torch.uint8 instead of torch.float32 after data loader

[greenyjwk]

Hello,

Thank you for your great work.

Q1
I have a quick question about the tensor type after data_loader.

I guess the main train Python script is default_trainer.py (BiomedParse/trainer/default_trainer.py)

So it printed out the tensor right after dataloader(), which is line #243.
for batch_idx, batch in enumerate(self.train_dataloaders):

And when I check the data type, it was torch.uint8, not torch.float32, which I expected.

Was it from my mistake, or is it meant to have a torch.uint8 instead of torch.float32. If it torch.uint8 on purpose, would you explain the reason why it is?

Q2
And also, I want to take a look at how the dataloader looks. Which Python scripts implement the dataloader?

Thank you.

[theodore-zhao]

Hi, thanks for the question! The data type in uint8 is because we loaded from the original file in PNG, which is efficient is storage. When going into the model pytorch modules, it needs to be converted to floating points. You may find the dataloader here: https://github.com/microsoft/BiomedParse/blob/main/datasets/dataset_mappers/biomed_dataset_mapper.py

Hope this helps!