Closed scottclowe closed 9 years ago
I already fixed this without making an issue for it, but thought it would be better to make a note of the changes in the issue log.
The input to augmentation_wrapper is called 'units'
, though perhaps it would have been better named 'dtype'
.
Many of the skimage transformations will convert a from uint8 to float64.
I'm pretty sure that the output from processing into CNN should be float, not uint8. At the moment this is happening just because we always use resize and resize always changes the image to float. We can't rely on this if we will use other shaping methods. We could end up in the situation where some images are uint8 [0,255] and others float [0,1] because some are rotated and some are not.
Therefore, all the image_processing functions should maintain the datatype of their input for consistency, and the target datatype to be returned should be specified by an input to the augmentation_wrapper.