Heterogeneous Domain Generalization via Domain Mixup

The code release of paper 'Heterogeneous Domain Generalization via Domain Mixup' ICASSP 2020.

A simple but effective way to improve the heterogeneous domain generalization performance. The core code is as follows.

The paper and slide can be found here.

Core Code

def mixall(x, y, beta=8, domain_num=6):
    '''Returns mixed inputs, pairs of targets, and lambda'''
    batch_size = x.size()[0]
    lam = torch.rand(domain_num, batch_size).cuda()
    lam = F.softmax(lam * beta, dim=0)
    index_list = []
    mixed_x = 0
    for i in range(6):
        index_list.append(torch.randperm(batch_size).cuda())
        if self.flags.mix_from =='image':
            mixed_x += x[index_list[i], :] * lam[i].reshape(batch_size, 1, 1, 1)
        else:
            mixed_x += x[index_list[i], :] * lam[i].reshape(batch_size, 1)
    return mixed_x, index_list, y, lam

def mixup_criterion(pred, lam, index_list, y, domain_num=6):
    loss = 0
    for i in range(domain_num):
        loss += lam[i] * F.cross_entropy(pred, y[index_list[i]], reduction="none")
    return loss.mean()

Workflow

The example steps are as follows:

for (x,y) in iterDomainBatch:
    mixed_x, index_list, mixed_y, lam  = mixall(x, y)
    pred_y = model(mixed_x)
    loss = mixup_criterion(pred_y, lam, index_list, mixed_y)
    loss.backward()
    optimizer.step()

Pretrained Model

The pretrained resnet-18 model can be downloaded at https://drive.google.com/file/d/12wLIh29bhBWxQZnpUoJghS5LEGrlSDsM/view?usp=sharing

from torchvison.models import resnet
model = resnet.resnet18()
para = torch.load('pretrained.pkt') # p[0] feature extractor p[1] classifier
model.load_state_dict(para[0],strict=False)

Please cite our paper if you find it is useful.

@inproceedings{wang2020heterogeneous,
  title={Heterogeneous Domain Generalization Via Domain Mixup},
  author={Wang, Yufei and Li, Haoliang and Kot, Alex C},
  booktitle={ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={3622--3626},
  year={2020},
  organization={IEEE}
}