mrusci / snatcher

[CVPR 2021] Few-shot Open-set Recognition by Transformation Consistency

Few-shot Open-set Recognition by Transformation Consistency [CVPR 2021]

Link to paper: arXiv

Description

This code executes SnaTCHer-F on miniImageNet and tieredImageNet

Prerequisites

• Ubuntu 18.04
• Python 3.7
• PyTorch 1.6
• CUDA 10.1
• Anaconda 4.9.2
• CUDNN v7.6.3

Usage

1. Prepare datasets and checkpoints from link
2. Move datasets under './data' (see main.py)
3. Move checkpoints under './checkpoints' with prefix (mini- or tiered-, see main.py)
4. Run main.py

Running example

• miniImageNet 5-way 1-shot python main.py --dataset MiniImageNet --shot 1

• tieredImageNet 5-way 5-shot python main.py --dataset TieredImageNet --shot 5

Results

	Acc (%)	Probability (%)	Distance (%)	SnaTCHer (%)
Mini1shot	66.15	59.37	68.74	69.39
Mini5shot	81.87	62.71	76.01	77.36
Tiered1shot	70.41	63.88	69.80	74.38
Tiered5shot	84.79	73.79	77.25	81.78

SnaTCHer details

See model/trainer/fsl_trainer_SnatCHerF.py

Acknowledgement

The code is based on github.com/Sha-Lab/FEAT