test_models.py

"""Tests for code related to model usage."""

from time import time
import os
import pathlib
import shutil
import cv2
import numpy as np
import pytest
import torch

from tiatoolbox import rcParam
from tiatoolbox.models.abc import ModelABC
from tiatoolbox.models.backbone import get_model
from tiatoolbox.models.classification import CNNPatchModel, CNNPatchPredictor
from tiatoolbox.models.dataset import (
    DatasetInfoABC,
    PatchDatasetABC,
    KatherPatchDataset,
    PatchDataset,
    WSIPatchDataset,
    predefined_preproc_func,
)
from tiatoolbox.utils.misc import download_data, unzip_data, imread, imwrite
from tiatoolbox.wsicore.wsireader import get_wsireader


ON_GPU = False


def _get_temp_folder_path():
    """Return unique temp folder path"""
    new_dir = os.path.join(
        rcParam["TIATOOLBOX_HOME"], f"test_model_patch_{int(time())}"
    )
    return new_dir


def test_create_backbone():
    """Test for creating backbone."""
    backbones = [
        "alexnet",
        "resnet18",
        "resnet34",
        "resnet50",
        "resnet101",
        "resnext50_32x4d",
        "resnext101_32x8d",
        "wide_resnet50_2",
        "wide_resnet101_2",
        "densenet121",
        "densenet161",
        "densenet169",
        "densenet201",
        "googlenet",
        "mobilenet_v2",
        "mobilenet_v3_large",
        "mobilenet_v3_small",
    ]
    for backbone in backbones:
        try:
            get_model(backbone, pretrained=False)
        except ValueError:
            raise AssertionError(f"Model {backbone} failed.")

    # test for model not defined
    with pytest.raises(ValueError, match=r".*not supported.*"):
        get_model("secret_model-kather100k", pretrained=False)


def test_DatasetInfo():
    """Test for kather patch dataset."""
    # test defining a subclass of dataset info but not defining
    # enforcing attributes - should crash
    with pytest.raises(TypeError):

        # intentionally create to check error
        # skipcq
        class Proto(DatasetInfoABC):
            def __init__(self):
                self.a = "a"

        # intentionally create to check error
        Proto()  # skipcq
    with pytest.raises(TypeError):

        # intentionally create to check error
        # skipcq
        class Proto(DatasetInfoABC):
            def __init__(self):
                self.inputs = "a"

        # intentionally create to check error
        Proto()  # skipcq
    with pytest.raises(TypeError):
        # intentionally create to check error
        # skipcq
        class Proto(DatasetInfoABC):
            def __init__(self):
                self.inputs = "a"
                self.labels = "a"

        # intentionally create to check error
        Proto()  # skipcq
    with pytest.raises(TypeError):

        # intentionally create to check error
        # skipcq
        class Proto(DatasetInfoABC):
            def __init__(self):
                self.inputs = "a"
                self.label_names = "a"

        # intentionally create to check error
        Proto()  # skipcq
    # test kather with default init
    dataset = KatherPatchDataset()
    # kather with default data path skip download
    dataset = KatherPatchDataset()
    # pytest for not exist dir
    with pytest.raises(
        ValueError,
        match=r".*not exist.*",
    ):
        _ = KatherPatchDataset(save_dir_path="unknown_place")

    # save to temporary location
    save_dir_path = _get_temp_folder_path()
    # remove previously generated data
    if os.path.exists(save_dir_path):
        shutil.rmtree(save_dir_path, ignore_errors=True)
    url = (
        "https://zenodo.org/record/53169/files/"
        "Kather_texture_2016_image_tiles_5000.zip"
    )
    save_zip_path = os.path.join(save_dir_path, "Kather.zip")
    download_data(url, save_zip_path)
    unzip_data(save_zip_path, save_dir_path)
    extracted_dir = os.path.join(save_dir_path, "Kather_texture_2016_image_tiles_5000/")
    dataset = KatherPatchDataset(save_dir_path=extracted_dir)
    assert dataset.inputs is not None
    assert dataset.labels is not None
    assert dataset.label_names is not None
    assert len(dataset.inputs) == len(dataset.labels)

    # to actually get the image, we feed it to PatchDataset
    actual_ds = PatchDataset(dataset.inputs, dataset.labels)
    sample_patch = actual_ds[100]
    assert isinstance(sample_patch["image"], np.ndarray)
    assert sample_patch["label"] is not None

    # remove generated data
    shutil.rmtree(save_dir_path, ignore_errors=True)
    shutil.rmtree(rcParam["TIATOOLBOX_HOME"])


def test_PatchDatasetpath_imgs(_sample_patch1, _sample_patch2):
    """Test for patch dataset with a list of file paths as input."""
    size = (224, 224, 3)

    dataset = PatchDataset([pathlib.Path(_sample_patch1), pathlib.Path(_sample_patch2)])

    for _, sample_data in enumerate(dataset):
        sampled_img_shape = sample_data["image"].shape
        assert (
            sampled_img_shape[0] == size[0]
            and sampled_img_shape[1] == size[1]
            and sampled_img_shape[2] == size[2]
        )


def test_PatchDatasetlist_imgs():
    """Test for patch dataset with a list of images as input."""
    size = (5, 5, 3)
    img = np.random.randint(0, 255, size=size)
    list_imgs = [img, img, img]
    dataset = PatchDataset(list_imgs)

    dataset.preproc_func = lambda x: x

    for _, sample_data in enumerate(dataset):
        sampled_img_shape = sample_data["image"].shape
        assert (
            sampled_img_shape[0] == size[0]
            and sampled_img_shape[1] == size[1]
            and sampled_img_shape[2] == size[2]
        )

    # test for changing to another preproc
    dataset.preproc_func = lambda x: x - 10
    item = dataset[0]
    assert np.sum(item["image"] - (list_imgs[0] - 10)) == 0

    # test for loading npy
    save_dir_path = _get_temp_folder_path()
    # remove previously generated data
    if os.path.exists(save_dir_path):
        shutil.rmtree(save_dir_path, ignore_errors=True)
    os.makedirs(save_dir_path)
    np.save(
        os.path.join(save_dir_path, "sample2.npy"), np.random.randint(0, 255, (4, 4, 3))
    )
    imgs = [
        os.path.join(save_dir_path, "sample2.npy"),
    ]
    _ = PatchDataset(imgs)
    assert imgs[0] is not None
    # test for path object
    imgs = [
        pathlib.Path(os.path.join(save_dir_path, "sample2.npy")),
    ]
    _ = PatchDataset(imgs)
    shutil.rmtree(save_dir_path)


def test_PatchDatasetarray_imgs():
    """Test for patch dataset with a numpy array of a list of images."""
    size = (5, 5, 3)
    img = np.random.randint(0, 255, size=size)
    list_imgs = [img, img, img]
    labels = [1, 2, 3]
    array_imgs = np.array(list_imgs)

    # test different setter for label
    dataset = PatchDataset(array_imgs, labels=labels)
    an_item = dataset[2]
    assert an_item["label"] == 3
    dataset = PatchDataset(array_imgs, labels=None)
    an_item = dataset[2]
    assert "label" not in an_item

    dataset = PatchDataset(array_imgs)
    for _, sample_data in enumerate(dataset):
        sampled_img_shape = sample_data["image"].shape
        assert (
            sampled_img_shape[0] == size[0]
            and sampled_img_shape[1] == size[1]
            and sampled_img_shape[2] == size[2]
        )


def test_PatchDataset_crash():
    """Test to make sure patch dataset crashes with incorrect input."""
    # all below examples below should fail when input to PatchDataset

    # not supported input type
    imgs = {"a": np.random.randint(0, 255, (4, 4, 4))}
    with pytest.raises(
        ValueError, match=r".*Input must be either a list/array of images.*"
    ):
        _ = PatchDataset(imgs)

    # ndarray of mixed dtype
    imgs = np.array([np.random.randint(0, 255, (4, 5, 3)), "Should crash"])
    with pytest.raises(ValueError, match="Provided input array is non-numerical."):
        _ = PatchDataset(imgs)

    # ndarrays of NHW images
    imgs = np.random.randint(0, 255, (4, 4, 4))
    with pytest.raises(ValueError, match=r".*array of images of the form NHWC.*"):
        _ = PatchDataset(imgs)

    # list of ndarrays with different sizes
    imgs = [
        np.random.randint(0, 255, (4, 4, 3)),
        np.random.randint(0, 255, (4, 5, 3)),
    ]
    with pytest.raises(ValueError, match="Images must have the same dimensions."):
        _ = PatchDataset(imgs)

    # list of ndarrays with HW and HWC mixed up
    imgs = [
        np.random.randint(0, 255, (4, 4, 3)),
        np.random.randint(0, 255, (4, 4)),
    ]
    with pytest.raises(
        ValueError, match="Each sample must be an array of the form HWC."
    ):
        _ = PatchDataset(imgs)

    # list of mixed dtype
    imgs = [np.random.randint(0, 255, (4, 4, 3)), "you_should_crash_here", 123, 456]
    with pytest.raises(
        ValueError,
        match="Input must be either a list/array of images or a list of "
        "valid image paths.",
    ):
        _ = PatchDataset(imgs)

    # list of mixed dtype
    imgs = ["you_should_crash_here", 123, 456]
    with pytest.raises(
        ValueError,
        match="Input must be either a list/array of images or a list of "
        "valid image paths.",
    ):
        _ = PatchDataset(imgs)

    # list not exist paths
    with pytest.raises(
        ValueError,
        match=r".*valid image paths.*",
    ):
        _ = PatchDataset(["img.npy"])

    # ** test different extenstion parser
    # save dummy data to temporary location
    save_dir_path = _get_temp_folder_path()
    # remove prev generated data
    if os.path.exists(save_dir_path):
        shutil.rmtree(save_dir_path, ignore_errors=True)
    os.makedirs(save_dir_path)
    torch.save({"a": "a"}, os.path.join(save_dir_path, "sample1.tar"))
    np.save(
        os.path.join(save_dir_path, "sample2.npy"), np.random.randint(0, 255, (4, 4, 3))
    )

    imgs = [
        os.path.join(save_dir_path, "sample1.tar"),
        os.path.join(save_dir_path, "sample2.npy"),
    ]
    with pytest.raises(
        ValueError,
        match=r"Can not load data of .*",
    ):
        _ = PatchDataset(imgs)
    shutil.rmtree(rcParam["TIATOOLBOX_HOME"])

    # preproc func for not defined dataset
    with pytest.raises(
        ValueError,
        match=r".* preprocessing .* does not exist.",
    ):
        predefined_preproc_func("secret-dataset")


def test_WSIPatchDataset(_sample_wsi_dict):
    """A test for creation and bare output."""
    # convert to pathlib Path to prevent wsireader complaint
    _mini_wsi_svs = pathlib.Path(_sample_wsi_dict["wsi2_4k_4k_svs"])
    _mini_wsi_jpg = pathlib.Path(_sample_wsi_dict["wsi2_4k_4k_jpg"])
    _mini_wsi_msk = pathlib.Path(_sample_wsi_dict["wsi2_4k_4k_msk"])

    def reuse_init(img_path=_mini_wsi_svs, **kwargs):
        """Testing function."""
        return WSIPatchDataset(img_path=_mini_wsi_svs, **kwargs)

    def reuse_init_wsi(**kwargs):
        """Testing function."""
        return reuse_init(mode="wsi", **kwargs)

    # test for ABC validate
    with pytest.raises(
        ValueError, match=r".*inputs should be a list of patch coordinates.*"
    ):
        # intentionally create to check error
        # skipcq
        class Proto(PatchDatasetABC):
            def __init__(self):
                super().__init__()
                self.inputs = "CRASH"
                self._check_input_integrity("wsi")

            # skipcq
            def __getitem__(self, idx):
                pass

        # intentionally create to check error
        Proto()  # skipcq

    # invalid path input
    with pytest.raises(ValueError, match=r".*`img_path` must be a valid file path.*"):
        WSIPatchDataset(
            img_path="aaaa",
            mode="wsi",
            patch_size=[512, 512],
            stride_size=[256, 256],
            auto_get_mask=False,
        )

    # invalid mask path input
    with pytest.raises(ValueError, match=r".*`mask_path` must be a valid file path.*"):
        WSIPatchDataset(
            img_path=_mini_wsi_svs,
            mask_path="aaaa",
            mode="wsi",
            patch_size=[512, 512],
            stride_size=[256, 256],
            resolution=1.0,
            units="mpp",
            auto_get_mask=False,
        )

    # invalid mode
    with pytest.raises(ValueError):
        reuse_init(mode="X")

    # invalid patch
    with pytest.raises(ValueError):
        reuse_init()
    with pytest.raises(ValueError):
        reuse_init_wsi(patch_size=[512, 512, 512])
    with pytest.raises(ValueError):
        reuse_init_wsi(patch_size=[512, "a"])
    with pytest.raises(ValueError):
        reuse_init_wsi(patch_size=512)
    # invalid stride
    with pytest.raises(ValueError):
        reuse_init_wsi(patch_size=[512, 512], stride_size=[512, "a"])
    with pytest.raises(ValueError):
        reuse_init_wsi(patch_size=[512, 512], stride_size=[512, 512, 512])
    # negative
    with pytest.raises(ValueError):
        reuse_init_wsi(patch_size=[512, -512], stride_size=[512, 512])
    with pytest.raises(ValueError):
        reuse_init_wsi(patch_size=[512, 512], stride_size=[512, -512])

    # * for wsi
    # dummy test for analysing the output
    # stride and patch size should be as expected
    patch_size = [512, 512]
    stride_size = [256, 256]
    ds = reuse_init_wsi(
        patch_size=patch_size,
        stride_size=stride_size,
        resolution=1.0,
        units="mpp",
        auto_get_mask=False,
    )
    reader = get_wsireader(_mini_wsi_svs)
    # tiling top to bottom, left to right
    ds_roi = ds[2]["image"]
    step_idx = 2  # manually calibrate
    start = (step_idx * stride_size[1], 0)
    end = (start[0] + patch_size[0], start[1] + patch_size[1])
    rd_roi = reader.read_bounds(
        start + end, resolution=1.0, units="mpp", coord_space="resolution"
    )
    correlation = np.corrcoef(
        cv2.cvtColor(ds_roi, cv2.COLOR_RGB2GRAY).flatten(),
        cv2.cvtColor(rd_roi, cv2.COLOR_RGB2GRAY).flatten(),
    )
    assert ds_roi.shape[0] == rd_roi.shape[0]
    assert ds_roi.shape[1] == rd_roi.shape[1]
    assert np.min(correlation) > 0.9, correlation

    # test creation with auto mask gen and input mask
    ds = reuse_init_wsi(
        patch_size=patch_size,
        stride_size=stride_size,
        resolution=1.0,
        units="mpp",
        auto_get_mask=True,
    )
    assert len(ds) > 0
    ds = WSIPatchDataset(
        img_path=_mini_wsi_svs,
        mask_path=_mini_wsi_msk,
        mode="wsi",
        patch_size=[512, 512],
        stride_size=[256, 256],
        auto_get_mask=False,
        resolution=1.0,
        units="mpp",
    )
    negative_mask = imread(_mini_wsi_msk)
    negative_mask = np.zeros_like(negative_mask)
    imwrite("negative_mask.png", negative_mask)
    with pytest.raises(ValueError, match=r".*No coordinate remain after tiling.*"):
        ds = WSIPatchDataset(
            img_path=_mini_wsi_svs,
            mask_path="negative_mask.png",
            mode="wsi",
            patch_size=[512, 512],
            stride_size=[256, 256],
            auto_get_mask=False,
            resolution=1.0,
            units="mpp",
        )
    shutil.rmtree("negative_mask.png", ignore_errors=True)

    # * for tile
    reader = get_wsireader(_mini_wsi_jpg)
    tile_ds = WSIPatchDataset(
        img_path=_mini_wsi_jpg,
        mode="tile",
        patch_size=patch_size,
        stride_size=stride_size,
        auto_get_mask=False,
    )
    step_idx = 3  # manually calibrate
    start = (step_idx * stride_size[1], 0)
    end = (start[0] + patch_size[0], start[1] + patch_size[1])
    roi2 = reader.read_bounds(
        start + end, resolution=1.0, units="baseline", coord_space="resolution"
    )
    roi1 = tile_ds[3]["image"]  # match with step_idx
    correlation = np.corrcoef(
        cv2.cvtColor(roi1, cv2.COLOR_RGB2GRAY).flatten(),
        cv2.cvtColor(roi2, cv2.COLOR_RGB2GRAY).flatten(),
    )
    assert roi1.shape[0] == roi2.shape[0]
    assert roi1.shape[1] == roi2.shape[1]
    assert np.min(correlation) > 0.9, correlation


def test_PatchDataset_abc():
    # test missing definition for abstract
    with pytest.raises(TypeError):

        # intentionally create to check error
        # skipcq
        class Proto(PatchDatasetABC):
            # skipcq
            def __init__(self):
                super().__init__()

        # crash due to not define __getitem__
        Proto()  # skipcq

    # skipcq
    class Proto(PatchDatasetABC):
        # skipcq
        def __init__(self):
            super().__init__()

        # skipcq
        def __getitem__(self, idx):
            pass

    ds = Proto()  # skipcq

    # test setter and getter
    assert ds.preproc_func(1) == 1
    ds.preproc_func = lambda x: x - 1
    assert ds.preproc_func(1) == 0
    assert ds.preproc(1) == 1, "Must be unchanged!"
    ds.preproc_func = None
    assert ds.preproc_func(2) == 2

    # test assign uncallable to preproc_func/postproc_func
    with pytest.raises(ValueError, match=r".*callable*"):
        ds.preproc_func = 1


def test_model_abc():
    """Test API in model ABC."""
    # test missing definition for abstract
    with pytest.raises(TypeError):

        # intentionally create to check error
        # skipcq
        class Proto(ModelABC):
            # skipcq
            def __init__(self):
                super().__init__()

        # crash due to not define forward and infer_batch
        Proto()  # skipcq

    with pytest.raises(TypeError):

        # intentionally create to check error
        # skipcq
        class Proto(ModelABC):
            # skipcq
            def __init__(self):
                super().__init__()

            @staticmethod
            # skipcq
            def infer_batch():
                pass

        # crash due to not define forward
        Proto()  # skipcq

    # intentionally create to check error
    # skipcq
    class Proto(ModelABC):
        # skipcq
        def __init__(self):
            super().__init__()

        @staticmethod
        def postproc(image):
            return image - 2

        # skipcq
        def forward(self):
            pass

        @staticmethod
        # skipcq
        def infer_batch():
            pass

    model = Proto()  # skipcq
    # test assign uncallable to preproc_func/postproc_func
    with pytest.raises(ValueError, match=r".*callable*"):
        model.postproc_func = 1
    with pytest.raises(ValueError, match=r".*callable*"):
        model.preproc_func = 1

    # test setter/getter/inital of preproc_func/postproc_func
    assert model.preproc_func(1) == 1
    model.preproc_func = lambda x: x - 1
    assert model.preproc_func(1) == 0
    assert model.preproc(1) == 1, "Must be unchanged!"
    assert model.postproc(1) == -1, "Must be unchanged!"
    model.preproc_func = None
    assert model.preproc_func(2) == 2

    # repeat the setter test for postproc
    assert model.postproc_func(2) == 0
    model.postproc_func = lambda x: x - 1
    assert model.postproc_func(1) == 0
    assert model.preproc(1) == 1, "Must be unchanged!"
    assert model.postproc(2) == 0, "Must be unchanged!"
    # coverage setter check
    model.postproc_func = None
    assert model.postproc_func(2) == 0


def test_predictor_crash():
    """Test for crash when making predictor."""
    # without providing any model
    with pytest.raises(ValueError, match=r"Must provide.*"):
        CNNPatchPredictor()

    # provide wrong unknown pretrained model
    with pytest.raises(ValueError, match=r"Pretrained .* does not exist"):
        CNNPatchPredictor(pretrained_model="secret_model-kather100k")

    # provide wrong model of unknown type, deprecated later with type hint
    with pytest.raises(ValueError, match=r".*must be a string.*"):
        CNNPatchPredictor(pretrained_model=123)

    # test predict crash
    predictor = CNNPatchPredictor(pretrained_model="resnet18-kather100k", batch_size=32)

    with pytest.raises(ValueError, match=r".*not a valid mode.*"):
        predictor.predict("aaa", mode="random")
    # remove previously generated data
    if os.path.exists("output"):
        shutil.rmtree("output", ignore_errors=True)
    with pytest.raises(ValueError, match=r".*must be a list of file paths.*"):
        predictor.predict("aaa", mode="wsi")
    # remove previously generated data
    if os.path.exists("output"):
        shutil.rmtree("output", ignore_errors=True)
    with pytest.raises(ValueError, match=r".*masks.*!=.*imgs.*"):
        predictor.predict([1, 2, 3], masks=[1, 2], mode="wsi")
    with pytest.raises(ValueError, match=r".*labels.*!=.*imgs.*"):
        predictor.predict([1, 2, 3], labels=[1, 2], mode="patch")
    # remove previously generated data
    if os.path.exists("output"):
        shutil.rmtree("output", ignore_errors=True)


def test_patch_predictor_api(_sample_patch1, _sample_patch2):
    """Helper function to get the model output using API 1."""
    # convert to pathlib Path to prevent reader complaint
    inputs = [pathlib.Path(_sample_patch1), pathlib.Path(_sample_patch2)]
    predictor = CNNPatchPredictor(pretrained_model="resnet18-kather100k", batch_size=1)
    # don't run test on GPU
    output = predictor.predict(
        inputs,
        on_gpu=ON_GPU,
    )
    assert sorted(list(output.keys())) == ["predictions"]
    assert len(output["predictions"]) == 2

    output = predictor.predict(
        inputs,
        labels=[1, "a"],
        return_labels=True,
        on_gpu=ON_GPU,
    )
    assert sorted(list(output.keys())) == sorted(["labels", "predictions"])
    assert len(output["predictions"]) == len(output["labels"])
    assert output["labels"] == [1, "a"]

    output = predictor.predict(
        inputs,
        return_probabilities=True,
        on_gpu=ON_GPU,
    )
    assert sorted(list(output.keys())) == sorted(["predictions", "probabilities"])
    assert len(output["predictions"]) == len(output["probabilities"])

    output = predictor.predict(
        inputs,
        return_probabilities=True,
        labels=[1, "a"],
        return_labels=True,
        on_gpu=ON_GPU,
    )
    assert sorted(list(output.keys())) == sorted(
        ["labels", "predictions", "probabilities"]
    )
    assert len(output["predictions"]) == len(output["labels"])
    assert len(output["predictions"]) == len(output["probabilities"])

    # test saving output, should have no effect
    output = predictor.predict(
        inputs,
        on_gpu=ON_GPU,
        save_dir="special_dir_not_exist",
    )
    assert not os.path.isdir("special_dir_not_exist")

    # test loading user weight
    pretrained_weight_url = (
        "https://tiatoolbox.dcs.warwick.ac.uk/models/pc/resnet18-kather100k.pth"
    )

    save_dir_path = _get_temp_folder_path()
    # remove prev generated data
    if os.path.exists(save_dir_path):
        shutil.rmtree(save_dir_path, ignore_errors=True)
    os.makedirs(save_dir_path)
    pretrained_weight = os.path.join(
        rcParam["TIATOOLBOX_HOME"],
        "tmp_pretrained_weigths",
        "resnet18-kather100k.pth",
    )
    download_data(pretrained_weight_url, pretrained_weight)

    predictor = CNNPatchPredictor(
        pretrained_model="resnet18-kather100k",
        pretrained_weight=pretrained_weight,
        batch_size=1,
    )

    # --- test different using user model
    model = CNNPatchModel(backbone="resnet18", num_classes=9)
    # test prediction
    predictor = CNNPatchPredictor(model=model, batch_size=1, verbose=False)
    output = predictor.predict(
        inputs,
        return_probabilities=True,
        labels=[1, "a"],
        return_labels=True,
        on_gpu=ON_GPU,
    )
    assert sorted(list(output.keys())) == sorted(
        ["labels", "predictions", "probabilities"]
    )
    assert len(output["predictions"]) == len(output["labels"])
    assert len(output["predictions"]) == len(output["probabilities"])


def test_wsi_predictor_api(_sample_wsi_dict):
    """Test normal run of wsi predictor."""
    # convert to pathlib Path to prevent wsireader complaint
    _mini_wsi_svs = pathlib.Path(_sample_wsi_dict["wsi2_4k_4k_svs"])
    _mini_wsi_jpg = pathlib.Path(_sample_wsi_dict["wsi2_4k_4k_jpg"])
    _mini_wsi_msk = pathlib.Path(_sample_wsi_dict["wsi2_4k_4k_msk"])

    patch_size = np.array([224, 224])
    predictor = CNNPatchPredictor(pretrained_model="resnet18-kather100k", batch_size=32)

    # wrapper to make this more clean
    kwargs = dict(
        return_probabilities=True,
        return_labels=True,
        on_gpu=ON_GPU,
        patch_size=patch_size,
        stride_size=patch_size,
        resolution=1.0,
        units="baseline",
    )
    # ! add this test back once the read at `baseline` is fixed
    # sanity check, both output should be the same with same resolution read args
    wsi_output = predictor.predict(
        [_mini_wsi_svs],
        masks=[_mini_wsi_msk],
        mode="wsi",
        **kwargs,
    )

    tile_output = predictor.predict(
        [_mini_wsi_jpg],
        masks=[_mini_wsi_msk],
        mode="tile",
        **kwargs,
    )

    wpred = np.array(wsi_output[0]["predictions"])
    tpred = np.array(tile_output[0]["predictions"])
    diff = tpred == wpred
    accuracy = np.sum(diff) / np.size(wpred)
    assert accuracy > 0.9, np.nonzero(~diff)

    # remove previously generated data
    save_dir = "model_wsi_output"
    if os.path.exists(save_dir):
        shutil.rmtree(save_dir, ignore_errors=True)

    kwargs = dict(
        return_probabilities=True,
        return_labels=True,
        on_gpu=ON_GPU,
        patch_size=patch_size,
        stride_size=patch_size,
        resolution=0.5,
        save_dir=save_dir,
        merge_predictions=True,  # to test the api coverage
        units="mpp",
    )

    import copy

    _kwargs = copy.deepcopy(kwargs)
    _kwargs["merge_predictions"] = False
    # test reading of multiple whole-slide images
    output = predictor.predict(
        [_mini_wsi_svs, _mini_wsi_svs],
        masks=[_mini_wsi_msk, _mini_wsi_msk],
        mode="wsi",
        **_kwargs,
    )
    for output_info in output.values():
        assert os.path.exists(output_info["raw"])
        assert "merged" not in output_info
    if os.path.exists(_kwargs["save_dir"]):
        shutil.rmtree(_kwargs["save_dir"], ignore_errors=True)

    # coverage test
    _kwargs = copy.deepcopy(kwargs)
    _kwargs["merge_predictions"] = True
    # test reading of multiple whole-slide images
    output = predictor.predict(
        [_mini_wsi_svs, _mini_wsi_svs],
        masks=[_mini_wsi_msk, _mini_wsi_msk],
        mode="wsi",
        **_kwargs,
    )
    with pytest.raises(ValueError, match=r".*save_dir.*exist.*"):
        _kwargs = copy.deepcopy(kwargs)
        predictor.predict(
            [_mini_wsi_svs, _mini_wsi_svs],
            masks=[_mini_wsi_msk, _mini_wsi_msk],
            mode="wsi",
            **_kwargs,
        )
    # remove previously generated data
    if os.path.exists(_kwargs["save_dir"]):
        shutil.rmtree(_kwargs["save_dir"], ignore_errors=True)
    if os.path.exists("output"):
        shutil.rmtree("output", ignore_errors=True)

    # test reading of multiple whole-slide images
    _kwargs = copy.deepcopy(kwargs)
    _kwargs["save_dir"] = None  # default coverage
    _kwargs["return_probabilities"] = False
    output = predictor.predict(
        [_mini_wsi_svs, _mini_wsi_svs],
        masks=[_mini_wsi_msk, _mini_wsi_msk],
        mode="wsi",
        **_kwargs,
    )
    assert os.path.exists("output")
    for output_info in output.values():
        assert os.path.exists(output_info["raw"])
        assert "merged" in output_info and os.path.exists(output_info["merged"])

    # remove previously generated data
    if os.path.exists("output"):
        shutil.rmtree("output", ignore_errors=True)


def test_wsi_predictor_merge_predictions(_sample_wsi_dict):
    """Test normal run of wsi predictor with merge predictions option."""
    # convert to pathlib Path to prevent reader complaint
    _mini_wsi_svs = pathlib.Path(_sample_wsi_dict["wsi2_4k_4k_svs"])
    _mini_wsi_jpg = pathlib.Path(_sample_wsi_dict["wsi2_4k_4k_jpg"])
    _mini_wsi_msk = pathlib.Path(_sample_wsi_dict["wsi2_4k_4k_msk"])

    predictor = CNNPatchPredictor(pretrained_model="resnet18-kather100k", batch_size=1)

    kwargs = dict(
        return_probabilities=True,
        return_labels=True,
        on_gpu=ON_GPU,
        patch_size=np.array([224, 224]),
        stride_size=np.array([224, 224]),
        resolution=1.0,
        units="baseline",
        merge_predictions=True,
    )
    # sanity check, both output should be the same with same resolution read args
    wsi_output = predictor.predict(
        [_mini_wsi_svs],
        masks=[_mini_wsi_msk],
        mode="wsi",
        **kwargs,
    )

    # mockup to change the preproc func and
    # force to use the default in merge function
    # stil should have the same resuls
    kwargs["merge_predictions"] = False
    tile_output = predictor.predict(
        [_mini_wsi_jpg],
        masks=[_mini_wsi_msk],
        mode="tile",
        **kwargs,
    )
    merged_tile_output = predictor.merge_predictions(
        _mini_wsi_jpg,
        tile_output[0],
        resolution=kwargs["resolution"],
        units=kwargs["units"],
    )
    tile_output.append(merged_tile_output)

    # first make sure nothing breaks with predictions
    wpred = np.array(wsi_output[0]["predictions"])
    tpred = np.array(tile_output[0]["predictions"])
    diff = tpred == wpred
    accuracy = np.sum(diff) / np.size(wpred)
    assert accuracy > 0.9, np.nonzero(~diff)

    merged_wsi = wsi_output[1]
    merged_tile = tile_output[1]
    # enure shape of merged predictions of tile and wsi input are the same
    assert merged_wsi.shape == merged_tile.shape
    # ensure consistent predictions between tile and wsi mode
    diff = merged_tile == merged_wsi
    accuracy = np.sum(diff) / np.size(merged_wsi)
    assert accuracy > 0.9, np.nonzero(~diff)


def _test_predictor_output(
    inputs,
    pretrained_model,
    probabilities_check=None,
    predictions_check=None,
    on_gpu=ON_GPU,
):
    """Test the predictions of multiple models included in tiatoolbox."""
    predictor = CNNPatchPredictor(
        pretrained_model=pretrained_model, batch_size=32, verbose=False
    )
    # don't run test on GPU
    output = predictor.predict(
        inputs,
        return_probabilities=True,
        return_labels=False,
        on_gpu=ON_GPU,
    )
    predictions = output["predictions"]
    probabilities = output["probabilities"]
    for idx, probabilities_ in enumerate(probabilities):
        probabilities_max = max(probabilities_)
        assert (
            np.abs(probabilities_max - probabilities_check[idx]) <= 1e-6
            and predictions[idx] == predictions_check[idx]
        ), (
            pretrained_model,
            probabilities_max,
            probabilities_check[idx],
            predictions[idx],
            predictions_check[idx],
        )


def test_patch_predictor_output(_sample_patch1, _sample_patch2):
    """Test the output of patch prediction models."""
    inputs = [pathlib.Path(_sample_patch1), pathlib.Path(_sample_patch2)]
    pretrained_info = {
        "alexnet-kather100k": [1.0, 0.9999735355377197],
        "resnet18-Kather100k": [1.0, 0.9999911785125732],
        "resnet34-kather100k": [1.0, 0.9979840517044067],
        "resnet50-kather100k": [1.0, 0.9999986886978149],
        "resnet101-kather100k": [1.0, 0.9999932050704956],
        "resnext50_32x4d-kather100k": [1.0, 0.9910059571266174],
        "resnext101_32x8d-kather100k": [1.0, 0.9999971389770508],
        "wide_resnet50_2-kather100k": [1.0, 0.9953408241271973],
        "wide_resnet101_2-kather100k": [1.0, 0.9999831914901733],
        "densenet121-kather100k": [1.0, 1.0],
        "densenet161-kather100k": [1.0, 0.9999959468841553],
        "densenet169-kather100k": [1.0, 0.9999934434890747],
        "densenet201-kather100k": [1.0, 0.9999983310699463],
        "mobilenet_v2-kather100k": [0.9999998807907104, 0.9999126195907593],
        "mobilenet_v3_large-kather100k": [0.9999996423721313, 0.9999878406524658],
        "mobilenet_v3_small-kather100k": [0.9999998807907104, 0.9999997615814209],
        "googlenet-kather100k": [1.0, 0.9999639987945557],
    }
    for pretrained_model, expected_prob in pretrained_info.items():
        _test_predictor_output(
            inputs,
            pretrained_model,
            probabilities_check=expected_prob,
            predictions_check=[6, 3],
        )