From e5ce7b9d2b1888bc3e285fc2eb65a5b717d54763 Mon Sep 17 00:00:00 2001
From: Christopher Rhodes <christopher.rhodes@embl.de>
Date: Mon, 30 Oct 2023 13:13:27 +0100
Subject: [PATCH] Re-organized object classifier generator
---
...fer_labels_to_ilastik_object_classifier.py | 141 ++++--------------
extensions/chaeo/models.py | 100 ++++++++++++-
2 files changed, 124 insertions(+), 117 deletions(-)
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 d9696e5a..97330c76 100644
--- a/extensions/chaeo/examples/transfer_labels_to_ilastik_object_classifier.py
+++ b/extensions/chaeo/examples/transfer_labels_to_ilastik_object_classifier.py
@@ -1,107 +1,13 @@
import shutil
from pathlib import Path
import h5py
-import json
import numpy as np
import pandas as pd
import skimage
-import uuid
-from extensions.chaeo.accessors import MonoPatchStack, MonoPatchStackFromFile
-from extensions.chaeo.h5util import get_dataset_info
-from extensions.chaeo.models import PatchStackObjectClassifier
-from model_server.accessors import generate_file_accessor, GenericImageDataAccessor, write_accessor_data_to_file
-
-
-def generate_ilastik_object_classifier(
- template_ilp: Path,
- target_ilp: Path,
- raw_tif: Path,
- mask_tif: Path,
- label_tif: Path,
- label_names: list,
- lane: int = 0,
-):
- """
- Starting with a template project file, transfer input data and labels to a new project file.
- :param template_ilp: path to existing ilastik object classifier to use as a template
- :param target_ilp: path to new classifier
- :param raw_tif: path to stack of patches containing raw data
- :param mask_tif: path to stack of patches containing object masks
- :param label_tif: path to stack of patches containing object labels
- :param label_names: list of label names
- :param lane: ilastik lane identifier
- :return:
- """
- new_ilp = shutil.copy(template_ilp, target_ilp)
-
- paths = {
- 'Raw Data': raw_tif,
- 'Segmentation Image': mask_tif,
- }
- accessors = {k: MonoPatchStackFromFile(root / pa) for k, pa in paths.items()}
-
- # get labels from label image
- acc_labels = MonoPatchStackFromFile(label_tif)
- labels = []
- for ii in range(0, acc_labels.count):
- unique = np.unique(acc_labels.iat(ii))
- assert len(unique) >= 2, 'Label image contains more than one non-zero value'
- assert unique[0] == 0, 'Label image does not contain unlabeled background'
- assert unique[-1] < len(label_names) + 1, f'Label ID {unique[-1]} exceeds number of label names: {len(label_names)}'
- labels.append(unique[-1])
-
- # write to new project file
- with h5py.File(new_ilp, 'r+') as h5:
-
- for gk in ['Raw Data', 'Segmentation Image']:
- group = f'Input Data/infos/lane{lane:04d}/{gk}'
-
- # set path to input image files
- del h5[f'{group}/filePath']
- h5[f'{group}/filePath'] = paths[gk].name
- assert not Path(h5[f'{group}/filePath'][()].decode()).is_absolute()
- assert h5[f'{group}/filePath'][()] == paths[gk].name.encode()
- assert h5[f'{group}/location'][()] == 'FileSystem'.encode()
-
- # set input nickname
- del h5[f'{group}/nickname']
- h5[f'{group}/nickname'] = paths[gk].stem
-
- # set input shape
- del h5[f'{group}/shape']
- shape_zyx = [accessors[gk].shape_dict[ax] for ax in ['Z', 'Y', 'X']]
- h5[f'{group}/shape'] = np.array(shape_zyx)
-
- # change key of label names
- if (k := 'ObjectClassification/LabelNames') in h5.keys():
- del h5[k]
- ln = np.array(label_names)
- h5.create_dataset(k, data=ln.astype('O'))
-
- if (k := 'ObjectClassification/MaxNumObj') in h5.keys():
- del h5[k]
- h5[k] = len(label_names) - 1
-
- del h5['currentApplet']
- h5['currentApplet'] = 1
-
- # change object labels
- if (k := f'ObjectClassification/LabelInputs/{lane:04d}') in h5.keys():
- del h5[k]
- lag = h5.create_group(k)
- # for zi in range(0, nz):
- # lag[f'{zi}'] = np.array([0., float(get_label(zi))])
- for zi, la in enumerate(labels):
- lag[f'{zi}'] = np.array([0., float(la)])
-
- # delete existing classification weights
- if (k := f'ObjectExtraction/RegionFeatures/{lane:04d}') in h5.keys():
- del h5[k]
- if (k := 'ObjectClassification/ClassifierForests') in h5.keys():
- del h5[k]
-
- return new_ilp
+from extensions.chaeo.accessors import MonoPatchStackFromFile
+from extensions.chaeo.models import generate_ilastik_object_classifier, PatchStackObjectClassifier
+from model_server.accessors import GenericImageDataAccessor, write_accessor_data_to_file
def compare_object_maps(truth: GenericImageDataAccessor, inferred: GenericImageDataAccessor) -> pd.DataFrame:
@@ -137,6 +43,17 @@ def compare_object_maps(truth: GenericImageDataAccessor, inferred: GenericImageD
labels.append(dd)
return pd.DataFrame(labels)
+def infer_and_compare(classifier: PatchStackObjectClassifier, prefix, raw, mask, labels):
+ result_acc, _ = classifier.infer(raw, mask)
+ write_accessor_data_to_file(root / f'zstack_train_result.tif', result_acc)
+
+ # write comparison tables
+ df_comp = compare_object_maps(labels, result_acc)
+ df_comp.to_csv(root / f'compare_{prefix}_result.csv', index=False)
+ print(f'Generated ilastik project {classifier_file.name}')
+ print('Truth and inferred labels match?')
+ print(pd.value_counts(df_comp['truth_label'] == df_comp['inferred_label']))
+
if __name__ == '__main__':
root = Path('c:/Users/rhodes/projects/proj0011-plankton-seg/exp0009/output/labeled_patches-20231030-0002')
template_ilp = Path('c:/Users/rhodes/projects/proj0011-plankton-seg/exp0014/template_obj.ilp')
@@ -147,7 +64,7 @@ if __name__ == '__main__':
assert len(label_names) >= 2
# auto-populate an object classifier
- new_ilp = generate_ilastik_object_classifier(
+ classifier_file = generate_ilastik_object_classifier(
template_ilp,
root / 'new_auto_obj.ilp',
root / 'zstack_train_raw.tif',
@@ -155,27 +72,23 @@ if __name__ == '__main__':
root / 'zstack_train_label.tif',
label_names,
)
+ classifier = PatchStackObjectClassifier({'project_file': classifier_file})
- auto_ilp = new_ilp
-
- def infer_and_compare(ilp, prefix, raw, mask, labels):
- mod = PatchStackObjectClassifier({'project_file': root / ilp})
- result_acc, _ = mod.infer(raw, mask)
- write_accessor_data_to_file(root / f'zstack_train_result.tif', result_acc)
-
- # write comparison tables
- df_comp = compare_object_maps(labels, result_acc)
- df_comp.to_csv(root / f'compare_{prefix}_result.csv', index=False)
- print(f'Generated ilastik project {ilp}')
- print('Truth and inferred labels match?')
- print(pd.value_counts(df_comp['truth_label'] == df_comp['inferred_label']))
-
- # infer_and_compare_training_set(auto_ilp, 'before')
+ # verify self-consistency of training set
infer_and_compare(
- auto_ilp,
+ classifier,
'train',
MonoPatchStackFromFile(root / 'zstack_train_raw.tif'),
MonoPatchStackFromFile(root / 'zstack_train_mask.tif'),
MonoPatchStackFromFile(root / 'zstack_train_label.tif')
)
+ # run test set
+ infer_and_compare(
+ classifier,
+ 'test',
+ MonoPatchStackFromFile(root / 'zstack_test_raw.tif'),
+ MonoPatchStackFromFile(root / 'zstack_test_mask.tif'),
+ MonoPatchStackFromFile(root / 'zstack_test_label.tif'),
+ )
+
diff --git a/extensions/chaeo/models.py b/extensions/chaeo/models.py
index 8a141b35..9e209be2 100644
--- a/extensions/chaeo/models.py
+++ b/extensions/chaeo/models.py
@@ -1,7 +1,11 @@
+from pathlib import Path
+import shutil
+
+import h5py
import numpy as np
import vigra
-from extensions.chaeo.accessors import MonoPatchStack
+from extensions.chaeo.accessors import MonoPatchStack, MonoPatchStackFromFile
from extensions.ilastik.models import IlastikObjectClassifierFromSegmentationModel
from model_server.accessors import InMemoryDataAccessor
@@ -12,7 +16,6 @@ class PatchStackObjectClassifier(IlastikObjectClassifierFromSegmentationModel):
as time-series images where each frame contains only one object.
"""
-
def infer(self, input_acc: MonoPatchStack, segmentation_acc: MonoPatchStack) -> (np.ndarray, dict):
assert segmentation_acc.is_mask()
assert input_acc.chroma == 1
@@ -41,4 +44,95 @@ class PatchStackObjectClassifier(IlastikObjectClassifierFromSegmentationModel):
assert yxz.shape[0:2] == input_acc.hw
assert yxz.shape[2] == input_acc.nz
- return MonoPatchStack(data=yxz), {'success': True}
\ No newline at end of file
+ return MonoPatchStack(data=yxz), {'success': True}
+
+
+def generate_ilastik_object_classifier(
+ template_ilp: Path,
+ target_ilp: Path,
+ raw_tif: Path,
+ mask_tif: Path,
+ label_tif: Path,
+ label_names: list,
+ lane: int = 0,
+) -> Path:
+ """
+ Starting with a template project file, transfer input data and labels to a new project file.
+ :param template_ilp: path to existing ilastik object classifier to use as a template
+ :param target_ilp: path to new classifier
+ :param raw_tif: path to stack of patches containing raw data
+ :param mask_tif: path to stack of patches containing object masks
+ :param label_tif: path to stack of patches containing object labels
+ :param label_names: list of label names
+ :param lane: ilastik lane identifier
+ :return: path to generated object classifier
+ """
+ new_ilp = shutil.copy(template_ilp, target_ilp)
+
+ paths = {
+ 'Raw Data': raw_tif,
+ 'Segmentation Image': mask_tif,
+ }
+ root = raw_tif.parent
+ accessors = {k: MonoPatchStackFromFile(root / pa) for k, pa in paths.items()}
+
+ # get labels from label image
+ acc_labels = MonoPatchStackFromFile(label_tif)
+ labels = []
+ for ii in range(0, acc_labels.count):
+ unique = np.unique(acc_labels.iat(ii))
+ assert len(unique) >= 2, 'Label image contains more than one non-zero value'
+ assert unique[0] == 0, 'Label image does not contain unlabeled background'
+ assert unique[-1] < len(label_names) + 1, f'Label ID {unique[-1]} exceeds number of label names: {len(label_names)}'
+ labels.append(unique[-1])
+
+ # write to new project file
+ with h5py.File(new_ilp, 'r+') as h5:
+
+ for gk in ['Raw Data', 'Segmentation Image']:
+ group = f'Input Data/infos/lane{lane:04d}/{gk}'
+
+ # set path to input image files
+ del h5[f'{group}/filePath']
+ h5[f'{group}/filePath'] = paths[gk].name
+ assert not Path(h5[f'{group}/filePath'][()].decode()).is_absolute()
+ assert h5[f'{group}/filePath'][()] == paths[gk].name.encode()
+ assert h5[f'{group}/location'][()] == 'FileSystem'.encode()
+
+ # set input nickname
+ del h5[f'{group}/nickname']
+ h5[f'{group}/nickname'] = paths[gk].stem
+
+ # set input shape
+ del h5[f'{group}/shape']
+ shape_zyx = [accessors[gk].shape_dict[ax] for ax in ['Z', 'Y', 'X']]
+ h5[f'{group}/shape'] = np.array(shape_zyx)
+
+ # change key of label names
+ if (k := 'ObjectClassification/LabelNames') in h5.keys():
+ del h5[k]
+ ln = np.array(label_names)
+ h5.create_dataset(k, data=ln.astype('O'))
+
+ if (k := 'ObjectClassification/MaxNumObj') in h5.keys():
+ del h5[k]
+ h5[k] = len(label_names) - 1
+
+ del h5['currentApplet']
+ h5['currentApplet'] = 1
+
+ # change object labels
+ if (k := f'ObjectClassification/LabelInputs/{lane:04d}') in h5.keys():
+ del h5[k]
+ lag = h5.create_group(k)
+ for zi, la in enumerate(labels):
+ lag[f'{zi}'] = np.array([0., float(la)])
+
+ # delete existing classification weights
+ if (k := f'ObjectExtraction/RegionFeatures/{lane:04d}') in h5.keys():
+ del h5[k]
+ if (k := 'ObjectClassification/ClassifierForests') in h5.keys():
+ del h5[k]
+
+ return Path(new_ilp)
+
--
GitLab