def extract_images_or_rois(parameters):
id_annotations = parse_domain_list(parameters.cytomine_roi_annotations)
# if ROI annotations are provided
if len(id_annotations) > 0:
image_cache = dict() # maps ImageInstance id with CytomineSlide object
zones = list()
for id_annot in id_annotations:
annotation = Annotation().fetch(id_annot)
if annotation.image not in image_cache:
image_cache[annotation.image] = CytomineSlide(
annotation.image, parameters.cytomine_zoom_level)
window = get_iip_window_from_annotation(
image_cache[annotation.image], annotation,
parameters.cytomine_zoom_level)
zones.append(window)
return zones
# work at image level or ROIs by term
images = ImageInstanceCollection()
if parameters.cytomine_id_images is not None:
id_images = parse_domain_list(parameters.cytomine_id_images)
images.extend([ImageInstance().fetch(_id) for _id in id_images])
else:
images = images.fetch_with_filter("project",
parameters.cytomine_id_project)
slides = [
CytomineSlide(img, parameters.cytomine_zoom_level) for img in images
]
if parameters.cytomine_id_roi_term is None:
return slides
# fetch ROI annotations
collection = AnnotationCollection(
terms=[parameters.cytomine_id_roi_term],
reviewed=parameters.cytomine_reviewed_roi,
showWKT=True)
collection.fetch_with_filter(project=parameters.cytomine_id_project)
slides_map = {slide.image_instance.id: slide for slide in slides}
regions = list()
for annotation in collection:
if annotation.image not in slides_map:
continue
slide = slides_map[annotation.image]
regions.append(
get_iip_window_from_annotation(slide, annotation,
parameters.cytomine_zoom_level))
return regions
def extract_images_or_rois(parameters):
# work at image level or ROIs by term
images = ImageInstanceCollection()
if parameters.cytomine_id_images is not None:
id_images = parse_domain_list(parameters.cytomine_id_images)
images.extend([ImageInstance().fetch(_id) for _id in id_images])
else:
images = images.fetch_with_filter("project",
parameters.cytomine_id_project)
slides = [
CytomineSlide(img, parameters.cytomine_zoom_level) for img in images
]
if parameters.cytomine_id_roi_term is None:
return slides
# fetch ROI annotations, all users
collection = AnnotationCollection(
terms=[parameters.cytomine_id_roi_term],
reviewed=parameters.cytomine_reviewed_roi,
project=parameters.cytomine_id_project,
showWKT=True,
includeAlgo=True).fetch()
slides_map = {slide.image_instance.id: slide for slide in slides}
regions = list()
for annotation in collection:
if annotation.image not in slides_map:
continue
slide = slides_map[annotation.image]
regions.append(
get_iip_window_from_annotation(slide, annotation,
parameters.cytomine_zoom_level))
return regions
def main(argv):
with CytomineJob.from_cli(argv) as cj:
# annotation filtering
cj.logger.info(str(cj.parameters))
# use only images from the current project
cj.parameters.cytomine_id_projects = "{}".format(cj.parameters.cytomine_id_project)
cj.job.update(progress=1, statuscomment="Preparing execution (creating folders,...).")
base_path, downloaded = setup_classify(
args=cj.parameters, logger=cj.job_logger(1, 40),
dest_pattern=os.path.join("{term}", "{image}_{id}.png"),
root_path=Path.home(), set_folder="train", showTerm=True
)
x = np.array([f for annotation in downloaded for f in annotation.filenames])
y = np.array([int(os.path.basename(os.path.dirname(filepath))) for filepath in x])
# transform classes
cj.job.update(progress=50, statusComment="Transform classes...")
classes = parse_domain_list(cj.parameters.cytomine_id_terms)
positive_classes = parse_domain_list(cj.parameters.cytomine_positive_terms)
classes = np.array(classes) if len(classes) > 0 else np.unique(y)
n_classes = classes.shape[0]
# filter unwanted terms
cj.logger.info("Size before filtering:")
cj.logger.info(" - x: {}".format(x.shape))
cj.logger.info(" - y: {}".format(y.shape))
keep = np.in1d(y, classes)
x, y = x[keep], y[keep]
cj.logger.info("Size after filtering:")
cj.logger.info(" - x: {}".format(x.shape))
cj.logger.info(" - y: {}".format(y.shape))
if cj.parameters.cytomine_binary:
cj.logger.info("Will be training on 2 classes ({} classes before binarization).".format(n_classes))
y = np.in1d(y, positive_classes).astype(np.int)
else:
cj.logger.info("Will be training on {} classes.".format(n_classes))
y = np.searchsorted(classes, y)
# build model
cj.job.update(progress=55, statusComment="Build model...")
_, pyxit = build_models(
n_subwindows=cj.parameters.pyxit_n_subwindows,
min_size=cj.parameters.pyxit_min_size,
max_size=cj.parameters.pyxit_max_size,
target_width=cj.parameters.pyxit_target_width,
target_height=cj.parameters.pyxit_target_height,
interpolation=cj.parameters.pyxit_interpolation,
transpose=cj.parameters.pyxit_transpose,
colorspace=cj.parameters.pyxit_colorspace,
fixed_size=cj.parameters.pyxit_fixed_size,
verbose=int(cj.logger.level == 10),
create_svm=cj.parameters.svm,
C=cj.parameters.svm_c,
random_state=cj.parameters.seed,
n_estimators=cj.parameters.forest_n_estimators,
min_samples_split=cj.parameters.forest_min_samples_split,
max_features=cj.parameters.forest_max_features,
n_jobs=cj.parameters.n_jobs
)
cj.job.update(progress=60, statusComment="Train model...")
pyxit.fit(x, y)
cj.job.update(progress=90, statusComment="Save model....")
model_filename = joblib.dump(pyxit, os.path.join(base_path, "model.joblib"), compress=3)[0]
AttachedFile(
cj.job,
domainIdent=cj.job.id,
filename=model_filename,
domainClassName="be.cytomine.processing.Job"
).upload()
Property(cj.job, key="classes", value=stringify(classes)).save()
Property(cj.job, key="binary", value=cj.parameters.cytomine_binary).save()
Property(cj.job, key="positive_classes", value=stringify(positive_classes)).save()
cj.job.update(status=Job.TERMINATED, status_comment="Finish", progress=100)
def main(argv):
with CytomineJob.from_cli(argv) as cj:
# use only images from the current project
cj.job.update(progress=1, statusComment="Preparing execution")
# extract images to process
if cj.parameters.cytomine_zoom_level > 0 and (
cj.parameters.cytomine_tile_size != 256
or cj.parameters.cytomine_tile_overlap != 0):
raise ValueError(
"when using zoom_level > 0, tile size should be 256 "
"(given {}) and overlap should be 0 (given {})".format(
cj.parameters.cytomine_tile_size,
cj.parameters.cytomine_tile_overlap))
cj.job.update(
progress=1,
statusComment="Preparing execution (creating folders,...).")
# working path
root_path = str(Path.home())
working_path = os.path.join(root_path, "images")
os.makedirs(working_path, exist_ok=True)
# load training information
cj.job.update(progress=5,
statusComment="Extract properties from training job.")
train_job = Job().fetch(cj.parameters.cytomine_id_job)
properties = PropertyCollection(train_job).fetch().as_dict()
binary = str2bool(properties["binary"].value)
classes = parse_domain_list(properties["classes"].value)
cj.job.update(progress=10, statusComment="Download the model file.")
attached_files = AttachedFileCollection(train_job).fetch()
model_file = attached_files.find_by_attribute("filename",
"model.joblib")
model_filepath = os.path.join(root_path, "model.joblib")
model_file.download(model_filepath, override=True)
pyxit = joblib.load(model_filepath)
# set n_jobs
pyxit.base_estimator.n_jobs = cj.parameters.n_jobs
pyxit.n_jobs = cj.parameters.n_jobs
cj.job.update(progress=45, statusComment="Build workflow.")
builder = SSLWorkflowBuilder()
builder.set_tile_size(cj.parameters.cytomine_tile_size,
cj.parameters.cytomine_tile_size)
builder.set_overlap(cj.parameters.cytomine_tile_overlap)
builder.set_tile_builder(
CytomineTileBuilder(working_path, n_jobs=cj.parameters.n_jobs))
builder.set_logger(StandardOutputLogger(level=Logger.INFO))
builder.set_n_jobs(1)
builder.set_background_class(0)
# value 0 will prevent merging but still requires to run the merging check
# procedure (inefficient)
builder.set_distance_tolerance(2 if cj.parameters.union_enabled else 0)
builder.set_segmenter(
ExtraTreesSegmenter(
pyxit=pyxit,
classes=classes,
prediction_step=cj.parameters.pyxit_prediction_step,
background=0,
min_std=cj.parameters.tile_filter_min_stddev,
max_mean=cj.parameters.tile_filter_max_mean))
workflow = builder.get()
area_checker = AnnotationAreaChecker(
min_area=cj.parameters.min_annotation_area,
max_area=cj.parameters.max_annotation_area)
def get_term(label):
if binary:
if "cytomine_id_predict_term" not in cj.parameters:
return []
else:
return [int(cj.parameters.cytomine_id_predict_term)]
# multi-class
return [label]
zones = extract_images_or_rois(cj.parameters)
for zone in cj.monitor(zones,
start=50,
end=90,
period=0.05,
prefix="Segmenting images/ROIs"):
results = workflow.process(zone)
annotations = AnnotationCollection()
for obj in results:
if not area_checker.check(obj.polygon):
continue
polygon = obj.polygon
if isinstance(zone, ImageWindow):
polygon = affine_transform(
polygon,
[1, 0, 0, 1, zone.abs_offset_x, zone.abs_offset_y])
polygon = change_referential(polygon, zone.base_image.height)
if cj.parameters.cytomine_zoom_level > 0:
zoom_mult = (2**cj.parameters.cytomine_zoom_level)
polygon = affine_transform(
polygon, [zoom_mult, 0, 0, zoom_mult, 0, 0])
annotations.append(
Annotation(location=polygon.wkt,
id_terms=get_term(obj.label),
id_project=cj.project.id,
id_image=zone.base_image.image_instance.id))
annotations.save()
cj.job.update(status=Job.TERMINATED,
status_comment="Finish",
progress=100)
def main(argv):
with CytomineJob.from_cli(argv) as cj:
# annotation filtering
cj.logger.info(str(cj.parameters))
cj.job.update(progress=1, statuscomment="Preparing execution (creating folders,...).")
base_path, downloaded = setup_classify(
args=cj.parameters, logger=cj.job_logger(1, 40),
dest_pattern=os.path.join("{term}", "{image}_{id}.png"),
root_path=Path.home(), set_folder="train", showTerm=True
)
x = np.array([f for annotation in downloaded for f in annotation.filenames])
y = np.array([int(os.path.basename(os.path.dirname(filepath))) for filepath in x])
# transform classes
cj.job.update(progress=50, statusComment="Transform classes...")
classes = parse_domain_list(cj.parameters.cytomine_id_terms)
positive_classes = parse_domain_list(cj.parameters.cytomine_positive_terms)
classes = np.array(classes) if len(classes) > 0 else np.unique(y)
n_classes = classes.shape[0]
# filter unwanted terms
cj.logger.info("Size before filtering:")
cj.logger.info(" - x: {}".format(x.shape))
cj.logger.info(" - y: {}".format(y.shape))
keep = np.in1d(y, classes)
x, y = x[keep], y[keep]
cj.logger.info("Size after filtering:")
cj.logger.info(" - x: {}".format(x.shape))
cj.logger.info(" - y: {}".format(y.shape))
labels = np.array([int(os.path.basename(f).split("_", 1)[0]) for f in x])
if cj.parameters.cytomine_binary:
cj.logger.info("Will be training on 2 classes ({} classes before binarization).".format(n_classes))
y = np.in1d(y, positive_classes).astype(np.int)
else:
cj.logger.info("Will be training on {} classes.".format(n_classes))
y = np.searchsorted(classes, y)
# build model
random_state = check_random_state(cj.parameters.seed)
cj.job.update(progress=55, statusComment="Build model...")
_, pyxit = build_models(
n_subwindows=cj.parameters.pyxit_n_subwindows,
min_size=cj.parameters.pyxit_min_size,
max_size=cj.parameters.pyxit_max_size,
target_width=cj.parameters.pyxit_target_width,
target_height=cj.parameters.pyxit_target_height,
interpolation=cj.parameters.pyxit_interpolation,
transpose=cj.parameters.pyxit_transpose,
colorspace=cj.parameters.pyxit_colorspace,
fixed_size=cj.parameters.pyxit_fixed_size,
verbose=int(cj.logger.level == 10),
create_svm=cj.parameters.svm,
C=cj.parameters.svm_c,
random_state=random_state,
n_estimators=cj.parameters.forest_n_estimators,
min_samples_split=cj.parameters.forest_min_samples_split,
max_features=cj.parameters.forest_max_features,
n_jobs=cj.parameters.n_jobs
)
cj.job.update(progress=60, statusComment="Start cross-validation...")
n_splits = cj.parameters.eval_k
cv = ShuffleSplit(n_splits, test_size=cj.parameters.eval_test_fraction)
if cj.parameters.folds == "group":
cv = GroupKFold(n_splits)
elif cj.parameters.folds == "stratified":
cv = StratifiedKFold(n_splits, shuffle=True, random_state=random_state)
elif cj.parameters.folds != "shuffle":
raise ValueError("Unknown folding policy '{}'.".format(cj.parameters.folds))
# Fit
accuracies = np.zeros(n_splits)
test_sizes = np.zeros(n_splits)
_x, _y = pyxit.extract_subwindows(x, y)
# CV loop
for i, (train, test) in cj.monitor(enumerate(cv.split(x, y, labels)), start=60, end=90, prefix="cross val. iteration"):
_pyxit = clone(pyxit)
w_train = window_indexes(x.shape[0], train, _pyxit.n_subwindows)
w_test = window_indexes(x.shape[0], test, _pyxit.n_subwindows)
_pyxit.fit(x[train], y[train], _X=_x[w_train], _y=_y[w_train])
y_pred = _pyxit.predict(x[test], _x[w_test])
accuracies[i] = accuracy_score(y[test], y_pred)
test_sizes[i] = test.shape[0] / float(x.shape[0])
del _pyxit
pyxit.fit(x, y)
accuracy = float(np.mean(test_sizes * accuracies))
cj.job.update(progress=90, statusComment="Accuracy: {}".format(accuracy))
cj.job.update(progress=90, statusComment="Save model...")
model_filename = joblib.dump(pyxit, os.path.join(base_path, "model.joblib"), compress=3)[0]
AttachedFile(
cj.job,
domainIdent=cj.job.id,
filename=model_filename,
domainClassName="be.cytomine.processing.Job"
).upload()
Property(cj.job, key="classes", value=stringify(classes)).save()
Property(cj.job, key="binary", value=cj.parameters.cytomine_binary).save()
Property(cj.job, key="positive_classes", value=stringify(positive_classes)).save()
Property(cj.job, key="accuracies", value=array2str(accuracies))
Property(cj.job, key="test_sizes", value=array2str(test_sizes))
Property(cj.job, key="accuracy", value=accuracy)
cj.job.update(status=Job.TERMINATED, status_comment="Finish", progress=100)
def main(argv):
with CytomineJob.from_cli(argv) as cj:
# annotation filtering
cj.logger.info(str(cj.parameters))
# use only images from the current project
cj.parameters.cytomine_id_projects = "{}".format(cj.parameters.cytomine_id_project)
cj.job.update(progress=1, statuscomment="Preparing execution (creating folders,...).")
root_path = "/data/" #Path.home()
image_path, downloaded = setup_classify(
args=cj.parameters, logger=cj.job_logger(1, 40),
dest_pattern="{image}_{id}.png", root_path=root_path,
set_folder="test", showWKT=True
)
annotations = [annotation for annotation in downloaded for f in annotation.filenames]
x = np.array([f for annotation in downloaded for f in annotation.filenames])
# extract model data from previous job
cj.job.update(progress=45, statusComment="Extract properties from training job.")
train_job = Job().fetch(cj.parameters.cytomine_id_job)
properties = PropertyCollection(train_job).fetch().as_dict()
binary = str2bool(properties["binary"].value)
if binary:
classes = np.array([cj.parameters.cytomine_id_term_negative, cj.parameters.cytomine_id_term_positive])
else:
classes = np.array(parse_domain_list(properties["classes"].value))
# extract model
cj.job.update(progress=50, statusComment="Download the model file.")
attached_files = AttachedFileCollection(train_job).fetch()
model_file = attached_files.find_by_attribute("filename", "model.joblib")
model_filepath = os.path.join(root_path, "model.joblib")
model_file.download(model_filepath, override=True)
pyxit = joblib.load(model_filepath)
# set n_jobs
pyxit.base_estimator.n_jobs = cj.parameters.n_jobs
pyxit.n_jobs = cj.parameters.n_jobs
cj.job.update(progress=55, statusComment="Predict...")
if hasattr(pyxit, "predict_proba"):
probas = pyxit.predict_proba(x)
y_pred = np.argmax(probas, axis=1)
else:
probas = [None] * x.shape[0]
y_pred = pyxit.predict(x)
predicted_terms = classes.take(y_pred, axis=0)
collection = AnnotationCollection()
for i in cj.monitor(range(x.shape[0]), start=80, end=99, period=0.005, prefix="Uploading predicted terms"):
annot, term, proba = annotations[i], predicted_terms[i], probas[i]
parameters = {
"location": annot.location,
"id_image": annot.image,
"id_project": cj.project.id,
"id_terms": [int(term)]
}
if proba is not None:
parameters["rate"] = float(np.max(proba))
collection.append(Annotation(**parameters))
collection.save()
cj.job.update(status=Job.TERMINATED, status_comment="Finish", progress=100)
def main(argv):
with CytomineJob.from_cli(argv) as cj:
# use only images from the current project
cj.job.update(
progress=1,
statusComment="Preparing execution (creating folders,...).")
# hardcode parameter for setup classify to fetch alphamask instead of plain crop.
cj.parameters.cytomine_download_alpha = True
cj.parameters.cytomine_id_projects = "{}".format(cj.project.id)
cj.job.update(progress=2, statusComment="Downloading crops.")
base_path, downloaded = setup_classify(args=cj.parameters,
logger=cj.job_logger(2, 25),
dest_pattern=os.path.join(
"{term}",
"{image}_{id}.png"),
root_path=str("tmp"),
set_folder="train",
showTerm=True)
x = np.array(
[f for annotation in downloaded for f in annotation.filenames])
y = np.array([
int(os.path.basename(os.path.dirname(filepath))) for filepath in x
])
# transform classes
cj.job.update(progress=25, statusComment="Transform classes...")
positive_terms = parse_domain_list(
cj.parameters.cytomine_id_positive_terms)
selected_terms = parse_domain_list(cj.parameters.cytomine_id_terms)
is_binary = len(selected_terms) > 0 and len(positive_terms) > 0
foreground_terms = np.unique(y) if len(
selected_terms) == 0 else np.array(selected_terms)
if len(positive_terms) == 0:
classes = np.hstack((np.zeros((1, ), dtype=int), foreground_terms))
else: # binary
foreground_terms = np.array(positive_terms)
classes = np.array([0, 1])
# cast to binary
fg_idx = np.in1d(y, list(foreground_terms))
bg_idx = np.in1d(
y, list(set(selected_terms).difference(foreground_terms)))
y[fg_idx] = 1
y[bg_idx] = 0
n_classes = classes.shape[0]
# filter unwanted terms
cj.logger.info("Size before filtering:")
cj.logger.info(" - x: {}".format(x.shape))
cj.logger.info(" - y: {}".format(y.shape))
keep = np.in1d(y, classes)
x, y = x[keep], y[keep]
cj.logger.info("Size after filtering:")
cj.logger.info(" - x: {}".format(x.shape))
cj.logger.info(" - y: {}".format(y.shape))
if x.shape[0] == 0:
raise ValueError("No training data")
if is_binary:
# 0 (background) vs 1 (classes in foreground )
cj.logger.info("Binary segmentation:")
cj.logger.info("> class '0': background & terms {}".format(
set(selected_terms).difference(positive_terms)))
cj.logger.info("> class '1': {}".format(set(foreground_terms)))
else:
# 0 (background vs 1 vs 2 vs ... n (n classes from cytomine_id_terms)
cj.logger.info("Multi-class segmentation:")
cj.logger.info("> background class '0'")
cj.logger.info("> term classes: {}".format(set(foreground_terms)))
# build model
cj.job.update(progress=27, statusComment="Build model...")
et, pyxit = build_models(
n_subwindows=cj.parameters.pyxit_n_subwindows,
min_size=cj.parameters.pyxit_min_size,
max_size=cj.parameters.pyxit_max_size,
target_width=cj.parameters.pyxit_target_width,
target_height=cj.parameters.pyxit_target_height,
interpolation=cj.parameters.pyxit_interpolation,
transpose=cj.parameters.pyxit_transpose,
colorspace=cj.parameters.pyxit_colorspace,
fixed_size=cj.parameters.pyxit_fixed_size,
verbose=int(cj.logger.level == 10),
random_state=cj.parameters.seed,
n_estimators=cj.parameters.forest_n_estimators,
min_samples_split=cj.parameters.forest_min_samples_split,
max_features=cj.parameters.forest_max_features,
n_jobs=cj.parameters.n_jobs)
# to extract the classes form the mask
pyxit.get_output = _get_output_from_mask
# extract subwindows manually to avoid class problem
cj.job.update(progress=30, statusComment="Extract subwindwos...")
_x, _y = pyxit.extract_subwindows(x, y)
actual_classes = np.unique(_y)
if actual_classes.shape[0] != classes.shape[0]:
raise ValueError(
"Some classes are missing from the dataset: actual='{}', expected='{}'"
.format(",".join(map(str, actual_classes)),
",".join(map(str, classes))))
cj.logger.info("Size of actual training data:")
cj.logger.info(" - x : {}".format(_x.shape))
cj.logger.info(" - y : {}".format(_y.shape))
cj.logger.info(" - dist: {}".format(", ".join([
"{}: {}".format(v, c)
for v, c in zip(*np.unique(_y, return_counts=True))
])))
cj.job.update(progress=60, statusComment="Train model...")
# "re-implement" pyxit.fit to avoid incorrect class handling
pyxit.classes_ = classes
pyxit.n_classes_ = n_classes
pyxit.base_estimator.fit(_x, _y)
cj.job.update(progress=90, statusComment="Save model....")
model_filename = joblib.dump(pyxit,
os.path.join(base_path, "model.joblib"),
compress=3)[0]
AttachedFile(cj.job,
domainIdent=cj.job.id,
filename=model_filename,
domainClassName="be.cytomine.processing.Job").upload()
Property(cj.job, key="classes", value=stringify(classes)).save()
Property(cj.job, key="binary", value=is_binary).save()
cj.job.update(status=Job.TERMINATED,
status_comment="Finish",
progress=100)