zjcv/zcls

Object Classification Training Framework

Language: 🇺🇸 🇨🇳

«ZCls» is a classification model training/inferring framework

A more faster training framework is under building: ZCls2

Supported Recognizers:

Refer to roadmap for details

Table of Contents

• Table of Contents
• Background
• Installation
• Usage
• Maintainers
• Thanks
• Contributing
• License

Background

In the fields of object detection/object segmentation/action recognition, there have been many training frameworks with high integration and perfect process, such as facebookresearch/detectron2, open-mmlab/mmaction2 ...

Object classification is the most developed and theoretically basic field in deeplearning. Referring to the existing training framework, a training/inferring framework based on object classification model is implemented. I hope ZCls can bring you a better realization.

Installation

See INSTALL

Usage

How to train, see Get Started with ZCls

Use builtin datasets, see Use Builtin Datasets

Use custom datasets, see Use Custom Datasets

Use pretrained model, see Use Pretrained Model

Maintainers

• zhujian - Initial work - zjykzj

Thanks

@misc{ding2021diverse,
      title={Diverse Branch Block: Building a Convolution as an Inception-like Unit}, 
      author={Xiaohan Ding and Xiangyu Zhang and Jungong Han and Guiguang Ding},
      year={2021},
      eprint={2103.13425},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@misc{ding2021repvgg,
      title={RepVGG: Making VGG-style ConvNets Great Again}, 
      author={Xiaohan Ding and Xiangyu Zhang and Ningning Ma and Jungong Han and Guiguang Ding and Jian Sun},
      year={2021},
      eprint={2101.03697},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@misc{fan2020pyslowfast,
  author =       {Haoqi Fan and Yanghao Li and Bo Xiong and Wan-Yen Lo and
                  Christoph Feichtenhofer},
  title =        {PySlowFast},
  howpublished = {\url{https://github.com/facebookresearch/slowfast}},
  year =         {2020}
}

@misc{zhang2020resnest,
      title={ResNeSt: Split-Attention Networks}, 
      author={Hang Zhang and Chongruo Wu and Zhongyue Zhang and Yi Zhu and Haibin Lin and Zhi Zhang and Yue Sun and Tong He and Jonas Mueller and R. Manmatha and Mu Li and Alexander Smola},
      year={2020},
      eprint={2004.08955},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@misc{han2020ghostnet,
      title={GhostNet: More Features from Cheap Operations}, 
      author={Kai Han and Yunhe Wang and Qi Tian and Jianyuan Guo and Chunjing Xu and Chang Xu},
      year={2020},
      eprint={1911.11907},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

For more thanks, check THANKS

Contributing

Anyone's participation is welcome! Open an issue or submit PRs.

Small note:

• Git submission specifications should be complied with Conventional Commits
• If versioned, please conform to the Semantic Versioning 2.0.0 specification
• If editing the README, please conform to the standard-readme specification.

License

Apache License 2.0 © 2020 zjykzj

zcls's Issues

Installation docs don't appear to be working

I clicked on the link for installation and the page does not have any instructions for install: https://zcls.readthedocs.io/install.html

Can you provide pretrained model for ResNet18/50-DBB?

or convert the pretrained model from the official repo?

any plans to release lightweight pretrained models？like mobilenetv3_x0.33 shufflenetv2_x0.25

ERROR: Could not find a version that satisfies the requirement ztransforms

test_ddbblock.py cannot pass

Very nice work!
I find that the test_resnet50_dbb() test cannot pass in test_resnetZCls/tests/test_model/test_layer/test_dbblock.py, while the test_dbb_helper() works fine. The diff is bigger than 1. Any ideas what can cause this bug?
Thanks a lot!

Your network is very great.I want to regard it as backbone.Need to load pre-training parameters？

Your network is very great.I want to regard it as backbone.Need to load pre-training parameters？
For Example？I use ResNet，load resnet50-19c8e357.pth
Can you help?Thanks！

Some questions with using your code

It's a nice work!! But I have a few questions might need you to answer:

1. For the ZCLS version of “resnest50_fast_2s1x64d” and the Official version, the only difference is the FEATURE_DIMS obtained by backbone? I observed that the ZCLS version is 2048 and the official version is 1024. Are there any other differences?
2. The “resnest50_fast_2s1x64d” pretrained model you provided in Baidu Cloud Disk is ZCLS version? Could you provide an official version pretrained model in this table?
3. The last and most important question, if I want to use your pretraining model ("resnest50_fast_2s1x64d" for me) to train custom data, how do I set the learning rate to get the best performance? (Multi-GPU training)

1% accuracy lower than expectation with default config

Hi,

I'm new to this fantastic work. When I tried to test the inference accuracy of resnet-18 with r18_torchvision_imagenet_224.yaml/r18_zcls_imagenet_224.yaml, the result accuracy is 68.136%, which is lower than your reported result, 69.22%. I wonder if I need to modify the config file or anything else?

Really looking forwards to your reply! Thanks a lot for your contribution to the open-source community!

Can't find implementation for structural reparameterization trick proposed in the paper

Hi,
I'm interested in your structural reparameterization implementation but cannot find it in your repo. Did I miss it? Can you help?
Thanks!

ERROR: Could not find a version that satisfies the requirement ztransforms