I want to reshape like reshaping an image. Like 32 * 32 to 16 * 16. (By using torch.tensor) Like decreasing the resolution? Can anyone help me?

CodePudding user response：

If you have an image (stored in a tensor) and you want to decrease it's resolution, then you are not reshaping it, but rather resizing it.
To that end, you can use pytorch's interpolate:

import torch
from torch.nn import functional as nnf

y = nnf.interpolate(x[None, None, ...], size=(16, 16), mode='bicubic', align_corners=False, antialias=True)

Notes:

1. nnf.interpolate operates on batches of multi-channel images, that is, it expects its input x to have 4 dimensions: batch-channels-height-width. So, if your x is a single image with a single channel (e.g., an MNIST digit) you'll have to create a singleton batch dimension and a singleton channel dimension.
2. Pay close attention to align_corners and antialias -- make sure you are using the right configuration for your needs.

For more information regarding aliasing and alignment when resizing images you can look at ResizeRight.

CodePudding user response：

you can use resize from torchvision transforms:

import torchvision.transform as T
T.Resize(16)(your_image)

you can also chain it to other transforms

preprocess = T.Compose([
   T.Resize(N),
   T.CenterCrop(N),
   T.ToTensor(),
   T.Normalize(
       mean=mean,
       std=std
   )
])
preprocess(your_image)