Code for this paper Long Live the Lottery: The Existence of Winning Tickets in Lifelong Learning
Tianlong Chen*, Zhenyu Zhang*, Sijia Liu, Shiyu Chang, Zhangyang Wang
We extend the lottery ticket hypothesis from one-shot task learning to class incremental learning scenario and propose top-down and bottom-up pruning strategies to identify winning tickets, which we call lifelong Tickets.
- Top-Down (TD) Pruning
We modify the iterative magnitude pruning approach and assign the pruning budget to each task based on an heuristic curriculum schedule.
- Bottom-Up (BU) Pruning
To tackle the greedy nature of Top-down pruning method, we propose Bottom-Up pruning. Once the current sparse network is too heavily pruned and has no more capacity for new tasks, BU pruning can make the sparse network to re-grow from the current sparsity.
class incremental learning with Top-Down pruning and Bottom-Up pruning
pytorch >= 1.4
torchvision
We reorganized the CIFAR10, CIFAR100 dataset into a dictionary, {key: value}, where the key is for labels, from 0-9 of CIFAR10 and values are the images. And the unlabel images are sampled from 80 Million Tiny Images dataset, which can be download from CIL_data
The pretrained models can be found at models, which contains:
- BU_ticket.pt # winning tickets found by Bottom-Up pruning method on CIFAR10
- full_model.pt # full model on CIFAR10
python -u main_TD.py # Top-Down Pruning
python -u main_BU.py # Bottom-Up Pruning
python -u main_CIL.py # Basic Class Incremental Learning
python -u main_train.py \
--weight [init_weight] \
--mask [init_sparse_structure] \
--state [task ID in CIL] # re-train the subnetwork
python -u test.py --pretrained BU_ticket.pt --pruned --state [taskID] # test prune model
python -u test.py --pretrained full_model.pt --state [taskID] # test full model
@inproceedings{
chen2021long,
title={Long Live the Lottery: The Existence of Winning Tickets in Lifelong Learning},
author={Tianlong Chen and Zhenyu Zhang and Sijia Liu and Shiyu Chang and Zhangyang Wang},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=LXMSvPmsm0g}
}