pytorch transforms. Use of lambda

When you want to set the image transforms strategy , Such as ：

from torchvision import transforms as T

normalize = T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
data_transforms = {
    
    'train': T.Compose([
        T.RandomResizedCrop(224),  #  Capture from picture center 
        T.RandomHorizontalFlip(),  #  Randomly flip a given... Horizontally PIL.Image, The turnover probability is 0.5
        T.ToTensor(),  #  Turn into Tensor Format , The size range is [0,1]
        normalize
    ]),

    'val': T.Compose([
        T.Resize(256),  #  Resize 
        T.CenterCrop(224),
        T.ToTensor(),
        normalize
    ]),
}

But sometimes official methods don't meet your needs , At this time, you need to customize your own transform Strategy

The way is to use transforms.Lambda

Illustrate with examples ：

For example, when we want to capture an image , But I don't want to intercept at random , Instead, you want to intercept at a designated location

Then you need to customize an interception function , And then use transforms.Lambda Just encapsulate it , Such as ：

# coding:utf-8
from torchvision import transforms as T

def __crop(img, pos, size):
    """ :param img:  Input image  :param pos:  Location of image capture , The type is tuple , contain (x, y) :param size:  Size of image capture  :return:  Returns the captured image  """
    ow, oh = img.size
    x1, y1 = pos
    tw = th = size
    #  There is enough size to intercept 
    # img.crop Coordinate representation  (left, upper, right, lower)
    if (ow > tw or oh > th):
        return img.crop((x1, y1, x1+tw, y1+th))
    return img

#  And then use transforms.Lambda Encapsulate it as transforms Strategy 
#  Then define new transforms by 
normalize = T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
data_transforms = T.Compose([
    T.Lambda(lambda img: __crop(img, (5,5), 224)),
    T.RandomHorizontalFlip(),  #  Randomly flip a given... Horizontally PIL.Image, The turnover probability is 0.5
    T.ToTensor(),  #  Turn into Tensor Format , The size range is [0,1]
    normalize
])