GitHub - NguyenTriQuan/SCCL: Over-parameterized expansion and Structure Compression for Continual Learning

NguyenTriQuan / SCCL Public

Notifications You must be signed in to change notification settings
Fork 0
Star 0

Over-parameterized expansion and Structure Compression for Continual Learning

Notifications

Name		Name	Last commit message	Last commit date
Latest commit History 473 Commits
approaches		approaches
dataloaders		dataloaders
layers		layers
networks		networks
.gitignore		.gitignore
README.txt		README.txt
arguments.py		arguments.py
main.py		main.py
requirements.txt		requirements.txt
utils.py		utils.py

Repository files navigation

Dynamic Architecture and past connection Dropout (DAD)

Reproduce results:

- Permuted MNIST:
python main.py --approach dad --experiment pmnist --arch MLP --lamb 2.3_0 --batch_size 256 --lr 0.001 --optimizer Adam --tasknum 10  --nepochs 100 --ensemble_drop 0

- Split CIFAR-100:
python main.py --approach dad --experiment split_cifar100 --arch VGG8 --lamb 7_2_0.5 --batch_size 32 --lr 0.001 --optimizer Adam --tasknum 10  --nepochs 150 --ensemble_drop 0

- Split CIAFR-10/100:
python main.py --approach dad --experiment split_cifar10_100 --arch VGG8 --lamb 8_2_0.5 --batch_size 32 --lr 0.001 --optimizer Adam --tasknum 11  --nepochs 150 --ensemble_drop 0

- Mixture:
python main.py --approach dad --experiment mixture --arch Alexnet --lamb 30_0 --batch_size 256 --lr 0.01 --optimizer SGD --tasknum 8  --nepochs 150 --ensemble_drop 0

- 10 Split mini ImageNet:
python main.py --approach dad --experiment 10_split_mini_imagenet --arch VGG16_small --lamb 0.18_0.06_0 --batch_size 32 --lr 0.01 --optimizer SGD --tasknum 10  --nepochs 100 --ensemble_drop 0.2 --norm_type affine_track

- 20 split min ImageNet:
python main.py --approach dad --experiment 20_split_mini_imagenet --arch VGG16_small --lamb 0.3_0.1_0 --batch_size 32 --lr 0.01 --optimizer SGD --tasknum 10  --nepochs 100 --ensemble_drop 0.2 --norm_type affine_track