llzlcl/continual_learning_cv

Continual Learning Toolbox for Computer Vision Tasks

Continual_Learning_CV (CLCV)

Continual Learning Toolbox for Computer Vision Tasks

This toolbox aims at prototyping current computer vision tasks, e.g., human gesture recognition, action localization/detection, object detection/segmentation, and person ReID in a continual/lifelong learning manner. It means most of the SOTAs can be updated with novel data without retraining from scratch, and at the same time, they are able to migrate from catastrophic forgetting problem, furthermore, the models can learn with few-shot samples and adapt quickly to the target domains. Since the CL strategies are quite complex and flexible, it has some intersections with recent few-shot/meta/multi-task learning work.

Datasets and Benchmarks

We are testing the performances based on OpenLORIS-Object dataset. The basic codes are the implementation of the following paper:

Qi She et al, OpenLORIS-Object: A Robotic Vision Dataset and Benchmark for Lifelong Deep Learning The paper has been accepted into ICRA 2020.

Also permutated MNIST and CIFAR-100 datasets are tested

Requirements [not hard constraints]

The current version of the code has been tested with following libs:

• pytorch 1.1.0
• torchvision 0.2.1
• tqdm 4.19.9
• visdom 0.1.8.9
• Pillow 6.2.0
• Pandas 1.0.3

Experimental platforms:

• Intel Core i9 CPU
• Nvidia RTX 2080 Ti GPU
• CUDA Toolkit 10.*

Install the required the packages inside the virtual environment:

$ conda create -n yourenvname python=3.7 anaconda
$ source activate yourenvname
$ pip install -r requirements.txt

Data Preparation

OpenLORIS-Object

For MNIST and CIFAR-100 datasets, Please refer to `./benchmarks/Readme.md`. 

Step 1: Download data (including RGB-D images, masks, and bounding boxes) following this instruction.

Step 2: Run following scripts:

 python3 benchmark1.py
 python3 benchmark2.py

Step 3: Put train/test/validation file under ./bechmakrs/data/OpenLORIS-Object. For more details, please follow note file under each sub-directories in ./img.

Step 4: Generate the .pkl files of data.

 python3 pk_gene.py
 python3 pk_gene_sequence.py

Quick Start

You can directly use scripts on 9 algorithms and 2 benchmarks stated in the paper (may need to modify arguments/parameters in .bash files if necessary):

bash clutter.bash
bash illumination.bash
bash pixel.bash
bash occlusion.bash
bash sequence.bash

Running Benchmark 1

Individual experiments can be run with main.py. Main option is:

python3 main.py --factor

which kind of experiment? (clutter|illumination|occlusion|pixel)

Running Benchmark 2

The main option to run benchmark2 is:

python3 main.py --factor=sequence

Running specific baseline methods

• Elastic weight consolidation (EWC):

main.py --ewc --savepath=ewc

• Online EWC:

main.py --ewc --online --savepath=ewconline

• Synaptic intelligence (SI):

main.py --si --savepath=si

• Learning without Forgetting (LwF):

main.py --replay=current --distill --savepath=lwf

• Deep Generative Replay (DGR):

main.py --replay=generative --savepath=dgr

• DGR with distillation:

main.py --replay=generative --distill --savepath=distilldgr

• Replay-trough-Feedback (RtF):

main.py --replay=generative --distill --feedback --savepath=rtf

• Cumulative:

main.py --cumulative=1 --savepath=cumulative

• Naive:

main.py --savepath=naive

Repository Structure

OpenLORISCode 
├── img
├── lib
│   ├── callbacks.py
│   ├── continual_learner.py
│   ├── encoder.py
│   ├── exemplars.py
│   ├── replayer.py
│   ├── train.py
│   ├── vae_models.py
│   ├── visual_plt.py
├── _compare.py
├── _compare_replay.py
├── _compare_taskID.py
├── data.py
├── evaluate.py
├── excitability_modules.py
├── main.py
├── linear_nets.py
├── param_stamp.py
├── pk_gene.py
├── visual_visdom.py
├── utils.py
└── README.md

Citation

Please consider citing our papers if you use this code in your research:

@article{she2020iros,
  title={IROS 2019 Lifelong Robotic Vision Challenge--Lifelong Object Recognition Report},
  author={She, Qi and Feng, Fan and Liu, Qi and Chan, Rosa HM and Hao, Xinyue and Lan, Chuanlin and Yang, Qihan and Lomonaco, Vincenzo and Parisi, German I and Bae, Heechul and others},
  journal={arXiv preprint arXiv:2004.14774},
  year={2020}
}

@article{she2019openlorisobject,
  title={Openlorisobject: A robotic vision dataset and benchmark for lifelong deep learning},
  author={She, Qi and Feng, Fan and Hao, Xinyue and Yang, Qihan and Lan, Chuanlin and Lomonaco, Vincenzo and Shi, Xuesong and Wang, Zhengwei and Guo, Yao and Zhang, Yimin and others},
  journal={International Conference on Robotics and Automation (ICRA)},
  year={2020}
}

Acknowledgements

Parts of code were borrowed from here.

Issue / Want to Contribute ?

Open a new issue or do a pull request in case you are facing any difficulty with the code base or if you want to contribute.

---

Features pending

OpenLORIS-Object base

• OpenLROIS-Object dataset configuration files;
• OpenLROIS-Object sample codes;

CL baseline

• SOTA CL methods;

CL benchmarks for image classification

• CL benchmarks: MNIST and CIFAR-100 datasets;

Visualization

• Visualization tools for 4 CL metrics;
• DL backbones: VGG-16, ResNet-18/50/101, EfficientNet;

Applications

• Ego-gesture recognition;
• Online action recognition;
• Constractive learning for self-supervised object segementation;
• Few-shot learning with object recognition;

Algorithms

• Robust adversial training with transfer learning;