如何在Pytorch中使用torchvision.transforms进行分割任务的数据扩充?

时间:2019-10-03 08:20:28

标签: python pytorch transformation torchvision

我对PyTorch中执行的数据扩充有些困惑。

因为我们正在处理分段任务,所以我们需要数据和掩码来进行相同的数据增强,但是其中一些是随机的,例如随机轮换。

Keras提供random seed保证数据和掩码执行相同的操作,如以下代码所示:

    data_gen_args = dict(featurewise_center=True,
                         featurewise_std_normalization=True,
                         rotation_range=25,
                         horizontal_flip=True,
                         vertical_flip=True)


    image_datagen = ImageDataGenerator(**data_gen_args)
    mask_datagen = ImageDataGenerator(**data_gen_args)

    seed = 1
    image_generator = image_datagen.flow(train_data, seed=seed, batch_size=1)
    mask_generator = mask_datagen.flow(train_label, seed=seed, batch_size=1)

    train_generator = zip(image_generator, mask_generator)

我在Pytorch官方文档中找不到类似的描述,所以我不知道如何确保数据和掩码可以同步处理。

Pytorch确实提供了这样的功能,但是我想将其应用于自定义的Dataloader。

例如:

def __getitem__(self, index):
    img = np.zeros((self.im_ht, self.im_wd, channel_size))
    mask = np.zeros((self.im_ht, self.im_wd, channel_size))

    temp_img = np.load(Image_path + '{:0>4}'.format(self.patient_index[index]) + '.npy')
    temp_label = np.load(Label_path + '{:0>4}'.format(self.patient_index[index]) + '.npy')

    for i in range(channel_size):
        img[:,:,i] = temp_img[self.count[index] + i]
        mask[:,:,i] = temp_label[self.count[index] + i]

    if self.transforms:
        img = np.uint8(img)
        mask = np.uint8(mask)
        img = self.transforms(img)
        mask = self.transforms(mask)

    return img, mask

在这种情况下,由于某些操作(例如随机旋转)是随机的,因此img和mask将分别转换,因此mask和image之间的对应关系可能会更改。换句话说,图像可能已经旋转了,但是遮罩却没有这样做。

编辑1

我在augmentations.py中使用了该方法,但出现错误::

Traceback (most recent call last):
  File "test_transform.py", line 87, in <module>
    for batch_idx, image, mask in enumerate(train_loader):
  File "/home/dirk/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 314, in __next__
    batch = self.collate_fn([self.dataset[i] for i in indices])
  File "/home/dirk/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 314, in <listcomp>
    batch = self.collate_fn([self.dataset[i] for i in indices])
  File "/home/dirk/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/utils/data/dataset.py", line 103, in __getitem__
    return self.dataset[self.indices[idx]]
  File "/home/dirk/home/data/dirk/segmentation_unet_pytorch/data.py", line 164, in __getitem__
    img, mask = self.transforms(img, mask)
  File "/home/dirk/home/data/dirk/segmentation_unet_pytorch/augmentations.py", line 17, in __call__
    img, mask = a(img, mask)
TypeError: __call__() takes 2 positional arguments but 3 were given

这是我的__getitem__()代码:

data_transforms = {
    'train': Compose([
        RandomHorizontallyFlip(),
        RandomRotate(degree=25),
        transforms.ToTensor()
    ]),
}

train_set = DatasetUnetForTestTransform(fold=args.fold, random_index=args.random_index,transforms=data_transforms['train'])

# __getitem__ in class DatasetUnetForTestTransform
def __getitem__(self, index):
    img = np.zeros((self.im_ht, self.im_wd, channel_size))
    mask = np.zeros((self.im_ht, self.im_wd, channel_size))
    temp_img = np.load(Label_path + '{:0>4}'.format(self.patient_index[index]) + '.npy')
    temp_label = np.load(Label_path + '{:0>4}'.format(self.patient_index[index]) + '.npy')
    temp_img, temp_label = crop_data_label_from_0(temp_img, temp_label)
    for i in range(channel_size):
        img[:,:,i] = temp_img[self.count[index] + i]
        mask[:,:,i] = temp_label[self.count[index] + i]

    if self.transforms:
        img = T.ToPILImage()(np.uint8(img))
        mask = T.ToPILImage()(np.uint8(mask))
        img, mask = self.transforms(img, mask)

    img = T.ToTensor()(img).copy()
    mask = T.ToTensor()(mask).copy()
    return img, mask

编辑2

我发现在ToTensor之后,相同标签之间的骰子变为255,而不是1,如何解决?

# Dice computation
def DSC_computation(label, pred):
    pred_sum = pred.sum()
    label_sum = label.sum()
    inter_sum = np.logical_and(pred, label).sum()
    return 2 * float(inter_sum) / (pred_sum + label_sum)

随时询问是否需要更多代码来解释问题。

3 个答案:

答案 0 :(得分:2)

需要输入参数(例如RandomCrop)的转换具有一种get_param方法,该方法将返回该特定转换的参数。然后可以使用transforms的功能接口将其应用于图像和蒙版:

from torchvision import transforms
import torchvision.transforms.functional as F

i, j, h, w = transforms.RandomCrop.get_params(input, (100, 100))
input = F.crop(input, i, j, h, w)
target = F.crop(target, i, j, h, w)

此处提供样品: https://github.com/pytorch/vision/releases/tag/v0.2.0

完整的示例可用于VOC和COCO: https://github.com/pytorch/vision/blob/master/references/segmentation/transforms.py https://github.com/pytorch/vision/blob/master/references/segmentation/train.py

关于错误,

ToTensor()未被覆盖以处理其他掩码参数,因此不能在data_transforms中。此外,__getitem__会同时返回ToTensorimg的{​​{1}}。

mask

答案 1 :(得分:1)

torchvision还提供了类似的功能[document]

这是一个简单的例子,

import torchvision
from torchvision import transforms

trans = transforms.Compose([transforms.CenterCrop((178, 178)),
                                    transforms.Resize(128),
                                    transforms.RandomRotation(20),
                                    transforms.ToTensor(),
                                    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
dset = torchvision.datasets.MNIST(data_root, transforms=trans)

编辑

自定义自己的CelebA数据集时的简短示例。请注意,要应用转换,您需要调用transform中的__getitem__列表。

class CelebADataset(Dataset):
    def __init__(self, root, transforms=None, num=None):
        super(CelebADataset, self).__init__()

        self.img_root = os.path.join(root, 'img_align_celeba')
        self.attr_root = os.path.join(root, 'Anno/list_attr_celeba.txt')
        self.transforms = transforms

        df = pd.read_csv(self.attr_root, sep='\s+', header=1, index_col=0)
        #print(df.columns.tolist())
        if num is None:
            self.labels = df.values
            self.img_name = df.index.values
        else:
            self.labels = df.values[:num]
            self.img_name = df.index.values[:num]

    def __getitem__(self, index):
        img = Image.open(os.path.join(self.img_root, self.img_name[index]))
        # only use blond_hair, eyeglass, male, smile
        indices = [9, 15, 20, 31]
        label = np.take(self.labels[index], indices)
        label[label==-1] = 0

        if self.transforms is not None:
            img = self.transforms(img)

        return np.asarray(img), label

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


编辑2

乍一看,我可能会错过一些东西。问题的重点是如何对img和标签应用“相同”的数据预处理。据我了解,没有可用的Pytorch内置功能。因此,我之前所做的就是自己实现增强。

class RandomRotate(object):
    def __init__(self, degree):
        self.degree = degree

    def __call__(self, img, mask):
        rotate_degree = random.random() * 2 * self.degree - self.degree
        return img.rotate(rotate_degree, Image.BILINEAR), 
                           mask.rotate(rotate_degree, Image.NEAREST)

请注意,输入应为PIL格式。有关更多信息,请参见this

答案 2 :(得分:1)

另一个想法是沿通道维度堆叠图像和蒙版,然后将它们一起转换。显然,这仅适用于几何类型转换,并且您需要对两者使用相同的 dtype。我使用这样的东西:

# Apply these to image and mask
affine_transforms = transforms.Compose([
    transforms.RandomAffine(degrees=180),
    ...
])

# Apply these to image only
image_transforms = transforms.Compose([
    transforms.GaussianBlur(),
    ...
])

# Loader...
def __getitem__(self, index: int):
    # Get the image and mask, here shape=(HxW) for both
    image = self.images[index]
    mask = self.masks[index]

    # Stack the image and mask together so they get the same geometric transformations
    stacked = torch.cat([image, mask], dim=0)  # shape=(2xHxW)
    stacked = self.affine_transforms(stacked)

    # Split them back up again
    image, mask = torch.chunk(stacked, chunks=2, dim=0)

    # Image transforms are only applied to the image
    image = self.image_transforms(image)

    return image, mask