Prompt-based Continual Learning Official Jax Implementation

This codebase contains the implementation of two continual learning methods:

• Learning to Prompt for Continual Learning (L2P) (CVPR2022) [Google AI Blog]
• DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning (ECCV2022)

Introduction

L2P is a novel continual learning technique which learns to dynamically prompt a pre-trained model to learn tasks sequentially under different task transitions. Different from mainstream rehearsal-based or architecture-based methods, L2P requires neither a rehearsal buffer nor test-time task identity. L2P can be generalized to various continual learning settings including the most challenging and realistic task-agnostic setting. L2P consistently outperforms prior state-of-the-art methods. Surprisingly, L2P achieves competitive results against rehearsal-based methods even without a rehearsal buffer.

DualPrompt improves L2P by attaching complementary prompts to the pre-trained backbone, and then formulates the objective as learning task-invariant and task-specific “instructions". With extensive experimental validation, DualPrompt consistently sets state-of-the-art performance under the challenging class-incremental setting. In particular, DualPrompt outperforms recent advanced continual learning methods with relatively large buffer sizes. We also introduce a more challenging benchmark, Split ImageNet-R, to help generalize rehearsal-free continual learning research.

Code is written by Zifeng Wang. Acknowledgement to https://github.com/google-research/nested-transformer.

This is not an officially supported Google product.

Novel CL benchmark: Split ImageNet-R

The Split ImageNet-R benchmark is build upon ImageNet-R by dividing the 200 classes into 10 tasks with 20 classes per task, see libml/input_pipeline.py for details. We believe the Split ImageNet-R is of great importance to the continual learning community, for the following reasons:

• Split ImageNet-R contains classes with different styles, which is closer to the complicated real-world problems.
• The significant intra-class diversity poses a great challenge for rehearsal-based methods to work effectively with a small buffer size, thus encouraging the development of more practical, rehearsal-free methods.
• Pre-trained vision models are useful in practical continual learning. However, their training set usually includes ImageNet. Thus, Split ImageNet-R serves as a relative fair and challenging benchmark, and an alternative to ImageNet-based benchmarks for continual learning that uses pre-trained models.

PyTorch Reimplementation

The codebase has been reimplemented in PyTorch by Jaeho Lee in l2p-pytorch and dualprompt-pytorch.

Enviroment setup

pip install -r requirements.txt

After this, you may need to adjust your jax version according to your cuda driver version so that jax correctly identifies your GPUs (see this issue for more details).

Note: The codebase has been throughly tested under the TPU enviroment using the newest JAX version. We are currently working on further verifying the GPU environment.

Dataset preparation

Before running experiments for 5-datasets and CORe50, additional dataset preparation step should be conducted as follows:

1. Download CORe50 classification benchmark here: https://vlomonaco.github.io/core50/ and download not-mnist here: http://yaroslavvb.com/upload/notMNIST/
2. Transform them into TFDS compatible form following the tutorial in https://www.tensorflow.org/datasets/add_dataset
3. Replace corresponding dataset paths "PATH_TO_CORE50" and "PATH_TO_NOT_MNIST" in libml/input_pipeline.py by the destination paths in step 2

Getting pretrained ViT model

ViT-B/16 model used in this paper can be downloaded at here. Note: Our codebase actually supports various sizes of ViTs. If you would like to try variations of ViTs, feel free to change the config.model_name in the config files, following the valid options defined in models/vit.py.

Instructions on running L2P and DualPrompt

We provide the configuration file to train and evaluate L2P and DualPrompt on multiple benchmarks in configs.

To run L2P on benchmark datasets:

python main.py --my_config configs/$L2P_CONFIG --workdir=./l2p --my_config.init_checkpoint=<ViT-saved-path/ViT-B_16.npz>

where $L2P_CONFIG can be one of the followings: [cifar100_l2p.py, five_datasets_l2p.py, core50_l2p.py, cifar100_gaussian_l2p.py].

Note: we run our experiments using 8 V100 GPUs or 4 TPUs, and we specify a per device batch size of 16 in the config files. This indicates that we use a total batch size of 128.

To run DualPrompt on benchmark datasets:

python main.py --my_config configs/$DUALPROMPT_CONFIG --workdir=./dualprompt --my_config.init_checkpoint=<ViT-saved-path/ViT-B_16.npz>

where $DUALPROMPT_CONFIG can be one of the followings: [imr_dualprompt.py, cifar100_dualprompt.py].

Visualize results

We use tensorboard to visualize the result. For example, if the working directory specified to run L2P is workdir=./cifar100_l2p, the command to check result is as follows:

tensorboard --logdir ./cifar100_l2p

Here are the important metrics to keep track of, and their corresponding meanings:

Metric	Description
accuracy_n	Accuracy of the n-th task
forgetting	Average forgetting up until the current task
avg_acc	Average evaluation accuracy up until the current task

Cite

@inproceedings{wang2022learning,
  title={Learning to prompt for continual learning},
  author={Wang, Zifeng and Zhang, Zizhao and Lee, Chen-Yu and Zhang, Han and Sun, Ruoxi and Ren, Xiaoqi and Su, Guolong and Perot, Vincent and Dy, Jennifer and Pfister, Tomas},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={139--149},
  year={2022}
}

@article{wang2022dualprompt,
  title={DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning},
  author={Wang, Zifeng and Zhang, Zizhao and Ebrahimi, Sayna and Sun, Ruoxi and Zhang, Han and Lee, Chen-Yu and Ren, Xiaoqi and Su, Guolong and Perot, Vincent and Dy, Jennifer and others},
  journal={European Conference on Computer Vision},
  year={2022}
}