Resizing? Any simple direct way?

1

Hello, is there a simple way, to resize an image? For example from (256,256) to (244,244)?
I looked at this thread Autogradable image resize and used the AvgPool2 method, but it seems quiet complicated for me, to resize an image from 256p to 244…
I am sitting on this problem for a quiet long time now…and I don’t find a way to fix it. Some help would be very appreciated!
thanks!

2 Likes
2

Hello,

You can use Torchvision transforms on your images. Pretty straight forward.

http://pytorch.org/docs/master/torchvision/transforms.html

3

Sorry, that is a quiet low level question, but I am quiet new to pytroch, but I tried resizing my image for ours now, also with .Resize.
Could you give me an simple example, how you use that? thanks!

4

The Data Loading and Processing Tutorial includes a session on image transforms.

I believe it would be best to follow/understand the tutorial than for me to try to summarize it.

3 Likes
5

I think an example, on how to simply apply a rescaling on my image would help me very much with understanding all that

6

First of all as Antonio stated read the transforms documentation it will teach you much more. Below is a simplified version of what is explained in the tutorials.

from skimage import io, transform

image = io.imread(PATH_TO_THE_IMAGE)
img = transform.resize(image, (desired_h, desired_w))

3 Likes
7

I use adaptive average pooling for resizing if I can / want move resizing step to GPU:

import torch
import torch.nn.functional as F
from torch.autograd import Variable

def resize2d(img, size):
    return (F.adaptive_avg_pool2d(Variable(img,volatile=True), size)).data

img = torch.Tensor(3,256,256).uniform_()
print(img.size())
img = resize2d(img, (224,224))
print(img.size())
8 Likes
8 
def resize_tensor(input_tensors, h, w):
  final_output = None
  batch_size, channel, height, width = input_tensors.shape
  input_tensors = torch.squeeze(input_tensors, 1)
  
  for img in input_tensors:
    img_PIL = transforms.ToPILImage()(img)
    img_PIL = torchvision.transforms.Resize([h,w])(img_PIL)
    img_PIL = torchvision.transforms.ToTensor()(img_PIL)
    if final_output is None:
      final_output = img_PIL
    else:
      final_output = torch.cat((final_output, img_PIL), 0)
  final_output = torch.unsqueeze(final_output, 1)
  return final_output

You have to convert to PIL and use transform functions there. Also, transform functions don’t support batch so I wrote a helper function. Hope this helps.

3 Likes
10

I would use this wonderful repo ResizeRight

11

Hi Antonio, the link is not working.

12

On the same lines, here’s a modified version directly using the transforms class. I lifted it directly from my training code.

transform = transforms.Compose([transforms.Resize((256, 256)), transforms.ToTensor(), 
                               transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
img = Image.open(os.path.join(input_path, file)) #PIL image
img = transform(img) #torch.FloatTensor