🩹 Support for np.ndarray and WSIReader in PatchPredictor

[blaginin]

This PR fixes #532 and improves PatchPredictor readability and intuitiveness.

What has changed

• PatchPredictor can now work with file paths and NumPy arrays, as declared in the docs. Now it's possible to do this:

>>> predictor = PatchPredictor(pretrained_model="resnet18-kather100k", batch_size=32)
>>> normalizer = get_normalizer("reinhard")
>>> normalizer.fit(wsi.slide_thumbnail())

>>> predictor.predict(
...    input_imgs=[normalizer.transform(tile.slide_thumbnail(1, "baseline"))],  # 👀
...    mode="tile",
...    merge_predictions=True,
...    patch_input_shape=[224, 224],
...    stride_shape=[224, 224],
...    resolution=1,
...    units="baseline",
...    return_probabilities=True,
...    on_gpu=False
... )

• Transformation to baseline ioconfig is now explicit. predict function used to disregard user-defined arguments, significantly impacting how the data is processed. Although it sent a warning, this warning can easily drown among the messages from pytorch/torchvision/wsireader/... especially when processing many files at once. Users' direct instructions should be changed only if we are 100% sure about the results. Since it's not the case here, instead of a warning user will receive an exception:

>>> W, H = 100, 200
>>> predictor = PatchPredictor(pretrained_model="resnet18-kather100k", batch_size=32)
>>> data = (np.arange(W * H * 3).reshape((W, H, 3)) % 255).astype(np.uint8)
>>> reader = WSIReader.open(data)  # we don't know MPP of the data

>>> predictor.predict(
...    input_imgs=[reader],
...    mode="wsi",
...    merge_predictions=True,
...    patch_input_shape=[224, 224],
...    stride_shape=[224, 224],
...    resolution=1,
...    units="mpp",  # 👀
...    return_probabilities=True,
...    on_gpu=False
... )

ValueError: MPP is None. Cannot determine scale in terms of MPP.

I also refactored the code:

• WSIPatchDataset and VirtualWSIReader both operated on image-level data via OpenCV. Duplicated logic is removed, and now WSIPatchDataset works with images through WSIReader abstraction.
• Removed a magic number.

Depends on #583

[codecov]

Codecov Report

Merging #576 (11674c4) into develop (1c42e39) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff            @@
##           develop     #576   +/-   ##
========================================
  Coverage    99.77%   99.77%           
========================================
  Files           63       63           
  Lines         6782     6794   +12     
  Branches      1352     1362   +10     
========================================
+ Hits          6767     6779   +12     
  Misses           7        7           
  Partials         8        8           

Impacted Files	Coverage Δ
tiatoolbox/models/dataset/classification.py	100.00% <100.00%> (ø)
tiatoolbox/models/dataset/dataset_abc.py	97.14% <100.00%> (+0.04%)	⬆️
tiatoolbox/models/engine/patch_predictor.py	100.00% <100.00%> (ø)
tiatoolbox/tools/patchextraction.py	100.00% <100.00%> (ø)
tiatoolbox/wsicore/wsireader.py	99.43% <100.00%> (+<0.01%)	⬆️

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[shaneahmed]

Hi @blaginin, Please can you break this down into smaller PRs? e.g., introducing blank sample can be a separate PR. Fixing documentation and support for np.ndarry can be separated as well. It will help with review.

[blaginin]

Hi @blaginin, Please can you break this down into smaller PRs? e.g., introducing blank sample can be a separate PR. Fixing documentation and support for np.ndarry can be separated as well. It will help with review.

@shaneahmed excluded extra stuff, as we discussed

[shaneahmed]

Added some initial comments.

[shaneahmed]

This is wsidicom dependency.

[blaginin]

Now requirements.txt is identical to develop

[shaneahmed]

Please do not update these as these need to be consistent with SemanticSegmentor and other engines. This will be fixed in redesigned engines #578

[shaneahmed]

These are minor edits which can be done in a separate PR.

[blaginin]

Moved to #598

[shaneahmed]

Same as above. These kind of edits can be suggested in a separate PR.

[blaginin]

Moved to #598

[shaneahmed]

#635 has restructured EngineABC to support numpy arrays, WSI and tile support will be added around this later. Is this what you intended with this PR? You can decouple changes to WSIPatchDataset and push those changes in a separate PR.