steinarlaenen/on-episodes-fsl

This repository contains the code for the paper On Episodes, Prototypical Networks, and Few-shot Learning

On Episodes, Prototypical Networks, and Few-Shot Learning

This is the codebase for the paper On Episodes, Prototypical Networks, and Few-Shot Learning, by Steinar Laenen and Luca Bertinetto.

A preliminary version of this work appeared as an oral presentation at NeurIPS 2020 workshop on meta-learning.

Usage

With conda

• conda env create -f environment.pytorch13.yml
• source activate few_shot_NCA
• pip install -e .

NOTE: if you are having issues with conda while re-installing the environment, e.g. old packages are used instead of the ones specified by the environment.yml file, try to remove (using conda) pytorch and torchvision and re-installing them with the appropriate cuda-toolkit, e.g. conda install pytorch torchvision cudatoolkit=10.0 -c pytorch.

Data preperation

For miniimagenet, the train/val/test split files need to be downloaded from the SimpleShot Repo, and should be copied to <REPO_ROOT>/split/.

2. Download Datasets

For any google drive link, if you want to download it from the command line, you can do that by using:

gdown --id FILEID -O FILENAME

The FILEID can be found in the link itself, and the FILENAME can be found by clicking the link.

--> When you are done downloading the dataset(s), remove example from ./dataset_configs/<dataset>.yaml.example and edit accordingly. This file won't be committed.

2.1 Mini-ImageNet

You can download the dataset from https://drive.google.com/open?id=0B3Irx3uQNoBMQ1FlNXJsZUdYWEE

When setting the data argument in ./dataset_configs/miniimagenet.yaml.example, make sure that it points to the images folder:path/to/miniimagenet/images/.

2.2 Tiered-ImageNet

You can download the dataset from https://drive.google.com/file/d/1g1aIDy2Ar_MViF2gDXFYDBTR-HYecV07/view. After downloading and unzipping this dataset, you have to run the follow script to generate split files.

python src/utils/tieredImagenet.py --data path-to-tiered --split split/tiered/

When setting the data argument in ./dataset_configs/tieredimagenet.yaml.example, make sure that it points to the data folder:path/to/tiered-imagenet/data/.

2.4 CIFARFS

You can download the dataset from https://drive.google.com/file/d/1pTsCCMDj45kzFYgrnO67BWVbKs48Q3NI/view.

After downloading and unzipping this dataset, you have to run the follow script to generate split files.

python ./scripts/cifarfs_split.py --data path-to-cifar100 --split split/cifar100/

When setting the data argument in ./dataset_configs/CIFARFS.yaml.example, make sure that it points to the images folder:path/to/cifar100/images/.

NOTE: If any of these download links are broken, these datasets can be recreated from the original papers that proposed them: miniImageNet. CIFAR-FS, and tieredImageNet. We put all the images into a single folder, and use .csv files for the train.csv, val.csv, and test.csv splits where there is one row for the image filename and a corresponding row for the label.

3 Loading and Evaluating Models

To train a model, first change directory: cd ./scripts and then run:

python ./start_training.py

./src/utils/configs/configuration.py contains all the arguments that can be specified.

4 Experimental results

Experiments scripts live under ./scripts/bash_scripts/

Batch-size experiments run:

• ./bash_scripts/batch_expm_cifar_proto.sh
• ./bash_scripts/batch_expm_miniimagenet_proto.sh
• ./bash_scripts/batch_expm_cifar_matching.sh
• ./bash_scripts/batch_expm_miniimagenet_matching.sh

Ablation experiments run:

• ./bash_scripts/abl_all_cifar_proto.sh
• ./bash_scripts/abl_all_miniimagenet_proto.sh

Subsampling experiment:

• ./bash_scripts/sample_NCA_pairs_proto.sh GPU='X,X' SAMPLE="0.01 0.05 0.1 0.2 0.4 0.8" SEED="0 1 2"
• ./bash_scripts/sample_NCA_pairs_matching.sh GPU='X,X' SAMPLE="0.01 0.05 0.1 0.2 0.4 0.8" SEED="0 1 2"

Comparisons for large table:

NCA results

• ./bash_scripts/nca_rn12_miniimagenet.sh GPU='X,X' SEED="0 1 2"
• ./bash_scripts/nca_rn12_miniimagenet_soft.sh GPU='X,X' SEED="0 1 2"
• ./bash_scripts/nca_rn12_cifar.sh GPU='X,X' SEED="0 1 2"
• ./bash_scripts/nca_rn12_cifar_soft.sh GPU='X,X' SEED="0 1 2"
• ./bash_scripts/nca_rn12_tiered.sh GPU='X,X,X,X' SEED="0 1 2"
• ./bash_scripts/nca_rn12_tiered_soft.sh GPU='X,X,X,X' SEED="0 1 2"

New (ours) episodic setups for matching and prototypical networks

• ./bash_scripts/protonew_rn12_cifar.sh GPU='X,X' SEED="0 1 2"
• ./bash_scripts/matchingnew_rn12_cifar.sh GPU='X,X' SEED="0 1 2"
• ./bash_scripts/protonew_rn12_tiered.sh GPU='X,X' SEED="0 1 2"
• ./bash_scripts/matchingnew_rn12_tiered.sh GPU='X,X' SEED="0 1 2"

Old episodic setups for matching and prototypical networks

• ./bash_scripts/protoold_miniimagenet.sh GPU='X,X' SEED="0 1 2"
• ./bash_scripts/matchingold_miniimagenet.sh GPU='X,X' SEED="0 1 2"
• ./bash_scripts/protoold_cifar.sh GPU='X,X' SEED="0 1 2"
• ./bash_scripts/matchingold_cifar.sh GPU='X,X' SEED="0 1 2"
• ./bash_scripts/protoold_tiered.sh GPU='X,X' SEED="0 1 2"
• ./bash_scripts/matchingold_tiered.sh GPU='X,X' SEED="0 1 2"

How to get accuracy + conf interval over 3 seeds

Results for all 30.000 episodes get saved in ./results/numpy_results. You can find an example script in ./scripts/extract_mean_conf.py, where the correct regex for the experiment needs to be written to extract the mean + conf interval for the test and validation set. For more details see the scripts.

The code in this repository is adapted from the code for SimpleShot. We thank the authors for making their code available.