unexpected data type conversion

[jxchen01]

I am looking at this segmentation tutorial: https://github.com/Project-MONAI/tutorials/blob/main/2d_segmentation/torch/unet_training_dict.py

It looks like the data is in uint8 (see: https://github.com/Project-MONAI/tutorials/blob/main/2d_segmentation/torch/unet_training_dict.py#L50)

But, after transformation, it becomes torch.float32 (I manually checked at this step: https://github.com/Project-MONAI/tutorials/blob/main/2d_segmentation/torch/unet_training_dict.py#L131)

I am curious how is this data conversion happening? Can we choose to not converting?

[Nic-Ma]

Hi @jxchen01 ,

I think maybe ScaleIntensityd transform converted it to float32?

Thanks.

[jxchen01]

@Nic-Ma , that was my guess, too. But, even I comment out all the transforms, except the LoadImage transform, the data returned from the dataloader is still torch.float32 ... Super strange to me ...

[jxchen01]

I guess this might be the root cause: https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/io/array.py#L102

I am not sure if this is good for LoadImage. In my opinion, the default behavior should be keeping the original data type instead of converting to float32