Best Practice: How to handle input data with multiple resolutions?

vision
1

Hi, I am looking for some information regarding the best practice in PyTorch for data inputs into a model.

Problem description:
I work with satellite channels of different resolutions. Because of this one input may be half the size of another input (i.e. a 0.5 km resolution channel vs. a 1 km resolution channel). When I access the data in a generator, I load a tensor of inputs from a netcdf as well as a tensor targets (i.e. x and y). In order to account for the difference of resolutions, I have previously used numpy to created a concatenated array of inputs and an initial upsampling step within my U-Net model (mode nearest). However, I want to get away from using numpy, instead using Torch tensors to help reduce i/o. Correct me if I am wrong, but I believe the Torch tensors need the data to be the same size for a Torch tensor with torch.cat(). I found the function torch.nn.functional.interpolate() which could be implemented before running the model (i.e. in the data generator), but I am not sure if it is proper to do this.

Thus, in the spirit of best practice, what is the best way to deal with data of multiple sizes?

Disclaimer:
I am just switching to PyTorch from Keras for the customizability with regards to GPU use, so forgive me if this is rather obvious.

2

You can think of torch tensors as np arrays.
I’m afraid tensors need to be same size to be concatenated of stacked.
Using numpy is a good practice as tensors can reuse the same memory address.
Usually data is padded. Pytorch doesn’t have a very optimized IO or there is no special tool.

If you want optimized IO you can try nvidia’s DALI.