improvement: reduce setup time of AbstractWriterCallback

[YoniSchirris]

Describe the bug
When running inference, AbstractWriterCallback loops over all datasets to construct the _dataset_size dict. This opens a slide from cache several times, which can take 1-3 seconds. For a dataset of 1500 wsis this often takes 20 minutes.

To Reproduce
Run inference on-the-fly (#87) with your data_dir and glob_pattern set up to find many whole-slide images.

Expected behavior
You'll find that after printing the dataset statistics, it takes a long time to start setting up callback workers.

In my case

[2024-06-07 12:24:32,332][ahcore.data.dataset.DlupDataModule][INFO] - Dataset for stage predict has 773079 samples and the following statistics:
 - Mean: 485.30
 - Std: 145.56
 - Min: 48.00
 - Max: 1056.00
[2024-06-07 12:29:30,294][ahcore.callbacks.converters.common][INFO] - Starting worker for TiffConverterCallback

Environment
dlup version: 0.3.38
How installed: unsure
Python version: 3.11.9
Operating System: linux

Quick solution to reduce time by half;
in

ahcore/ahcore/callbacks/abstract_writer_callback.py

Line 181 in 93274e5

for current_dataset in self._total_dataset.datasets: # type: ignore

change

assert current_dataset.slide_image.identifier
self._dataset_sizes[current_dataset.slide_image.identifier] = len(current_dataset)

to

current_dataset_slide_id = current_dataset.slide_image.identifier
assert current_dataset_slide_id
self._dataset_sizes[current_dataset_slide_id] = len(current_dataset)

which will likely reduce the time by half

[YoniSchirris]

I thought about this a bit more:

• AbstractFileWriterCallback._dataset_sizes is only used internally to track size of datasets and being in the last batch or not
• AbstractFileWriterCallback._dataset_sizes keys are curnetly being set using teh slide_identifier, which requires opening the slide, which is slow
• Generally, the slide_identifier in a SlideImage is the filepath, if it is not explicitly set.
• Later in AbstractFileWriterCallback, this is even entirely expected, e.g. on line 302, we see if self._tile_counter[curr_filename] == self._dataset_sizes[curr_filename]:....
• Hence, it is currently of no value to set it to the identifier. If the identifier WOULD be set, this part of the code would break since it tries to use the filename
• Hence, it is best to simply use the _path from the dataset class to set the _dataset_sizes keys, which will be faster en not lose any unctionality.

If, in the future, we want to support identifier WITHIN this class, this can be considered a feature request that requires some more refactoring.

[YoniSchirris]

[2024-06-12 12:00:19,387][ahcore.data.dataset.DlupDataModule][INFO] - Dataset for stage predict has 773079 samples and the following statistics:
 - Mean: 485.30
 - Std: 145.56
 - Min: 48.00
 - Max: 1056.00
[2024-06-12 12:00:19,393][ahcore.callbacks.abstract_writer_callback][DEBUG] - Prediction epoch start
[2024-06-12 12:00:19,416][ahcore.callbacks.converters.common][INFO] - Starting worker for TiffConverterCallback
[2024-06-12 12:00:19,432][ahcore.callbacks.converters.common][INFO] - Starting worker for TiffConverterCallback
[2024-06-12 12:00:19,442][ahcore.callbacks.converters.common][DEBUG] - Workers started.
[2024-06-12 12:00:19,447][ahcore.callbacks.converters.common][INFO] - Starting worker for TiffConverterCallback

this fixes this slowness as seen above. Whenever the dataset is loaded, the tiffwriter is immediately ready to go and inference starts

[YoniSchirris]

fixed here b1f747e in #87