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
Model Train Test Top-1 Top-5 BS Iters Scripts
ResNet50 train val 77.54 93.59 512 30W ResNet50
ResNet101 train val 78.94 94.56 512 30W ResNet101
ShuffleNetV2x0.5 train val 60.90 82.54 1024 40W ShuffleNetV2x0.5
ShuffleNetV2x1.0 train val 69.71 88.91 1024 40W ShuffleNetV2x1.0
DFNetV1 train val 70.99 89.68 1024 40W DFNetV1
DFNetV2 train val 74.22 91.61 1024 40W DFNetV2

Semantic Segmentation

  • Cityscapes (Single Scale Whole Image Test): Base LR 0.01, Crop Size 769
Model Backbone Train Test mIOU BS Iters Scripts
PSPNet 3x3-Res101 train val 78.20 8 4W PSPNet
DeepLabV3 3x3-Res101 train val 79.13 8 4W DeepLabV3
  • ADE20K (Single Scale Whole Image Test): Base LR 0.02, Crop Size 520
Model Backbone Train Test mIOU PixelACC BS Iters Scripts
PSPNet 3x3-Res50 train val 41.52 80.09 16 15W PSPNet
DeepLabv3 3x3-Res50 train val 42.16 80.36 16 15W DeepLabV3
PSPNet 3x3-Res101 train val 43.60 81.30 16 15W PSPNet
DeepLabv3 3x3-Res101 train val 44.13 81.42 16 15W DeepLabV3

Object Detection

  • Pascal VOC2007/2012 (Single Scale Test): 20 Classes
Model Backbone Train Test mAP BS Epochs Scripts
SSD300 VGG16 07+12_trainval 07_test 0.786 32 235 SSD300
SSD512 VGG16 07+12_trainval 07_test 0.808 32 235 SSD512
Faster R-CNN VGG16 07_trainval 07_test 0.706 1 15 Faster R-CNN

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

GitHub

https://github.com/donnyyou/torchcv