21Fall\ Machine Learning Transforms

import numpy

import os

from torch.utils.data import Dataset

from PIL import Image

class Mydata(Dataset):

    def __init__(self,root_dir,label_dir):

        self.root_dir = root_dir

        self.label_dir= label_dir

        self . path = os.path.join(self.root_dir, self.label_dir)

        self.img_path = os.listdir(self.path)

    def __getitem__(self, index):

        img_name = self.img_path[index]

        img_item_path = os.path.join(self.root_dir,self.label_dir,img_name)

        img = Image.open(img_item_path)

        label = self.label_dir

        return img, label

    def __len__(self):

        return len(self.img_path)

 

 

在上述的代码之中，是把整个文件夹中的所有关于ant的图片作为一个整体，（类似于数组），

然后就可以用类似于数列的取地址的方式来进行访问了。

差不多就是自己定义了一个数据的类型，（参考struct），通过数组式的访问有图像PIL和label的tuple类型返回值

 

 

 

通过这一种方式来在tensorboard之中打开Image

Shape:

img_tensor:Defaultis:math:`(3,H,W)`.Youcanuse``torchvision.utils.make_grid()``to

convertabatchoftensorinto3xHxWformatorcall``add_images``andletusdothejob.

Tensorwith:math:`(1,H,W)`,:math:`(H,W)`,:math:`(H,W,3)`isalsosuitableaslongas

corresponding``dataformats``argumentispassed,e.g.``CHW``,``HWC``,``HW``.

transforms.py工具箱

输入：一些特定格式的图片

--> totensor   #转化为 tensor 文件

--> resize

从而实现预期的一些变化

tensor数据类型

 

'''

        img(PIL format)

>>>     via method...transforms.Resize(512,512) or transforms.Resize(512) # resize proportionately

            ...Methods Objectified from Class Resize, the function, too, can be a kind of object able to                                 be objectified

-->     img_resized(PIL format)

...     via method...transforms.ToTensor()

            ...Objectified from Class: ToTensor.

            >>> tensor_transform = transforms.ToTensor()

            >>> img_tensor = tensor_transform(img_PIL)





-->     img_tensor(Tensor format)

...     via Tensorboard to print the image      

'''

'''

        use class Compose to shorten the procedure like bellow:

        img(PIL format)



...     activate an instance of transforms.Compose Class

        compose can sum up a series of transforms objects together and operate them orderly.

            >>> trans_compose = transforms.Compose([transforms.Resize(512),transforms.ToTensor()])

            >>> img_tensor_resize = trans_compose(img)

       

-->     output: resized image in Tensor Format

'''



'''

        note::

            One significant thing: we've learnt in OOP(Object Oriented Programming) that we can objectify a class

            and created a precise instance related to the class.

            In this code, we can find out that objects can sometimes be the abstractions of a certain function. The

            function is first created and later called.

           

'''



from torchvision import transforms

from PIL import Image

from torch.utils.tensorboard import SummaryWriter



img_path = r"dataset/val/bees/2506114833_90a41c5267.jpg"

img_PIL = Image.open(img_path)

tensor_transform = transforms.ToTensor()

# Objectify the tools in the module transform

img_tensor = tensor_transform(img_PIL)

print(img_tensor)

writer = SummaryWriter("logs")

writer.add_image("bees", img_tensor)

trans_normalize = transforms.Normalize([0.5, 0.5, 0.7], [2, 2, 2])

img_norm = trans_normalize(img_tensor)

writer.add_image("new normalized", img_norm)

writer.close()