HyProMeta, Mitigating Label Noise using Prompt-Based Hyperbolic Meta-Learning in Open-Set Domain Generalization

1. Introduction

This repository provides the source code for Mitigating Label Noise using Prompt-Based Hyperbolic Meta-Learning in Open-Set Domain Generalization:

PACS dataset is available at https://huggingface.co/datasets/flwrlabs/pacs

DigitsDG dataset is available at https://csip.fzu.edu.cn/files/datasets/SSDG/digits_dg.zip

The corrupted labels are inside osdg_nl_files folder

# PACS
Known classes: ['dog', 'elephant', 'giraffe', 'guitar', 'horse', 'house']
Unknown classes: ['person']

2. Dataset Construction

The dataset needs to be divided into two folders for training and validation. We provide reference code for automatically dividing data using official split in data_list/split_kfold.py.

root_dir = "path/to/PACS"
instr_dir = "path/to/PACS_data_list"

3. Train

To run the training code, please update the path of the dataset in ml_open.py:

if dataset == 'PACS':	
    train_dir = 'path/to/PACS_train' # the folder of training data 
	val_dir = 'path/to/PACS_val' # the folder of validation data 
	test_dir = 'path/to/PACS_all' or ['path/to/PACS_train', 'path/to/PACS_val']

then simply run:

python train_file.py