Image transforms

[geekswaroop]

Image transforms

[Shree987]

@geekswaroop Should the original image also be changed, or only resizing the input image before converting to array is sufficient?

[geekswaroop]

What's the difference between original image and input image?

[Shree987]

Input image could be the resized image (resized only once, not each time we run the code)

[Shree987]

I think PIL.Image.resize() does the work. Can you cross-check once if that is what is needed? @geekswaroop

[geekswaroop]

PIL functions don't work well with torch Tensors. Refer to the documentation linked above

[Shree987]

@geekswaroop Why should the transformation be applied to tensors? Can't we directly do it for images and save them as part of preprocessing?
Doing for tensors might be hectic as each time the code is run, tensors are changed .

[geekswaroop]

Hi, I wasn't clear in my description. I want you to compose 2 transformations for both ground truth and images.

Basically, return a dictionary containing keys as img and gt whose corresponding values in the composition object whose documentation I have linked above.

For the transformation object, you first need to resize the image to 256x256 with 3 channels. Then convert to tensor. Next, normalise the image

For ground truth, just resize image to 256x256 with 1 channel.

Hope this is clear. Send in a PR, I will review it. Thanks!