Resizing Seismic Data for deep learning

I am trying to build a UNET model for some fault prediction using seismic data. So, my training datasets contains 2 sets:
A seismic and fault mask: Both are 3D volumes with following shapes:101, 589, 751
Here, 101 is number of inlines, 589 is number of xlines
and 751 is vertical sample.
So, the idea is to treat the volume as stack of 2D arrays, in either xline or inline direction.

There is no RGB, so we treat them as single channel.
i.e, along Inline direction: shape : 589 x 751 and there is 101 inlines stacked. Or we can look into xline in similar manner.

Now, the main issue I am having is to resize this into proper sample intervals of 256, which I believe is required for U-Net.
Can anyone help me how I can achieve this with pytorch transforms.
Thanks and much appreciated.

Below is an example of way the data is loaded:

You could pad your volume with F.pad. I am not familiar with the seismic data, so you would have to see the padding value and make sure it represents the background. For example in image segmentation I always pad with zeros giving it a thick black boundary which can be considered as the background.
You could read about F.pad here- torch.nn.functional — PyTorch 1.8.1 documentation