def test_download_azureml_dataset(test_output_dirs: OutputFolderForTests) -> None:
dataset_name = "test-dataset"
config = DummyModel()
config.local_dataset = None
config.azure_dataset_id = ""
azure_config = get_default_azure_config()
runner = MLRunner(config, azure_config=azure_config)
# If the model has neither local_dataset or azure_dataset_id, mount_or_download_dataset should fail.
# This mounting call must happen before any other operations on the container, because already the model
# creation may need access to the dataset.
with pytest.raises(ValueError) as ex:
runner.setup()
assert ex.value.args[0] == "The model must contain either local_dataset or azure_dataset_id."
runner.project_root = test_output_dirs.root_dir
# Pointing the model to a dataset folder that does not exist should raise an Exception
fake_folder = runner.project_root / "foo"
runner.container.local_dataset = fake_folder
with pytest.raises(FileNotFoundError):
runner.mount_or_download_dataset(runner.container.azure_dataset_id, runner.container.local_dataset)
# If the local dataset folder exists, mount_or_download_dataset should not do anything.
fake_folder.mkdir()
local_dataset = runner.mount_or_download_dataset(runner.container.azure_dataset_id, runner.container.local_dataset)
assert local_dataset == fake_folder
# Pointing the model to a dataset in Azure should trigger a download
runner.container.local_dataset = None
runner.container.azure_dataset_id = dataset_name
with logging_section("Starting download"):
result_path = runner.mount_or_download_dataset(runner.container.azure_dataset_id,
runner.container.local_dataset)
# Download goes into <project_root> / "datasets" / "test_dataset"
expected_path = runner.project_root / fixed_paths.DATASETS_DIR_NAME / dataset_name
assert result_path == expected_path
assert result_path.is_dir()
dataset_csv = Path(result_path) / DATASET_CSV_FILE_NAME
assert dataset_csv.is_file()
# Check that each individual file in the dataset is present
for folder in [1, *range(10, 20)]:
sub_folder = result_path / str(folder)
sub_folder.is_dir()
for file in ["ct", "esophagus", "heart", "lung_l", "lung_r", "spinalcord"]:
f = (sub_folder / file).with_suffix(".nii.gz")
assert f.is_file()
def _test_mount_for_lightning_container(test_output_dirs: OutputFolderForTests,
is_offline_run: bool,
local_dataset: Optional[Path],
azure_dataset: str,
is_lightning_model: bool) -> LightningContainer:
config: Optional[DeepLearningConfig] = None
container: Optional[LightningContainer] = None
if is_lightning_model:
container = DummyContainerWithDatasets()
container.azure_dataset_id = azure_dataset
container.local_dataset = local_dataset
else:
config = DummyModel()
config.azure_dataset_id = azure_dataset
config.local_dataset = local_dataset
# The legacy InnerEye models require an existing dataset_csv file present in the dataset folder. Create that.
download_path = test_output_dirs.root_dir / "downloaded"
mount_path = test_output_dirs.root_dir / "mounted"
if not is_lightning_model:
train_and_test_data = "train_and_test_data"
for path in [download_path, mount_path, test_output_dirs.root_dir]:
# If destination folder exists, delete content to ensure consistency and avoid 'FileExistsError'
if (path / train_and_test_data).is_dir():
shutil.rmtree(path / train_and_test_data)
# Creates directory structure and copy data
shutil.copytree(full_ml_test_data_path(train_and_test_data), path / train_and_test_data)
# Copy 'data.csv' file
shutil.copy(full_ml_test_data_path(DATASET_CSV_FILE_NAME), path / DATASET_CSV_FILE_NAME)
with mock.patch("InnerEye.ML.run_ml.MLRunner.is_offline_run", is_offline_run):
with mock.patch("InnerEye.ML.run_ml.download_dataset", return_value=download_path):
with mock.patch("InnerEye.ML.run_ml.try_to_mount_input_dataset", return_value=mount_path):
runner = MLRunner(config, container=container,
azure_config=None, project_root=test_output_dirs.root_dir)
runner.setup()
return runner.container
def test_model_test(test_output_dirs: OutputFolderForTests) -> None:
train_and_test_data_dir = full_ml_test_data_path("train_and_test_data")
config = DummyModel()
config.set_output_to(test_output_dirs.root_dir)
epoch = 1
config.num_epochs = epoch
assert config.get_test_epochs() == [epoch]
placeholder_dataset_id = "place_holder_dataset_id"
config.azure_dataset_id = placeholder_dataset_id
transform = config.get_full_image_sample_transforms().test
df = pd.read_csv(full_ml_test_data_path(DATASET_CSV_FILE_NAME))
df = df[df.subject.isin([1, 2])]
# noinspection PyTypeHints
config._datasets_for_inference = \
{ModelExecutionMode.TEST: FullImageDataset(config, df, full_image_sample_transforms=transform)} # type: ignore
execution_mode = ModelExecutionMode.TEST
checkpoint_handler = get_default_checkpoint_handler(model_config=config,
project_root=test_output_dirs.root_dir)
# Mimic the behaviour that checkpoints are downloaded from blob storage into the checkpoints folder.
stored_checkpoints = full_ml_test_data_path("checkpoints")
shutil.copytree(str(stored_checkpoints), str(config.checkpoint_folder))
checkpoint_handler.additional_training_done()
inference_results = model_testing.segmentation_model_test(config,
data_split=execution_mode,
checkpoint_handler=checkpoint_handler)
epoch_dir = config.outputs_folder / get_epoch_results_path(epoch, execution_mode)
assert inference_results.epochs[epoch] == pytest.approx(0.66606902, abs=1e-6)
assert config.outputs_folder.is_dir()
assert epoch_dir.is_dir()
patient1 = io_util.load_nifti_image(train_and_test_data_dir / "id1_channel1.nii.gz")
patient2 = io_util.load_nifti_image(train_and_test_data_dir / "id2_channel1.nii.gz")
assert_file_contains_string(epoch_dir / DATASET_ID_FILE, placeholder_dataset_id)
assert_file_contains_string(epoch_dir / GROUND_TRUTH_IDS_FILE, "region")
assert_text_files_match(epoch_dir / model_testing.METRICS_FILE_NAME,
train_and_test_data_dir / model_testing.METRICS_FILE_NAME)
assert_text_files_match(epoch_dir / model_testing.METRICS_AGGREGATES_FILE,
train_and_test_data_dir / model_testing.METRICS_AGGREGATES_FILE)
# Plotting results vary between platforms. Can only check if the file is generated, but not its contents.
assert (epoch_dir / model_testing.BOXPLOT_FILE).exists()
assert_nifti_content(epoch_dir / "001" / "posterior_region.nii.gz", get_image_shape(patient1),
patient1.header,
[136], np.ubyte)
assert_nifti_content(epoch_dir / "002" / "posterior_region.nii.gz", get_image_shape(patient2),
patient2.header,
[136], np.ubyte)
assert_nifti_content(epoch_dir / "001" / DEFAULT_RESULT_IMAGE_NAME, get_image_shape(patient1),
patient1.header,
[1], np.ubyte)
assert_nifti_content(epoch_dir / "002" / DEFAULT_RESULT_IMAGE_NAME, get_image_shape(patient2),
patient2.header,
[1], np.ubyte)
assert_nifti_content(epoch_dir / "001" / "posterior_background.nii.gz", get_image_shape(patient1),
patient1.header,
[118], np.ubyte)
assert_nifti_content(epoch_dir / "002" / "posterior_background.nii.gz", get_image_shape(patient2),
patient2.header,
[118], np.ubyte)
thumbnails_folder = epoch_dir / model_testing.THUMBNAILS_FOLDER
assert thumbnails_folder.is_dir()
png_files = list(thumbnails_folder.glob("*.png"))
overlays = [f for f in png_files if "_region_slice_" in str(f)]
assert len(overlays) == len(df.subject.unique()), "There should be one overlay/contour file per subject"
# Writing dataset.csv normally happens at the beginning of training,
# but this test reads off a saved checkpoint file.
# Dataset.csv must be present for plot_cross_validation.
config.write_dataset_files()
# Test if the metrics files can be picked up correctly by the cross validation code
config_and_files = get_config_and_results_for_offline_runs(config)
result_files = config_and_files.files
assert len(result_files) == 1
for file in result_files:
assert file.execution_mode == execution_mode
assert file.dataset_csv_file is not None
assert file.dataset_csv_file.exists()
assert file.metrics_file is not None
assert file.metrics_file.exists()
def test_model_test(test_output_dirs: OutputFolderForTests,
use_partial_ground_truth: bool,
allow_partial_ground_truth: bool) -> None:
"""
Check the CSVs (and image files) output by InnerEye.ML.model_testing.segmentation_model_test
:param test_output_dirs: The fixture in conftest.py
:param use_partial_ground_truth: Whether to remove some ground truth labels from some test users
:param allow_partial_ground_truth: What to set the allow_incomplete_labels flag to
"""
train_and_test_data_dir = full_ml_test_data_path("train_and_test_data")
seed_everything(42)
config = DummyModel()
config.allow_incomplete_labels = allow_partial_ground_truth
config.set_output_to(test_output_dirs.root_dir)
placeholder_dataset_id = "place_holder_dataset_id"
config.azure_dataset_id = placeholder_dataset_id
transform = config.get_full_image_sample_transforms().test
df = pd.read_csv(full_ml_test_data_path(DATASET_CSV_FILE_NAME))
if use_partial_ground_truth:
config.check_exclusive = False
config.ground_truth_ids = ["region", "region_1"]
# As in Tests.ML.pipelines.test.inference.test_evaluate_model_predictions patients 3, 4,
# and 5 are in the test dataset with:
# Patient 3 has one missing ground truth channel: "region"
df = df[df["subject"].ne(3) | df["channel"].ne("region")]
# Patient 4 has all missing ground truth channels: "region", "region_1"
df = df[df["subject"].ne(4) | df["channel"].ne("region")]
df = df[df["subject"].ne(4) | df["channel"].ne("region_1")]
# Patient 5 has no missing ground truth channels.
config.dataset_data_frame = df
df = df[df.subject.isin([3, 4, 5])]
config.train_subject_ids = ['1', '2']
config.test_subject_ids = ['3', '4', '5']
config.val_subject_ids = ['6', '7']
else:
df = df[df.subject.isin([1, 2])]
if use_partial_ground_truth and not allow_partial_ground_truth:
with pytest.raises(ValueError) as value_error:
# noinspection PyTypeHints
config._datasets_for_inference = {
ModelExecutionMode.TEST:
FullImageDataset(config,
df,
full_image_sample_transforms=transform)
} # type: ignore
assert "Patient 3 does not have channel 'region'" in str(
value_error.value)
return
else:
# noinspection PyTypeHints
config._datasets_for_inference = {
ModelExecutionMode.TEST:
FullImageDataset(config,
df,
full_image_sample_transforms=transform)
} # type: ignore
execution_mode = ModelExecutionMode.TEST
checkpoint_handler = get_default_checkpoint_handler(
model_config=config, project_root=test_output_dirs.root_dir)
# Mimic the behaviour that checkpoints are downloaded from blob storage into the checkpoints folder.
create_model_and_store_checkpoint(
config,
config.checkpoint_folder / LAST_CHECKPOINT_FILE_NAME_WITH_SUFFIX)
checkpoint_handler.additional_training_done()
inference_results = model_testing.segmentation_model_test(
config,
execution_mode=execution_mode,
checkpoint_paths=checkpoint_handler.get_checkpoints_to_test())
epoch_dir = config.outputs_folder / get_best_epoch_results_path(
execution_mode)
total_num_patients_column_name = f"total_{MetricsFileColumns.Patient.value}".lower(
)
if not total_num_patients_column_name.endswith("s"):
total_num_patients_column_name += "s"
if use_partial_ground_truth:
num_subjects = len(pd.unique(df["subject"]))
if allow_partial_ground_truth:
assert csv_column_contains_value(
csv_file_path=epoch_dir / METRICS_AGGREGATES_FILE,
column_name=total_num_patients_column_name,
value=num_subjects,
contains_only_value=True)
assert csv_column_contains_value(
csv_file_path=epoch_dir / SUBJECT_METRICS_FILE_NAME,
column_name=MetricsFileColumns.Dice.value,
value='',
contains_only_value=False)
else:
aggregates_df = pd.read_csv(epoch_dir / METRICS_AGGREGATES_FILE)
assert total_num_patients_column_name not in aggregates_df.columns # Only added if using partial ground truth
assert not csv_column_contains_value(
csv_file_path=epoch_dir / SUBJECT_METRICS_FILE_NAME,
column_name=MetricsFileColumns.Dice.value,
value='',
contains_only_value=False)
assert inference_results.metrics == pytest.approx(0.66606902, abs=1e-6)
assert config.outputs_folder.is_dir()
assert epoch_dir.is_dir()
patient1 = io_util.load_nifti_image(train_and_test_data_dir /
"id1_channel1.nii.gz")
patient2 = io_util.load_nifti_image(train_and_test_data_dir /
"id2_channel1.nii.gz")
assert_file_contains_string(epoch_dir / DATASET_ID_FILE,
placeholder_dataset_id)
assert_file_contains_string(epoch_dir / GROUND_TRUTH_IDS_FILE,
"region")
assert_text_files_match(
epoch_dir / model_testing.SUBJECT_METRICS_FILE_NAME,
train_and_test_data_dir / model_testing.SUBJECT_METRICS_FILE_NAME)
assert_text_files_match(
epoch_dir / model_testing.METRICS_AGGREGATES_FILE,
train_and_test_data_dir / model_testing.METRICS_AGGREGATES_FILE)
# Plotting results vary between platforms. Can only check if the file is generated, but not its contents.
assert (epoch_dir / model_testing.BOXPLOT_FILE).exists()
assert_nifti_content(epoch_dir / "001" / "posterior_region.nii.gz",
get_image_shape(patient1), patient1.header, [137],
np.ubyte)
assert_nifti_content(epoch_dir / "002" / "posterior_region.nii.gz",
get_image_shape(patient2), patient2.header, [137],
np.ubyte)
assert_nifti_content(epoch_dir / "001" / DEFAULT_RESULT_IMAGE_NAME,
get_image_shape(patient1), patient1.header, [1],
np.ubyte)
assert_nifti_content(epoch_dir / "002" / DEFAULT_RESULT_IMAGE_NAME,
get_image_shape(patient2), patient2.header, [1],
np.ubyte)
assert_nifti_content(epoch_dir / "001" / "posterior_background.nii.gz",
get_image_shape(patient1), patient1.header, [117],
np.ubyte)
assert_nifti_content(epoch_dir / "002" / "posterior_background.nii.gz",
get_image_shape(patient2), patient2.header, [117],
np.ubyte)
thumbnails_folder = epoch_dir / model_testing.THUMBNAILS_FOLDER
assert thumbnails_folder.is_dir()
png_files = list(thumbnails_folder.glob("*.png"))
overlays = [f for f in png_files if "_region_slice_" in str(f)]
assert len(overlays) == len(df.subject.unique(
)), "There should be one overlay/contour file per subject"
# Writing dataset.csv normally happens at the beginning of training,
# but this test reads off a saved checkpoint file.
# Dataset.csv must be present for plot_cross_validation.
config.write_dataset_files()
# Test if the metrics files can be picked up correctly by the cross validation code
config_and_files = get_config_and_results_for_offline_runs(config)
result_files = config_and_files.files
assert len(result_files) == 1
for file in result_files:
assert file.execution_mode == execution_mode
assert file.dataset_csv_file is not None
assert file.dataset_csv_file.exists()
assert file.metrics_file is not None
assert file.metrics_file.exists()
