B-SMALL: A BAYESIAN NEURAL NETWORK APPROACH TO SPARSE MAML

This repository contains code for B-SMALL Link to Paper (Accepted at ICASSP '21)
It includes PyTorch code to train and evaluate both MAML[1] and B-SMALL for few-shot classification tasks on CIFAR-FS and MiniImageNet. The models were trained on a single 2080Ti but use less memory than its full capacity.

Dependencies

This code requires the following key dependencies:

• Pytorch 1.4
• Python 3.6
• Tensorboard

Setup

Download MiniImagenet dataset and put the images in data/mini-imagenet/images/. The splits according to [1] are already provided in the repo. For CIFAR-FS, see data/get_cifarfs.py script to setup.

Usage

Training 5-way 5-shot MAML from a checkpoint

python train.py --n_way 5 --k_spt 5 --restore_model <Path_to_model>

Training 5-way 5-shot B-SMALL

python train.py --n_way 5 --k_spt 5 --svdo 

There are a number of parameters given as flags in the script which can be easily changed.

Key Results

We show some results obtained on the MiniImagenet dataset after training for 60,000 iterations and with the same hyperparameter settings as MAML[1]. Detailed results on CIFAR-FS and inferences in paper.

Model/Experiment	5-way 1-shot	5-way 5-shot
MAML[1]	48.70 ± 1.84%	63.11 ± 0.92%
CAVIA[2]	47.24 ± 0.65%	61.87 ± 0.93%
MAML(Ours)	46.30 ± 0.29%	66.3 ± 0.21%
B-SMALL	49.12 ± 0.30%	66.97 ± 0.3%
Sparsity	76%	44%

To do

• Sinusoid Regression Tasks

Acknowledgement

Special thanks to Jackie Loong's implementation, of which some parts are directly taken for quick prototyping : https://github.com/dragen1860/MAML-Pytorch

References

[1] Finn, Chelsea, Pieter Abbeel, and Sergey Levine. "Model-agnostic meta-learning for fast adaptation of deep networks." arXiv preprint arXiv:1703.03400 (2017).
[2] Zintgraf, Luisa, et al. "Fast context adaptation via meta-learning." International Conference on Machine Learning. PMLR, 2019.