Add color profile to training #22
Conversation
| # This only works when the mode is 'overflow' and in 'C' order. | ||
| metadata = { | ||
| "mpp": self._mpp, | ||
| "dtype": str(batch_dtype), |
There was a problem hiding this comment.
Removing dtype from metadata causes the reader to fail in the _open_file function.
There was a problem hiding this comment.
Two further comments:
- Applying the ICC profiles during training works (tested with TCGA images and Openslide backend).
- When using the CLI functionality of
tiling.py, a type error indata.pyis raised. Simply ignoring the pylinting with a comment is not enough. I believe addingfrom __future__ import annotationsat the top would solve the problem.
There was a problem hiding this comment.
Should be fixed?
| @@ -33,18 +57,23 @@ def __init__( | |||
| tile_size: tuple[int, int], | |||
| tile_overlap: tuple[int, int], | |||
| num_samples: int, | |||
| is_binary: bool = False, | |||
| is_compressed_image: bool = False, | |||
There was a problem hiding this comment.
Should this be a bool, i.e., is the behavior completely the same irrespective of the compression type? Or should we make an Enum somewhere with the known compression types?
There was a problem hiding this comment.
It should be independent as the compression type is included in the binary blob.
| # This only works when the mode is 'overflow' and in 'C' order. | ||
| metadata = { | ||
| "mpp": self._mpp, | ||
| "dtype": str(batch_dtype), |
There was a problem hiding this comment.
Why is this removed? It is being used in readers, for example here:
Line 113 in 0ae5c1b
We could add it with first_batch.dtype if necessary.
| else: | ||
| _mode = "ARRAY" | ||
| _format = "RAW" | ||
| _num_channels = first_batch.shape[-1] |
There was a problem hiding this comment.
I think there is a typo here, should be first_batch.shape[1], right?
There was a problem hiding this comment.
I think it's correct as it's the conversion of a PIL Image to an array. If it would be a tensor it would indeed be channels first:
import PIL.Image
import numpy as np
print(np.asarray(PIL.Image.open("image.jpg")).shape)
(1635, 1966, 3)
|
@jonasteuwen.. this PR also fixes {#16}. It would be nice to add it in the PR description. |
Made this commit to make ahcore compatible with upstream changes made to dlup.
This is the PR that allows ICC profiles to be used during training.
Furthermore it also fixes #43.