def main(base_dir, base_filename):
rc1 = ResultsContainer("", "")
rc1.ar = np.empty(shape=(2,), dtype=object)
rc1.axis_list = ["classifiers"]
rc1.labels_dict["classifiers"] = ["svm", "dt"]
from environment.structures import ClassifierResult
c1 = ClassifierResult("", "")
c1.scores["accuracy"] = 0.8
c2 = ClassifierResult("", "")
c2.scores["accuracy"] = 0.95
rc1.ar[0] = c1
rc1.ar[1] = c2
rc2 = copy.deepcopy(rc1)
rc3 = copy.deepcopy(rc1)
rc4 = copy.deepcopy(rc1)
d2_rc = ResultsContainer("", "")
d2_rc.add_dim_layer([rc1, rc2, rc3, rc4], "cv_folds", ["f1", "f2", "f3", "f4"])
d2rc1 = copy.deepcopy(d2_rc)
d2rc2 = copy.deepcopy(d2_rc)
d2rc3 = copy.deepcopy(d2_rc)
d3_rc = ResultsContainer(base_dir, base_filename)
d3_rc.add_dim_layer([d2rc1, d2rc2, d2rc3], "fenotype_features",
["age", "sex", "tissue"])
return d3_rc
def aggregate_prediction_vectors(self, axis_list_to_preserve):
"""
Produce new np.array be merging `y_true`, `y_predicted` fields of ClassifierResult objects
in axis that not present in axis_list_to_preserve. New array is reshaped to complain
with axis order in axis_list_to_preserve.
@return: Array of len(axis_axis_list_to_preserve) dimensions
each element [ClassifierResult] would have joined y_true and y_predicted vectors
@rtype: np.array
"""
if len(axis_list_to_preserve) == len(self.axis_list):
return np.transpose(
self.ar,
[self.axis_list.index(axis) for axis in axis_list_to_preserve])
new_shape = tuple([
len(self.labels_dict[axis_name])
for axis_name in axis_list_to_preserve
])
result = np.empty(shape=new_shape, dtype=object)
index_labels = [
self.labels_dict[axis] for axis in axis_list_to_preserve
]
for row_def in product(*index_labels):
# log.debug("processing row: %s", row_def)
spec_def = {
axis: val
for val, axis in zip(row_def, axis_list_to_preserve)
}
key = np.array(self.build_axis_mask(spec_def))
key_for_result = tuple([
self.inverse_labels_dict[axis_name][label]
for label, axis_name in zip(row_def, axis_list_to_preserve)
])
sliced = self.ar[tuple(key)]
if hasattr(sliced, 'flatten'):
flatten = sliced.flatten()
new_cr = ClassifierResult("", "")
new_cr.classifier = "aggregated_result"
for cr in flatten:
# new_cr.labels_encode_vector.extend(cr.labels_encode_vector)
if cr is not None:
new_cr.y_true.extend(cr.y_true)
new_cr.y_predicted.extend(cr.y_predicted)
# import ipdb; ipdb.set_trace()
result[key_for_result] = new_cr
return result
def apply_classifier(exp,
block,
train_es,
test_es,
classifier_name,
classifier_options=None,
fit_options=None,
base_folder="/tmp",
base_filename="cl"):
"""
@type train_es: ExpressionSet
@type test_es: ExpressionSet
"""
if not classifier_options:
classifier_options = {}
if not fit_options:
fit_options = {}
target_class_column = train_es.pheno_metadata["user_class_title"]
# Unpack data
x_train = train_es.get_assay_data_frame().as_matrix().transpose()
y_train = train_es.get_pheno_data_frame()[target_class_column].as_matrix()
x_test = test_es.get_assay_data_frame().as_matrix().transpose()
y_test = test_es.get_pheno_data_frame()[target_class_column].as_matrix()
# Unfortunately svm can't operate with string labels as a target classes
# so we need to preprocess labels
le = preprocessing.LabelEncoder()
le.fit(y_train)
y_train_fixed = le.transform(y_train)
y_test_fixed = le.transform(y_test)
# Classifier initialization
fabric, apply_func = classifiers_map[classifier_name]
log.debug("Classifier options: %s", classifier_options)
if apply_func is None:
cl = fabric(**classifier_options)
cl.fit(x_train, y_train_fixed, **fit_options)
else:
raise NotImplementedError()
# Applying on test partition
y_test_predicted = cl.predict(x_test)
# Here we build result object
cr = ClassifierResult(base_folder, base_filename)
cr.labels_encode_vector = le.classes_ # Store target class labels
cr.y_true = y_test_fixed
cr.y_predicted = y_test_predicted
cr.classifier = classifier_name
cr.store_model(cl)
return [cr], {}
def aggregate_prediction_vectors(self, axis_list_to_preserve):
"""
Produce new np.array be merging `y_true`, `y_predicted` fields of ClassifierResult objects
in axis that not present in axis_list_to_preserve. New array is reshaped to complain
with axis order in axis_list_to_preserve.
@return: Array of len(axis_axis_list_to_preserve) dimensions
each element [ClassifierResult] would have joined y_true and y_predicted vectors
@rtype: np.array
"""
if len(axis_list_to_preserve) == len(self.axis_list):
return np.transpose(self.ar, [
self.axis_list.index(axis)
for axis in axis_list_to_preserve
])
new_shape = tuple([
len(self.labels_dict[axis_name])
for axis_name in axis_list_to_preserve
])
result = np.empty(shape=new_shape, dtype=object)
index_labels = [self.labels_dict[axis] for axis in axis_list_to_preserve]
for row_def in product(*index_labels):
# log.debug("processing row: %s", row_def)
spec_def = {axis: val for val, axis in zip(row_def, axis_list_to_preserve)}
key = np.array(self.build_axis_mask(spec_def))
key_for_result = tuple([
self.inverse_labels_dict[axis_name][label]
for label, axis_name
in zip(row_def, axis_list_to_preserve)
])
sliced = self.ar[tuple(key)]
if hasattr(sliced, 'flatten'):
flatten = sliced.flatten()
new_cr = ClassifierResult("", "")
new_cr.classifier = "aggregated_result"
for cr in flatten:
# new_cr.labels_encode_vector.extend(cr.labels_encode_vector)
if cr is not None:
new_cr.y_true.extend(cr.y_true)
new_cr.y_predicted.extend(cr.y_predicted)
# import ipdb; ipdb.set_trace()
result[key_for_result] = new_cr
return result
def apply_classifier(
exp, block,
train_es, test_es,
classifier_name, classifier_options=None, fit_options=None,
base_folder="/tmp", base_filename="cl"
):
"""
@type train_es: ExpressionSet
@type test_es: ExpressionSet
"""
if not classifier_options:
classifier_options = {}
if not fit_options:
fit_options = {}
target_class_column = train_es.pheno_metadata["user_class_title"]
# Unpack data
x_train = train_es.get_assay_data_frame().as_matrix().transpose()
y_train = train_es.get_pheno_data_frame()[target_class_column].as_matrix()
x_test = test_es.get_assay_data_frame().as_matrix().transpose()
y_test = test_es.get_pheno_data_frame()[target_class_column].as_matrix()
# Unfortunately svm can't operate with string labels as a target classes
# so we need to preprocess labels
le = preprocessing.LabelEncoder()
le.fit(y_train)
y_train_fixed = le.transform(y_train)
y_test_fixed = le.transform(y_test)
# Classifier initialization
fabric, apply_func = classifiers_map[classifier_name]
log.debug("Classifier options: %s", classifier_options)
if apply_func is None:
cl = fabric(**classifier_options)
cl.fit(x_train, y_train_fixed, **fit_options)
else:
raise NotImplementedError()
# Applying on test partition
y_test_predicted = cl.predict(x_test)
# Here we build result object
cr = ClassifierResult(base_folder, base_filename)
cr.labels_encode_vector = le.classes_ # Store target class labels
cr.y_true = y_test_fixed
cr.y_predicted = y_test_predicted
cr.classifier = classifier_name
cr.store_model(cl)
return [cr], {}
def main(base_dir, base_filename):
rc1 = ResultsContainer("", "")
rc1.ar = np.empty(shape=(2, ), dtype=object)
rc1.axis_list = ["classifiers"]
rc1.labels_dict["classifiers"] = ["svm", "dt"]
from environment.structures import ClassifierResult
c1 = ClassifierResult("", "")
c1.scores["accuracy"] = 0.8
c2 = ClassifierResult("", "")
c2.scores["accuracy"] = 0.95
rc1.ar[0] = c1
rc1.ar[1] = c2
rc2 = copy.deepcopy(rc1)
rc3 = copy.deepcopy(rc1)
rc4 = copy.deepcopy(rc1)
d2_rc = ResultsContainer("", "")
d2_rc.add_dim_layer([rc1, rc2, rc3, rc4], "cv_folds",
["f1", "f2", "f3", "f4"])
d2rc1 = copy.deepcopy(d2_rc)
d2rc2 = copy.deepcopy(d2_rc)
d2rc3 = copy.deepcopy(d2_rc)
d3_rc = ResultsContainer(base_dir, base_filename)
d3_rc.add_dim_layer([d2rc1, d2rc2, d2rc3], "fenotype_features",
["age", "sex", "tissue"])
return d3_rc
def apply_classifier(
exp, block,
train_es, test_es,
classifier_name, classifier_options=None, fit_options=None,
base_folder="/tmp", base_filename="cl"
):
"""
@type train_es: ExpressionSet
@type test_es: ExpressionSet
@type exp: Experiment
@type block: GenericBlock
"""
if settings.CELERY_DEBUG:
import sys
sys.path.append('/Migration/skola/phd/projects/miXGENE/mixgene_project/wrappers/pycharm-debug.egg')
import pydevd
pydevd.settrace('localhost', port=6901, stdoutToServer=True, stderrToServer=True)
if not classifier_options:
classifier_options = {}
if not fit_options:
fit_options = {}
target_class_column = train_es.pheno_metadata["user_class_title"]
tr_es = train_es.get_assay_data_frame()
cols = tr_es.columns
te_es = test_es.get_assay_data_frame()[list(cols)]
# Unpack data
x_train = tr_es.as_matrix()
# x_train = train_es.get_assay_data_frame().as_matrix().transpose()
y_train = train_es.get_pheno_data_frame()[target_class_column].as_matrix()
x_test = te_es.as_matrix()
# x_test = test_es.get_assay_data_frame().as_matrix().transpose()
y_test = test_es.get_pheno_data_frame()[target_class_column].as_matrix()
# Unfortunately svm can't operate with string labels as a target classes
# so we need to preprocess labels
le = preprocessing.LabelEncoder()
le.fit(y_train)
y_train_fixed = le.transform(y_train)
y_test_fixed = le.transform(y_test)
# Classifier initialization
fabric, apply_func = classifiers_map[classifier_name]
# log.debug("Classifier options: %s", classifier_options)
if apply_func is None:
cl = get_classifier(fabric, classifier_options, classifier_name, block)
log.debug("Fitting classifier.")
cl.fit(x_train, y_train_fixed)
log.debug("Finished fitting classifier.")
else:
raise NotImplementedError()
log.debug("Applying on test.")
# Applying on test partition
y_test_predicted = cl.predict(x_test)
log.debug("Building result.")
# Here we build result object
cr = ClassifierResult(base_folder, base_filename)
log.debug("Storing labels.")
cr.labels_encode_vector = le.classes_ # Store target class labels
log.debug("Storing y.")
cr.y_true = y_test_fixed
cr.y_predicted = y_test_predicted
cr.classifier = classifier_name
log.debug("Storing model.")
# TODO Why to store model?
# cr.store_model(cl)
log.debug("Finished apply_classifier.")
return [cr], {}
def apply_classifier(
exp, block,
train_es, test_es,
classifier_name, classifier_options=None, fit_options=None,
base_folder="/tmp", base_filename="cl"
):
"""
@type train_es: ExpressionSet
@type test_es: ExpressionSet
@type exp: Experiment
@type block: GenericBlock
"""
if settings.CELERY_DEBUG:
import sys
sys.path.append('/Migration/skola/phd/projects/miXGENE/mixgene_project/wrappers/pycharm-debug.egg')
import pydevd
pydevd.settrace('localhost', port=6901, stdoutToServer=True, stderrToServer=True)
if not classifier_options:
classifier_options = {}
if not fit_options:
fit_options = {}
target_class_column = train_es.pheno_metadata["user_class_title"]
tr_es = train_es.get_assay_data_frame()
cols = tr_es.columns
te_es = test_es.get_assay_data_frame()[list(cols)]
# Unpack data
x_train = tr_es.as_matrix()
# x_train = train_es.get_assay_data_frame().as_matrix().transpose()
y_train = train_es.get_pheno_data_frame()[target_class_column].as_matrix()
x_test = te_es.as_matrix()
# x_test = test_es.get_assay_data_frame().as_matrix().transpose()
y_test = test_es.get_pheno_data_frame()[target_class_column].as_matrix()
# Unfortunately svm can't operate with string labels as a target classes
# so we need to preprocess labels
le = preprocessing.LabelEncoder()
le.fit(y_train)
y_train_fixed = le.transform(y_train)
y_test_fixed = le.transform(y_test)
# Classifier initialization
fabric, apply_func = classifiers_map[classifier_name]
# log.debug("Classifier options: %s", classifier_options)
if apply_func is None:
cl = get_classifier(fabric, classifier_options, classifier_name, block)
log.debug("Fitting classifier.")
try:
log.debug(str(x_train))
cl.fit(x_train, y_train_fixed)
except ValueError:
# if settings.CELERY_DEBUG:
# import sys
# sys.path.append('/Migration/skola/phd/projects/miXGENE/mixgene_project/wrappers/pycharm-debug.egg')
# import pydevd
# pydevd.settrace('localhost', port=6901, stdoutToServer=True, stderrToServer=True)
log.debug(str(x_train))
raise
log.debug("Finished fitting classifier.")
else:
raise NotImplementedError()
log.debug("Applying on test.")
# Applying on test partition
y_test_predicted = cl.predict(x_test)
log.debug("Building result.")
# Here we build result object
cr = ClassifierResult(base_folder, base_filename)
log.debug("Storing labels.")
cr.labels_encode_vector = le.classes_ # Store target class labels
log.debug("Storing y.")
cr.y_true = y_test_fixed
cr.y_predicted = y_test_predicted
cr.classifier = classifier_name
log.debug("Storing model.")
# TODO Why to store model?
# cr.store_model(cl)
log.debug("Finished apply_classifier.")
return [cr], {}