Here is a small demo of the problem.
import torch
import torchvision
import time
from PIL import Image
ToTensor=torchvision.transforms.ToTensor()
start_time=time.time()
for i in range(10000):
img=Image.open('out.png')
img=ToTensor(img)
print ('time:',time.time()-start_time)
V1.0 takes more time than 0.3.1 (48s vs. 39s)