012
    |_ JPEGImages
    |_ Annotations
    |_ ImageSets
    |_ SegmentationClass
|__ ...

Where VOC_ROOT default is datasets folder in current project, you can create symlinks to datasets or export VOC_ROOT="/path/to/voc_root".


For COCO dataset, make the folder structure like this:

|__ annotations
    |_ instances_valminusminival2014.json
    |_ instances_minival2014.json
    |_ instances_train2014.json
    |_ instances_val2014.json
    |_ ...
|__ train2014
    |_ <im-1-name>.jpg
    |_ ...
    |_ <im-N-name>.jpg
|__ val2014
    |_ <im-1-name>.jpg
    |_ ...
    |_ <im-N-name>.jpg
|__ ...

Where COCO_ROOT default is datasets folder in current project, you can create symlinks to datasets or export COCO_ROOT="/path/to/coco_root".

Single GPU training

# for example, train SSD300:
python --config-file configs/vgg_ssd300_voc0712.yaml

Multi-GPU training

# for example, train SSD300 with 4 GPUs:
export NGPUS=4
python -m torch.distributed.launch --nproc_per_node=$NGPUS --config-file configs/vgg_ssd300_voc0712.yaml SOLVER.WARMUP_FACTOR 0.03333 SOLVER.WARMUP_ITERS 1000

The configuration files that I provide assume that we are running on single GPU. When changing number of GPUs, hyper-parameter (lr, max_iter, …) will also changed according to this paper: Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour.


Single GPU evaluating

# for example, evaluate SSD300:
python --config-file configs/vgg_ssd300_voc0712.yaml

Multi-GPU evaluating

# for example, evaluate SSD300 with 4 GPUs:
export NGPUS=4
python -m torch.distributed.launch --nproc_per_node=$NGPUS --config-file configs/vgg_ssd300_voc0712.yaml


Predicting image in a folder is simple:

python --config-file configs/vgg_ssd300_voc0712.yaml --images_dir demo --ckpt

Then it will download and cache vgg_ssd300_voc0712.pth automatically and predicted images with boxes, scores and label names will saved to demo/result folder by default.

You will see a similar output:

(0001/0005) 004101.jpg: objects 01 | load 010ms | inference 033ms | FPS 31
(0002/0005) 003123.jpg: objects 05 | load 009ms | inference 019ms | FPS 53
(0003/0005) 000342.jpg: objects 02 | load 009ms | inference 019ms | FPS 51
(0004/0005) 008591.jpg: objects 02 | load 008ms | inference 020ms | FPS 50
(0005/0005) 000542.jpg: objects 01 | load 011ms | inference 019ms | FPS 53


Origin Paper:

VOC2007 test coco test-dev2015
SSD300* 77.2 25.1
SSD512* 79.8 28.8


Backbone Input Size box AP Model Size Download
VGG16 300 25.2 262MB model
VGG16 512 29.0 275MB model


Backbone Input Size mAP Model Size Download
VGG16 300 77.7 201MB model
VGG16 512 80.7 207MB model
Mobilenet V2 320 68.9 25.5MB model
EfficientNet-B3 300 73.9 97.1MB model

Develop Guide

If you want to add your custom components, please see for more details.


If you have issues running or compiling this code, we have compiled a list of common issues in If your issue is not present there, please feel free to open a new issue.


If you use this project in your research, please cite this project.

    author = {Congcong Li},
    title = {{High quality, fast, modular reference implementation of SSD in PyTorch}},
    year = {2018},
    howpublished = {\url{}}
comments powered by Disqus