def get_experiment_name_from_environment() -> str:
"""
Reads the name of the present branch from environment variable "BUILD_BRANCH". This must be set in the YML file.
If the variable is not found, return an empty string.
With this setup, all AML runs that belong to a given pull request are listed at the same place.
"""
env_branch = "BUILD_BRANCH"
build_branch = os.environ.get(env_branch, None)
if not build_branch:
if is_running_on_azure_agent():
raise ValueError(f"Environment variable {env_branch} should be set when running on Azure agents.")
return ""
return to_azure_friendly_string(build_branch) or ""
def test_visualize_patch_sampling_2d(
test_output_dirs: OutputFolderForTests) -> None:
"""
Tests if patch sampling works for 2D images.
:param test_output_dirs:
"""
set_random_seed(0)
shape = (1, 20, 30)
foreground_classes = ["fg"]
class_weights = equally_weighted_classes(foreground_classes)
config = SegmentationModelBase(should_validate=False,
crop_size=(1, 5, 10),
class_weights=class_weights)
image = np.random.rand(1, *shape).astype(np.float32) * 1000
mask = np.ones(shape)
labels = np.zeros((len(class_weights), ) + shape)
labels[1, 0, 8:12, 5:25] = 1
labels[0] = 1 - labels[1]
output_folder = Path(test_output_dirs.root_dir)
image_header = None
sample = Sample(image=image,
mask=mask,
labels=labels,
metadata=PatientMetadata(patient_id='123',
image_header=image_header))
heatmap = visualize_random_crops(sample,
config,
output_folder=output_folder)
expected_folder = full_ml_test_data_path("patch_sampling")
expected_heatmap = expected_folder / "sampling_2d.npy"
# To update the stored results, uncomment this line:
# np.save(str(expected_heatmap), heatmap)
assert np.allclose(heatmap, np.load(
str(expected_heatmap))), "Patch sampling created a different heatmap."
assert len(list(output_folder.rglob("*.nii.gz"))) == 0
assert len(list(output_folder.rglob("*.png"))) == 1
actual_file = output_folder / "123_sampled_patches.png"
assert_file_exists(actual_file)
expected = expected_folder / "sampling_2d.png"
# To update the stored results, uncomment this line:
# expected.write_bytes(actual_file.read_bytes())
if not is_running_on_azure_agent():
# When running on the Azure build agents, it appears that the bounding box of the images
# is slightly different than on local runs, even with equal dpi settings.
# It says: Image sizes don't match: actual (685, 469), expected (618, 424)
# Not able to figure out how to make the run results consistent, hence disable in cloud runs.
assert_binary_files_match(actual_file, expected)
def get_most_recent_run() -> str:
"""
Gets the string name of the most recently executed AzureML run. This is picked up from the `most_recent_run.txt`
file when running on the cloud. For execution on the local dev box, a fixed path is returned.
:return:
"""
run_recovery_file = Path(RUN_RECOVERY_FILE)
if is_running_on_azure_agent():
assert run_recovery_file.is_file(), "When running in cloud builds, this should pick up the ID of a previous " \
"training run"
print("Reading run information from file.")
return run_recovery_file.read_text().strip()
else:
# When executing on the local box, we usually don't have any recent runs. Use a hardcoded run ID here.
# Consequently, local execution of tests that use this run may fail, while executing in the cloud passes.
# In this case, modify the run here to something more recent.
print("Using hardcoded run ID.")
# This is Run 89 in refs_pull_276_merge
return "refs_pull_276_merge:refs_pull_276_merge_1605139760_18b251e1"
def get_most_recent_run_id(fallback_run_id_for_local_execution: str = FALLBACK_SINGLE_RUN) -> str:
"""
Gets the string name of the most recently executed AzureML run. This is picked up from the `most_recent_run.txt`
file when running on the cloud. For execution on the local dev box, use a hardcoded run ID.
Consequently, local execution of tests that use this run may fail, while executing in the cloud passes.
In this case, modify the run here to something more recent.
:param fallback_run_id_for_local_execution: A hardcoded AzureML run ID that is used when executing this code
on a local box, outside of Azure build agents.
:return:
"""
run_recovery_file = Path(RUN_RECOVERY_FILE)
if is_running_on_azure_agent():
assert run_recovery_file.is_file(), "When running in cloud builds, this should pick up the ID of a previous " \
"training run"
run_id = run_recovery_file.read_text().strip()
print(f"Read this run ID from file: {run_id}")
return run_id
else:
assert fallback_run_id_for_local_execution, "When running on local box, a hardcoded run ID must be given."
print(f"Using this hardcoded run ID: {fallback_run_id_for_local_execution}")
return fallback_run_id_for_local_execution
def test_visualize_patch_sampling(test_output_dirs: OutputFolderForTests,
labels_to_boundary: bool) -> None:
"""
Tests if patch sampling and producing diagnostic images works as expected.
:param test_output_dirs:
:param labels_to_boundary: If true, the ground truth labels are placed close to the image boundary, so that
crops have to be adjusted inwards. If false, ground truth labels are all far from the image boundaries.
"""
set_random_seed(0)
shape = (10, 30, 30)
foreground_classes = ["fg"]
class_weights = equally_weighted_classes(foreground_classes)
config = SegmentationModelBase(should_validate=False,
crop_size=(2, 10, 10),
class_weights=class_weights)
image = np.random.rand(1, *shape).astype(np.float32) * 1000
mask = np.ones(shape)
labels = np.zeros((len(class_weights),) + shape)
if labels_to_boundary:
# Generate foreground labels in such a way that a patch centered around a foreground pixel would
# reach outside of the image.
labels[1, 4:8, 3:27, 3:27] = 1
else:
labels[1, 4:8, 15:18, 15:18] = 1
labels[0] = 1 - labels[1]
output_folder = Path(test_output_dirs.root_dir)
image_header = get_unit_image_header()
sample = Sample(image=image,
mask=mask,
labels=labels,
metadata=PatientMetadata(patient_id='123',
image_header=image_header))
expected_folder = full_ml_test_data_path("patch_sampling")
heatmap = visualize_random_crops(sample, config, output_folder=output_folder)
expected_heatmap = expected_folder / ("sampled_to_boundary.npy" if labels_to_boundary else "sampled_center.npy")
# To update the stored results, uncomment this line:
# np.save(str(expected_heatmap), heatmap)
assert np.allclose(heatmap, np.load(str(expected_heatmap))), "Patch sampling created a different heatmap."
f1 = output_folder / "123_ct.nii.gz"
assert_file_exists(f1)
f2 = output_folder / "123_sampled_patches.nii.gz"
assert_file_exists(f2)
thumbnails = [
"123_sampled_patches_dim0.png",
"123_sampled_patches_dim1.png",
"123_sampled_patches_dim2.png",
]
for f in thumbnails:
assert_file_exists(output_folder / f)
expected = expected_folder / ("sampled_to_boundary.nii.gz" if labels_to_boundary else "sampled_center.nii.gz")
# To update test results:
# shutil.copy(str(f2), str(expected))
expected_image = io_util.load_nifti_image(expected)
actual_image = io_util.load_nifti_image(f2)
np.allclose(expected_image.image, actual_image.image)
if labels_to_boundary:
for f in thumbnails:
# Uncomment this line to update test results
# (expected_folder / f).write_bytes((output_folder / f).read_bytes())
if not is_running_on_azure_agent():
# When running on the Azure build agents, it appears that the bounding box of the images
# is slightly different than on local runs, even with equal dpi settings.
# Not able to figure out how to make the run results consistent, hence disable in cloud runs.
assert_binary_files_match(output_folder / f, expected_folder / f)