Binary Unet Segmentation **Without Metadata file**

Hello Experts,

I am new to fastai and machine learning. I just started learning some weeks ago. I have gone through most of the lessons in the fastbook.

I have a problem I am trying to solve: Image segmentation.
I am trying to apply the DataBlock Api or the SegmentationDataloaders for the extraction of building footprints from satelite imagery. I get errors when I donot pprovide the codes argument in the SegmentationDataloaders() such as Could not do one pass in yor dataloader, ther is something wrong in it and when I try calling dls.one_batch() I get RuntimeError: cuDNN error: CUDNN_STATUS_NOT_INITIALIZED .

I would like to write my own custom dataloder which doesn’t require the code argument, because the training data does not contain and metadata file. I have not found an example implementation where the dataset doesn’t have one.

Can anyone here help me out with some guide or an example?


Are you running it in Colab? If yes, did you change your runtime to gpu? This looks like a CUDA error.