layumi / adaboost_seg

TIP2022 Adaptive Boosting (AdaBoost) for Domain Adaptation ? :woman_shrugging: Why not ! :ok_woman:

Hi there 👋

Contact Me:

✉ Email: zhedongzheng AT um.edu.mo

✧ Website: http://zdzheng.xyz

✧ Linkedin: https://www.linkedin.com/in/zhedongzheng/

✧ Google Scholar: https://scholar.google.com/citations?hl=en&user=XT17oUEAAAAJ

adaboost_seg's Issues

The pth is not found

RESTORE_FROM = 'http://vllab.ucmerced.edu/ytsai/CVPR18/DeepLab_resnet_pretrained_init-f81d91e8.pth'
in trainer_ms.py ,this pth "404 not found"

the update_class_critertion , this function comes from

 def update_class_criterion(self, labels):
            weight = torch.FloatTensor(self.num_classes).zero_().cuda()
            weight += 1
            count = torch.FloatTensor(self.num_classes).zero_().cuda()
            often = torch.FloatTensor(self.num_classes).zero_().cuda()
            often += 1
            print(labels.shape)
            n, h, w = labels.shape
            for i in range(self.num_classes):
                count[i] = torch.sum(labels==i)
                if count[i] < 64*64*n: #small objective
                    weight[i] = self.max_value
            if self.often_balance:
                often[count == 0] = self.max_value

            self.often_weight = 0.9 * self.often_weight + 0.1 * often 
            self.class_weight = weight * self.often_weight
            print(self.class_weight)
            if self.focal:
                print('Using Focal loss')
                return FocalLoss(weight = self.class_weight, ignore_index=255, gamma = self.gamma)
            else:
                return nn.CrossEntropyLoss(weight = self.class_weight, ignore_index=255)

I find the update_class_criterion function in your trainer_ms.py.
I want to know who created it in the first place because I want to change it somewhere, I need quote.
I think it's a very come class_criterion If you know, could you tell me.

Thank you for your work. Has the number of lines of core code changed because of the modification?