Meta-Neighborhoods

This is the official repository for our NeurIPS 2020 paper Meta-Neighborhoods. A short introduction to this paper can be found in this talk.

@article{shan2019meta,
  title={Meta-Neighborhoods},
  author={Shan, Siyuan and Li, Yang and Oliva, Junier},
  journal={arXiv preprint arXiv:1909.09140},
  year={2019}
}

Prerequisites

clone our repository and install TensorFlow 1.12.0

Prepare datasets

the script below help to prepare CIFAR100 and save it to path_to_save_data, based on which you can prepare other datasets.

python prepare_cifar100.py --save_dir=path_to_save_data

Experiments

Set --visualize=True if you want to visualize dictionary values and attention similarity as images in TensorBoard.

MN+iFilm model

(Train) python main.py --logdir=path_to_save_log --data_dir=path_to_save_data --dataset=cifar100 --meta_batch_size=128 --dropout_ratio=0.5 --vanilla=False --dict_size=10000 --update_lr=0.1 --num_updates=1 --fix_v=False --alpha=5.0 --meta_lr=1e-3 --backbone=resnet56 --scalar_lr=True --dot=False --modulate=all --film_dict_size=10 --visualize=False

(Test) python main.py --logdir=path_to_save_log --data_dir=path_to_save_data --dataset=cifar100 --meta_batch_size=128 --dropout_ratio=0.5 --vanilla=False --dict_size=10000 --update_lr=0.1 --num_updates=1 --fix_v=False --alpha=5.0 --meta_lr=1e-3 --backbone=resnet56 --scalar_lr=True --dot=False --modulate=all --film_dict_size=10 --train=False --test_set=True

MN model

(Train) python main.py --logdir=path_to_save_log --data_dir=path_to_save_data --dataset=cifar100 --meta_batch_size=128 --dropout_ratio=0.5 --vanilla=False --dict_size=10000 --update_lr=0.1 --num_updates=1 --fix_v=False --alpha=5.0 --meta_lr=1e-3 --backbone=resnet56 --scalar_lr=True --dot=False --modulate=None --visualize=False

(Test) python main.py --logdir=path_to_save_log --data_dir=path_to_save_data --dataset=cifar100 --meta_batch_size=128 --dropout_ratio=0.5 --vanilla=False --dict_size=10000 --update_lr=0.1 --num_updates=1 --fix_v=False --alpha=5.0 --meta_lr=1e-3 --backbone=resnet56 --scalar_lr=True --dot=False --modulate=None --train=False --test_set=True

vanilla model

(Train) python main.py --logdir=path_to_save_log --data_dir=path_to_save_data --dataset=cifar100 --meta_batch_size=128 --vanilla=True --meta_lr=1e-3 --backbone=resnet56 --dot=False --modulate=None --visualize=False

(Test) python main.py --logdir=path_to_save_log --data_dir=path_to_save_data --dataset=cifar100 --meta_batch_size=128 --vanilla=True --meta_lr=1e-3 --backbone=resnet56 --dot=False --modulate=None --train=False --test_set=True

Spiral Toy Example

Please refer to https://github.com/lupalab/MetaNeighborhood-Spiral