IntToFloatTensor in the MaskBlock?

jc-denton · September 10, 2021, 9:58pm

Hi there,

I noticed that the MaskBlock also contains an IntToFloatTensor batch transformation and I wonder why it is there. Batch transforms are not applied to label masks if you build a U-Net learner with fastai (after all you want the masks as int tensors). Is this a bug or am I missing something?

def MaskBlock(codes=None):
    "A `TransformBlock` for segmentation masks, potentially with `codes`"
    return TransformBlock(type_tfms=PILMask.create, item_tfms=AddMaskCodes(codes=codes), **batch_tfms=IntToFloatTensor**)

github.com

fastai/fastai/blob/master/fastai/vision/data.py#L69

    
      
              for i in range(2):
                  ctxs[i::2] = [b.show(ctx=c, **kwargs) for b,c,_ in zip(samples.itemgot(i),ctxs[i::2],range(max_n))]
              return ctxs
          
          
# Cell
          def ImageBlock(cls=PILImage):
              "A `TransformBlock` for images of `cls`"
              return TransformBlock(type_tfms=cls.create, batch_tfms=IntToFloatTensor)
          
          
# Cell
          def MaskBlock(codes=None):
              "A `TransformBlock` for segmentation masks, potentially with `codes`"
              return TransformBlock(type_tfms=PILMask.create, item_tfms=AddMaskCodes(codes=codes), batch_tfms=IntToFloatTensor)
          
          
# Cell
          PointBlock = TransformBlock(type_tfms=TensorPoint.create, item_tfms=PointScaler)
          BBoxBlock = TransformBlock(type_tfms=TensorBBox.create, item_tfms=PointScaler, dls_kwargs = {'before_batch': bb_pad})
          
          
PointBlock.__doc__ = "A `TransformBlock` for points in an image"
          BBoxBlock.__doc__  = "A `TransformBlock` for bounding boxes in an image"

bilalUWE · July 20, 2022, 7:22pm

Just found the need for using the batch_tfms with masks. If your masks are binary with two classes say background and foreground denoted by pixel values of 0 and 255, you will need to change masks to only contain 0 and 1 for pixel values denoting the segmentation classes. Transforms like IntToFloatTensor for masks are life savers. All you need to do is to add IntToFloatTensor(div_mask=255) to your batch_tfms and you are all set. One use case for performing batch transformation on masks.