WSI reader improvements #218
Conversation
This reverts commit 438e38c.
| @@ -79,6 +79,16 @@ def __init__(self, path: str, tilesize: int = 512): | |||
| self._writeable = False | |||
| self._erasable = False | |||
|
|
|||
| def __contains__(self, key: str): | |||
There was a problem hiding this comment.
Why is this needed here?
| @@ -36,9 +35,9 @@ def wsi( | |||
| scale_image = DataArray( | |||
| img[key].transpose("S", f"Y{suffix}", f"X{suffix}"), | |||
| dims=("c", "y", "x"), | |||
| ).chunk(chunks) | |||
| ) | |||
There was a problem hiding this comment.
Do we still have chunks (since you removed .chunk(chunks))? Do we sometimes have images without any chunks?
|
While trying to answer your questions I noticed that even if this branch passes all tests it fails when used for normal WSIs (not the small ones I used for testing). It seems that the numpy._core._exceptions._ArrayMemoryError: Unable to allocate 67.0 GiB for an array with shape (3, 45402, 66000) and data type int64 I will take a closer look at this because there is a ~7-10x drop in reading time when using sopa vs the native |
Added
slideiobackend for WSI is now available and allows loading among others the Olympus.vsiformatnn.DataParallelfor WSI inference now allows using multiple GPUs and larger batch_sizesChanged
io.wsireader (leads x10 speedup when reading tiles from the sdata)