Do Your Best and Get Enough Rest for Continual Learning

This folder contains official pyTorch implementations for "Do Your Best and Get Enough Rest for Continual Learning" accepted in CVPR'25. (see our paper, slides, poster).

Illustration of View-Batch Model

1. Tutorial

1. Clone this repository and install the requirements.

   git clone https://github.com/hankyul2/ViewBatchModel.git
   cd ViewBatchModel
   pip install -r requirements.txt
   

2. Train ResNet18 on S-CIFAR-10 using iCaRL as baseline methods with 200 buffers.

   iCaRL

   CUDA_VISIBLE_DEVICES=0 python utils/main.py --model icarl --load_best_args --dataset seq-cifar10 --buffer_size 200 --seed 1993 --savecheck 1 --ckpt_name icarl_r1_s1993
   

   Ours-iCaRL

   CUDA_VISIBLE_DEVICES=4 python utils/main.py --model icarl --load_best_args --dataset seq-cifar10 --buffer_size 200 --aug-repeat 4 --prog-aug 5 --seed 1993 --flag hard_aug --savecheck 1 --ckpt_name icarl_r4_hard_aug_s1993
   

3. Validate the trained network using the saved checkpoint.

   iCaRL

   CUDA_VISIBLE_DEVICES=0 python utils/main.py --model icarl --load_best_args --dataset seq-cifar10 --buffer_size 200 --seed 1993 --loadcheck checkpoints/icarl_r1_s1993_cifar10_t0.pth --start_from 0 --stop_after 0 --inference_only 1
   

   Ours-iCaRL

   CUDA_VISIBLE_DEVICES=4 python utils/main.py --model icarl --load_best_args --dataset seq-cifar10 --buffer_size 200 --seed 1997 --loadcheck checkpoints/icarl_r4_hard_aug_s1997_cifar10_t0.pth --start_from 0 --stop_after 0 --inference_only 1
   

4. See scripts/icarl for more commands to reproduce Table 6 in the paper. Also, check datasets/utils/continual_dataset.py#L24 for view-batch replay and models/icarl.py#L78 for view-batch SSL.

2. Reproduced Results

After the paper has been accepted, we rerun everything to provide complete logs and checkpoints for our Table 6 in the paper. Our exact environments are:

• torch==1.12.1+cu113
• torchvision==0.13.1+cu113
• timm==1.0.7
• numpy==1.24.4

Experimental Results

The table below reproduces Table 6 of our paper, which contains the main ablation study for the proposed method.

Method	View-batch Replay	Strong Augment	View-batch SSL	Forgetting(⬇️)	CIL(⬆️)	TIL(⬆️)	AVG	∆
iCaRL	❌	❌	❌	28.05±4.21	63.58±2.64	90.32±3.19	76.95	-
iCaRL	❌	✅	❌	22.16±0.91	65.33±1.05	89.33±0.58	77.33	+0.38
iCaRL	✅	❌	❌	18.72±1.76	67.21±0.42	91.63±0.98	79.42	+2.47
iCaRL	✅	✅	❌	18.29±0.91	67.16±0.75	91.02±0.97	79.09	+2.14
iCaRL	✅	✅	✅	13.81±1.58	69.25±0.41	92.73±0.57	80.99	+4.04

Log

Below, the WanDB project link provides the complete logs that are made during the training of the above tables. It includes:

• command line
• metrics
• console outputs
• environments

WanDB Project Link: https://wandb.ai/gregor99/view_batch_model.

Checkpoint

The tables below provide the checkpoints saved at the end of tasks during the training of the above tables.

seed=1993

method	View-batch Replay	Strong Augmentation	View-batch SSL	task 1	task 2	task 3	task 4	task 5
iCaRL	-	-	-	ckpt	ckpt	ckpt	ckpt	ckpt
iCaRL	-	v	-	ckpt	ckpt	ckpt	ckpt	ckpt
iCaRL	v	-	-	ckpt	ckpt	ckpt	ckpt	ckpt
iCaRL	v	v	-	ckpt	ckpt	ckpt	ckpt	ckpt
iCaRL	v	v	v	ckpt	ckpt	ckpt	ckpt	ckpt

seed=1996

method	View-batch Replay	Strong Augmentation	View-batch SSL	task 1	task 2	task 3	task 4	task 5
iCaRL	-	-	-	ckpt	ckpt	ckpt	ckpt	ckpt
iCaRL	-	v	-	ckpt	ckpt	ckpt	ckpt	ckpt
iCaRL	v	-	-	ckpt	ckpt	ckpt	ckpt	ckpt
iCaRL	v	v	-	ckpt	ckpt	ckpt	ckpt	ckpt
iCaRL	v	v	v	ckpt	ckpt	ckpt	ckpt	ckpt

seed=1997

method	View-batch Replay	Strong Augmentation	View-batch SSL	task 1	task 2	task 3	task 4	task 5
iCaRL	-	-	-	ckpt	ckpt	ckpt	ckpt	ckpt
iCaRL	-	v	-	ckpt	ckpt	ckpt	ckpt	ckpt
iCaRL	v	-	-	ckpt	ckpt	ckpt	ckpt	ckpt
iCaRL	v	v	-	ckpt	ckpt	ckpt	ckpt	ckpt
iCaRL	v	v	v	ckpt	ckpt	ckpt	ckpt	ckpt

3. Acknowledgement

This project is heavily based on Mammoth. We sincerely appreciate the authors of the mentioned works for sharing such great library as open-source project.