MAML Algorithms for Meta Learning

What is this ?

This is a package to quickly run the following Meta-Learning algorithms:

• MAML
• PLATIPUS
• BMAML
• CLV
• Baseline (classical supervised learning)

Getting Started

1. (Optional) create and activate a conda environment

   conda create --name meta_learning
   conda activate meta_learning
   conda install pip
   

2. clone this repository and run

   pip install .
   

   from the source directory

Usage

run in vscode

We defined some default scenarios in the .vscode/launch.json you can simply run if you open the project in vscode.

• QuickRun
• Full Run (might take some hours depending on your machine)
• Adaptation Sanity Check
• Meta-Update Sanity Check

import into another project

Install this package in your own project using pip

pip install git+https://github.com/LuisAugenstein/Bayesian-Meta-Learning

Then import the Learner into your script

from bayesian_meta_learning import Learner

config = Learner.get_default_config()
config['algorithm'] = 'maml'
config['benchmark'] = 'Sinusoid1D'
config['num_epochs'] = 10
config['num_test_tasks'] = 4
config['minibatch'] = 4

Learner.run(config)

Comparison Runs

The runs we compare in the final report were created using the script slurm/comparison_run.sh.

This script requires a slurm Workload Manager to be installed. Make sure to change to the appropriate Conda environment before starting the jobs.

At the time of testing, all jobs were run on CPU. To change to GPU execution, go to slurm/start_job.sh and change the required resources to GPU. If installed correctly, the GPU should be recognized automatically, so no further changes are necessary.

To start the full comparison, run:

sh slurm/comparison_run.sh

Starting the full comparison requires significant compute power. With the current configuration, you can expect around 2000 CPU hours.

Acknowledgement

This implementation depends heavily on few_shot_meta_learning. If you do not need our comparison suite and only need to train meta learning algorithms, this repository might be more suitable.