ToTensor - torchvision.transforms.ToTensor

 

torchvision.transforms는 파이토치의 이미지 전처리 및 증강을 위한 모듈.

그 중에서 ToTensor는 PIL image나 numpy array를 tensor화 시켜주는 클래스.

변환 후 Channel x Height x Width로 변경

 

ToTensor — Torchvision main documentation (pytorch.org)

 

 

하단의 이미지를 바탕으로 ToTensor의 적용 후 결과를 확인.

 

https://commons.wikimedia.org/wiki/File:Cloud_with_long_silver_lining.jpg#/media/File:Cloud_with_long_silver_lining.jpg\

File:Cloud with long silver lining.jpg - Wikimedia Commons

 

 

from torchvision.transforms import ToTensor
from PIL import Image
import numpy as np

img  = Image.open("/content/silver_lining.jpg")
img_array  = np.array(img)

transform = ToTensor()

# ToTensor : turn an img to tensor, 픽셀 값 0~255 -> 0~1으로 변환
img_tensor = transform(img)
print(img_tensor) 
print(img_tensor.size()) # Channel x Height x Width
print(img_tensor.min(), img_tensor.max()) # 0~1


'''
tensor([[[0.4235, 0.4235, 0.4235,  ..., 0.2627, 0.2667, 0.2667],
         [0.4275, 0.4275, 0.4275,  ..., 0.2627, 0.2667, 0.2667],
         [0.4275, 0.4275, 0.4275,  ..., 0.2667, 0.2667, 0.2667],
         ...,
         [0.0157, 0.0118, 0.0118,  ..., 0.0941, 0.0902, 0.0863],
         [0.0196, 0.0157, 0.0118,  ..., 0.0902, 0.0902, 0.0902],
         [0.0235, 0.0196, 0.0157,  ..., 0.0980, 0.0980, 0.0980]],

        [[0.5922, 0.5922, 0.5922,  ..., 0.4275, 0.4314, 0.4314],
         [0.5922, 0.5922, 0.5922,  ..., 0.4275, 0.4314, 0.4314],
         [0.5922, 0.5922, 0.5922,  ..., 0.4314, 0.4314, 0.4314],
         ...,
         [0.0078, 0.0039, 0.0039,  ..., 0.1373, 0.1333, 0.1294],
         [0.0118, 0.0078, 0.0039,  ..., 0.1333, 0.1333, 0.1333],
         [0.0157, 0.0118, 0.0078,  ..., 0.1294, 0.1294, 0.1294]],

        [[0.7569, 0.7569, 0.7569,  ..., 0.6235, 0.6275, 0.6275],
         [0.7569, 0.7569, 0.7569,  ..., 0.6235, 0.6275, 0.6275],
         [0.7569, 0.7569, 0.7569,  ..., 0.6275, 0.6275, 0.6275],
         ...,
         [0.0118, 0.0078, 0.0078,  ..., 0.2078, 0.2039, 0.2000],
         [0.0157, 0.0118, 0.0078,  ..., 0.2039, 0.2039, 0.2039],
         [0.0196, 0.0157, 0.0118,  ..., 0.2039, 0.2039, 0.2039]]])
채널 x 높이 x 너비 -> torch.Size([3, 4016, 6016]) 
min: tensor(0.) max: tensor(1.)
'''