September 16, 2019

840 words 4 mins read



Fast, modular reference implementation and easy training of Semantic Segmentation algorithms in PyTorch.

repo name ycszen/TorchSeg
repo link
language Python
size (curr.) 2507 kB
stars (curr.) 1038
created 2019-01-10
license MIT License


This project aims at providing a fast, modular reference implementation for semantic segmentation models using PyTorch.

demo image


  • Modular Design: easily construct customized semantic segmentation models by combining different components.
  • Distributed Training: >60% faster than the multi-thread parallel method(nn.DataParallel), we use the multi-processing parallel method.
  • Multi-GPU training and inference: support different manners of inference.
  • Provides pre-trained models and implement different semantic segmentation models.


  • PyTorch 1.0
    • pip3 install torch torchvision
  • Easydict
    • pip3 install easydict
  • Apex
  • Ninja
    • sudo apt-get install ninja-build
  • tqdm
    • pip3 install tqdm


v0.1.1 (05/14/2019)

  • Release the pre-trained models and all trained models
  • Add PSANet for ADE20K
  • Add support for CamVid, PASCAL-Context datasets
  • Start only supporting the distributed training manner

Model Zoo

Pretrained Model

Supported Model

Performance and Benchmarks

SS:Single Scale MSF:Multi-scale + Flip


Methods Backbone TrainSet EvalSet Mean IoU(ss) Mean IoU(msf) Model
FCN-32s R101_v1c train_aug val 71.26 -
DFN(paper) R101_v1c train_aug val 79.67 80.6*
DFN(ours) R101_v1c train_aug val 79.40 81.40 GoogleDrive

80.6*: this result reported in paper is further finetuned on train dataset.


Non-real-time Methods

Methods Backbone OHEM TrainSet EvalSet Mean IoU(ss) Mean IoU(msf) Model
DFN(paper) R101_v1c train_fine val 78.5 79.3
DFN(ours) R101_v1c train_fine val 79.09 80.41 GoogleDrive
DFN(ours) R101_v1c train_fine val 79.16 80.53 GoogleDrive
BiSeNet(paper) R101_v1c train_fine val - 80.3
BiSeNet(ours) R101_v1c train_fine val 79.09 80.39 GoogleDrive
BiSeNet(paper) R18 train_fine val 76.21 78.57
BiSeNet(ours) R18 train_fine val 76.28 78.00 GoogleDrive
BiSeNet(paper) X39 train_fine val 70.1 72
BiSeNet(ours)* X39 train_fine val 70.32 72.06 GoogleDrive

Real-time Methods

Methods Backbone OHEM TrainSet EvalSet Mean IoU Model
BiSeNet(paper) R18 train_fine val 74.8
BiSeNet(ours) R18 train_fine val 74.83 GoogleDrive
BiSeNet(paper) X39 train_fine val 69
BiSeNet(ours)* X39 train_fine val 68.51 GoogleDrive

BiSeNet(ours)*: because we didn’t pre-train the Xception39 model on ImageNet in PyTorch, we train this experiment from scratch. We will release the pre-trained Xception39 model in PyTorch and the corresponding experiment.


Methods Backbone TrainSet EvalSet Mean IoU(ss) Accuracy(ss) Model
PSPNet(paper) R50_v1c train val 41.68 80.04
PSPNet(ours) R50_v1c train val 41.65 79.74 GoogleDrive
PSPNet(paper) R101_v1c train val 41.96 80.64
PSPNet(ours) R101_v1c train val 42.89 80.55 GoogleDrive
PSANet(paper) R50_v1c train val 41.92 80.17
PSANet(ours)* R50_v1c train val 41.67 80.09 GoogleDrive
PSANet(paper) R101_v1c train val 42.75 80.71
PSANet(ours) R101_v1c train val 43.04 80.56 GoogleDrive

PSANet(ours)*: The original PSANet in the paper constructs the attention map with over-parameters, while we only predict the attention map with the same size of the feature map. The performance is almost similar to the original one.

To Do

  • offer comprehensive documents
  • support more semantic segmentation models
    • Deeplab v3 / Deeplab v3+
    • DenseASPP
    • EncNet
    • OCNet


  1. create the config file of dataset:train.txt, val.txt, test.txt
    file structure:(split with tab)
    path-of-the-image   path-of-the-groundtruth
  2. modify the according to your requirements
  3. train a network:

Distributed Training

We use the official torch.distributed.launch in order to launch multi-gpu training. This utility function from PyTorch spawns as many Python processes as the number of GPUs we want to use, and each Python process will only use a single GPU.

For each experiment, you can just run this script:

export NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS


In the evaluator, we have implemented the multi-gpu inference base on the multi-process. In the inference phase, the function will spawns as many Python processes as the number of GPUs we want to use, and each Python process will handle a subset of the whole evaluation dataset on a single GPU.

  1. evaluate a trained network on the validation set:
  2. input arguments:
    usage: -e epoch_idx -d device_idx [--verbose ] 
    [--show_image] [--save_path Pred_Save_Path]


This project is under active development. So things that are currently working might break in a future release. However, feel free to open issue if you get stuck anywhere.


The following are BibTeX references. The BibTeX entry requires the url LaTeX package.

Please consider citing this project in your publications if it helps your research.

  author =       {Yu, Changqian},
  title =        {TorchSeg},
  howpublished = {\url{}},
  year =         {2019}

Please consider citing the DFN in your publications if it helps your research.

  title={Learning a Discriminative Feature Network for Semantic Segmentation},
  author={Yu, Changqian and Wang, Jingbo and Peng, Chao and Gao, Changxin and Yu, Gang and Sang, Nong},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},

Please consider citing the BiSeNet in your publications if it helps your research.

  title={Bisenet: Bilateral segmentation network for real-time semantic segmentation},
  author={Yu, Changqian and Wang, Jingbo and Peng, Chao and Gao, Changxin and Yu, Gang and Sang, Nong},
  booktitle={European Conference on Computer Vision},

Why this name, Furnace?

Furnace means the Alchemical Furnace. We all are the Alchemist, so I hope everyone can have a good alchemical furnace to practice the Alchemy. Hope you can be a excellent alchemist.

comments powered by Disqus