lxhtaciturnity / iccv2019-inc

ICCV 2019 Overcoming Catastrophic Forgetting with Unlabeled Data in the Wild

Introduction

This repository implements Lee et al. Overcoming Catastrophic Forgetting with Unlabeled Data in the Wild. In ICCV, 2019 in PyTorch.

@inproceedings{lee2019overcoming,
  title={Overcoming Catastrophic Forgetting with Unlabeled Data in the Wild},
  author={Lee, Kibok and Lee, Kimin and Shin, Jinwoo and Lee, Honglak},
  booktitle={ICCV},
  year={2019}
}

This implementation also includes the state-of-the-art distillation-based methods for class-incremental learning (a.k.a. single-head continual learning):

• Learning without forgetting [LwF, ECCV 2016]
• Distillation and retrospection [DR, ECCV 2018]
• End-to-end incremental learning [E2E, ECCV 2018]

Please see [training recipes] for replicating them.

Dependencies

• Python 3.6.8
• NumPy 1.16.2
• PyTorch 0.4.1
• torchvision 0.2.1
• h5py 2.7.1
• tqdm 4.25.0
• tensorboardx 1.4
• SciPy 1.1.0 for sample_tiny.py
• matplotlib 2.2.2 for plotter.py
• seaborn 0.9.0 for plotter.py
• pandas 0.23.0 for plotter.py

Data

You may either generate datasets by yourself or download h5 files in the following links. You may not download external data if you don't want to use them. All data are assumed to be in data/{dataset}/. ({dataset} = cifar100, tiny, imagenet)

CIFAR-100 (Training data)

This will be automatically downloaded.

TinyImages (External data)

• DIY
  • Download [images (227GB)] [metadata (57GB)] [words] [cifar indexes] and place them in data/tiny/.
  • Run python sample_tiny.py -s {seed}. {seed} corresponds to the stage number in incremental learning.
    • This takes a long time, so running in parallel is recommended.
• Don't DIY
  • Download [here (20 files; each takes 3GB)] and place them in data/tiny/.
    • Download [0 (3GB)] only, if you use --ex-static for training.

ImageNet (Training and external data)

• DIY
  • Download [ImageNet ILSVRC 2012 train (154.6GB)] and place them in data/imagenet/ilsvrc2012.
  • Run the following command. This takes a long time.

    python image_resizer_imagenet.py -i 'imagenet/ilsvrc2012' -o 'imagenet/ilsvrc2012_resized' -s 32 -a box -r -j 16
    

  • Download [ImageNet 2011 Fall (1.3TB)] and place them in data/imagenet/fall11_whole.
  • Run the following command. This takes a long time.

    python image_resizer_imagenet.py -i 'imagenet/fall11_whole' -o 'imagenet/fall11_whole_resized' -s 32 -a box -r -j 16
    

  • Run python sample_imagenet.py -s {seed}. {seed} corresponds to the stage number in incremental learning.
    • Training and test data will be generated at seed=0.
    • This takes a long time, so running in parallel is recommended.
• Don't DIY
  • Download [train (1.5GB)], [test (0.3GB)] and place them in data/imagenet/.
  • Download [here (20 files; each takes 3GB)] and place them in data/imagenet/.
    • Download [0 (3GB)] only, if you use --ex-static for training.

Task splits

• DIY
  • Run python shuffle_task.py.
• Don't DIY
  • Task splits are already in split/.

Train and test

• Run python main.py -h to see the general usage.
• With --ex-static, only 0-th external dataset is used for all stages.
• Please see [training recipes] for replicating the models compared in our paper.
• Examples on CIFAR-100, task size 10, seed 0, gpu 0:
  • GD (Ours) without external data

    python main.py --gpu 0 --seed 0 -d cifar100 -s res -t 10 10 -r PC  -b dw -f cls
    

  • GD (Ours) with external data

    python main.py --gpu 0 --seed 0 -d cifar100 -e tiny -s res -t 10 10 -r PCQ -b dw -f cls
    

Evaluation

• Run python plotter.py -h to see the general usage. bar and time replicate Figure 2(a,b) and (c,d), and the others replicate tables.
• Examples:
  • Replicate CIFAR100 with task size 10 in Table 1

    python plotter.py -d cifar100 -e tiny -s res -t 10 10 --exp t1
    

  • Replicate bar graphs in Figure 2(a,b)

    python plotter.py -d cifar100 -e tiny -s res --exp bar
    

  • Compare arbitrary models you want

    python plotter.py --exp custom
    

Note

• image_resizer_imagenet.py is adapted from [here]
• models is adapted from [here]
• datasets is adapted from [here]