train_with_da.py

class SegDataset(Dataset):
    def __init__(self, csv_loc, data_dir, augments=200):
        self.data_dir = data_dir
        self.images_data = read_csv(csv_loc)
        self.images = self.prepare_images()

    def transform(self, raw, seg):
        i, j, h, w = transforms.RandomCrop.get_params(
            raw, output_size=(128, 128))
        raw = trans_f.crop(raw, i, j, h, w)
        seg = trans_f.crop(seg, i, j, h, w)

        if random.random() > 0.5:
            raw = trans_f.hflip(raw)
            seg = trans_f.hflip(seg)

        if random.random() > 0.5:
            raw = trans_f.vflip(raw)
            seg = trans_f.vflip(seg)

        raw = trans_f.to_tensor(raw).mul(255).float().to(device)
        seg = trans_f.to_tensor(seg).mul(255).long().to(device)

        return {'raw': raw, 'seg': seg}

    def prepare_images(self):
        images = []

        def read_image_by_id(idx, raw_image=True):
            img_name = os.path.join(self.data_dir,
                                self.images_data[idx][int(not raw_image)])
            return Image.open(img_name)

        for idx in range(len(self.images_data)):
            raw = read_image_by_id(idx)
            seg = read_image_by_id(idx, False)
            for _ in range(self.augments):
                images.append(self.transform(raw, seg))
        random.shuffle(images)
        return images

    def __len__(self):
        return len(self.images)

    def __getitem__(self, idx):
        return self.images[idx]