TorchCV: A PyTorch-Based Framework for Deep Learning in Computer Vision

@misc{you2019torchcv,
    author = {Ansheng You and Xiangtai Li and Zhen Zhu and Yunhai Tong},
    title = {TorchCV: A PyTorch-Based Framework for Deep Learning in Computer Vision},
    howpublished = {\url{https://github.com/donnyyou/torchcv}},
    year = {2019}
}

This repository provides source code for most deep learning based cv problems. We'll do our best to keep this repository up-to-date. If you do find a problem about this repository, please raise an issue or submit a pull request.

- Semantic Flow for Fast and Accurate Scene Parsing
- Code and models: https://github.com/lxtGH/SFSegNets

Implemented Papers

• Image Classification

  • VGG: Very Deep Convolutional Networks for Large-Scale Image Recognition
  • ResNet: Deep Residual Learning for Image Recognition
  • DenseNet: Densely Connected Convolutional Networks
  • ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices
  • ShuffleNet V2: Practical Guidelines for Ecient CNN Architecture Design
  • Partial Order Pruning: for Best Speed/Accuracy Trade-off in Neural Architecture Search

• Semantic Segmentation

  • DeepLabV3: Rethinking Atrous Convolution for Semantic Image Segmentation
  • PSPNet: Pyramid Scene Parsing Network
  • DenseASPP: DenseASPP for Semantic Segmentation in Street Scenes
  • Asymmetric Non-local Neural Networks for Semantic Segmentation
  • Semantic Flow for Fast and Accurate Scene Parsing

• Object Detection

  • SSD: Single Shot MultiBox Detector
  • Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
  • YOLOv3: An Incremental Improvement
  • FPN: Feature Pyramid Networks for Object Detection

• Pose Estimation

  • CPM: Convolutional Pose Machines
  • OpenPose: Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields

• Instance Segmentation

  • Mask R-CNN

• Generative Adversarial Networks

  • Pix2pix: Image-to-Image Translation with Conditional Adversarial Nets
  • CycleGAN: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks.

QuickStart with TorchCV

Now only support Python3.x, pytorch 1.3.

pip3 install -r requirements.txt
cd lib/exts
sh make.sh

Performances with TorchCV

All the performances showed below fully reimplemented the papers' results.

Image Classification

• ImageNet (Center Crop Test): 224x224

Semantic Segmentation

• Cityscapes (Single Scale Whole Image Test): Base LR 0.01, Crop Size 769

• ADE20K (Single Scale Whole Image Test): Base LR 0.02, Crop Size 520

Object Detection

• Pascal VOC2007/2012 (Single Scale Test): 20 Classes

Pose Estimation

• OpenPose: Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields

Instance Segmentation

• Mask R-CNN

Generative Adversarial Networks

• Pix2pix
• CycleGAN

DataSets with TorchCV

TorchCV has defined the dataset format of all the tasks which you could check in the subdirs of data. Following is an example dataset directory trees for training semantic segmentation. You could preprocess the open datasets with the scripts in folder data/seg/preprocess

Dataset
    train
        image
            00001.jpg/png
            00002.jpg/png
            ...
        label
            00001.png
            00002.png
            ...
    val
        image
            00001.jpg/png
            00002.jpg/png
            ...
        label
            00001.png
            00002.png
            ...

Commands with TorchCV

Take PSPNet as an example. ("tag" could be any string, include an empty one.)

• Training

  cd scripts/seg/cityscapes/
  bash run_fs_pspnet_cityscapes_seg.sh train tag

• Resume Training

  cd scripts/seg/cityscapes/
  bash run_fs_pspnet_cityscapes_seg.sh train tag

• Validate

  cd scripts/seg/cityscapes/
  bash run_fs_pspnet_cityscapes_seg.sh val tag

• Testing:

  cd scripts/seg/cityscapes/
  bash run_fs_pspnet_cityscapes_seg.sh test tag

Demos with TorchCV

Example output of VGG19-OpenPose

Example output of VGG19-OpenPose