Skip to content

Latest commit

 

History

History
executable file
·
34 lines (25 loc) · 3.12 KB

README.md

File metadata and controls

executable file
·
34 lines (25 loc) · 3.12 KB

GATE - Generalization After Transfer Evaluation Benchmark Engine

Welcome to GATE - A benchmark framework built to evaluate a learning process on its ability to learn and generalize on previously unseen Tasks, Data domains and Modalities. GATE ensures fair benchmarking by enforcing separation between learners, models, tasks and data, enabling multiple approaches (learners) to be fairly compared against each other by using identical network backbones (models) for all learners on well defined tasks and datasets.

This repository utilises E-GATE to benchmark Partial Observation Experts Modelling (POEM) at self-supervised few-shot learning from partial observations, as introduced and described in the paper Contrastive Meta-Learning for Partially Observable Few-Shot Learning (accepted for publication at ICLR 2023).

Installation

Create a new environment and run install_env_nvidia_gpu.sh or install_dependencies.sh for a GPU or CPU machine respectively, and optionally install_dev_tools.sh to install addional development tools. Create a .env file from .env.template by adding your personal environment variables.

Results

The final results of the GATE evaluation on Partially Observable Meta-Dataset (as defined in the paper Contrastive Meta-Learning for Partially Observable Few-Shot Learning) achieved using the learners in this repository (with ResNet-18 backbones) are:

Test Source Finetune ProtoNet MAML POEM
Aircraft 46.5+/-0.6 48.5+/-1.0 37.5+/-0.3 55.3+/-0.7
Birds 62.6+/-0.7 67.4+/-1.2 52.5+/-0.6 71.1+/-0.1
Flowers 48.5+/-0.4 46.4+/-0.7 33.5+/-0.3 49.2+/-1.5
Fungi 61.0+/-0.2 61.4+/-0.4 46.1+/-0.4 64.8+/-0.3
Omniglot 71.3+/-0.1 87.8+/-0.1 47.4+/-1.0 89.2+/-0.7
Textures 83.2+/-0.4 76.7+/-1.6 73.1+/-0.4 81.4+/-0.6

Ablating to the standard Meta-Dataset evaluation procedure (with fully observed images for standard few-shot learning) as defined in the original Meta-Dataset paper the results final results achieved using the learners in this repository (with ResNet-18 backbones) are:

Test Source Finetune ProtoNet MAML POEM
Aircraft 56.2+/-1.1 47.2+/-1.2 35.9+/-1.8 46.5+/-1.5
Birds 52.6+/-1.8 78.3+/-0.5 65.2+/-0.3 79.4+/-0.3
Flowers 80.1+/-2.0 84.2+/-0.7 70.4+/-0.4 83.6+/-1.3
Fungi 33.6+/-1.7 84.7+/-0.2 18.9+/-0.2 81.0+/-0.1
Omniglot 89.6+/-3.3 98.7+/-0.1 94.7+/-0.1 98.6+/-0.1
Textures 60.4+/-1.0 65.3+/-1.2 56.1+/-0.3 65.7+/-0.8

This demonstrates that POEM provides additional benefits for unifying representations when dealing with partial observability, while remaining competitive at standard few-shot learning.