Few shot learning by features adaptation with Graph Neural Networks

PyTorch implementation for Few shot learning by features adaptation with Graph Neural Networks, part of my Bachelor’s thesis.

Work presented at Eastern European Summer School (EEML 2020), winning a Best Poster Award.

Installation

conda create -n few-shot python=3.8
conda activate few-shot
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch
pip install higher
pip install torchmeta
pip install tensorboard

Running

This will load the hyperparameters saved in config and start a new training process. You can stop it any time with CTRL+C.

python main.py

To start the process in the background, run the following command:

./start_experiment.sh [experiment_name]

It will also create a new experiment directory with all the logs, config and current code changes. CTRL+C now will only stop the log from displaying, but the process continues running.

Evaluate

Run python main.py --test from an experiment's copy of the project directory.

cd experiments/experiment__21_november_2021__15_43_36
cd project
python main.py --test