README

Dataset

For CIFAR-FS dataset please go this Download link

For Tieredimagenet dataset please go this Download link

For Toy dataset please go these Download link (Aircraft) Download link (Bird) Download link (Car) Download link (Fungi) and partition the train and test set according to our appendix. In order to accelerate training phase, please collect this dataset into a pkl cache file using import pickle

Environment

Tensorflow == 1.13

cuda == 10.0

cudnn == 7.6

python == 3.6.8

numpy == 1.18

opencv==3.4

pillow == 7.1

scikit-learn

tqdm

Put data folder struction like this

root
│   README.md
│   train.py    
│
└───data
│   │   cifar-fs
│   │   tiered-imagenet
|   |   miniimagenet
│   │   ...
│   
└───logs
    │   maml_cifar5w1s
    │   tsa_maml_cifar5w1s
    |   ...

Quick test via loading trained model

To quick evalute with our trained model, please download model from this Download link (Cifar-100 models).

Test CIFAR-FS 5-way 1-shot MAML

python train_maml.py --train=False --dataset cifar100 --meta_batch_size 4 --update_batch_size 1 \
	--update_lr 0.01 --num_updates 5 --num_classes 5 --logdir logs/maml_cifar5w1s \
	--num_test_tasks 600 --num_filters 32 --max_pool True

Test CIFAR-FS 5-way 1-shot TSA-MAML

python train_tsamaml.py --train=False --dataset cifar100 --metatrain_iterations 40000 \
	--meta_batch_size 4 --update_batch_size 1 --update_lr 0.01 --num_updates 5 --num_classes 5 \
	--logdir logs/tsa_maml_cifar5w1s --premaml logs/maml_cifar5w1s/bestmodel --num_test_tasks 600 \
	--num_filters 32 --max_pool True --num_groups 5

For other settings, you can quickly access the commands in train.sh and test.sh.

Train

At first, train a vallina MAML for task solution clustering.

CIFAR100 For N-way K-shot tasks (default group number is 5)

python train_maml.py --train=False --dataset cifar100 --metatrain_iterations 60000 \
	--meta_batch_size 4 --update_batch_size <K> --update_lr 0.01 --num_updates 5 \
	--num_classes <N> --logdir <logdir> --num_test_tasks 600 --num_filters 32 \
	--max_pool True --num_groups 5

TieredImageNet For N-way K-shot tasks (default group number is 5)

python train_maml.py --train=False --dataset tiered --metatrain_iterations 100000 \
	--meta_batch_size 4 --update_batch_size <K> --update_lr 0.01 --num_updates 5 \
	--num_classes <N> --logdir <logdir> --num_test_tasks 600 --num_filters 32 \
	--max_pool True --num_groups 5

Then load this pretrained MAML and do training for TSA-MAML.

CIFAR100 For N-way K-shot tasks (default group number is 5)

python train_tsamaml.py --train=False --dataset cifar100 --metatrain_iterations 40000 \
	--meta_batch_size 4 --update_batch_size <K> --update_lr 0.01 --num_updates 5 \
	--num_classes <N> --logdir <logdir> --premaml <ModelPathOfMAML> --num_test_tasks 600 \
	--num_filters 32 --max_pool True --num_groups 5 --cosann=True

TieredImageNet For N-way K-shot tasks (default group number is 5)

python train_tsamaml.py --train=False --dataset tiered --metatrain_iterations 60000 \
	--meta_batch_size 4 --update_batch_size <K> --update_lr 0.01 --num_updates 5 \
	--num_classes <N> --logdir <logdir>  --premaml <ModelPathOfMAML> --num_test_tasks 600 \
	--num_filters 32 --max_pool True --num_groups 5 --cosann=True

Test

Same as the Quick test section.

Citation

@inproceedings{zhou2020task,
	title={Task Similarity Aware Meta Learning: Theory-inspired Improvement on MAML},
	author={Zhou, Pan and Zou, Yingtian and Yuan, Xiaotong and Feng, Jiashi and Xiong, Caiming and Hoi, SC},
	booktitle={4th Workshop on Meta-Learning at NeurIPS},
	year={2020}
}

Carbonaraa / TSA-MAML