WSIReader, cucim , tifffile, performance regressions

[myron]

with pathology MIL classification tutorial, as an example
https://github.com/Project-MONAI/tutorials/blob/main/pathology/multiple_instance_learning/panda_mil_train_evaluate_pytorch_gpu.py

WSIReader is now imported from from monai.data.wsi_reader import WSIReader. If changing the import to the old way (now deprecated) from monai.data.image_reader import WSIReader, and setting backend=tifffile, then training is ~10% faster. seems like a performance regression.

Training time on 8gpu, 1 epoch

• from monai.data.wsi_reader import WSIReader, (backend=cucim) 313 seconds
• from monai.data.wsi_reader import WSIReader, (backend=tiffile) 303 seconds
• from monai.data.image_reader import WSIReader, (backend=tiffile) 283 seconds

[drbeh]

Hi @myron, thank you very much for reporting this. I have investigated this issue and realized that the root cause is an "unnecessary" array copy in the wsi_reader.WSIReader. After fixing this, wsi_reader.WSIReader seems to be faster than image_reader.WSIReader (see blow). I will submit the PR momentarily. It would be great if you can check the run time of MIL tutorial after the PR is merged. Thanks

from monai.data.wsi_reader import WSIReader
from monai.data.image_reader import WSIReader as WSIReader2

filename = "temp_CMU-1.tiff"
reader1 = WSIReader(backend="TiffFile")
reader2 = WSIReader2(backend="TiffFile")
obj1 = reader1.read(filename)
obj2 = reader2.read(filename)

%%timeit
reader1.get_data(obj1, level=1)

current performance: 2.32 s ± 97.7
after fixing the issue: 1.45 s ± 12.9 ms

%%timeit
reader2.get_data(obj2, level=1)

current performance: 1.97 s ± 50.2 ms
after fixing the issue:. 1.93 s ± 141 ms

[myron]

okay, there was a very short time window for me to test it before merging. but since it's already merged, I'll just test it a bit later, when I get to it. But thank you for debugging and fixing it!