Official pytorch implementation of the paper:

"Accelerating Convergence in Bayesian Few-Shot Classification"

Requirements

1. Python >= 3.x
2. Numpy >= 1.17
3. pyTorch >= 1.2.0
4. GPyTorch >= 0.3.5
5. (optional) TensorboardX

Installation

pip install numpy torch torchvision gpytorch h5py pillow

We confirm that the following configuration worked for us: numpy 1.22.4, torch 1.11.0, torchvision 0.12.0, gpytorch 1.9.0, h5py 3.4.0, pillow 9.1.1.

MD-BSFC: code of our method

Classification. The code for the classification case is accessible in mirror.py, with most of the important pieces contained in the train_loop() method (training), and in the correct() method (testing).

Experiments

These are the instructions to train and test the methods reported in the paper in the various conditions.

Download and prepare a dataset. This is an example of how to download and prepare a dataset for training/testing. Here we assume the current directory is the project root folder:

cd filelists/DATASET_NAME/
sh download_DATASET_NAME.sh

Replace DATASET_NAME with one of the following: CUB, miniImagenet, omniglot, emnist. Notice that the link to mini-Imagenet is no longer available on the official website. We refer to mini-imagenet-tools for various alternatives to downloading the dataset.

Classification

Train classification. Our methods can be trained using the following syntax:

python train.py --dataset="CUB" --method="MD" --train_n_way=5 --test_n_way=5 --n_shot=1 --seed=1 --train_aug --steps=3 --tau=1 --loss="ELBO"

This will train MD 5-way 1-shot on the CUB dataset with seed 1 and the ELBO loss. The dataset string can be one of the following: CUB, miniImagenet, cross, cross_char. At training time the best model is evaluated on the validation set and stored as best_model.tar in the folder ./save/checkpoints/DATASET_NAME. The parameter --train_aug enables data augmentation. The parameter seed set the seed for pytorch, numpy, and random. Set --seed=0 or remove the parameter for a random seed. The parameter steps controls the task-level update steps for inner-loop update steps of mirror descent. The parameter loss can be ELBO or PL, corresponding to the ELBO loss and the predictive likelihood loss (only used in converging experiments). Additional parameters are provided in the file io_utils.py, such as mean and kernel, where mean sets the prior mean for GP (default is 0) and the choices of kernel include linear, rbf, matern, poli1, poli2, bncossim.

Test classification. For testing our methods it is enough to repeat the train command replacing the call to train.py with the call to test.py as follows:

python test.py --dataset="CUB" --method="MD" --train_n_way=5 --test_n_way=5 --n_shot=1 --seed=1 --train_aug --steps=3 --tau=1 --loss="ELBO"

Calibration. For calibrating our methods it is enough to repeat the train command replacing the call to train.py with the call to calibrate.py as follows:

python calibrate.py --dataset="CUB" --method="MD" --train_n_way=5 --test_n_way=5 --n_shot=5 --seed=53 --train_aug --steps=3 --tau=1 --loss="ELBO"

Reproduction of our results. You can simply run

sh classification.sh

and

sh calibrate.sh

to reproduce the results presented in the paper.