/meta-curvature

Code for "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks"

Primary LanguagePythonMIT LicenseMIT

Meta-Curvature

This repo contains code for reproducing the experimental results (sinusoid regression and Omniglot classification) in the paper, Meta-Curvature (Park et al., NeurIPS 2019). This is basd on original MAML implementation.

Usage

For sinuosoid experiments,

python main.py --datasource=sinusoid --logdir=logs/sine/ --metatrain_iterations=70000 --norm=None --update_batch_size=10 --mc=True

For Omniglot experiments,

# 5-way 1-shot
python main.py --datasource=omniglot --metatrain_iterations=60000 --meta_batch_size=32 --update_batch_size=1 --num_classes=5 --update_lr=0.4 --num_updates=1 --logdir=logs/omniglot5way/ --mc=True
# 5-way 5-shot
python main.py --datasource=omniglot --metatrain_iterations=60000 --meta_batch_size=32 --update_batch_size=5 --num_classes=5 --update_lr=0.4 --num_updates=1 --logdir=logs/omniglot5way/ --mc=True
# 20-way 1-shot
python main.py --datasource=omniglot --metatrain_iterations=60000 --meta_batch_size=32 --update_batch_size=1 --num_classes=20 --update_lr=0.4 --num_updates=1 --logdir=logs/omniglot20way/ --mc=True
# 20-way 5-shot
python main.py --datasource=omniglot --metatrain_iterations=60000 --meta_batch_size=32 --update_batch_size=5 --num_classes=20 --update_lr=0.4 --num_updates=1 --logdir=logs/omniglot20way/ --mc=True