clovaai / eclipse

(CVPR 2024) ECLIPSE: Efficient Continual Learning in Panoptic Segmentation with Visual Prompt Tuning

ECLIPSE (CVPR 2024)

ECLIPSE: Efficient Continual Learning in Panoptic Segmentation with Visual Prompt Tuning
Beomyoung Kim^1,2, Joonsang Yu¹, Sung Ju Hwang²

¹ _{NAVER Cloud, ImageVision}
² _KAIST

Introduction

Panoptic segmentation, combining semantic and instance segmentation, stands as a cutting-edge computer vision task. Despite recent progress with deep learning models, the dynamic nature of real-world applications necessitates continual learning, where models adapt to new classes (plasticity) over time without forgetting old ones (catastrophic forgetting). Current continual segmentation methods often rely on distillation strategies like knowledge distillation and pseudo-labeling, which are effective but result in increased training complexity and computational overhead. In this paper, we introduce a novel and efficient method for continual panoptic segmentation based on Visual Prompt Tuning, dubbed ECLIPSE. Our approach involves freezing the base model parameters and fine-tuning only a small set of prompt embeddings, addressing both catastrophic forgetting and plasticity and significantly reducing the trainable parameters. To mitigate inherent challenges such as error propagation and semantic drift in continual segmentation, we propose logit manipulation to effectively leverage common knowledge across the classes. Experiments on ADE20K continual panoptic segmentation benchmark demonstrate the superiority of ECLIPSE, notably its robustness against catastrophic forgetting and its reasonable plasticity, achieving a new state-of-the-art.

Updates

2024-04-29 First Commit, We release the official implementation of ECLIPSE.

Installation

Our implementation is based on CoMFormer and Mask2Former.

Please check the installation instructions and dataset preparation.

You can see our core implementation from

• mask2former/maskformer_model.py
• mask2former/modeling/transformer_decoder/mask2former_transformer_decoder.py

Quick Start

1. Step t=0: Training the model for base classes (you can skip this process if you use pre-trained weights.)
2. Step t>1: Training the model for novel classes with ECLIPSE

Scenario	Script	Step-0 Weight	Final Weight
ADE20K-Panoptic 100-5	bash script/ade_ps/100_5.sh	step0	step10
ADE20K-Panoptic 100-10	bash script/ade_ps/100_10.sh	step0	step5
ADE20K-Panoptic 100-50	bash script/ade_ps/100_50.sh	step0	step1
ADE20K-Panoptic 50-10	bash script/ade_ps/50_10.sh	step0	step10
ADE20K-Panoptic 50-20	bash script/ade_ps/50_20.sh	step0	step5
ADE20K-Panoptic 50-50	bash script/ade_ps/50_50.sh	step0	step2
ADE20K-Semantic 100-5	bash script/ade_ss/100_5.sh	step0	step10
ADE20K-Semantic 100-10	bash script/ade_ss/100_10.sh	step0	step5
ADE20K-Semantic 100-50	reproduce error	step0	step1
COCO-Panoptic 83-5	bash script/coco_ps/83_5.sh	step0	step10
COCO-Panoptic 83-10	bash script/coco_ps/83_10.sh	step0	step5

How to Cite

@article{kim2024eclipse,
  title={ECLIPSE: Efficient Continual Learning in Panoptic Segmentation with Visual Prompt Tuning},
  author={Kim, Beomyoung and Yu, Joonsang and Hwang, Sung Ju},
  journal={arXiv preprint arXiv:2403.20126},
  year={2024}
}

License

ECLIPSE
Copyright (c) 2024-present NAVER Cloud Corp.
CC BY-NC 4.0 (https://creativecommons.org/licenses/by-nc/4.0/)

eclipse's Issues

environment and package version

Hi! May you provide the precise packages & environment version? Such as python, torch, torchvision?

Looking for the code

Thank you for your great work!
Can't wait to follow your work!
May I ask when you plan to release your code?

delta settings are different in training and evaluation

Hi! I found delta used in training and evaluation sometimes sat different.
Such as delta in ade_ps/100_5.sh, ade_ps/100_50.sh.

May I ask why? Different delta settings would result in very different accuracy between the last training step and the evaluation.

'MaskFormer' object has no attribute 'module'

Hi! I got a error below: 'MaskFormer' object has no attribute 'module' and tried the weights from self-training step0 and which download from your page.
Do you know what's wrong is it? Or this is a version problem?
(Sorry I asked your version problem in the previous issue, but the cost for changing cuda version is too heavy to me.)

in train_inc.py L140

if self.cfg.CONT.NUM_PROMPTS > 0:
    self.model.module.copy_prompt_embed_weights()# error occur
    self.model.module.copy_mask_embed_weights()
    self.model.module.copy_no_obj_weights()
    self.model.module.copy_prompt_trans_decoder_weights()

ERROR message

[05/03 17:47:44 d2.checkpoint.detection_checkpoint]: [DetectionCheckpointer] Loading from results/ade_ss_100_step0.pth ...
[05/03 17:47:44 fvcore.common.checkpoint]: [Checkpointer] Loading from results/ade_ss_100_step0.pth ...
WARNING [05/03 17:47:44 fvcore.common.checkpoint]: Some model parameters or buffers are not found in the checkpoint:
sem_seg_head.predictor.class_embed.cls.2.layers.0.{bias, weight}
sem_seg_head.predictor.class_embed.cls.2.layers.1.{bias, weight}
sem_seg_head.predictor.class_embed.cls.2.layers.2.{bias, weight}
sem_seg_head.predictor.prompt_embed.0.0.weight
sem_seg_head.predictor.prompt_embed.0.1.weight
sem_seg_head.predictor.prompt_embed.0.2.weight
sem_seg_head.predictor.prompt_embed.0.3.weight
sem_seg_head.predictor.prompt_embed.0.4.weight
sem_seg_head.predictor.prompt_embed.0.5.weight
sem_seg_head.predictor.prompt_embed.0.6.weight
sem_seg_head.predictor.prompt_embed.0.7.weight
sem_seg_head.predictor.prompt_embed.0.8.weight
sem_seg_head.predictor.prompt_feat.0.weight
sem_seg_head.predictor.prompt_mask_embed.0.layers.0.{bias, weight}
sem_seg_head.predictor.prompt_mask_embed.0.layers.1.{bias, weight}
sem_seg_head.predictor.prompt_mask_embed.0.layers.2.{bias, weight}
Traceback (most recent call last):
  File "train_inc.py", line 782, in <module>
    launch(
  File "/home/vllab/anaconda3/envs/mask2former/lib/python3.8/site-packages/detectron2/engine/launch.py", line 84, in launch
    main_func(*args)
  File "train_inc.py", line 772, in main
    trainer.resume_or_load(resume=args.resume)
  File "train_inc.py", line 140, in resume_or_load
    self.model.module.copy_prompt_embed_weights()
  File "/home/vllab/anaconda3/envs/mask2former/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1695, in __getattr__
    raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")

About training cost

Hi, thanks for your work! I want to reproduce the result and wonder the number of GPUs used in these experiments and the training time, hope you can help me!