Fix for stuck Linux build: Move pytest to Windows

.github/workflows/check_changelog.yml

@@ -6,7 +6,8 @@ name: Check Changelog
 on:
   pull_request:
 jobs:
-  check:
+  check_changelog:
+    name: Check Changelog
     runs-on: ubuntu-latest
     if: ${{ contains(github.event.pull_request.labels.*.name, 'no changelog needed') == 0 }}
     steps:

.github/workflows/codeql-analysis.yml

@@ -10,8 +10,8 @@ on:
     - cron: '45 4 * * 1'
 
 jobs:
-  analyze:
-    name: Analyze
+  codeql_analyze:
+    name: CodeQL Analyze
     runs-on: ubuntu-latest
 
     strategy:

.github/workflows/issues_to_ado.yml

@@ -6,7 +6,8 @@ on:
       [opened, edited, deleted, closed, reopened, labeled, unlabeled, assigned]
 
 jobs:
-  alert:
+  issues_to_ado:
+    name: Sync issues with Azure DevOps
     runs-on: ubuntu-latest
     steps:
       - uses: danhellem/github-actions-issue-to-work-item@master

.github/workflows/linting_and_hello_world.yml

@@ -7,7 +7,8 @@ on:
   pull_request:
 
 jobs:
-  linux:
+  flake_mypy_helloworld_linux:
+    name: Flake8, MyPy, HelloWorld on Linux
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v2
@@ -53,7 +54,8 @@ jobs:
           PYTHONPATH: ${{ github.workspace }}
         if: always()
 
-  windows:
+  hello_world_windows:
+    name: HelloWorld on Windows
     runs-on: windows-latest
     steps:
       - uses: actions/checkout@v2

CHANGELOG.md

@@ -119,6 +119,7 @@ in inference-only runs when using lightning containers.
 - ([#628](https://github.com/microsoft/InnerEye-DeepLearning/pull/628)) SSL SimCLR using the wrong LR schedule when running on multiple nodes
 - ([#638](https://github.com/microsoft/InnerEye-DeepLearning/pull/638)) SimClr cosine LR scheduler was using wrong length information when using with long linear head datasets
 - ([#612](https://github.com/microsoft/InnerEye-DeepLearning/pull/612)) SSL online evaluator was not doing distributed training
+- ([#652](https://github.com/microsoft/InnerEye-DeepLearning/pull/652)) Run pytest build on Windows after Linux agent version upgrade
 
 ### Removed
 

InnerEye/ML/Histopathology/datasets/panda_dataset.py

@@ -2,19 +2,23 @@
 #  Copyright (c) Microsoft Corporation. All rights reserved.
 #  Licensed under the MIT License (MIT). See LICENSE in the repo root for license information.
 #  ------------------------------------------------------------------------------------------
-
+import logging
 from pathlib import Path
 from typing import Any, Dict, Union, Optional
 
 import pandas as pd
-from cucim import CuImage
 from health_ml.utils import box_utils
 from monai.config import KeysCollection
 from monai.data.image_reader import ImageReader, WSIReader
 from monai.transforms import MapTransform
 
 from InnerEye.ML.Histopathology.datasets.base_dataset import SlidesDataset
 
+try:
+    from cucim import CuImage
+except:
+    logging.warning("cucim library not available, code may fail.")
+
 
 class PandaDataset(SlidesDataset):
     """Dataset class for loading files from the PANDA challenge dataset.
@@ -48,6 +52,7 @@ def __init__(self,
 # MONAI's convention is that dictionary transforms have a 'd' suffix in the class name
 class ReadImaged(MapTransform):
     """Basic transform to read image files."""
+
     def __init__(self, reader: ImageReader, keys: KeysCollection,
                  allow_missing_keys: bool = False, **kwargs: Any) -> None:
         super().__init__(keys, allow_missing_keys=allow_missing_keys)
@@ -71,6 +76,7 @@ class LoadPandaROId(MapTransform):
     - `'level'` (int): chosen magnification level
     - `'scale'` (float): corresponding scale, loaded from the file
     """
+
     def __init__(self, reader: WSIReader, image_key: str = 'image', mask_key: str = 'mask',
                  level: int = 0, margin: int = 0, **kwargs: Any) -> None:
         """
@@ -88,7 +94,7 @@ def __init__(self, reader: WSIReader, image_key: str = 'image', mask_key: str =
         self.margin = margin
         self.kwargs = kwargs
 
-    def _get_bounding_box(self, mask_obj: CuImage) -> box_utils.Box:
+    def _get_bounding_box(self, mask_obj: 'CuImage') -> box_utils.Box:
         # Estimate bounding box at the lowest resolution (i.e. highest level)
         highest_level = mask_obj.resolutions['level_count'] - 1
         scale = mask_obj.resolutions['level_downsamples'][highest_level]

InnerEye/ML/Histopathology/preprocessing/loading.py

@@ -1,14 +1,19 @@
+import logging
 from typing import Dict, Optional, Tuple
 
 import numpy as np
 import skimage.filters
-from cucim import CuImage
 from health_ml.utils import box_utils
 from monai.data.image_reader import WSIReader
 from monai.transforms import MapTransform
 
 from InnerEye.ML.Histopathology.utils.naming import SlideKey
 
+try:
+    from cucim import CuImage
+except:
+    logging.warning("cucim library not available, code may fail.")
+
 
 def get_luminance(slide: np.ndarray) -> np.ndarray:
     """Compute a grayscale version of the input slide.
@@ -35,7 +40,7 @@ def segment_foreground(slide: np.ndarray, threshold: Optional[float] = None) \
     return luminance < threshold, threshold
 
 
-def load_slide_at_level(reader: WSIReader, slide_obj: CuImage, level: int) -> np.ndarray:
+def load_slide_at_level(reader: WSIReader, slide_obj: 'CuImage', level: int) -> np.ndarray:
     """Load full slide array at the given magnification level.
 
     This is a manual workaround for a MONAI bug (https://github.com/Project-MONAI/MONAI/issues/3415)
@@ -60,6 +65,7 @@ class LoadROId(MapTransform):
     - `SlideKey.SCALE` (float): corresponding scale, loaded from the file
     - `SlideKey.FOREGROUND_THRESHOLD` (float): threshold used to segment the foreground
     """
+
     def __init__(self, reader: WSIReader, image_key: str = SlideKey.IMAGE, level: int = 0,
                  margin: int = 0, foreground_threshold: Optional[float] = None) -> None:
         """
@@ -77,7 +83,7 @@ def __init__(self, reader: WSIReader, image_key: str = SlideKey.IMAGE, level: in
         self.margin = margin
         self.foreground_threshold = foreground_threshold
 
-    def _get_bounding_box(self, slide_obj: CuImage) -> Tuple[box_utils.Box, float]:
+    def _get_bounding_box(self, slide_obj: 'CuImage') -> Tuple[box_utils.Box, float]:
         # Estimate bounding box at the lowest resolution (i.e. highest level)
         highest_level = slide_obj.resolutions['level_count'] - 1
         scale = slide_obj.resolutions['level_downsamples'][highest_level]
@@ -88,6 +94,7 @@ def _get_bounding_box(self, slide_obj: CuImage) -> Tuple[box_utils.Box, float]:
         return bbox, threshold
 
     def __call__(self, data: Dict) -> Dict:
+        from cucim import CuImage
         image_obj: CuImage = self.reader.read(data[self.image_key])
 
         level0_bbox, threshold = self._get_bounding_box(image_obj)

Tests/Azure/test_azure_config.py

@@ -9,6 +9,7 @@
 
 from InnerEye.Azure.azure_config import AzureConfig
 from InnerEye.Azure.azure_runner import create_dataset_configs
+from InnerEye.Common.common_util import is_linux
 from InnerEye.ML.deep_learning_config import DatasetParams
 from Tests.ML.util import get_default_azure_config
 
@@ -65,8 +66,10 @@ def test_dataset_consumption2() -> None:
     assert datasets[1].name == "2"
     assert datasets[0].local_folder == Path("l1")
     assert datasets[1].local_folder == Path("l2")
-    assert datasets[0].target_folder == PosixPath("mp1")
-    assert datasets[1].target_folder == PosixPath("mp2")
+    if is_linux():
+        # PosixPath cannot be instantiated on Windows
+        assert datasets[0].target_folder == PosixPath("mp1")
+        assert datasets[1].target_folder == PosixPath("mp2")
 
 
 def test_dataset_consumption3() -> None:

Tests/ML/histopathology/models/test_deepmil.py

@@ -7,6 +7,7 @@
 from typing import Callable, Dict, List, Type  # noqa
 
 import pytest
+import torch
 from torch import Tensor, argmax, nn, rand, randint, randn, round, stack, allclose
 from torchvision.models import resnet18
 
@@ -29,7 +30,7 @@
 )
 from InnerEye.ML.Histopathology.models.deepmil import DeepMILModule
 from InnerEye.ML.Histopathology.models.encoders import ImageNetEncoder, TileEncoder
-from InnerEye.ML.Histopathology.utils.naming import ResultsKey
+from InnerEye.ML.Histopathology.utils.naming import MetricsKey, ResultsKey
 
 
 def get_supervised_imagenet_encoder() -> TileEncoder:
@@ -38,10 +39,10 @@ def get_supervised_imagenet_encoder() -> TileEncoder:
 
 @pytest.mark.parametrize("n_classes", [1, 3])
 @pytest.mark.parametrize("pooling_layer", [AttentionLayer, GatedAttentionLayer])
-@pytest.mark.parametrize("batch_size", [1, 15])
-@pytest.mark.parametrize("max_bag_size", [1, 7])
-@pytest.mark.parametrize("pool_hidden_dim", [1, 5])
-@pytest.mark.parametrize("pool_out_dim", [1, 6])
+@pytest.mark.parametrize("batch_size", [1, 2])
+@pytest.mark.parametrize("max_bag_size", [1, 3])
+@pytest.mark.parametrize("pool_hidden_dim", [1, 4])
+@pytest.mark.parametrize("pool_out_dim", [1, 5])
 def test_lightningmodule(
     n_classes: int,
     pooling_layer: Callable[[int, int, int], nn.Module],
@@ -108,9 +109,12 @@ def test_lightningmodule(
     assert preds.shape[0] == batch_size
 
     for metric_name, metric_object in module.train_metrics.items():
-        if (batch_size > 1) or (not metric_name == "auroc"):
+        if metric_name == MetricsKey.CONF_MATRIX or metric_name == MetricsKey.AUROC:
+            continue
+        if batch_size > 1:
             score = metric_object(preds.view(-1, 1), bag_labels.view(-1, 1))
-            assert score >= 0 and score <= 1
+            assert torch.all(score >= 0)
+            assert torch.all(score <= 1)
 
 
 def move_batch_to_expected_device(batch: Dict[str, List], use_gpu: bool) -> Dict:

Tests/ML/histopathology/preprocessing/test_slide_loading.py

@@ -2,21 +2,24 @@
 
 import numpy as np
 import pytest
-from cucim import CuImage
 from monai.data.image_reader import WSIReader
 
+from InnerEye.Common.common_util import is_windows
 from InnerEye.Common.fixed_paths_for_tests import tests_root_directory
 from InnerEye.ML.Histopathology.preprocessing.tiling import tile_array_2d
-from InnerEye.ML.Histopathology.preprocessing.loading import LoadROId, get_luminance, load_slide_at_level, segment_foreground
+from InnerEye.ML.Histopathology.preprocessing.loading import (LoadROId, get_luminance, load_slide_at_level,
+                                                              segment_foreground)
 from InnerEye.ML.Histopathology.utils.naming import SlideKey
 from Tests.ML.histopathology.datasets.test_slides_dataset import MockSlidesDataset
 
 TEST_IMAGE_PATH = str(tests_root_directory("ML/histopathology/test_data/panda_wsi_example.tiff"))
 
 
+@pytest.mark.skipif(is_windows(), reason="cucim package is not available on Windows")
 def test_load_slide() -> None:
     level = 2
     reader = WSIReader('cuCIM')
+    from cucim import CuImage
     slide_obj: CuImage = reader.read(TEST_IMAGE_PATH)
     dims = slide_obj.resolutions['level_dimensions'][level][::-1]
 
@@ -39,9 +42,11 @@ def test_load_slide() -> None:
     assert np.array_equiv(larger_slide[:, :, dims[1]:], empty_fill_value)
 
 
+@pytest.mark.skipif(is_windows(), reason="cucim package is not available on Windows")
 def test_get_luminance() -> None:
     level = 2  # here we only need to test at a single resolution
     reader = WSIReader('cuCIM')
+    from cucim import CuImage
     slide_obj: CuImage = reader.read(TEST_IMAGE_PATH)
 
     slide = load_slide_at_level(reader, slide_obj, level)
@@ -61,9 +66,11 @@ def test_get_luminance() -> None:
     assert np.array_equal(slide_luminance_tiles.squeeze(1), tiles_luminance)
 
 
+@pytest.mark.skipif(is_windows(), reason="cucim package is not available on Windows")
 def test_segment_foreground() -> None:
     level = 2  # here we only need to test at a single resolution
     reader = WSIReader('cuCIM')
+    from cucim import CuImage
     slide_obj: CuImage = reader.read(TEST_IMAGE_PATH)
     slide = load_slide_at_level(reader, slide_obj, level)
 
@@ -95,11 +102,13 @@ def test_segment_foreground() -> None:
 
 @pytest.mark.parametrize('level', [1, 2])
 @pytest.mark.parametrize('foreground_threshold', [None, 215])
+@pytest.mark.skipif(is_windows(), reason="cucim package is not available on Windows")
 def test_get_bounding_box(level: int, foreground_threshold: Optional[float]) -> None:
     margin = 0
     reader = WSIReader('cuCIM')
     loader = LoadROId(reader, image_key=SlideKey.IMAGE, level=level, margin=margin,
                       foreground_threshold=foreground_threshold)
+    from cucim import CuImage
     slide_obj: CuImage = reader.read(TEST_IMAGE_PATH)
     level0_bbox, _ = loader._get_bounding_box(slide_obj)
 
@@ -130,6 +139,7 @@ def test_get_bounding_box(level: int, foreground_threshold: Optional[float]) ->
 @pytest.mark.parametrize('level', [1, 2])
 @pytest.mark.parametrize('margin', [0, 42])
 @pytest.mark.parametrize('foreground_threshold', [None, 215])
+@pytest.mark.skipif(is_windows(), reason="cucim package is not available on Windows")
 def test_load_roi(level: int, margin: int, foreground_threshold: Optional[float]) -> None:
     dataset = MockSlidesDataset()
     sample = dataset[0]