diff --git a/extensions/chaeo/examples/transfer_labels_to_ilastik_object_classifier.py b/extensions/chaeo/examples/transfer_labels_to_ilastik_object_classifier.py
index d79127c39825cafc56866cc6c4c3711801b404a1..a02840be65598df36065a37a750eb9597a707ee2 100644
--- a/extensions/chaeo/examples/transfer_labels_to_ilastik_object_classifier.py
+++ b/extensions/chaeo/examples/transfer_labels_to_ilastik_object_classifier.py
@@ -13,9 +13,13 @@ from extensions.ilastik.models import IlastikObjectClassifierFromSegmentationMod
from model_server.accessors import generate_file_accessor, GenericImageDataAccessor, InMemoryDataAccessor, write_accessor_data_to_file
class PatchStackObjectClassifier(IlastikObjectClassifierFromSegmentationModel):
+ """
+ Wrap ilastik object classification for inputs comprising raw image and binary segmentation masks, both represented
+ as time-series images where each frame contains only one object.
+ """
@staticmethod
- def make_tczyx(acc):
+ def make_tczyx(acc: GenericImageDataAccessor):
assert acc.chroma == 1
tyx = np.moveaxis(
acc.data[:, :, 0, :], # YX(C)Z
@@ -23,7 +27,6 @@ class PatchStackObjectClassifier(IlastikObjectClassifierFromSegmentationModel):
[0, 1, 2]
)
return np.expand_dims(tyx, (1, 2))
- # return tyx
def infer(self, input_img: GenericImageDataAccessor, segmentation_img: GenericImageDataAccessor) -> (np.ndarray, dict):
assert segmentation_img.is_mask()
@@ -39,11 +42,11 @@ class PatchStackObjectClassifier(IlastikObjectClassifierFromSegmentationModel):
}
]
- obmaps = self.shell.workflow.batchProcessingApplet.run_export(dsi, export_to_array=True) # [z x h x w x n]
+ obmaps = self.shell.workflow.batchProcessingApplet.run_export(dsi, export_to_array=True) # [z x h x w x n]
assert len(obmaps) == 1, 'ilastik generated more than one object map'
- # for some reason these axes get scrambled to Z(1)YX(1)
+ # for some reason ilastik scrambles these axes to Z(1)YX(1)
assert obmaps[0].shape == (input_img.nz, 1, input_img.hw[0], input_img.hw[1], 1)
yxcz = np.moveaxis(
obmaps[0][:, :, :, :, 0],
@@ -54,7 +57,13 @@ class PatchStackObjectClassifier(IlastikObjectClassifierFromSegmentationModel):
assert yxcz.shape == input_img.shape
return InMemoryDataAccessor(data=yxcz), {'success': True}
-def get_dataset_info(h5, lane=0):
+def get_dataset_info(h5: h5py.File, lane : int = 0):
+ """
+ Report out specific datasets in ilastik project file HDF5
+ :param h5: handle to ilastik project file, as h5py.File object
+ :param lane: ilastik lane identifier
+ :return: (dict) selected data values from project file
+ """
lns = f'{lane:04d}'
lane = f'Input Data/infos/lane{lns}'
info = {}
@@ -79,13 +88,22 @@ def get_dataset_info(h5, lane=0):
return info
-def generate_ilastik_object_classifier(template_ilp, where: str, lane=0):
+def generate_ilastik_object_classifier(template_ilp: str, where: str, stack_name: str = 'train', lane: int = 0):
+ """
+ Starting with a template project file, transfer input data and labels to a duplicate project file.
+
+ :param template_ilp: absolute path to existing ilastik object classifier to use as a template
+ :param where: location of folder containing input data, segmentation maps, labels, and label descriptions
+ :poram stack_name: prefix of .tif and .csv files that contain classifier training data (e.g. train, test)
+ :param lane: ilastik lane identifier
+ :return: (str) name of new ilastik classifier project file
+ """
# validate z-stack input data
root = Path(where)
paths = {
- 'Raw Data': root / 'zstack_train_raw.tif',
- 'Segmentation Image': root / 'zstack_train_mask.tif',
+ 'Raw Data': root / f'zstack_{stack_name}_raw.tif',
+ 'Segmentation Image': root / f'zstack_{stack_name}_mask.tif',
}
accessors = {k: generate_file_accessor(pa) for k, pa in paths.items()}
@@ -97,9 +115,9 @@ def generate_ilastik_object_classifier(template_ilp, where: str, lane=0):
nz = accessors['Raw Data'].nz
# now load CSV
- csv_path = root / 'train_stack.csv'
+ csv_path = root / f'{stack_name}_stack.csv'
assert csv_path.exists()
- df_patches = pd.read_csv(root / 'train_stack.csv')
+ df_patches = pd.read_csv(csv_path)
assert np.all(
df_patches['zi'].sort_values().to_numpy() == np.arange(0, nz)
)
@@ -142,7 +160,6 @@ def generate_ilastik_object_classifier(template_ilp, where: str, lane=0):
# change object labels
la_groupname = f'ObjectClassification/LabelInputs/{lns}'
-
del h5[la_groupname]
lag = h5.create_group(la_groupname)
for zi in range(0, nz):
@@ -151,6 +168,12 @@ def generate_ilastik_object_classifier(template_ilp, where: str, lane=0):
return new_ilp
def compare_object_maps(truth: GenericImageDataAccessor, inferred: GenericImageDataAccessor) -> pd.DataFrame:
+ """
+ Compare two object maps to assess classification results
+ :param truth: t-stack of truth objects
+ :param inferred: t-stack of inferred objects, presumably with same segmentation boundaries as truth
+ :return: DataFrame comparing results for each frame in truth and inferred stacks
+ """
assert truth.shape == inferred.shape
assert np.all((truth.data == 0) == (inferred.data == 0))
assert inferred.chroma == 1
@@ -180,25 +203,43 @@ def compare_object_maps(truth: GenericImageDataAccessor, inferred: GenericImageD
if __name__ == '__main__':
root = Path('c:/Users/rhodes/projects/proj0011-plankton-seg/')
template_ilp = root / 'exp0014/template_obj.ilp'
- # template_ilp = root / 'exp0014/test_obj_from_seg.ilp'
- where_patch_stack = root / 'exp0009/output/labeled_patches-20231016-0002'
+ where_patch_stack = root / 'exp0009/output/labeled_patches-20231018-0000'
+ # auto-populate an object classifier
new_ilp = generate_ilastik_object_classifier(
template_ilp,
where_patch_stack,
+ stack_name='train'
)
- train_zstack_raw = generate_file_accessor(where_patch_stack / 'zstack_train_raw.tif')
- train_zstack_mask = generate_file_accessor(where_patch_stack / 'zstack_train_mask.tif')
+ def infer_and_compare(suffix):
+ # infer object labels from the same data used to train the classifier
+ train_zstack_raw = generate_file_accessor(where_patch_stack / 'zstack_train_raw.tif')
+ train_zstack_mask = generate_file_accessor(where_patch_stack / 'zstack_train_mask.tif')
+ mod = PatchStackObjectClassifier({'project_file': where_patch_stack / new_ilp})
+ result_acc, _ = mod.infer(train_zstack_raw, train_zstack_mask)
+ write_accessor_data_to_file(where_patch_stack / f'zstack_train_result_{suffix}.tif', result_acc)
+
+ # write comparison tables
+ train_truth_labels = generate_file_accessor(where_patch_stack / f'zstack_train_label.tif')
+ df_comp = compare_object_maps(train_truth_labels, result_acc)
+ df_comp.to_csv(
+ where_patch_stack / autonumber_new_file(
+ where_patch_stack, f'compare_train_result_{suffix}', 'csv'
+ ),
+ index=False
+ )
+ print('Truth and inferred labels match?')
+ print(pd.value_counts(df_comp['truth_label'] == df_comp['inferred_label']))
+
+ # infer object labels from the same data used to train the classifier
+ infer_and_compare('before')
- mod = PatchStackObjectClassifier({'project_file': new_ilp})
+ # prompt user input when ilastik file has been modified in-app
+ print(f'Press enter when project file {new_ilp} has been updated in ilastik')
+ input()
- result_acc, _ = mod.infer(train_zstack_raw, train_zstack_mask)
- write_accessor_data_to_file(where_patch_stack / 'result.tif', result_acc)
- print(where_patch_stack / 'result.tif')
+ # repeat inference with the same project file, but a fresh model handle
+ infer_and_compare('after')
- # write comparison
- train_labels = generate_file_accessor(where_patch_stack / 'zstack_train_label.tif')
- df_comp = compare_object_maps(train_labels, result_acc)
- df_comp.to_csv(where_patch_stack / autonumber_new_file(where_patch_stack, 'comp', 'csv'), index=False)