Implementation of the Meta-Learning-by-Adjusting-Priors algorithm in PyTorch 1.0

Implementation the paper R. Amit, R. Meir, “Meta-Learning by Adjusting Priors Based on Extended PAC-Bayes Theory”, ICML 2018 [paper] [Slides-Short] [Slides-long] [video]

Prerequisites

• Python 3.5+
• PyTorch 1.0+ with CUDA
• NumPy and Matplotlib

Data

The data sets are downloaded automatically. Specify the main data path in the file 'Data_Path.py'

Reproducing experiments in the paper:

• PriorMetaLearning/run_MPB_*.py - Learns a prior from the obsereved (meta-training) tasks and use it to learn new (meta-test) tasks.

• Toy_Examples/Toy_Main.py - Toy example of 2D estimation.

• Single_Task/main_TwoTaskTransfer_PermuteLabels and Single_Task/main_TwoTaskTransfer_PermutePixels.py - run alternative tranfer methods.

• PriorMetaLearning/Analyze_Prior.py - Analysis of the weight uncertainty ine each layer of the learned prior (run after creating a prior with main_Meta_Bayes.py)

Other experiments:

• Single_Task/main_single_standard.py - Learn standard neural network in a single task.
• Single_Task/main_single_Bayes.py - Learn stochastic neural network in a single task.

MAML code is based on: https://github.com/katerakelly/pytorch-maml