Skip to content

Latest commit

 

History

History
86 lines (60 loc) · 4.7 KB

README_bak.md

File metadata and controls

86 lines (60 loc) · 4.7 KB
Raw

CVPODS

Introduction

A versatile and efficient codebase for many computer vision tasks: classification, segmentation, detection, self-supervised learning, keypoints and 3D, etc.

Features

  • Clean & simple & flexible development: When using detectron2, if you want to implement a new module such as CustomRetinanet, you need to register it to meta_arch, then specify in xxx_config.yaml, and you still need to esplicitly invoke 'from net import CustomRetinanet' to allow registry to retrieve your module successfully. It means you need to copy train_net.py from common tools directory and insert the line above;
  • Flexible and easy-to-use configuration system: When add new config options in Detectron2, you need to add it into config/defaults.py first and then modify the config.yaml. But now in cvpods you just need to add to config.py once. When you need to debug a component, you may need to set SOLVER.IMS_PER_BATCH to 2, before you need to modify it in config, after it starts running correctly, you need to modify it to 16. That's unconvenient too. So ​cvpods allow you to dynamicly update config, for example: pods_train --num-gpus 1 SOLVER.IMS_PER_BATCH 2.
  • Task specific incremental updating.: For example, if you need to modify Retinanet relative configurations, you just need to modify retinanet_config.py and don't need care other common configs. On the other hand, we avoid putting all kinds of methods' configuration all in one base config file(such as detectron2/config/defaults.py) like detectron2, maskrcnn_benchmark and mmdetection. So retinanet will not include ROI_HEADS, MASK_HEADS configurations, but only has all necessary component.
  • Efficient experiments management: When you need to implement a new model, you can either copy a project from examples and inheritate some kind of networks such as RetinaNet of FasterRCNN, then define your custom functions; or you can add a new base / commonly used model(such as FCOS) into `cvpods/modeling/meta_arch' and using it like a library.
  • Versatile tasks & datasets support:
    • Detection, Segmentation (Semantic, Panoptic, Instance), Keypoint, Self-supervised Learning, 3D Detection & Segmentation, etc.
    • COCO, Objects365, WiderFace, VOC, LVIS, CityPersons, ImageNet, CrowdHuman, CityScapes, ModelNet40, ScanNet, KITTI, nuScenes, etc.
  • Global training / testing scripts.: you just need to invoke pods_train/test --num-gpus x in your playground; and your projects only need to include all project-specific configs and network modules.
  • Compatible with detectron2: All models in detectron2 can be easily migrated into cvpods.

News

  • 2020.07 cvpods is released.

Requirements

  • CUDA 10.1 & cuDNN 7.6.3 & nccl 2.4.8 (optional)
  • Python >= 3.6
  • PyTorch >= 1.3
  • torchvision >= 0.4.2
  • OpenCV
  • pycocotools
  • GCC >= 4.9

Get Started

# Install cvpods (requires GPU) locally
python -m pip install 'git+https://github.com/Megvii-BaseDetection/cvpods.git'
# (add --user if you don't have permission)

# Or, to install it from a local clone:
git clone https://github.com/Megvii-BaseDetection/cvpods.git
python -m pip install -e cvpods 

# Or,
pip install -r requirements.txt
python setup.py build develop

# Preprare data path
ln -s /path/to/your/coco/dataset datasets/coco

# Enter a specific experiment dir 
cd playground/retinanet/retinanet.res50.fpn.coco.multiscale.1x

# Train
pods_train --num-gpus 8
# Test
pods_test --num-gpus 8 \
    MODEL.WEIGHTS /path/to/your/save_dir/ckpt.pth # optional
    OUTPUT_DIR /path/to/your/save_dir # optional

# Multi node training
## sudo apt install net-tools ifconfig
pods_train --num-gpus 8 --num-machines N --machine-rank 0/1/.../N-1 --dist-url "tcp://MASTER_IP:port"

Model ZOO

For all the models supported by cvpods, please refer to MODEL_ZOO.

Projects based on cvpods

Acknowledgement

cvpods is developed based on Detectron2. For more details about official detectron2, please check DETECTRON2