Descriptor Analysis for Domain Generalization

This repository contains code for understanding descriptors in domain generalization algorithms quantitatively and qualitatively. The baseline algorithms and datasets have been been adapted from DomainBed.

Currently, 2 algorithms have been included:

• Empirical Risk Minimization (ERM, Vapnik, 1998)
• Interdomain Mixup (Mixup, Yan et al., 2020)

In the final version, we add more:

• Domain Adversarial Neural Network (DANN, Ganin et al., 2015)
• Group Distributionally Robust Optimization (GroupDRO, Sagawa et al., 2020)

The dataset being used currently is:

• DomainNet (Peng et al., 2019)

In the final version, we use:

• PACS (Li et al., 2017)

• Checkout commands.sh on how to train and generate descriptors.
• Requirements for pip can be found inside requirements.txt

For the above algorithms and datasets, the descriptors can be found in the extracted folder.

This work has been done as part of the CS 444 Deep Learning for Computer Vision project.

Loading PACS Descriptors

def extract_data(data, merge_groups=True, transform=None, ):
    zs, ys, preds, gs, logits = data['feature'], data['label'], data['pred'], data['group'], data['logits']
    if transform is not None:
        zs = transform(zs)
    #     gs = gs % 2
    return zs, ys, gs, preds, logits
    
PATH_TO_FILE_TRAIN = "put path to .p train file here"
PATH_TO_FILE_TEST = "put path to .p test file here"

train_data = pickle.load(open(PATH_TO_FILE_TRAIN, 'rb'))
test_data = pickle.load(open(PATH_TO_FILE_TEST', 'rb'))

train_gs = train_data['group']
n_train = len(train_gs)
groups, counts = np.unique(train_data['group'], return_counts=True, axis=0)
n_groups = len(groups)
n_classes = len(np.unique(train_data['label']))
  
  
# Load train and test data according to extract_data function
  
zs, ys, gs, preds, logits = extract_data(train_data)

test_zs, test_ys, test_gs, test_preds, test_logits = extract_data(test_data)