MEDIC++

2025.11.06

Newly released updates to the v2 branch:

1. Fix the incompatibility issue between model_urls and newer versions of torchvision.
2. Address several runtime warnings.

2025.10.31

1. Add the algorithm Balanced Direction from Multifarious Choices: Arithmetic Meta-Learning for Domain Generalization to the v2 branch of our repository. You can set the algorithm in config.py to arith to enable it. We recommend using a smaller learning rate than MEDIC.
2. Remove the evaluation metric H-score, as we are still uncertain about its validity. However, you can still find it in our previous repository.
3. Refactor parts of the code and fix some bugs.

1. Introduction

This repository contains the implementation of the under-review paper Exploring Dualistic Meta-Learning to Enhance Domain Generalization in Open Set Scenarios, entended from our ICCV 2023 paper Generalizable Decision Boundaries: Dualistic Meta-Learning for Open Set Domain Generalization.

2. Dataset Construction (Optional)

You can divide the dataset 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 main.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 main.py --source-domain ... --target-domain ... --save-name ... --gpu 0

If there is no validation folder, please use the option --random-split to create the validation set.

You can use --save-later to save the model in the last 15% iterations.

4. DomainBed

To simulate the DomainBed benchmark, please set transforms in dataset/dataloader.py as follows:

if small_img == False:
        img_tr = [transforms.RandomResizedCrop((224, 224), (0.7, 1.0))] # 0.7 for DomainBed, otherwise 0.8
        
if color_jitter:
        img_tr.append(transforms.ColorJitter(brightness=0.3, contrast=0.3, saturation=0.3, hue=0.3)) # 0.3 for DomainBed, otherwise 0.4

then add options:

• --batch-size 12 (18 for DomainNet, the total batch size is this value times the number of tasks 9)
• --optimize-method Adam
• --num-epoch 5100 (15000 for DomainNet)
• --eval-step 300
• --lr 6e-6 (divide the default learning rate 5e-5 by the number of tasks 9)
• --without-bcls

python main.py --source-domain ... --target-domain ... --save-name ... --gpu 0 --batch-size 12 --optimize-method Adam --num-epoch 5100 --eval-step 300 --lr 6e-6 --without-bcls

5. Evalution

To run the evaluation code, please update the path of the dataset in eval.py:

if dataset == 'PACS':
        root_dir = 'path/to/PACS_all' or ['path/to/PACS_train', 'path/to/PACS_val']

then simply run:

python eval.py --hits ... --save-name ... --gpu 0

• --hits (check points for H-score)