MAML

Intro

This repository implements the paper, Model-Agnostic Meta-Leanring for Fast Adaptation of Deep Networks.

Requirements

• Tensorflow (v1.3 or higher)
• better_exceptions, tqdm, Pillow, etc.

Training statisitics

• Sinusoid

  Red line - ground truth, Red dots - given observation, Blue line - predicted line after 1 sgd step

• Omniglot

  • Omniglot Testing (multiple descent steps)

  

  • valid_acc_{0,1} means accuracy after 1 and 2 SGD steps. The valid_acc is the accuracy after the weights are trained with 3 SGD steps.

Training

Download datasets

Downlaod Omniglot dataset from the link. Only images_background.zip and images_evalueation.zip are required.

Unzip on the directory (repository)/datasets/omniglot/, so the directory shoud looks like (repo)/datasets/omniglot/{images_background,images_evaluation}.

Run train

• Run sinusoide: python sinusoide.py
• Run omniglot: python omniglot.py

Change the hyperparameters accordingly as you want. Please check at the bottom of each script.

TODO

• Mini Imagenet Training
• Robotic Simulation.

Acknowledgement

• Author's original implementation: link