def test_apply_mask_to_posteriors(image_size: List[int]) -> None:
"""
Test to make sure masks are being applied as expected.
"""
image = np.ones(image_size)
mask_size = image_size[:1] + image_size[-3:] if len(
image_size) == 5 else image_size[-3:]
mask = np.zeros(shape=mask_size)
mask[0] = 1
with pytest.raises(Exception):
# noinspection PyTypeChecker
image_util.apply_mask_to_posteriors(posteriors=None, mask=None)
with pytest.raises(Exception):
# noinspection PyTypeChecker
image_util.apply_mask_to_posteriors(posteriors=image, mask=None)
with pytest.raises(Exception):
# noinspection PyTypeChecker
image_util.apply_mask_to_posteriors(posteriors=None, mask=image)
with pytest.raises(Exception):
# noinspection PyTypeChecker
image_util.apply_mask_to_posteriors(posteriors=image, mask=image)
image = image_util.apply_mask_to_posteriors(posteriors=image, mask=mask)
assert np.all(image[:, 0, ...] == 1)
assert np.all(image[1:, 1:, ...] == 0)
def post_process(self) -> InferenceBatch:
"""
Perform post processing on the computed outputs of the a single pass of the pipelines.
Currently the following operations are performed:
-------------------------------------------------------------------------------------
1) the mask is applied to the posteriors (if required).
2) the final posteriors are used to perform an argmax to generate a multi-label segmentation.
3) extract the largest foreground connected component in the segmentation if required
"""
mask = self.get_component(InferenceBatch.Components.Mask)
posteriors = self.get_component(InferenceBatch.Components.Posteriors)
if mask is not None:
posteriors = image_util.apply_mask_to_posteriors(
posteriors=posteriors, mask=mask)
# create segmentation using an argmax over the posterior probabilities
segmentation = image_util.posteriors_to_segmentation(posteriors)
# update the post-processed posteriors and save the segmentation
self.set_component(component=InferenceBatch.Components.Posteriors,
data=posteriors)
self.set_component(component=InferenceBatch.Components.Segmentation,
data=segmentation)
return self
def _forward_pass(
self,
patches: torch.Tensor,
mask: torch.Tensor = None,
labels: torch.Tensor = None) -> SegmentationForwardPass.Result:
# ensure that we always have float tensors as the model is defined over floats
# and transfer the tensors to the GPU if possible before the forward pass
patches = self.config.get_gpu_tensor_if_possible(patches)
if mask is not None:
mask = self.config.get_gpu_tensor_if_possible(mask)
logits, loss = self._compute_loss(patches, labels)
if self.in_training_mode:
if loss is None:
raise ValueError(
"When running training, the labels must be present for loss computation."
)
assert self.optimizer is not None # for mypy
single_optimizer_step(loss, self.optimizer, self.gradient_scaler)
# Aggregate data parallel logits if multiple hardware are used in forward pass
if isinstance(logits, list):
# When using multiple GPUs, logits is a list of tensors. Gather will concatenate them
# across the first dimension, and move them to GPU0.
logits = torch.nn.parallel.gather(logits, target_device=0)
# apply Softmax on dimension 1 (Class) to map model output into a posterior probability distribution [0,1]
posteriors = torch.nn.functional.softmax(logits, dim=1)
# apply mask if required
if not self.in_training_mode and mask is not None:
posteriors = image_util.apply_mask_to_posteriors(
posteriors=posteriors, mask=mask)
# post process posteriors to compute result
segmentations = image_util.posteriors_to_segmentation(
posteriors=posteriors).data.cpu().numpy()
posteriors = posteriors.data.cpu().numpy()
return SegmentationForwardPass.Result(
posteriors=posteriors,
segmentations=segmentations,
loss=loss.item() if loss is not None else None)
def post_process_posteriors(
self,
posteriors: np.ndarray,
mask: np.ndarray = None) -> Tuple[np.ndarray, np.ndarray]:
"""
Perform post processing on the computed outputs of the a single pass of the pipelines.
Currently the following operations are performed:
-------------------------------------------------------------------------------------
1) the mask is applied to the posteriors (if required).
2) the final posteriors are used to perform an argmax to generate a multi-label segmentation.
3) extract the largest foreground connected component in the segmentation if required
"""
if mask is not None:
posteriors = image_util.apply_mask_to_posteriors(
posteriors=posteriors, mask=mask)
# create segmentation using an argmax over the posterior probabilities
segmentation = image_util.posteriors_to_segmentation(posteriors)
return posteriors, segmentation
def training_or_validation_step(self, sample: Dict[str,
Any], batch_index: int,
is_training: bool) -> torch.Tensor:
"""
Runs training for a single minibatch of training or validation data, and computes all metrics.
:param is_training: If true, the method is called from `training_step`, otherwise it is called from
`validation_step`.
:param sample: The batched sample on which the model should be trained.
:param batch_index: The index of the present batch (supplied only for diagnostics).
"""
cropped_sample: CroppedSample = CroppedSample.from_dict(sample=sample)
# Forward propagation can lead to a model output that is smaller than the input image (crop).
# labels_center_crop is the relevant part of the labels tensor that the model will actually produce.
labels = cropped_sample.labels_center_crop
mask = cropped_sample.mask_center_crop if is_training else None
if is_training:
logits = self.model(cropped_sample.image)
else:
with torch.no_grad():
logits = self.model(cropped_sample.image)
loss = self.loss_fn(logits, labels)
# apply Softmax on dimension 1 (Class) to map model output into a posterior probability distribution [0,1]
posteriors = self.logits_to_posterior(logits)
# apply mask if required
if mask is not None:
posteriors = image_util.apply_mask_to_posteriors(
posteriors=posteriors, mask=mask) # type: ignore
# post process posteriors to compute result
segmentation = image_util.posteriors_to_segmentation(
posteriors=posteriors) # type: ignore
self.compute_metrics(cropped_sample, segmentation,
is_training) # type: ignore
self.write_loss(is_training, loss)
return loss
def inference_identity(
test_output_dirs: OutputFolderForTests,
image_size: Any = (4, 5, 8),
crop_size: Any = (5, 5, 5),
shrink_by: Any = (0, 0, 0),
num_classes: int = 5,
create_mask: bool = True,
extract_largest_foreground_connected_component: bool = False,
is_ensemble: bool = False,
posterior_smoothing_mm: Any = None) -> None:
"""
Test to make sure inference pipeline is identity preserving, ie: we can recreate deterministic
model output, ensuring the patching and stitching is robust.
"""
# fix random seed
np.random.seed(0)
ground_truth_ids = list(map(str, range(num_classes)))
# image to run inference on: The mock model passes the input image through, hence the input
# image must have as many channels as we have classes (plus background), such that the output is
# also a valid posterior.
num_channels = num_classes + 1
image_channels = np.random.randn(num_channels, *list(image_size))
# create a random mask if required
mask = np.round(np.random.uniform(
size=image_size)).astype(np.int) if create_mask else None
config = InferenceIdentityModel(shrink_by=shrink_by)
config.crop_size = crop_size
config.test_crop_size = crop_size
config.image_channels = list(map(str, range(num_channels)))
config.ground_truth_ids = ground_truth_ids
config.posterior_smoothing_mm = posterior_smoothing_mm
# We have to set largest_connected_component_foreground_classes after creating the model config,
# because this parameter is not overridable and hence will not be set by GenericConfig's constructor.
if extract_largest_foreground_connected_component:
config.largest_connected_component_foreground_classes = [
(c, None) for c in ground_truth_ids
]
# set expected posteriors
expected_posteriors = torch.nn.functional.softmax(
torch.tensor(image_channels), dim=0).numpy()
# apply the mask if required
if mask is not None:
expected_posteriors = image_util.apply_mask_to_posteriors(
expected_posteriors, mask)
if posterior_smoothing_mm is not None:
expected_posteriors = image_util.gaussian_smooth_posteriors(
posteriors=expected_posteriors,
kernel_size_mm=posterior_smoothing_mm,
voxel_spacing_mm=(1, 1, 1))
# compute expected segmentation
expected_segmentation = image_util.posteriors_to_segmentation(
expected_posteriors)
if extract_largest_foreground_connected_component:
largest_component = image_util.extract_largest_foreground_connected_component(
multi_label_array=expected_segmentation)
# make sure the test data is accurate by checking if more than one component exists
assert not np.array_equal(largest_component, expected_segmentation)
expected_segmentation = largest_component
# instantiate the model
checkpoint = test_output_dirs.root_dir / "checkpoint.ckpt"
create_model_and_store_checkpoint(config, checkpoint_path=checkpoint)
# create single or ensemble inference pipeline
inference_pipeline = InferencePipeline.create_from_checkpoint(
path_to_checkpoint=checkpoint, model_config=config)
assert inference_pipeline is not None
full_image_inference_pipeline = EnsemblePipeline([inference_pipeline], config) \
if is_ensemble else inference_pipeline
# compute full image inference results
inference_result = full_image_inference_pipeline \
.predict_and_post_process_whole_image(image_channels=image_channels, mask=mask, voxel_spacing_mm=(1, 1, 1))
# Segmentation must have same size as input image
assert inference_result.segmentation.shape == image_size
assert inference_result.posteriors.shape == (num_classes +
1, ) + image_size
# check that the posteriors and segmentations are as expected. Flatten to a list so that the error
# messages are more informative.
assert np.allclose(inference_result.posteriors, expected_posteriors)
assert np.array_equal(inference_result.segmentation, expected_segmentation)
