def mouseMoveEvent(self, event):
if event.buttons() & Qt.LeftButton:
downPos = event.buttonDownPos(Qt.LeftButton)
if not self.__tmpLine and self.__dragStartItem and \
(downPos - event.pos()).manhattanLength() > \
QApplication.instance().startDragDistance():
# Start a line drag
line = QGraphicsLineItem(self)
start = self.__dragStartItem.boundingRect().center()
start = self.mapFromItem(self.__dragStartItem, start)
line.setLine(start.x(), start.y(),
event.pos().x(),
event.pos().y())
pen = QPen(self.palette().color(QPalette.Foreground), 4)
pen.setCapStyle(Qt.RoundCap)
line.setPen(pen)
line.show()
self.__tmpLine = line
if self.__tmpLine:
# Update the temp line
line = self.__tmpLine.line()
line.setP2(event.pos())
self.__tmpLine.setLine(line)
QGraphicsWidget.mouseMoveEvent(self, event)
def _draw_border(point_1, point_2, border_width, parent):
pen = QPen(QColor(self.border_color))
pen.setCosmetic(True)
pen.setWidth(border_width)
line = QGraphicsLineItem(QLineF(point_1, point_2), parent)
line.setPen(pen)
return line
def mouseMoveEvent(self, event):
if event.buttons() & Qt.LeftButton:
downPos = event.buttonDownPos(Qt.LeftButton)
if not self.__tmpLine and self.__dragStartItem and \
(downPos - event.pos()).manhattanLength() > \
QApplication.instance().startDragDistance():
# Start a line drag
line = QGraphicsLineItem(self)
start = self.__dragStartItem.boundingRect().center()
start = self.mapFromItem(self.__dragStartItem, start)
line.setLine(start.x(), start.y(),
event.pos().x(), event.pos().y())
pen = QPen(Qt.black, 4)
pen.setCapStyle(Qt.RoundCap)
line.setPen(pen)
line.show()
self.__tmpLine = line
if self.__tmpLine:
# Update the temp line
line = self.__tmpLine.line()
line.setP2(event.pos())
self.__tmpLine.setLine(line)
QGraphicsWidget.mouseMoveEvent(self, event)
def _draw_border(point_1, point_2, border_width, parent):
pen = QPen(QColor(self.border_color))
pen.setCosmetic(True)
pen.setWidth(border_width)
line = QGraphicsLineItem(QLineF(point_1, point_2), parent)
line.setPen(pen)
return line
class LinePlotViewBox(ViewBox):
def __init__(self, graph, enable_menu=False):
ViewBox.__init__(self, enableMenu=enable_menu)
self.graph = graph
self.setMouseMode(self.PanMode)
pen = mkPen(LinePlotStyle.SELECTION_LINE_COLOR,
width=LinePlotStyle.SELECTION_LINE_WIDTH)
self.selection_line = QGraphicsLineItem()
self.selection_line.setPen(pen)
self.selection_line.setZValue(1e9)
self.addItem(self.selection_line, ignoreBounds=True)
def update_selection_line(self, button_down_pos, current_pos):
p1 = self.childGroup.mapFromParent(button_down_pos)
p2 = self.childGroup.mapFromParent(current_pos)
self.selection_line.setLine(QLineF(p1, p2))
self.selection_line.resetTransform()
self.selection_line.show()
def mouseDragEvent(self, event, axis=None):
if self.graph.state == SELECT and axis is None:
event.accept()
if event.button() == Qt.LeftButton:
self.update_selection_line(event.buttonDownPos(), event.pos())
if event.isFinish():
self.selection_line.hide()
p1 = self.childGroup.mapFromParent(
event.buttonDownPos(event.button()))
p2 = self.childGroup.mapFromParent(event.pos())
self.graph.select_by_line(p1, p2)
else:
self.update_selection_line(
event.buttonDownPos(), event.pos())
elif self.graph.state == ZOOMING or self.graph.state == PANNING:
event.ignore()
super().mouseDragEvent(event, axis=axis)
else:
event.ignore()
def mouseClickEvent(self, event):
if event.button() == Qt.RightButton:
self.autoRange()
else:
event.accept()
self.graph.deselect_all()
def addLink(self, output, input):
"""
Add a link between `output` (:class:`OutputSignal`) and `input`
(:class:`InputSignal`).
"""
if not compatible_channels(output, input):
return
if output not in self.source.output_channels():
raise ValueError("%r is not an output channel of %r" % \
(output, self.source))
if input not in self.sink.input_channels():
raise ValueError("%r is not an input channel of %r" % \
(input, self.sink))
if input.single:
# Remove existing link if it exists.
for s1, s2, _ in self.__links:
if s2 == input:
self.removeLink(s1, s2)
line = QGraphicsLineItem(self)
source_anchor = self.sourceNodeWidget.anchor(output)
sink_anchor = self.sinkNodeWidget.anchor(input)
source_pos = source_anchor.boundingRect().center()
source_pos = self.mapFromItem(source_anchor, source_pos)
sink_pos = sink_anchor.boundingRect().center()
sink_pos = self.mapFromItem(sink_anchor, sink_pos)
line.setLine(source_pos.x(), source_pos.y(), sink_pos.x(),
sink_pos.y())
pen = QPen(self.palette().color(QPalette.Foreground), 4)
pen.setCapStyle(Qt.RoundCap)
line.setPen(pen)
self.__links.append(_Link(output, input, line))
def addLink(self, output, input):
"""
Add a link between `output` (:class:`OutputSignal`) and `input`
(:class:`InputSignal`).
"""
if not compatible_channels(output, input):
return
if output not in self.source.output_channels():
raise ValueError("%r is not an output channel of %r" % \
(output, self.source))
if input not in self.sink.input_channels():
raise ValueError("%r is not an input channel of %r" % \
(input, self.sink))
if input.single:
# Remove existing link if it exists.
for s1, s2, _ in self.__links:
if s2 == input:
self.removeLink(s1, s2)
line = QGraphicsLineItem(self)
source_anchor = self.sourceNodeWidget.anchor(output)
sink_anchor = self.sinkNodeWidget.anchor(input)
source_pos = source_anchor.boundingRect().center()
source_pos = self.mapFromItem(source_anchor, source_pos)
sink_pos = sink_anchor.boundingRect().center()
sink_pos = self.mapFromItem(sink_anchor, sink_pos)
line.setLine(source_pos.x(), source_pos.y(),
sink_pos.x(), sink_pos.y())
pen = QPen(Qt.black, 4)
pen.setCapStyle(Qt.RoundCap)
line.setPen(pen)
self.__links.append(_Link(output, input, line))
class OWNomogram(OWWidget):
name = "Nomogram"
description = " Nomograms for Visualization of Naive Bayesian" \
" and Logistic Regression Classifiers."
icon = "icons/Nomogram.svg"
priority = 2000
inputs = [("Classifier", Model, "set_classifier"),
("Data", Table, "set_instance")]
MAX_N_ATTRS = 1000
POINT_SCALE = 0
ALIGN_LEFT = 0
ALIGN_ZERO = 1
ACCEPTABLE = (NaiveBayesModel, LogisticRegressionClassifier)
settingsHandler = ClassValuesContextHandler()
target_class_index = ContextSetting(0)
normalize_probabilities = Setting(False)
scale = Setting(1)
display_index = Setting(1)
n_attributes = Setting(10)
sort_index = Setting(SortBy.ABSOLUTE)
cont_feature_dim_index = Setting(0)
graph_name = "scene"
class Error(OWWidget.Error):
invalid_classifier = Msg("Nomogram accepts only Naive Bayes and "
"Logistic Regression classifiers.")
def __init__(self):
super().__init__()
self.instances = None
self.domain = None
self.data = None
self.classifier = None
self.align = OWNomogram.ALIGN_ZERO
self.log_odds_ratios = []
self.log_reg_coeffs = []
self.log_reg_coeffs_orig = []
self.log_reg_cont_data_extremes = []
self.p = None
self.b0 = None
self.points = []
self.feature_items = []
self.feature_marker_values = []
self.scale_back = lambda x: x
self.scale_forth = lambda x: x
self.nomogram = None
self.nomogram_main = None
self.vertical_line = None
self.hidden_vertical_line = None
self.old_target_class_index = self.target_class_index
self.markers_set = False
self.repaint = False
# GUI
box = gui.vBox(self.controlArea, "Target class")
self.class_combo = gui.comboBox(box,
self,
"target_class_index",
callback=self._class_combo_changed,
contentsLength=12)
self.norm_check = gui.checkBox(
box,
self,
"normalize_probabilities",
"Normalize probabilities",
hidden=True,
callback=self._norm_check_changed,
tooltip="For multiclass data 1 vs. all probabilities do not"
" sum to 1 and therefore could be normalized.")
self.scale_radio = gui.radioButtons(
self.controlArea,
self,
"scale", ["Point scale", "Log odds ratios"],
box="Scale",
callback=self._radio_button_changed)
box = gui.vBox(self.controlArea, "Display features")
grid = QGridLayout()
self.display_radio = gui.radioButtonsInBox(
box,
self,
"display_index", [],
orientation=grid,
callback=self._display_radio_button_changed)
radio_all = gui.appendRadioButton(self.display_radio,
"All:",
addToLayout=False)
radio_best = gui.appendRadioButton(self.display_radio,
"Best ranked:",
addToLayout=False)
spin_box = gui.hBox(None, margin=0)
self.n_spin = gui.spin(spin_box,
self,
"n_attributes",
1,
self.MAX_N_ATTRS,
label=" ",
controlWidth=60,
callback=self._n_spin_changed)
grid.addWidget(radio_all, 1, 1)
grid.addWidget(radio_best, 2, 1)
grid.addWidget(spin_box, 2, 2)
self.sort_combo = gui.comboBox(box,
self,
"sort_index",
label="Sort by: ",
items=SortBy.items(),
orientation=Qt.Horizontal,
callback=self._sort_combo_changed)
self.cont_feature_dim_combo = gui.comboBox(
box,
self,
"cont_feature_dim_index",
label="Continuous features: ",
items=["1D projection", "2D curve"],
orientation=Qt.Horizontal,
callback=self._cont_feature_dim_combo_changed)
gui.rubber(self.controlArea)
self.scene = QGraphicsScene()
self.view = QGraphicsView(
self.scene,
horizontalScrollBarPolicy=Qt.ScrollBarAlwaysOff,
renderHints=QPainter.Antialiasing | QPainter.TextAntialiasing
| QPainter.SmoothPixmapTransform,
alignment=Qt.AlignLeft)
self.view.viewport().installEventFilter(self)
self.view.viewport().setMinimumWidth(300)
self.view.sizeHint = lambda: QSize(600, 500)
self.mainArea.layout().addWidget(self.view)
def _class_combo_changed(self):
values = [item.dot.value for item in self.feature_items]
self.feature_marker_values = self.scale_back(values)
coeffs = [
np.nan_to_num(p[self.target_class_index] /
p[self.old_target_class_index]) for p in self.points
]
points = [p[self.old_target_class_index] for p in self.points]
self.feature_marker_values = [
self.get_points_from_coeffs(v, c, p)
for (v, c, p) in zip(self.feature_marker_values, coeffs, points)
]
self.update_scene()
self.old_target_class_index = self.target_class_index
def _norm_check_changed(self):
values = [item.dot.value for item in self.feature_items]
self.feature_marker_values = self.scale_back(values)
self.update_scene()
def _radio_button_changed(self):
values = [item.dot.value for item in self.feature_items]
self.feature_marker_values = self.scale_back(values)
self.update_scene()
def _display_radio_button_changed(self):
self.__hide_attrs(self.n_attributes if self.display_index else None)
def _n_spin_changed(self):
self.display_index = 1
self.__hide_attrs(self.n_attributes)
def __hide_attrs(self, n_show):
if self.nomogram_main is None:
return
self.nomogram_main.hide(n_show)
if self.vertical_line:
x = self.vertical_line.line().x1()
y = self.nomogram_main.layout.preferredHeight() + 30
self.vertical_line.setLine(x, -6, x, y)
self.hidden_vertical_line.setLine(x, -6, x, y)
rect = QRectF(self.scene.sceneRect().x(),
self.scene.sceneRect().y(),
self.scene.itemsBoundingRect().width(),
self.nomogram.preferredSize().height())
self.scene.setSceneRect(rect.adjusted(0, 0, 70, 70))
def _sort_combo_changed(self):
if self.nomogram_main is None:
return
self.nomogram_main.hide(None)
self.nomogram_main.sort(self.sort_index)
self.__hide_attrs(self.n_attributes if self.display_index else None)
def _cont_feature_dim_combo_changed(self):
values = [item.dot.value for item in self.feature_items]
self.feature_marker_values = self.scale_back(values)
self.update_scene()
def eventFilter(self, obj, event):
if obj is self.view.viewport() and event.type() == QEvent.Resize:
self.repaint = True
values = [item.dot.value for item in self.feature_items]
self.feature_marker_values = self.scale_back(values)
self.update_scene()
return super().eventFilter(obj, event)
def update_controls(self):
self.class_combo.clear()
self.norm_check.setHidden(True)
self.cont_feature_dim_combo.setEnabled(True)
if self.domain:
self.class_combo.addItems(self.domain.class_vars[0].values)
if len(self.domain.attributes) > self.MAX_N_ATTRS:
self.display_index = 1
if len(self.domain.class_vars[0].values) > 2:
self.norm_check.setHidden(False)
if not self.domain.has_continuous_attributes():
self.cont_feature_dim_combo.setEnabled(False)
self.cont_feature_dim_index = 0
model = self.sort_combo.model()
item = model.item(SortBy.POSITIVE)
item.setFlags(item.flags() | Qt.ItemIsEnabled)
item = model.item(SortBy.NEGATIVE)
item.setFlags(item.flags() | Qt.ItemIsEnabled)
self.align = OWNomogram.ALIGN_ZERO
if self.classifier and isinstance(self.classifier,
LogisticRegressionClassifier):
self.align = OWNomogram.ALIGN_LEFT
item = model.item(SortBy.POSITIVE)
item.setFlags(item.flags() & ~Qt.ItemIsEnabled)
item = model.item(SortBy.NEGATIVE)
item.setFlags(item.flags() & ~Qt.ItemIsEnabled)
if self.sort_index in (SortBy.POSITIVE, SortBy.POSITIVE):
self.sort_index = SortBy.NO_SORTING
def set_instance(self, data):
self.instances = data
self.feature_marker_values = []
self.set_feature_marker_values()
def set_classifier(self, classifier):
self.closeContext()
self.classifier = classifier
self.Error.clear()
if self.classifier and not isinstance(self.classifier,
self.ACCEPTABLE):
self.Error.invalid_classifier()
self.classifier = None
self.domain = self.classifier.domain if self.classifier else None
self.data = None
self.calculate_log_odds_ratios()
self.calculate_log_reg_coefficients()
self.update_controls()
self.target_class_index = 0
self.openContext(self.domain and self.domain.class_var)
self.points = self.log_odds_ratios or self.log_reg_coeffs
self.feature_marker_values = []
self.old_target_class_index = self.target_class_index
self.update_scene()
def calculate_log_odds_ratios(self):
self.log_odds_ratios = []
self.p = None
if self.classifier is None or self.domain is None:
return
if not isinstance(self.classifier, NaiveBayesModel):
return
log_cont_prob = self.classifier.log_cont_prob
class_prob = self.classifier.class_prob
for i in range(len(self.domain.attributes)):
ca = np.exp(log_cont_prob[i]) * class_prob[:, None]
_or = (ca / (1 - ca)) / (class_prob / (1 - class_prob))[:, None]
self.log_odds_ratios.append(np.log(_or))
self.p = class_prob
def calculate_log_reg_coefficients(self):
self.log_reg_coeffs = []
self.log_reg_cont_data_extremes = []
self.b0 = None
if self.classifier is None or self.domain is None:
return
if not isinstance(self.classifier, LogisticRegressionClassifier):
return
self.domain = self.reconstruct_domain(self.classifier.original_domain,
self.domain)
self.data = self.classifier.original_data.transform(self.domain)
attrs, ranges, start = self.domain.attributes, [], 0
for attr in attrs:
stop = start + len(attr.values) if attr.is_discrete else start + 1
ranges.append(slice(start, stop))
start = stop
self.b0 = self.classifier.intercept
coeffs = self.classifier.coefficients
if len(self.domain.class_var.values) == 2:
self.b0 = np.hstack((self.b0 * (-1), self.b0))
coeffs = np.vstack((coeffs * (-1), coeffs))
self.log_reg_coeffs = [coeffs[:, ranges[i]] for i in range(len(attrs))]
self.log_reg_coeffs_orig = self.log_reg_coeffs.copy()
min_values = nanmin(self.data.X, axis=0)
max_values = nanmax(self.data.X, axis=0)
for i, min_t, max_t in zip(range(len(self.log_reg_coeffs)), min_values,
max_values):
if self.log_reg_coeffs[i].shape[1] == 1:
coef = self.log_reg_coeffs[i]
self.log_reg_coeffs[i] = np.hstack(
(coef * min_t, coef * max_t))
self.log_reg_cont_data_extremes.append(
[sorted([min_t, max_t], reverse=(c < 0)) for c in coef])
else:
self.log_reg_cont_data_extremes.append([None])
def update_scene(self):
if not self.repaint:
return
self.clear_scene()
if self.domain is None or not len(self.points[0]):
return
name_items = [
QGraphicsTextItem(a.name) for a in self.domain.attributes
]
point_text = QGraphicsTextItem("Points")
probs_text = QGraphicsTextItem("Probabilities (%)")
all_items = name_items + [point_text, probs_text]
name_offset = -max(t.boundingRect().width() for t in all_items) - 50
w = self.view.viewport().rect().width()
max_width = w + name_offset - 100
points = [pts[self.target_class_index] for pts in self.points]
minimums = [min(p) for p in points]
if self.align == OWNomogram.ALIGN_LEFT:
points = [p - m for m, p in zip(minimums, points)]
max_ = np.nan_to_num(max(max(abs(p)) for p in points))
d = 100 / max_ if max_ else 1
if self.scale == OWNomogram.POINT_SCALE:
points = [p * d for p in points]
if self.scale == OWNomogram.POINT_SCALE and \
self.align == OWNomogram.ALIGN_LEFT:
self.scale_back = lambda x: [
p / d + m for m, p in zip(minimums, x)
]
self.scale_forth = lambda x: [(p - m) * d
for m, p in zip(minimums, x)]
if self.scale == OWNomogram.POINT_SCALE and \
self.align != OWNomogram.ALIGN_LEFT:
self.scale_back = lambda x: [p / d for p in x]
self.scale_forth = lambda x: [p * d for p in x]
if self.scale != OWNomogram.POINT_SCALE and \
self.align == OWNomogram.ALIGN_LEFT:
self.scale_back = lambda x: [p + m for m, p in zip(minimums, x)]
self.scale_forth = lambda x: [p - m for m, p in zip(minimums, x)]
if self.scale != OWNomogram.POINT_SCALE and \
self.align != OWNomogram.ALIGN_LEFT:
self.scale_back = lambda x: x
self.scale_forth = lambda x: x
point_item, nomogram_head = self.create_main_nomogram(
name_items, points, max_width, point_text, name_offset)
probs_item, nomogram_foot = self.create_footer_nomogram(
probs_text, d, minimums, max_width, name_offset)
for item in self.feature_items:
item.dot.point_dot = point_item.dot
item.dot.probs_dot = probs_item.dot
item.dot.vertical_line = self.hidden_vertical_line
self.nomogram = nomogram = NomogramItem()
nomogram.add_items([nomogram_head, self.nomogram_main, nomogram_foot])
self.scene.addItem(nomogram)
self.set_feature_marker_values()
rect = QRectF(self.scene.itemsBoundingRect().x(),
self.scene.itemsBoundingRect().y(),
self.scene.itemsBoundingRect().width(),
self.nomogram.preferredSize().height())
self.scene.setSceneRect(rect.adjusted(0, 0, 70, 70))
def create_main_nomogram(self, name_items, points, max_width, point_text,
name_offset):
cls_index = self.target_class_index
min_p = min(min(p) for p in points)
max_p = max(max(p) for p in points)
values = self.get_ruler_values(min_p, max_p, max_width)
min_p, max_p = min(values), max(values)
diff_ = np.nan_to_num(max_p - min_p)
scale_x = max_width / diff_ if diff_ else max_width
nomogram_header = NomogramItem()
point_item = RulerItem(point_text, values, scale_x, name_offset,
-scale_x * min_p)
point_item.setPreferredSize(point_item.preferredWidth(), 35)
nomogram_header.add_items([point_item])
self.nomogram_main = SortableNomogramItem()
cont_feature_item_class = ContinuousFeature2DItem if \
self.cont_feature_dim_index else ContinuousFeatureItem
self.feature_items = [
DiscreteFeatureItem(name_items[i], [val for val in att.values],
points[i], scale_x, name_offset, -scale_x *
min_p, self.points[i][cls_index])
if att.is_discrete else cont_feature_item_class(
name_items[i], self.log_reg_cont_data_extremes[i][cls_index],
self.get_ruler_values(
np.min(points[i]), np.max(points[i]),
scale_x * (np.max(points[i]) - np.min(points[i])),
False), scale_x, name_offset, -scale_x *
min_p, self.log_reg_coeffs_orig[i][cls_index][0])
for i, att in enumerate(self.domain.attributes)
]
self.nomogram_main.add_items(
self.feature_items, self.sort_index,
self.n_attributes if self.display_index else None)
x = -scale_x * min_p
y = self.nomogram_main.layout.preferredHeight() + 30
self.vertical_line = QGraphicsLineItem(x, -6, x, y)
self.vertical_line.setPen(QPen(Qt.DotLine))
self.vertical_line.setParentItem(point_item)
self.hidden_vertical_line = QGraphicsLineItem(x, -6, x, y)
pen = QPen(Qt.DashLine)
pen.setBrush(QColor(Qt.red))
self.hidden_vertical_line.setPen(pen)
self.hidden_vertical_line.setParentItem(point_item)
return point_item, nomogram_header
def create_footer_nomogram(self, probs_text, d, minimums, max_width,
name_offset):
eps, d_ = 0.05, 1
k = -np.log(self.p / (1 - self.p)) if self.p is not None else -self.b0
min_sum = k[self.target_class_index] - np.log((1 - eps) / eps)
max_sum = k[self.target_class_index] - np.log(eps / (1 - eps))
if self.align == OWNomogram.ALIGN_LEFT:
max_sum = max_sum - sum(minimums)
min_sum = min_sum - sum(minimums)
for i in range(len(k)):
k[i] = k[i] - sum(
[min(q) for q in [p[i] for p in self.points]])
if self.scale == OWNomogram.POINT_SCALE:
min_sum *= d
max_sum *= d
d_ = d
values = self.get_ruler_values(min_sum, max_sum, max_width)
min_sum, max_sum = min(values), max(values)
diff_ = np.nan_to_num(max_sum - min_sum)
scale_x = max_width / diff_ if diff_ else max_width
cls_var, cls_index = self.domain.class_var, self.target_class_index
nomogram_footer = NomogramItem()
def get_normalized_probabilities(val):
if not self.normalize_probabilities:
return 1 / (1 + np.exp(k[cls_index] - val / d_))
totals = self.__get_totals_for_class_values(minimums)
p_sum = np.sum(1 / (1 + np.exp(k - totals / d_)))
return 1 / (1 + np.exp(k[cls_index] - val / d_)) / p_sum
def get_points(prob):
if not self.normalize_probabilities:
return (k[cls_index] - np.log(1 / prob - 1)) * d_
totals = self.__get_totals_for_class_values(minimums)
p_sum = np.sum(1 / (1 + np.exp(k - totals / d_)))
return (k[cls_index] - np.log(1 / (prob * p_sum) - 1)) * d_
self.markers_set = False
probs_item = ProbabilitiesRulerItem(
probs_text,
values,
scale_x,
name_offset,
-scale_x * min_sum,
get_points=get_points,
title="{}='{}'".format(cls_var.name, cls_var.values[cls_index]),
get_probabilities=get_normalized_probabilities)
self.markers_set = True
nomogram_footer.add_items([probs_item])
return probs_item, nomogram_footer
def __get_totals_for_class_values(self, minimums):
cls_index = self.target_class_index
marker_values = [item.dot.value for item in self.feature_items]
if not self.markers_set:
marker_values = self.scale_forth(marker_values)
totals = np.empty(len(self.domain.class_var.values))
totals[cls_index] = sum(marker_values)
marker_values = self.scale_back(marker_values)
for i in range(len(self.domain.class_var.values)):
if i == cls_index:
continue
coeffs = [np.nan_to_num(p[i] / p[cls_index]) for p in self.points]
points = [p[cls_index] for p in self.points]
total = sum([
self.get_points_from_coeffs(v, c, p)
for (v, c, p) in zip(marker_values, coeffs, points)
])
if self.align == OWNomogram.ALIGN_LEFT:
points = [p - m for m, p in zip(minimums, points)]
total -= sum([min(p) for p in [p[i] for p in self.points]])
d = 100 / max(max(abs(p)) for p in points)
if self.scale == OWNomogram.POINT_SCALE:
total *= d
totals[i] = total
return totals
def set_feature_marker_values(self):
if not (len(self.points) and len(self.feature_items)):
return
if not len(self.feature_marker_values):
self._init_feature_marker_values()
self.feature_marker_values = self.scale_forth(
self.feature_marker_values)
item = self.feature_items[0]
for i, item in enumerate(self.feature_items):
item.dot.move_to_val(self.feature_marker_values[i])
item.dot.probs_dot.move_to_sum()
def _init_feature_marker_values(self):
self.feature_marker_values = []
cls_index = self.target_class_index
instances = Table(self.domain, self.instances) \
if self.instances else None
for i, attr in enumerate(self.domain.attributes):
value, feature_val = 0, None
if len(self.log_reg_coeffs):
if attr.is_discrete:
ind, n = unique(self.data.X[:, i], return_counts=True)
feature_val = np.nan_to_num(ind[np.argmax(n)])
else:
feature_val = mean(self.data.X[:, i])
inst_in_dom = instances and attr in instances.domain
if inst_in_dom and not np.isnan(instances[0][attr]):
feature_val = instances[0][attr]
if feature_val is not None:
value = self.points[i][cls_index][int(feature_val)] \
if attr.is_discrete else \
self.log_reg_coeffs_orig[i][cls_index][0] * feature_val
self.feature_marker_values.append(value)
def clear_scene(self):
self.feature_items = []
self.scale_back = lambda x: x
self.scale_forth = lambda x: x
self.nomogram = None
self.nomogram_main = None
self.vertical_line = None
self.hidden_vertical_line = None
self.scene.clear()
def send_report(self):
self.report_plot()
@staticmethod
def reconstruct_domain(original, preprocessed):
# abuse dict to make "in" comparisons faster
attrs = OrderedDict()
for attr in preprocessed.attributes:
cv = attr._compute_value.variable._compute_value
var = cv.variable if cv else original[attr.name]
if var in attrs: # the reason for OrderedDict
continue
attrs[var] = None # we only need keys
attrs = list(attrs.keys())
return Domain(attrs, original.class_var, original.metas)
@staticmethod
def get_ruler_values(start, stop, max_width, round_to_nearest=True):
if max_width == 0:
return [0]
diff = np.nan_to_num((stop - start) / max_width)
if diff <= 0:
return [0]
decimals = int(np.floor(np.log10(diff)))
if diff > 4 * pow(10, decimals):
step = 5 * pow(10, decimals + 2)
elif diff > 2 * pow(10, decimals):
step = 2 * pow(10, decimals + 2)
elif diff > 1 * pow(10, decimals):
step = 1 * pow(10, decimals + 2)
else:
step = 5 * pow(10, decimals + 1)
round_by = int(-np.floor(np.log10(step)))
r = start % step
if not round_to_nearest:
_range = np.arange(start + step, stop + r, step) - r
start, stop = np.floor(start * 100) / 100, np.ceil(
stop * 100) / 100
return np.round(np.hstack((start, _range, stop)), 2)
return np.round(np.arange(start, stop + r + step, step) - r, round_by)
@staticmethod
def get_points_from_coeffs(current_value, coefficients, possible_values):
if any(np.isnan(possible_values)):
return 0
indices = np.argsort(possible_values)
sorted_values = possible_values[indices]
sorted_coefficients = coefficients[indices]
for i, val in enumerate(sorted_values):
if current_value < val:
break
diff = sorted_values[i] - sorted_values[i - 1]
k = 0 if diff < 1e-6 else (sorted_values[i] - current_value) / \
(sorted_values[i] - sorted_values[i - 1])
return sorted_coefficients[i - 1] * sorted_values[i - 1] * k + \
sorted_coefficients[i] * sorted_values[i] * (1 - k)
def __init__(self, name, data_extremes, values, scale, name_offset, offset,
coef):
super().__init__()
data_start, data_stop = data_extremes[0], data_extremes[1]
self.name = name.toPlainText()
self.diff = (data_stop - data_start) * coef
labels = [
str(
np.round(
data_start + (data_stop - data_start) * i /
(self.n_tck - 1), 1)) for i in range(self.n_tck)
]
# leading label
font = name.document().defaultFont()
name.setFont(font)
name.setPos(name_offset, -10)
name.setParentItem(self)
# labels
ascending = data_start < data_stop
y_start, y_stop = (self.y_diff, 0) if ascending else (0, self.y_diff)
for i in range(self.n_tck):
text = QGraphicsSimpleTextItem(labels[i])
w = text.boundingRect().width()
y = y_start + (y_stop - y_start) / (self.n_tck - 1) * i
text.setPos(-5 - w, y - 8)
text.setParentItem(self)
tick = QGraphicsLineItem(-2, y, 2, y)
tick.setParentItem(self)
# prediction marker
self.dot = Continuous2DMovableDotItem(self.dot_r, scale, offset,
values[0], values[-1], y_start,
y_stop)
self.dot.tooltip_labels = labels
self.dot.tooltip_values = values
self.dot.setParentItem(self)
h_line = QGraphicsLineItem(values[0] * scale + offset, self.y_diff / 2,
values[-1] * scale + offset,
self.y_diff / 2)
pen = QPen(Qt.DashLine)
pen.setBrush(QColor(Qt.red))
h_line.setPen(pen)
h_line.setParentItem(self)
self.dot.horizontal_line = h_line
# line
line = QGraphicsLineItem(values[0] * scale + offset, y_start,
values[-1] * scale + offset, y_stop)
line.setParentItem(self)
# ticks
for value in values:
diff_ = np.nan_to_num(values[-1] - values[0])
k = (value - values[0]) / diff_ if diff_ else 0
y_tick = (y_stop - y_start) * k + y_start - self.tick_height / 2
x_tick = value * scale - self.tick_width / 2 + offset
tick = QGraphicsRectItem(x_tick, y_tick, self.tick_width,
self.tick_height)
tick.setBrush(QColor(Qt.black))
tick.setParentItem(self)
# rect
rect = QGraphicsRectItem(values[0] * scale + offset,
-self.y_diff * 0.125,
values[-1] * scale + offset,
self.y_diff * 1.25)
pen = QPen(Qt.DotLine)
pen.setBrush(QColor(50, 150, 200, 255))
rect.setPen(pen)
rect.setParentItem(self)
self.setPreferredSize(self.preferredWidth(), self.y_diff * 1.5)
class LinePlotViewBox(ViewBox):
selection_changed = Signal(np.ndarray)
def __init__(self):
super().__init__(enableMenu=False)
self._profile_items = None
self._can_select = True
self._graph_state = SELECT
self.setMouseMode(self.PanMode)
pen = mkPen(LinePlotStyle.SELECTION_LINE_COLOR,
width=LinePlotStyle.SELECTION_LINE_WIDTH)
self.selection_line = QGraphicsLineItem()
self.selection_line.setPen(pen)
self.selection_line.setZValue(1e9)
self.addItem(self.selection_line, ignoreBounds=True)
def update_selection_line(self, button_down_pos, current_pos):
p1 = self.childGroup.mapFromParent(button_down_pos)
p2 = self.childGroup.mapFromParent(current_pos)
self.selection_line.setLine(QLineF(p1, p2))
self.selection_line.resetTransform()
self.selection_line.show()
def set_graph_state(self, state):
self._graph_state = state
def enable_selection(self, enable):
self._can_select = enable
def get_selected(self, p1, p2):
if self._profile_items is None:
return np.array(False)
return line_intersects_profiles(np.array([p1.x(), p1.y()]),
np.array([p2.x(), p2.y()]),
self._profile_items)
def add_profiles(self, y):
if sp.issparse(y):
y = y.todense()
self._profile_items = np.array(
[np.vstack((np.full((1, y.shape[0]), i + 1), y[:, i].flatten())).T
for i in range(y.shape[1])])
def remove_profiles(self):
self._profile_items = None
def mouseDragEvent(self, ev, axis=None):
if self._graph_state == SELECT and axis is None and self._can_select:
ev.accept()
if ev.button() == Qt.LeftButton:
self.update_selection_line(ev.buttonDownPos(), ev.pos())
if ev.isFinish():
self.selection_line.hide()
p1 = self.childGroup.mapFromParent(
ev.buttonDownPos(ev.button()))
p2 = self.childGroup.mapFromParent(ev.pos())
self.selection_changed.emit(self.get_selected(p1, p2))
elif self._graph_state == ZOOMING or self._graph_state == PANNING:
ev.ignore()
super().mouseDragEvent(ev, axis=axis)
else:
ev.ignore()
def mouseClickEvent(self, ev):
if ev.button() == Qt.RightButton:
self.autoRange()
self.enableAutoRange()
else:
ev.accept()
self.selection_changed.emit(np.array(False))
def reset(self):
self._profile_items = None
self._can_select = True
self._graph_state = SELECT
class OWNomogram(OWWidget):
name = "Nomogram"
description = " Nomograms for Visualization of Naive Bayesian" \
" and Logistic Regression Classifiers."
icon = "icons/Nomogram.svg"
priority = 2000
class Inputs:
classifier = Input("Classifier", Model)
data = Input("Data", Table)
MAX_N_ATTRS = 1000
POINT_SCALE = 0
ALIGN_LEFT = 0
ALIGN_ZERO = 1
ACCEPTABLE = (NaiveBayesModel, LogisticRegressionClassifier)
settingsHandler = ClassValuesContextHandler()
target_class_index = ContextSetting(0)
normalize_probabilities = Setting(False)
scale = Setting(1)
display_index = Setting(1)
n_attributes = Setting(10)
sort_index = Setting(SortBy.ABSOLUTE)
cont_feature_dim_index = Setting(0)
graph_name = "scene"
class Error(OWWidget.Error):
invalid_classifier = Msg("Nomogram accepts only Naive Bayes and "
"Logistic Regression classifiers.")
def __init__(self):
super().__init__()
self.instances = None
self.domain = None
self.data = None
self.classifier = None
self.align = OWNomogram.ALIGN_ZERO
self.log_odds_ratios = []
self.log_reg_coeffs = []
self.log_reg_coeffs_orig = []
self.log_reg_cont_data_extremes = []
self.p = None
self.b0 = None
self.points = []
self.feature_items = {}
self.feature_marker_values = []
self.scale_marker_values = lambda x: x
self.nomogram_main = None
self.vertical_line = None
self.hidden_vertical_line = None
self.old_target_class_index = self.target_class_index
self.repaint = False
# GUI
box = gui.vBox(self.controlArea, "Target class")
self.class_combo = gui.comboBox(
box, self, "target_class_index", callback=self._class_combo_changed,
contentsLength=12)
self.norm_check = gui.checkBox(
box, self, "normalize_probabilities", "Normalize probabilities",
hidden=True, callback=self.update_scene,
tooltip="For multiclass data 1 vs. all probabilities do not"
" sum to 1 and therefore could be normalized.")
self.scale_radio = gui.radioButtons(
self.controlArea, self, "scale", ["Point scale", "Log odds ratios"],
box="Scale", callback=self.update_scene)
box = gui.vBox(self.controlArea, "Display features")
grid = QGridLayout()
radio_group = gui.radioButtonsInBox(
box, self, "display_index", [], orientation=grid,
callback=self.update_scene)
radio_all = gui.appendRadioButton(
radio_group, "All", addToLayout=False)
radio_best = gui.appendRadioButton(
radio_group, "Best ranked:", addToLayout=False)
spin_box = gui.hBox(None, margin=0)
self.n_spin = gui.spin(
spin_box, self, "n_attributes", 1, self.MAX_N_ATTRS, label=" ",
controlWidth=60, callback=self._n_spin_changed)
grid.addWidget(radio_all, 1, 1)
grid.addWidget(radio_best, 2, 1)
grid.addWidget(spin_box, 2, 2)
self.sort_combo = gui.comboBox(
box, self, "sort_index", label="Rank by:", items=SortBy.items(),
orientation=Qt.Horizontal, callback=self.update_scene)
self.cont_feature_dim_combo = gui.comboBox(
box, self, "cont_feature_dim_index", label="Numeric features: ",
items=["1D projection", "2D curve"], orientation=Qt.Horizontal,
callback=self.update_scene)
gui.rubber(self.controlArea)
class _GraphicsView(QGraphicsView):
def __init__(self, scene, parent, **kwargs):
for k, v in dict(verticalScrollBarPolicy=Qt.ScrollBarAlwaysOff,
horizontalScrollBarPolicy=Qt.ScrollBarAlwaysOff,
viewportUpdateMode=QGraphicsView.BoundingRectViewportUpdate,
renderHints=(QPainter.Antialiasing |
QPainter.TextAntialiasing |
QPainter.SmoothPixmapTransform),
alignment=(Qt.AlignTop |
Qt.AlignLeft),
sizePolicy=QSizePolicy(QSizePolicy.MinimumExpanding,
QSizePolicy.MinimumExpanding)).items():
kwargs.setdefault(k, v)
super().__init__(scene, parent, **kwargs)
class GraphicsView(_GraphicsView):
def __init__(self, scene, parent):
super().__init__(scene, parent,
verticalScrollBarPolicy=Qt.ScrollBarAlwaysOn,
styleSheet='QGraphicsView {background: white}')
self.viewport().setMinimumWidth(300) # XXX: This prevents some tests failing
self._is_resizing = False
w = self
def resizeEvent(self, resizeEvent):
# Recompute main scene on window width change
if resizeEvent.size().width() != resizeEvent.oldSize().width():
self._is_resizing = True
self.w.update_scene()
self._is_resizing = False
return super().resizeEvent(resizeEvent)
def is_resizing(self):
return self._is_resizing
def sizeHint(self):
return QSize(400, 200)
class FixedSizeGraphicsView(_GraphicsView):
def __init__(self, scene, parent):
super().__init__(scene, parent,
sizePolicy=QSizePolicy(QSizePolicy.MinimumExpanding,
QSizePolicy.Minimum))
def sizeHint(self):
return QSize(400, 85)
scene = self.scene = QGraphicsScene(self)
top_view = self.top_view = FixedSizeGraphicsView(scene, self)
mid_view = self.view = GraphicsView(scene, self)
bottom_view = self.bottom_view = FixedSizeGraphicsView(scene, self)
for view in (top_view, mid_view, bottom_view):
self.mainArea.layout().addWidget(view)
def _class_combo_changed(self):
with np.errstate(invalid='ignore'):
coeffs = [np.nan_to_num(p[self.target_class_index] /
p[self.old_target_class_index])
for p in self.points]
points = [p[self.old_target_class_index] for p in self.points]
self.feature_marker_values = [
self.get_points_from_coeffs(v, c, p) for (v, c, p) in
zip(self.feature_marker_values, coeffs, points)]
self.feature_marker_values = np.asarray(self.feature_marker_values)
self.update_scene()
self.old_target_class_index = self.target_class_index
def _n_spin_changed(self):
self.display_index = 1
self.update_scene()
def update_controls(self):
self.class_combo.clear()
self.norm_check.setHidden(True)
self.cont_feature_dim_combo.setEnabled(True)
if self.domain is not None:
self.class_combo.addItems(self.domain.class_vars[0].values)
if len(self.domain.attributes) > self.MAX_N_ATTRS:
self.display_index = 1
if len(self.domain.class_vars[0].values) > 2:
self.norm_check.setHidden(False)
if not self.domain.has_continuous_attributes():
self.cont_feature_dim_combo.setEnabled(False)
self.cont_feature_dim_index = 0
model = self.sort_combo.model()
item = model.item(SortBy.POSITIVE)
item.setFlags(item.flags() | Qt.ItemIsEnabled)
item = model.item(SortBy.NEGATIVE)
item.setFlags(item.flags() | Qt.ItemIsEnabled)
self.align = OWNomogram.ALIGN_ZERO
if self.classifier and isinstance(self.classifier,
LogisticRegressionClassifier):
self.align = OWNomogram.ALIGN_LEFT
@Inputs.data
def set_data(self, data):
self.instances = data
self.feature_marker_values = []
self.set_feature_marker_values()
self.update_scene()
@Inputs.classifier
def set_classifier(self, classifier):
self.closeContext()
self.classifier = classifier
self.Error.clear()
if self.classifier and not isinstance(self.classifier, self.ACCEPTABLE):
self.Error.invalid_classifier()
self.classifier = None
self.domain = self.classifier.domain if self.classifier else None
self.data = None
self.calculate_log_odds_ratios()
self.calculate_log_reg_coefficients()
self.update_controls()
self.target_class_index = 0
self.openContext(self.domain.class_var if self.domain is not None
else None)
self.points = self.log_odds_ratios or self.log_reg_coeffs
self.feature_marker_values = []
self.old_target_class_index = self.target_class_index
self.update_scene()
def calculate_log_odds_ratios(self):
self.log_odds_ratios = []
self.p = None
if self.classifier is None or self.domain is None:
return
if not isinstance(self.classifier, NaiveBayesModel):
return
log_cont_prob = self.classifier.log_cont_prob
class_prob = self.classifier.class_prob
for i in range(len(self.domain.attributes)):
ca = np.exp(log_cont_prob[i]) * class_prob[:, None]
_or = (ca / (1 - ca)) / (class_prob / (1 - class_prob))[:, None]
self.log_odds_ratios.append(np.log(_or))
self.p = class_prob
def calculate_log_reg_coefficients(self):
self.log_reg_coeffs = []
self.log_reg_cont_data_extremes = []
self.b0 = None
if self.classifier is None or self.domain is None:
return
if not isinstance(self.classifier, LogisticRegressionClassifier):
return
self.domain = self.reconstruct_domain(self.classifier.original_domain,
self.domain)
self.data = self.classifier.original_data.transform(self.domain)
attrs, ranges, start = self.domain.attributes, [], 0
for attr in attrs:
stop = start + len(attr.values) if attr.is_discrete else start + 1
ranges.append(slice(start, stop))
start = stop
self.b0 = self.classifier.intercept
coeffs = self.classifier.coefficients
if len(self.domain.class_var.values) == 2:
self.b0 = np.hstack((self.b0 * (-1), self.b0))
coeffs = np.vstack((coeffs * (-1), coeffs))
self.log_reg_coeffs = [coeffs[:, ranges[i]] for i in range(len(attrs))]
self.log_reg_coeffs_orig = self.log_reg_coeffs.copy()
min_values = nanmin(self.data.X, axis=0)
max_values = nanmax(self.data.X, axis=0)
for i, min_t, max_t in zip(range(len(self.log_reg_coeffs)),
min_values, max_values):
if self.log_reg_coeffs[i].shape[1] == 1:
coef = self.log_reg_coeffs[i]
self.log_reg_coeffs[i] = np.hstack((coef * min_t, coef * max_t))
self.log_reg_cont_data_extremes.append(
[sorted([min_t, max_t], reverse=(c < 0)) for c in coef])
else:
self.log_reg_cont_data_extremes.append([None])
def update_scene(self):
self.clear_scene()
if self.domain is None or not len(self.points[0]):
return
n_attrs = self.n_attributes if self.display_index else int(1e10)
attr_inds, attributes = zip(*self.get_ordered_attributes()[:n_attrs])
name_items = [QGraphicsTextItem(attr.name) for attr in attributes]
point_text = QGraphicsTextItem("Points")
probs_text = QGraphicsTextItem("Probabilities (%)")
all_items = name_items + [point_text, probs_text]
name_offset = -max(t.boundingRect().width() for t in all_items) - 10
w = self.view.viewport().rect().width()
max_width = w + name_offset - 30
points = [self.points[i][self.target_class_index]
for i in attr_inds]
if self.align == OWNomogram.ALIGN_LEFT:
points = [p - p.min() for p in points]
max_ = np.nan_to_num(max(max(abs(p)) for p in points))
d = 100 / max_ if max_ else 1
minimums = [p[self.target_class_index].min() for p in self.points]
if self.scale == OWNomogram.POINT_SCALE:
points = [p * d for p in points]
if self.align == OWNomogram.ALIGN_LEFT:
self.scale_marker_values = lambda x: (x - minimums) * d
else:
self.scale_marker_values = lambda x: x * d
else:
if self.align == OWNomogram.ALIGN_LEFT:
self.scale_marker_values = lambda x: x - minimums
else:
self.scale_marker_values = lambda x: x
point_item, nomogram_head = self.create_main_nomogram(
attributes, attr_inds,
name_items, points, max_width, point_text, name_offset)
probs_item, nomogram_foot = self.create_footer_nomogram(
probs_text, d, minimums, max_width, name_offset)
for item in self.feature_items.values():
item.dot.point_dot = point_item.dot
item.dot.probs_dot = probs_item.dot
item.dot.vertical_line = self.hidden_vertical_line
self.nomogram = nomogram = NomogramItem()
nomogram.add_items([nomogram_head, self.nomogram_main, nomogram_foot])
self.scene.addItem(nomogram)
self.set_feature_marker_values()
rect = QRectF(self.scene.itemsBoundingRect().x(),
self.scene.itemsBoundingRect().y(),
self.scene.itemsBoundingRect().width(),
self.nomogram.preferredSize().height()).adjusted(10, 0, 20, 0)
self.scene.setSceneRect(rect)
# Clip top and bottom (60 and 150) parts from the main view
self.view.setSceneRect(rect.x(), rect.y() + 80, rect.width() - 10, rect.height() - 160)
self.view.viewport().setMaximumHeight(rect.height() - 160)
# Clip main part from top/bottom views
# below point values are imprecise (less/more than required) but this
# is not a problem due to clipped scene content still being drawn
self.top_view.setSceneRect(rect.x(), rect.y() + 3, rect.width() - 10, 20)
self.bottom_view.setSceneRect(rect.x(), rect.height() - 110, rect.width() - 10, 30)
def create_main_nomogram(self, attributes, attr_inds, name_items, points,
max_width, point_text, name_offset):
cls_index = self.target_class_index
min_p = min(p.min() for p in points)
max_p = max(p.max() for p in points)
values = self.get_ruler_values(min_p, max_p, max_width)
min_p, max_p = min(values), max(values)
diff_ = np.nan_to_num(max_p - min_p)
scale_x = max_width / diff_ if diff_ else max_width
nomogram_header = NomogramItem()
point_item = RulerItem(point_text, values, scale_x, name_offset,
- scale_x * min_p)
point_item.setPreferredSize(point_item.preferredWidth(), 35)
nomogram_header.add_items([point_item])
self.nomogram_main = NomogramItem()
cont_feature_item_class = ContinuousFeature2DItem if \
self.cont_feature_dim_index else ContinuousFeatureItem
feature_items = [
DiscreteFeatureItem(
name_item, attr.values, point,
scale_x, name_offset, - scale_x * min_p)
if attr.is_discrete else
cont_feature_item_class(
name_item, self.log_reg_cont_data_extremes[i][cls_index],
self.get_ruler_values(
point.min(), point.max(),
scale_x * point.ptp(), False),
scale_x, name_offset, - scale_x * min_p)
for i, attr, name_item, point in zip(attr_inds, attributes, name_items, points)]
self.nomogram_main.add_items(feature_items)
self.feature_items = OrderedDict(sorted(zip(attr_inds, feature_items)))
x = - scale_x * min_p
y = self.nomogram_main.layout().preferredHeight() + 10
self.vertical_line = QGraphicsLineItem(x, -6, x, y)
self.vertical_line.setPen(QPen(Qt.DotLine))
self.vertical_line.setParentItem(point_item)
self.hidden_vertical_line = QGraphicsLineItem(x, -6, x, y)
pen = QPen(Qt.DashLine)
pen.setBrush(QColor(Qt.red))
self.hidden_vertical_line.setPen(pen)
self.hidden_vertical_line.setParentItem(point_item)
return point_item, nomogram_header
def get_ordered_attributes(self):
"""Return (in_domain_index, attr) pairs, ordered by method in SortBy combo"""
if self.domain is None or not self.domain.attributes:
return []
attrs = self.domain.attributes
sort_by = self.sort_index
class_value = self.target_class_index
if sort_by == SortBy.NO_SORTING:
return list(enumerate(attrs))
elif sort_by == SortBy.NAME:
def key(x):
_, attr = x
return attr.name.lower()
elif sort_by == SortBy.ABSOLUTE:
def key(x):
i, attr = x
if attr.is_discrete:
ptp = self.points[i][class_value].ptp()
else:
coef = np.abs(self.log_reg_coeffs_orig[i][class_value]).mean()
ptp = coef * np.ptp(self.log_reg_cont_data_extremes[i][class_value])
return -ptp
elif sort_by == SortBy.POSITIVE:
def key(x):
i, attr = x
max_value = (self.points[i][class_value].max()
if attr.is_discrete else
np.mean(self.log_reg_cont_data_extremes[i][class_value]))
return -max_value
elif sort_by == SortBy.NEGATIVE:
def key(x):
i, attr = x
min_value = (self.points[i][class_value].min()
if attr.is_discrete else
np.mean(self.log_reg_cont_data_extremes[i][class_value]))
return min_value
return sorted(enumerate(attrs), key=key)
def create_footer_nomogram(self, probs_text, d, minimums,
max_width, name_offset):
eps, d_ = 0.05, 1
k = - np.log(self.p / (1 - self.p)) if self.p is not None else - self.b0
min_sum = k[self.target_class_index] - np.log((1 - eps) / eps)
max_sum = k[self.target_class_index] - np.log(eps / (1 - eps))
if self.align == OWNomogram.ALIGN_LEFT:
max_sum = max_sum - sum(minimums)
min_sum = min_sum - sum(minimums)
for i in range(len(k)):
k[i] = k[i] - sum([min(q) for q in [p[i] for p in self.points]])
if self.scale == OWNomogram.POINT_SCALE:
min_sum *= d
max_sum *= d
d_ = d
values = self.get_ruler_values(min_sum, max_sum, max_width)
min_sum, max_sum = min(values), max(values)
diff_ = np.nan_to_num(max_sum - min_sum)
scale_x = max_width / diff_ if diff_ else max_width
cls_var, cls_index = self.domain.class_var, self.target_class_index
nomogram_footer = NomogramItem()
def get_normalized_probabilities(val):
if not self.normalize_probabilities:
return 1 / (1 + np.exp(k[cls_index] - val / d_))
totals = self.__get_totals_for_class_values(minimums)
p_sum = np.sum(1 / (1 + np.exp(k - totals / d_)))
return 1 / (1 + np.exp(k[cls_index] - val / d_)) / p_sum
def get_points(prob):
if not self.normalize_probabilities:
return (k[cls_index] - np.log(1 / prob - 1)) * d_
totals = self.__get_totals_for_class_values(minimums)
p_sum = np.sum(1 / (1 + np.exp(k - totals / d_)))
return (k[cls_index] - np.log(1 / (prob * p_sum) - 1)) * d_
probs_item = ProbabilitiesRulerItem(
probs_text, values, scale_x, name_offset, - scale_x * min_sum,
get_points=get_points,
title="{}='{}'".format(cls_var.name, cls_var.values[cls_index]),
get_probabilities=get_normalized_probabilities)
nomogram_footer.add_items([probs_item])
return probs_item, nomogram_footer
def __get_totals_for_class_values(self, minimums):
cls_index = self.target_class_index
marker_values = self.scale_marker_values(self.feature_marker_values)
totals = np.full(len(self.domain.class_var.values), np.nan)
totals[cls_index] = marker_values.sum()
for i in range(len(self.domain.class_var.values)):
if i == cls_index:
continue
coeffs = [np.nan_to_num(p[i] / p[cls_index]) for p in self.points]
points = [p[cls_index] for p in self.points]
total = sum([self.get_points_from_coeffs(v, c, p) for (v, c, p)
in zip(self.feature_marker_values, coeffs, points)])
if self.align == OWNomogram.ALIGN_LEFT:
points = [p - m for m, p in zip(minimums, points)]
total -= sum([min(p) for p in [p[i] for p in self.points]])
d = 100 / max(max(abs(p)) for p in points)
if self.scale == OWNomogram.POINT_SCALE:
total *= d
totals[i] = total
assert not np.any(np.isnan(totals))
return totals
def set_feature_marker_values(self):
if not (len(self.points) and len(self.feature_items)):
return
if not len(self.feature_marker_values):
self._init_feature_marker_values()
marker_values = self.scale_marker_values(self.feature_marker_values)
invisible_sum = 0
for i in range(len(marker_values)):
try:
item = self.feature_items[i]
except KeyError:
invisible_sum += marker_values[i]
else:
item.dot.move_to_val(marker_values[i])
item.dot.probs_dot.move_to_sum(invisible_sum)
def _init_feature_marker_values(self):
self.feature_marker_values = []
cls_index = self.target_class_index
instances = Table(self.domain, self.instances) \
if self.instances else None
values = []
for i, attr in enumerate(self.domain.attributes):
value, feature_val = 0, None
if len(self.log_reg_coeffs):
if attr.is_discrete:
ind, n = unique(self.data.X[:, i], return_counts=True)
feature_val = np.nan_to_num(ind[np.argmax(n)])
else:
feature_val = nanmean(self.data.X[:, i])
# If data is provided on a separate signal, use the first data
# instance to position the points instead of the mean
inst_in_dom = instances and attr in instances.domain
if inst_in_dom and not np.isnan(instances[0][attr]):
feature_val = instances[0][attr]
if feature_val is not None:
value = (self.points[i][cls_index][int(feature_val)]
if attr.is_discrete else
self.log_reg_coeffs_orig[i][cls_index][0] * feature_val)
values.append(value)
self.feature_marker_values = np.asarray(values)
def clear_scene(self):
self.feature_items = {}
self.scale_marker_values = lambda x: x
self.nomogram = None
self.nomogram_main = None
self.vertical_line = None
self.hidden_vertical_line = None
self.scene.clear()
def send_report(self):
self.report_plot()
@staticmethod
def reconstruct_domain(original, preprocessed):
# abuse dict to make "in" comparisons faster
attrs = OrderedDict()
for attr in preprocessed.attributes:
cv = attr._compute_value.variable._compute_value
var = cv.variable if cv else original[attr.name]
if var in attrs: # the reason for OrderedDict
continue
attrs[var] = None # we only need keys
attrs = list(attrs.keys())
return Domain(attrs, original.class_var, original.metas)
@staticmethod
def get_ruler_values(start, stop, max_width, round_to_nearest=True):
if max_width == 0:
return [0]
diff = np.nan_to_num((stop - start) / max_width)
if diff <= 0:
return [0]
decimals = int(np.floor(np.log10(diff)))
if diff > 4 * pow(10, decimals):
step = 5 * pow(10, decimals + 2)
elif diff > 2 * pow(10, decimals):
step = 2 * pow(10, decimals + 2)
elif diff > 1 * pow(10, decimals):
step = 1 * pow(10, decimals + 2)
else:
step = 5 * pow(10, decimals + 1)
round_by = int(- np.floor(np.log10(step)))
r = start % step
if not round_to_nearest:
_range = np.arange(start + step, stop + r, step) - r
start, stop = np.floor(start * 100) / 100, np.ceil(stop * 100) / 100
return np.round(np.hstack((start, _range, stop)), 2)
return np.round(np.arange(start, stop + r + step, step) - r, round_by)
@staticmethod
def get_points_from_coeffs(current_value, coefficients, possible_values):
if np.isnan(possible_values).any():
return 0
indices = np.argsort(possible_values)
sorted_values = possible_values[indices]
sorted_coefficients = coefficients[indices]
for i, val in enumerate(sorted_values):
if current_value < val:
break
diff = sorted_values[i] - sorted_values[i - 1]
k = 0 if diff < 1e-6 else (sorted_values[i] - current_value) / \
(sorted_values[i] - sorted_values[i - 1])
return sorted_coefficients[i - 1] * sorted_values[i - 1] * k + \
sorted_coefficients[i] * sorted_values[i] * (1 - k)
def __init__(self, name, data_extremes, values, scale, name_offset, offset):
super().__init__()
data_start, data_stop = data_extremes[0], data_extremes[1]
labels = [str(np.round(data_start + (data_stop - data_start) * i /
(self.n_tck - 1), 1)) for i in range(self.n_tck)]
# leading label
font = name.document().defaultFont()
name.setFont(font)
name.setPos(name_offset, -10)
name.setParentItem(self)
# labels
ascending = data_start < data_stop
y_start, y_stop = (self.y_diff, 0) if ascending else (0, self.y_diff)
for i in range(self.n_tck):
text = QGraphicsSimpleTextItem(labels[i], self)
w = text.boundingRect().width()
y = y_start + (y_stop - y_start) / (self.n_tck - 1) * i
text.setPos(-5 - w, y - 8)
tick = QGraphicsLineItem(-2, y, 2, y, self)
# prediction marker
self.dot = Continuous2DMovableDotItem(
self.DOT_RADIUS, scale, offset, values[0], values[-1], y_start, y_stop)
self.dot.tooltip_labels = labels
self.dot.tooltip_values = values
self.dot.setParentItem(self)
h_line = QGraphicsLineItem(values[0] * scale + offset, self.y_diff / 2,
values[-1] * scale + offset, self.y_diff / 2,
self)
pen = QPen(Qt.DashLine)
pen.setBrush(QColor(Qt.red))
h_line.setPen(pen)
self.dot.horizontal_line = h_line
# pylint: disable=unused-variable
# line
line = QGraphicsLineItem(values[0] * scale + offset, y_start,
values[-1] * scale + offset, y_stop,
self)
# ticks
for value in values:
diff_ = np.nan_to_num(values[-1] - values[0])
k = (value - values[0]) / diff_ if diff_ else 0
y_tick = (y_stop - y_start) * k + y_start - self.tick_height / 2
x_tick = value * scale - self.tick_width / 2 + offset
tick = QGraphicsRectItem(
x_tick, y_tick, self.tick_width, self.tick_height,
self)
tick.setBrush(QColor(Qt.black))
# rect
rect = QGraphicsRectItem(
values[0] * scale + offset, -self.y_diff * 0.125,
values[-1] * scale + offset, self.y_diff * 1.25,
self)
pen = QPen(Qt.DotLine)
pen.setBrush(QColor(50, 150, 200, 255))
rect.setPen(pen)
self.setPreferredSize(self.preferredWidth(), self.y_diff * 1.5)
def _offseted_line(ax, ay):
r = QGraphicsLineItem(x + ax, y + ay, x + (ax or w),
y + (ay or h))
self.canvas.addItem(r)
r.setPen(pen)
class FeaturesPlot(QGraphicsWidget):
BOTTOM_AXIS_LABEL = "Feature Importance"
LABEL_COLUMN, ITEM_COLUMN = range(2)
ITEM_COLUMN_WIDTH, OFFSET = 300, 80
selection_cleared = Signal()
selection_changed = Signal(object)
resized = Signal()
def __init__(self):
super().__init__()
self._item_column_width = self.ITEM_COLUMN_WIDTH
self._range: Optional[Tuple[float, float]] = None
self._items: List[FeatureItem] = []
self._variable_items: List[VariableItem] = []
self._bottom_axis = AxisItem(parent=self,
orientation="bottom",
maxTickLength=7,
pen=QPen(Qt.black))
self._bottom_axis.setLabel(self.BOTTOM_AXIS_LABEL)
self._vertical_line = QGraphicsLineItem(self._bottom_axis)
self._vertical_line.setPen(QPen(Qt.gray))
self._layout = QGraphicsGridLayout()
self._layout.setVerticalSpacing(0)
self.setLayout(self._layout)
self.parameter_setter = BaseParameterSetter(self)
@property
def item_column_width(self) -> int:
return self._item_column_width
@item_column_width.setter
def item_column_width(self, view_width: int):
j = FeaturesPlot.LABEL_COLUMN
w = max([
self._layout.itemAt(i, j).item.boundingRect().width()
for i in range(len(self._items))
] + [0])
width = view_width - self.OFFSET - w
self._item_column_width = max(self.ITEM_COLUMN_WIDTH, width)
@property
def x0_scaled(self) -> float:
min_max = self._range[1] - self._range[0]
return -self._range[0] * self.item_column_width / min_max
@property
def bottom_axis(self) -> AxisItem:
return self._bottom_axis
@property
def labels(self) -> List[VariableItem]:
return self._variable_items
def set_data(self, x: np.ndarray, names: List[str], n_attrs: int,
view_width: int, *plot_args):
self.item_column_width = view_width
self._set_range(x, *plot_args)
self._set_items(x, names, *plot_args)
self._set_labels(names)
self._set_bottom_axis()
self.set_n_visible(n_attrs)
def _set_range(self, *_):
raise NotImplementedError
def _set_items(self, *_):
raise NotImplementedError
def set_n_visible(self, n: int):
for i in range(len(self._items)):
item = self._layout.itemAt(i, FeaturesPlot.ITEM_COLUMN)
item.setVisible(i < n)
text_item = self._layout.itemAt(i, FeaturesPlot.LABEL_COLUMN).item
text_item.setVisible(i < n)
self.set_vertical_line()
def rescale(self, view_width: int):
self.item_column_width = view_width
for item in self._items:
item.rescale(self.item_column_width)
self._bottom_axis.setWidth(self.item_column_width)
x = self.x0_scaled
self._vertical_line.setLine(x, 0, x, self._vertical_line.line().y2())
self.updateGeometry()
def set_height(self, height: float):
for i in range(len(self._items)):
item = self._layout.itemAt(i, FeaturesPlot.ITEM_COLUMN)
item.set_height(height)
self.set_vertical_line()
self.updateGeometry()
def _set_labels(self, labels: List[str]):
for i, (label, _) in enumerate(zip(labels, self._items)):
text = VariableItem(self, label)
item = SimpleLayoutItem(text)
item.setSizePolicy(QSizePolicy.Fixed, QSizePolicy.Fixed)
self._layout.addItem(item, i, FeaturesPlot.LABEL_COLUMN,
Qt.AlignRight | Qt.AlignVCenter)
self._variable_items.append(item)
def _set_bottom_axis(self):
self._bottom_axis.setRange(*self._range)
self._layout.addItem(self._bottom_axis, len(self._items),
FeaturesPlot.ITEM_COLUMN)
def set_vertical_line(self):
height = 0
for i in range(len(self._items)):
item = self._layout.itemAt(i, FeaturesPlot.ITEM_COLUMN)
text_item = self._layout.itemAt(i, FeaturesPlot.LABEL_COLUMN).item
if item.isVisible():
height += max(text_item.boundingRect().height(),
item.preferredSize().height())
self._vertical_line.setLine(self.x0_scaled, 0, self.x0_scaled, -height)
def deselect(self):
self.selection_cleared.emit()
def select(self, *args):
self.selection_changed.emit(*args)
def select_from_settings(self, *_):
raise NotImplementedError
def apply_visual_settings(self, settings: Dict):
for key, value in settings.items():
self.parameter_setter.set_parameter(key, value)
def line(x1, y1, x2, y2):
r = QGraphicsLineItem(x1, y1, x2, y2, None)
self.canvas.addItem(r)
r.setPen(QPen(Qt.white, 2))
r.setZValue(30)
class OWAxis(QGraphicsItem):
Role = OWPalette.Axis
def __init__(self, id, title='', title_above=False, title_location=AxisMiddle,
line=None, arrows=0, plot=None, bounds=None):
QGraphicsItem.__init__(self)
self.setFlag(QGraphicsItem.ItemHasNoContents)
self.setZValue(AxisZValue)
self.id = id
self.title = title
self.title_location = title_location
self.data_line = line
self.plot = plot
self.graph_line = None
self.size = None
self.scale = None
self.tick_length = (10, 5, 0)
self.arrows = arrows
self.title_above = title_above
self.line_item = QGraphicsLineItem(self)
self.title_item = QGraphicsTextItem(self)
self.end_arrow_item = None
self.start_arrow_item = None
self.show_title = False
self.scale = None
path = QPainterPath()
path.setFillRule(Qt.WindingFill)
path.moveTo(0, 3.09)
path.lineTo(0, -3.09)
path.lineTo(9.51, 0)
path.closeSubpath()
self.arrow_path = path
self.label_items = []
self.label_bg_items = []
self.tick_items = []
self._ticks = []
self.zoom_transform = QTransform()
self.labels = None
self.values = None
self._bounds = bounds
self.auto_range = None
self.auto_scale = True
self.zoomable = False
self.update_callback = None
self.max_text_width = 50
self.text_margin = 5
self.always_horizontal_text = False
@staticmethod
def compute_scale(min, max):
magnitude = int(3 * log10(abs(max - min)) + 1)
if magnitude % 3 == 0:
first_place = 1
elif magnitude % 3 == 1:
first_place = 2
else:
first_place = 5
magnitude = magnitude // 3 - 1
step = first_place * pow(10, magnitude)
first_val = ceil(min / step) * step
return first_val, step
def update_ticks(self):
self._ticks = []
major, medium, minor = self.tick_length
if self.labels is not None and not self.auto_scale:
values = self.values or range(len(self.labels))
for i, text in zip(values, self.labels):
self._ticks.append((i, text, medium, 1))
else:
if self.scale and not self.auto_scale:
min, max, step = self.scale
elif self.auto_range:
min, max = self.auto_range
if min is not None and max is not None:
step = (max - min) / 10
else:
return
else:
return
if max == min:
return
val, step = self.compute_scale(min, max)
while val <= max:
self._ticks.append((val, "%.4g" % val, medium, step))
val += step
def update_graph(self):
if self.update_callback:
self.update_callback()
def update(self, zoom_only=False):
self.update_ticks()
line_color = self.plot.color(OWPalette.Axis)
text_color = self.plot.color(OWPalette.Text)
if not self.graph_line or not self.scene():
return
self.line_item.setLine(self.graph_line)
self.line_item.setPen(line_color)
if self.title:
self.title_item.setHtml('<b>' + self.title + '</b>')
self.title_item.setDefaultTextColor(text_color)
if self.title_location == AxisMiddle:
title_p = 0.5
elif self.title_location == AxisEnd:
title_p = 0.95
else:
title_p = 0.05
title_pos = self.graph_line.pointAt(title_p)
v = self.graph_line.normalVector().unitVector()
dense_text = False
if hasattr(self, 'title_margin'):
offset = self.title_margin
elif self._ticks:
if self.should_be_expanded():
offset = 55
dense_text = True
else:
offset = 35
else:
offset = 10
if self.title_above:
title_pos += (v.p2() - v.p1()) * (offset + QFontMetrics(self.title_item.font()).height())
else:
title_pos -= (v.p2() - v.p1()) * offset
## TODO: Move it according to self.label_pos
self.title_item.setVisible(self.show_title)
self.title_item.setRotation(-self.graph_line.angle())
c = self.title_item.mapToParent(self.title_item.boundingRect().center())
tl = self.title_item.mapToParent(self.title_item.boundingRect().topLeft())
self.title_item.setPos(title_pos - c + tl)
## Arrows
if not zoom_only:
if self.start_arrow_item:
self.scene().removeItem(self.start_arrow_item)
self.start_arrow_item = None
if self.end_arrow_item:
self.scene().removeItem(self.end_arrow_item)
self.end_arrow_item = None
if self.arrows & AxisStart:
if not zoom_only or not self.start_arrow_item:
self.start_arrow_item = QGraphicsPathItem(self.arrow_path, self)
self.start_arrow_item.setPos(self.graph_line.p1())
self.start_arrow_item.setRotation(-self.graph_line.angle() + 180)
self.start_arrow_item.setBrush(line_color)
self.start_arrow_item.setPen(line_color)
if self.arrows & AxisEnd:
if not zoom_only or not self.end_arrow_item:
self.end_arrow_item = QGraphicsPathItem(self.arrow_path, self)
self.end_arrow_item.setPos(self.graph_line.p2())
self.end_arrow_item.setRotation(-self.graph_line.angle())
self.end_arrow_item.setBrush(line_color)
self.end_arrow_item.setPen(line_color)
## Labels
n = len(self._ticks)
resize_plot_item_list(self.label_items, n, QGraphicsTextItem, self)
resize_plot_item_list(self.label_bg_items, n, QGraphicsRectItem, self)
resize_plot_item_list(self.tick_items, n, QGraphicsLineItem, self)
test_rect = QRectF(self.graph_line.p1(), self.graph_line.p2()).normalized()
test_rect.adjust(-1, -1, 1, 1)
n_v = self.graph_line.normalVector().unitVector()
if self.title_above:
n_p = n_v.p2() - n_v.p1()
else:
n_p = n_v.p1() - n_v.p2()
l_v = self.graph_line.unitVector()
l_p = l_v.p2() - l_v.p1()
for i in range(n):
pos, text, size, step = self._ticks[i]
hs = 0.5 * step
tick_pos = self.map_to_graph(pos)
if not test_rect.contains(tick_pos):
self.tick_items[i].setVisible(False)
self.label_items[i].setVisible(False)
continue
item = self.label_items[i]
item.setVisible(True)
if not zoom_only:
if self.id in XAxes or getattr(self, 'is_horizontal', False):
item.setHtml('<center>' + Qt.escape(text.strip()) + '</center>')
else:
item.setHtml(Qt.escape(text.strip()))
item.setTextWidth(-1)
text_angle = 0
if dense_text:
w = min(item.boundingRect().width(), self.max_text_width)
item.setTextWidth(w)
if self.title_above:
label_pos = tick_pos + n_p * (w + self.text_margin) + l_p * item.boundingRect().height() / 2
else:
label_pos = tick_pos + n_p * self.text_margin + l_p * item.boundingRect().height() / 2
text_angle = -90 if self.title_above else 90
else:
w = min(item.boundingRect().width(),
QLineF(self.map_to_graph(pos - hs), self.map_to_graph(pos + hs)).length())
label_pos = tick_pos + n_p * self.text_margin + l_p * item.boundingRect().height() / 2
item.setTextWidth(w)
if not self.always_horizontal_text:
if self.title_above:
item.setRotation(-self.graph_line.angle() - text_angle)
else:
item.setRotation(self.graph_line.angle() - text_angle)
item.setPos(label_pos)
item.setDefaultTextColor(text_color)
self.label_bg_items[i].setRect(item.boundingRect())
self.label_bg_items[i].setPen(QPen(Qt.NoPen))
self.label_bg_items[i].setBrush(self.plot.color(OWPalette.Canvas))
item = self.tick_items[i]
item.setVisible(True)
tick_line = QLineF(v)
tick_line.translate(-tick_line.p1())
tick_line.setLength(size)
if self.title_above:
tick_line.setAngle(tick_line.angle() + 180)
item.setLine(tick_line)
item.setPen(line_color)
item.setPos(self.map_to_graph(pos))
@staticmethod
def make_title(label, unit=None):
lab = '<i>' + label + '</i>'
if unit:
lab = lab + ' [' + unit + ']'
return lab
def set_line(self, line):
self.graph_line = line
self.update()
def set_title(self, title):
self.title = title
self.update()
def set_show_title(self, b):
self.show_title = b
self.update()
def set_labels(self, labels, values):
self.labels = labels
self.values = values
self.graph_line = None
self.auto_scale = False
self.update_ticks()
self.update_graph()
def set_scale(self, min, max, step_size):
self.scale = (min, max, step_size)
self.graph_line = None
self.auto_scale = False
self.update_ticks()
self.update_graph()
def set_tick_length(self, minor, medium, major):
self.tick_length = (minor, medium, major)
self.update()
def map_to_graph(self, x):
min, max = self.plot.bounds_for_axis(self.id)
if min == max:
return QPointF()
line_point = self.graph_line.pointAt((x - min) / (max - min))
end_point = line_point * self.zoom_transform
return self.projection(end_point, self.graph_line)
@staticmethod
def projection(point, line):
norm = line.normalVector()
norm.translate(point - norm.p1())
p = QPointF()
type = line.intersect(norm, p)
return p
def continuous_labels(self):
min, max, step = self.scale
magnitude = log10(abs(max - min))
def paint(self, painter, option, widget):
pass
def boundingRect(self):
return QRectF()
def ticks(self):
if not self._ticks:
self.update_ticks()
return self._ticks
def bounds(self):
if self._bounds:
return self._bounds
if self.labels:
return -0.2, len(self.labels) - 0.8
elif self.scale:
min, max, _step = self.scale
return min, max
elif self.auto_range:
return self.auto_range
else:
return 0, 1
def set_bounds(self, value):
self._bounds = value
def should_be_expanded(self):
self.update_ticks()
return self.id in YAxes or self.always_horizontal_text or sum(
len(t[1]) for t in self._ticks) * 12 > self.plot.width()
def line(x1, y1, x2, y2):
r = QGraphicsLineItem(x1, y1, x2, y2, None)
self.canvas.addItem(r)
r.setPen(QPen(Qt.white, 2))
r.setZValue(30)
class ViolinPlot(QGraphicsWidget):
LABEL_COLUMN, VIOLIN_COLUMN, LEGEND_COLUMN = range(3)
VIOLIN_COLUMN_WIDTH, OFFSET = 300, 250
MAX_N_ITEMS = 100
MAX_ATTR_LEN = 20
selection_cleared = Signal()
selection_changed = Signal(float, float, str)
def __init__(self):
super().__init__()
self.__violin_column_width = self.VIOLIN_COLUMN_WIDTH # type: int
self.__range = None # type: Optional[Tuple[float, float]]
self.__violin_items = [] # type: List[ViolinItem]
self.__bottom_axis = pg.AxisItem(parent=self, orientation="bottom",
maxTickLength=7, pen=QPen(Qt.black))
self.__bottom_axis.setLabel("Impact on model output")
self.__vertical_line = QGraphicsLineItem(self.__bottom_axis)
self.__vertical_line.setPen(QPen(Qt.gray))
self.__legend = Legend(self)
self.__layout = QGraphicsGridLayout()
self.__layout.addItem(self.__legend, 0, ViolinPlot.LEGEND_COLUMN)
self.__layout.setVerticalSpacing(0)
self.setLayout(self.__layout)
@property
def violin_column_width(self):
return self.__violin_column_width
@violin_column_width.setter
def violin_column_width(self, view_width: int):
self.__violin_column_width = max(self.VIOLIN_COLUMN_WIDTH,
view_width - self.OFFSET)
@property
def bottom_axis(self):
return self.__bottom_axis
def set_data(self, x: np.ndarray, colors: np.ndarray,
names: List[str], n_attrs: float, view_width: int):
self.violin_column_width = view_width
abs_max = np.max(np.abs(x)) * 1.05
self.__range = (-abs_max, abs_max)
self._set_violin_items(x, colors, names)
self._set_labels(names)
self._set_bottom_axis()
self._set_vertical_line()
self.set_n_visible(n_attrs)
def set_n_visible(self, n: int):
for i in range(len(self.__violin_items)):
violin_item = self.__layout.itemAt(i, ViolinPlot.VIOLIN_COLUMN)
violin_item.setVisible(i < n)
text_item = self.__layout.itemAt(i, ViolinPlot.LABEL_COLUMN).item
text_item.setVisible(i < n)
x = self.__vertical_line.line().x1()
n = min(n, len(self.__violin_items))
self.__vertical_line.setLine(x, 0, x, -ViolinItem.HEIGHT * n)
def rescale(self, view_width: int):
self.violin_column_width = view_width
with temp_seed(0):
for item in self.__violin_items:
item.rescale(self.violin_column_width)
self.__bottom_axis.setWidth(self.violin_column_width)
x = self.violin_column_width / 2
self.__vertical_line.setLine(x, 0, x, self.__vertical_line.line().y2())
def show_legend(self, show: bool):
self.__legend.setVisible(show)
self.__bottom_axis.setWidth(self.violin_column_width)
x = self.violin_column_width / 2
self.__vertical_line.setLine(x, 0, x, self.__vertical_line.line().y2())
def _set_violin_items(self, x: np.ndarray, colors: np.ndarray,
labels: List[str]):
with temp_seed(0):
for i in range(x.shape[1]):
item = ViolinItem(self, labels[i], self.__range,
self.violin_column_width)
item.set_data(x[:, i], colors[:, i])
item.selection_changed.connect(self.select)
self.__violin_items.append(item)
self.__layout.addItem(item, i, ViolinPlot.VIOLIN_COLUMN)
if i == self.MAX_N_ITEMS:
break
def _set_labels(self, labels: List[str]):
for i, (label, _) in enumerate(zip(labels, self.__violin_items)):
short = f"{label[:self.MAX_ATTR_LEN - 1]}..." \
if len(label) > self.MAX_ATTR_LEN else label
text = QGraphicsSimpleTextItem(short, self)
text.setToolTip(label)
item = SimpleLayoutItem(text)
item.setSizePolicy(QSizePolicy.Fixed, QSizePolicy.Fixed)
self.__layout.addItem(item, i, ViolinPlot.LABEL_COLUMN,
Qt.AlignRight | Qt.AlignVCenter)
def _set_bottom_axis(self):
self.__bottom_axis.setRange(*self.__range)
self.__layout.addItem(self.__bottom_axis,
len(self.__violin_items),
ViolinPlot.VIOLIN_COLUMN)
def _set_vertical_line(self):
x = self.violin_column_width / 2
n = len(self.__violin_items)
self.__vertical_line.setLine(x, 0, x, -ViolinItem.HEIGHT * n)
def deselect(self):
self.selection_cleared.emit()
def select(self, *args):
self.selection_changed.emit(*args)
def select_from_settings(self, x1: float, x2: float, attr_name: str):
point_r_diff = 2 * self.__range[1] / (self.violin_column_width / 2)
for item in self.__violin_items:
if item.attr_name == attr_name:
item.add_selection_rect(x1 - point_r_diff, x2 + point_r_diff)
break
self.select(x1, x2, attr_name)
class ViolinPlot(QGraphicsWidget):
LABEL_COLUMN, VIOLIN_COLUMN, LEGEND_COLUMN = range(3)
VIOLIN_COLUMN_WIDTH, OFFSET = 300, 80
MAX_N_ITEMS = 100
selection_cleared = Signal()
selection_changed = Signal(float, float, str)
resized = Signal()
def __init__(self):
super().__init__()
self.__violin_column_width = self.VIOLIN_COLUMN_WIDTH # type: int
self.__range = None # type: Optional[Tuple[float, float]]
self.__violin_items = [] # type: List[ViolinItem]
self.__variable_items = [] # type: List[VariableItem]
self.__bottom_axis = AxisItem(parent=self, orientation="bottom",
maxTickLength=7, pen=QPen(Qt.black))
self.__bottom_axis.setLabel("Impact on model output")
self.__vertical_line = QGraphicsLineItem(self.__bottom_axis)
self.__vertical_line.setPen(QPen(Qt.gray))
self.__legend = Legend(self)
self.__layout = QGraphicsGridLayout()
self.__layout.addItem(self.__legend, 0, ViolinPlot.LEGEND_COLUMN)
self.__layout.setVerticalSpacing(0)
self.setLayout(self.__layout)
self.parameter_setter = ParameterSetter(self)
@property
def violin_column_width(self):
return self.__violin_column_width
@violin_column_width.setter
def violin_column_width(self, view_width: int):
j = ViolinPlot.LABEL_COLUMN
w = max([self.__layout.itemAt(i, j).item.boundingRect().width()
for i in range(len(self.__violin_items))] + [0])
width = view_width - self.legend.sizeHint().width() - self.OFFSET - w
self.__violin_column_width = max(self.VIOLIN_COLUMN_WIDTH, width)
@property
def bottom_axis(self):
return self.__bottom_axis
@property
def labels(self):
return self.__variable_items
@property
def legend(self):
return self.__legend
def set_data(self, x: np.ndarray, colors: np.ndarray,
names: List[str], n_attrs: float, view_width: int):
self.violin_column_width = view_width
abs_max = np.max(np.abs(x)) * 1.05
self.__range = (-abs_max, abs_max)
self._set_violin_items(x, colors, names)
self._set_labels(names)
self._set_bottom_axis()
self.set_n_visible(n_attrs)
def set_n_visible(self, n: int):
for i in range(len(self.__violin_items)):
violin_item = self.__layout.itemAt(i, ViolinPlot.VIOLIN_COLUMN)
violin_item.setVisible(i < n)
text_item = self.__layout.itemAt(i, ViolinPlot.LABEL_COLUMN).item
text_item.setVisible(i < n)
self.set_vertical_line()
def rescale(self, view_width: int):
self.violin_column_width = view_width
with temp_seed(0):
for item in self.__violin_items:
item.rescale(self.violin_column_width)
self.__bottom_axis.setWidth(self.violin_column_width)
x = self.violin_column_width / 2
self.__vertical_line.setLine(x, 0, x, self.__vertical_line.line().y2())
def show_legend(self, show: bool):
self.__legend.setVisible(show)
self.__bottom_axis.setWidth(self.violin_column_width)
x = self.violin_column_width / 2
self.__vertical_line.setLine(x, 0, x, self.__vertical_line.line().y2())
def _set_violin_items(self, x: np.ndarray, colors: np.ndarray,
labels: List[str]):
with temp_seed(0):
for i in range(x.shape[1]):
item = ViolinItem(self, labels[i], self.__range,
self.violin_column_width)
item.set_data(x[:, i], colors[:, i])
item.selection_changed.connect(self.select)
self.__violin_items.append(item)
self.__layout.addItem(item, i, ViolinPlot.VIOLIN_COLUMN)
if i == self.MAX_N_ITEMS:
break
def _set_labels(self, labels: List[str]):
for i, (label, _) in enumerate(zip(labels, self.__violin_items)):
text = VariableItem(self, label)
item = SimpleLayoutItem(text)
item.setSizePolicy(QSizePolicy.Fixed, QSizePolicy.Fixed)
self.__layout.addItem(item, i, ViolinPlot.LABEL_COLUMN,
Qt.AlignRight | Qt.AlignVCenter)
self.__variable_items.append(item)
def _set_bottom_axis(self):
self.__bottom_axis.setRange(*self.__range)
self.__layout.addItem(self.__bottom_axis,
len(self.__violin_items),
ViolinPlot.VIOLIN_COLUMN)
def set_vertical_line(self):
x = self.violin_column_width / 2
height = 0
for i in range(len(self.__violin_items)):
violin_item = self.__layout.itemAt(i, ViolinPlot.VIOLIN_COLUMN)
text_item = self.__layout.itemAt(i, ViolinPlot.LABEL_COLUMN).item
if violin_item.isVisible():
height += max(text_item.boundingRect().height(),
violin_item.preferredSize().height())
self.__vertical_line.setLine(x, 0, x, -height)
def deselect(self):
self.selection_cleared.emit()
def select(self, *args):
self.selection_changed.emit(*args)
def select_from_settings(self, x1: float, x2: float, attr_name: str):
point_r_diff = 2 * self.__range[1] / (self.violin_column_width / 2)
for item in self.__violin_items:
if item.attr_name == attr_name:
item.add_selection_rect(x1 - point_r_diff, x2 + point_r_diff)
break
self.select(x1, x2, attr_name)
def apply_visual_settings(self, settings: Dict):
for key, value in settings.items():
self.parameter_setter.set_parameter(key, value)
class AnchorItem(pg.GraphicsObject):
def __init__(self, parent=None, line=QLineF(), text="", **kwargs):
super().__init__(parent, **kwargs)
self._text = text
self.setFlag(pg.GraphicsObject.ItemHasNoContents)
self._spine = QGraphicsLineItem(line, self)
angle = line.angle()
self._arrow = pg.ArrowItem(parent=self, angle=0)
self._arrow.setPos(self._spine.line().p2())
self._arrow.setRotation(angle)
self._arrow.setStyle(headLen=10)
self._label = TextItem(text=text, color=(10, 10, 10))
self._label.setParentItem(self)
self._label.setPos(*self.get_xy())
if parent is not None:
self.setParentItem(parent)
def get_xy(self):
point = self._spine.line().p2()
return point.x(), point.y()
def setText(self, text):
if text != self._text:
self._text = text
self._label.setText(text)
self._label.setVisible(bool(text))
def text(self):
return self._text
def setLine(self, *line):
line = QLineF(*line)
if line != self._spine.line():
self._spine.setLine(line)
self.__updateLayout()
def line(self):
return self._spine.line()
def setPen(self, pen):
self._spine.setPen(pen)
def setArrowVisible(self, visible):
self._arrow.setVisible(visible)
def paint(self, painter, option, widget):
pass
def boundingRect(self):
return QRectF()
def viewTransformChanged(self):
self.__updateLayout()
def __updateLayout(self):
T = self.sceneTransform()
if T is None:
T = QTransform()
# map the axis spine to scene coord. system.
viewbox_line = T.map(self._spine.line())
angle = viewbox_line.angle()
assert not np.isnan(angle)
# note in Qt the y axis is inverted (90 degree angle 'points' down)
left_quad = 270 < angle <= 360 or -0.0 <= angle < 90
# position the text label along the viewbox_line
label_pos = self._spine.line().pointAt(0.90)
if left_quad:
# Anchor the text under the axis spine
anchor = (0.5, -0.1)
else:
# Anchor the text over the axis spine
anchor = (0.5, 1.1)
self._label.setPos(label_pos)
self._label.setAnchor(pg.Point(*anchor))
self._label.setRotation(-angle if left_quad else 180 - angle)
self._arrow.setPos(self._spine.line().p2())
self._arrow.setRotation(180 - angle)
class LinePlotViewBox(ViewBox):
selection_changed = Signal(np.ndarray)
def __init__(self):
super().__init__(enableMenu=False)
self._profile_items = None
self._can_select = True
self._graph_state = SELECT
self.setMouseMode(self.PanMode)
pen = mkPen(LinePlotStyle.SELECTION_LINE_COLOR,
width=LinePlotStyle.SELECTION_LINE_WIDTH)
self.selection_line = QGraphicsLineItem()
self.selection_line.setPen(pen)
self.selection_line.setZValue(1e9)
self.addItem(self.selection_line, ignoreBounds=True)
def update_selection_line(self, button_down_pos, current_pos):
p1 = self.childGroup.mapFromParent(button_down_pos)
p2 = self.childGroup.mapFromParent(current_pos)
self.selection_line.setLine(QLineF(p1, p2))
self.selection_line.resetTransform()
self.selection_line.show()
def set_graph_state(self, state):
self._graph_state = state
def enable_selection(self, enable):
self._can_select = enable
def get_selected(self, p1, p2):
if self._profile_items is None:
return np.array(False)
return line_intersects_profiles(np.array([p1.x(), p1.y()]),
np.array([p2.x(), p2.y()]),
self._profile_items)
def add_profiles(self, y):
if sp.issparse(y):
y = y.todense()
self._profile_items = np.array(
[np.vstack((np.full((1, y.shape[0]), i + 1), y[:, i].flatten())).T
for i in range(y.shape[1])])
def remove_profiles(self):
self._profile_items = None
def mouseDragEvent(self, event, axis=None):
if self._graph_state == SELECT and axis is None and self._can_select:
event.accept()
if event.button() == Qt.LeftButton:
self.update_selection_line(event.buttonDownPos(), event.pos())
if event.isFinish():
self.selection_line.hide()
p1 = self.childGroup.mapFromParent(
event.buttonDownPos(event.button()))
p2 = self.childGroup.mapFromParent(event.pos())
self.selection_changed.emit(self.get_selected(p1, p2))
elif self._graph_state == ZOOMING or self._graph_state == PANNING:
event.ignore()
super().mouseDragEvent(event, axis=axis)
else:
event.ignore()
def mouseClickEvent(self, event):
if event.button() == Qt.RightButton:
self.autoRange()
self.enableAutoRange()
else:
event.accept()
self.selection_changed.emit(np.array(False))
def reset(self):
self._profile_items = None
self._can_select = True
self._graph_state = SELECT
class OWAxis(QGraphicsItem):
Role = OWPalette.Axis
def __init__(self,
id,
title='',
title_above=False,
title_location=AxisMiddle,
line=None,
arrows=0,
plot=None,
bounds=None):
QGraphicsItem.__init__(self)
self.setFlag(QGraphicsItem.ItemHasNoContents)
self.setZValue(AxisZValue)
self.id = id
self.title = title
self.title_location = title_location
self.data_line = line
self.plot = plot
self.graph_line = None
self.size = None
self.scale = None
self.tick_length = (10, 5, 0)
self.arrows = arrows
self.title_above = title_above
self.line_item = QGraphicsLineItem(self)
self.title_item = QGraphicsTextItem(self)
self.end_arrow_item = None
self.start_arrow_item = None
self.show_title = False
self.scale = None
path = QPainterPath()
path.setFillRule(Qt.WindingFill)
path.moveTo(0, 3.09)
path.lineTo(0, -3.09)
path.lineTo(9.51, 0)
path.closeSubpath()
self.arrow_path = path
self.label_items = []
self.label_bg_items = []
self.tick_items = []
self._ticks = []
self.zoom_transform = QTransform()
self.labels = None
self.values = None
self._bounds = bounds
self.auto_range = None
self.auto_scale = True
self.zoomable = False
self.update_callback = None
self.max_text_width = 50
self.text_margin = 5
self.always_horizontal_text = False
@staticmethod
def compute_scale(min, max):
magnitude = int(3 * log10(abs(max - min)) + 1)
if magnitude % 3 == 0:
first_place = 1
elif magnitude % 3 == 1:
first_place = 2
else:
first_place = 5
magnitude = magnitude // 3 - 1
step = first_place * pow(10, magnitude)
first_val = ceil(min / step) * step
return first_val, step
def update_ticks(self):
self._ticks = []
major, medium, minor = self.tick_length
if self.labels is not None and not self.auto_scale:
values = self.values or range(len(self.labels))
for i, text in zip(values, self.labels):
self._ticks.append((i, text, medium, 1))
else:
if self.scale and not self.auto_scale:
min, max, step = self.scale
elif self.auto_range:
min, max = self.auto_range
if min is not None and max is not None:
step = (max - min) / 10
else:
return
else:
return
if max == min:
return
val, step = self.compute_scale(min, max)
while val <= max:
self._ticks.append((val, "%.4g" % val, medium, step))
val += step
def update_graph(self):
if self.update_callback:
self.update_callback()
def update(self, zoom_only=False):
self.update_ticks()
line_color = self.plot.color(OWPalette.Axis)
text_color = self.plot.color(OWPalette.Text)
if not self.graph_line or not self.scene():
return
self.line_item.setLine(self.graph_line)
self.line_item.setPen(line_color)
if self.title:
self.title_item.setHtml('<b>' + self.title + '</b>')
self.title_item.setDefaultTextColor(text_color)
if self.title_location == AxisMiddle:
title_p = 0.5
elif self.title_location == AxisEnd:
title_p = 0.95
else:
title_p = 0.05
title_pos = self.graph_line.pointAt(title_p)
v = self.graph_line.normalVector().unitVector()
dense_text = False
if hasattr(self, 'title_margin'):
offset = self.title_margin
elif self._ticks:
if self.should_be_expanded():
offset = 55
dense_text = True
else:
offset = 35
else:
offset = 10
if self.title_above:
title_pos += (v.p2() - v.p1()) * (
offset + QFontMetrics(self.title_item.font()).height())
else:
title_pos -= (v.p2() - v.p1()) * offset
## TODO: Move it according to self.label_pos
self.title_item.setVisible(self.show_title)
self.title_item.setRotation(-self.graph_line.angle())
c = self.title_item.mapToParent(
self.title_item.boundingRect().center())
tl = self.title_item.mapToParent(
self.title_item.boundingRect().topLeft())
self.title_item.setPos(title_pos - c + tl)
## Arrows
if not zoom_only:
if self.start_arrow_item:
self.scene().removeItem(self.start_arrow_item)
self.start_arrow_item = None
if self.end_arrow_item:
self.scene().removeItem(self.end_arrow_item)
self.end_arrow_item = None
if self.arrows & AxisStart:
if not zoom_only or not self.start_arrow_item:
self.start_arrow_item = QGraphicsPathItem(
self.arrow_path, self)
self.start_arrow_item.setPos(self.graph_line.p1())
self.start_arrow_item.setRotation(-self.graph_line.angle() + 180)
self.start_arrow_item.setBrush(line_color)
self.start_arrow_item.setPen(line_color)
if self.arrows & AxisEnd:
if not zoom_only or not self.end_arrow_item:
self.end_arrow_item = QGraphicsPathItem(self.arrow_path, self)
self.end_arrow_item.setPos(self.graph_line.p2())
self.end_arrow_item.setRotation(-self.graph_line.angle())
self.end_arrow_item.setBrush(line_color)
self.end_arrow_item.setPen(line_color)
## Labels
n = len(self._ticks)
resize_plot_item_list(self.label_items, n, QGraphicsTextItem, self)
resize_plot_item_list(self.label_bg_items, n, QGraphicsRectItem, self)
resize_plot_item_list(self.tick_items, n, QGraphicsLineItem, self)
test_rect = QRectF(self.graph_line.p1(),
self.graph_line.p2()).normalized()
test_rect.adjust(-1, -1, 1, 1)
n_v = self.graph_line.normalVector().unitVector()
if self.title_above:
n_p = n_v.p2() - n_v.p1()
else:
n_p = n_v.p1() - n_v.p2()
l_v = self.graph_line.unitVector()
l_p = l_v.p2() - l_v.p1()
for i in range(n):
pos, text, size, step = self._ticks[i]
hs = 0.5 * step
tick_pos = self.map_to_graph(pos)
if not test_rect.contains(tick_pos):
self.tick_items[i].setVisible(False)
self.label_items[i].setVisible(False)
continue
item = self.label_items[i]
item.setVisible(True)
if not zoom_only:
if self.id in XAxes or getattr(self, 'is_horizontal', False):
item.setHtml('<center>' + Qt.escape(text.strip()) +
'</center>')
else:
item.setHtml(Qt.escape(text.strip()))
item.setTextWidth(-1)
text_angle = 0
if dense_text:
w = min(item.boundingRect().width(), self.max_text_width)
item.setTextWidth(w)
if self.title_above:
label_pos = tick_pos + n_p * (
w + self.text_margin
) + l_p * item.boundingRect().height() / 2
else:
label_pos = tick_pos + n_p * self.text_margin + l_p * item.boundingRect(
).height() / 2
text_angle = -90 if self.title_above else 90
else:
w = min(
item.boundingRect().width(),
QLineF(self.map_to_graph(pos - hs),
self.map_to_graph(pos + hs)).length())
label_pos = tick_pos + n_p * self.text_margin + l_p * item.boundingRect(
).height() / 2
item.setTextWidth(w)
if not self.always_horizontal_text:
if self.title_above:
item.setRotation(-self.graph_line.angle() - text_angle)
else:
item.setRotation(self.graph_line.angle() - text_angle)
item.setPos(label_pos)
item.setDefaultTextColor(text_color)
self.label_bg_items[i].setRect(item.boundingRect())
self.label_bg_items[i].setPen(QPen(Qt.NoPen))
self.label_bg_items[i].setBrush(self.plot.color(OWPalette.Canvas))
item = self.tick_items[i]
item.setVisible(True)
tick_line = QLineF(v)
tick_line.translate(-tick_line.p1())
tick_line.setLength(size)
if self.title_above:
tick_line.setAngle(tick_line.angle() + 180)
item.setLine(tick_line)
item.setPen(line_color)
item.setPos(self.map_to_graph(pos))
@staticmethod
def make_title(label, unit=None):
lab = '<i>' + label + '</i>'
if unit:
lab = lab + ' [' + unit + ']'
return lab
def set_line(self, line):
self.graph_line = line
self.update()
def set_title(self, title):
self.title = title
self.update()
def set_show_title(self, b):
self.show_title = b
self.update()
def set_labels(self, labels, values):
self.labels = labels
self.values = values
self.graph_line = None
self.auto_scale = False
self.update_ticks()
self.update_graph()
def set_scale(self, min, max, step_size):
self.scale = (min, max, step_size)
self.graph_line = None
self.auto_scale = False
self.update_ticks()
self.update_graph()
def set_tick_length(self, minor, medium, major):
self.tick_length = (minor, medium, major)
self.update()
def map_to_graph(self, x):
min, max = self.plot.bounds_for_axis(self.id)
if min == max:
return QPointF()
line_point = self.graph_line.pointAt((x - min) / (max - min))
end_point = line_point * self.zoom_transform
return self.projection(end_point, self.graph_line)
@staticmethod
def projection(point, line):
norm = line.normalVector()
norm.translate(point - norm.p1())
p = QPointF()
type = line.intersect(norm, p)
return p
def continuous_labels(self):
min, max, step = self.scale
magnitude = log10(abs(max - min))
def paint(self, painter, option, widget):
pass
def boundingRect(self):
return QRectF()
def ticks(self):
if not self._ticks:
self.update_ticks()
return self._ticks
def bounds(self):
if self._bounds:
return self._bounds
if self.labels:
return -0.2, len(self.labels) - 0.8
elif self.scale:
min, max, _step = self.scale
return min, max
elif self.auto_range:
return self.auto_range
else:
return 0, 1
def set_bounds(self, value):
self._bounds = value
def should_be_expanded(self):
self.update_ticks()
return self.id in YAxes or self.always_horizontal_text or sum(
len(t[1]) for t in self._ticks) * 12 > self.plot.width()
class AnchorItem(pg.GraphicsWidget):
def __init__(self, parent=None, line=QLineF(), text="", **kwargs):
super().__init__(parent, **kwargs)
self._text = text
self.setFlag(pg.GraphicsObject.ItemHasNoContents)
self._spine = QGraphicsLineItem(line, self)
angle = line.angle()
self._arrow = pg.ArrowItem(parent=self, angle=0)
self._arrow.setPos(self._spine.line().p2())
self._arrow.setRotation(angle)
self._arrow.setStyle(headLen=10)
self._label = TextItem(text=text, color=(10, 10, 10))
self._label.setParentItem(self)
self._label.setPos(*self.get_xy())
self._label.setColor(self.palette().color(QPalette.Text))
if parent is not None:
self.setParentItem(parent)
def get_xy(self):
point = self._spine.line().p2()
return point.x(), point.y()
def setFont(self, font):
self._label.setFont(font)
def setText(self, text):
if text != self._text:
self._text = text
self._label.setText(text)
self._label.setVisible(bool(text))
def text(self):
return self._text
def setLine(self, *line):
line = QLineF(*line)
if line != self._spine.line():
self._spine.setLine(line)
self.__updateLayout()
def line(self):
return self._spine.line()
def setPen(self, pen):
self._spine.setPen(pen)
def setArrowVisible(self, visible):
self._arrow.setVisible(visible)
def paint(self, painter, option, widget):
pass
def boundingRect(self):
return QRectF()
def viewTransformChanged(self):
self.__updateLayout()
def __updateLayout(self):
T = self.sceneTransform()
if T is None:
T = QTransform()
# map the axis spine to scene coord. system.
viewbox_line = T.map(self._spine.line())
angle = viewbox_line.angle()
assert not np.isnan(angle)
# note in Qt the y axis is inverted (90 degree angle 'points' down)
left_quad = 270 < angle <= 360 or -0.0 <= angle < 90
# position the text label along the viewbox_line
label_pos = self._spine.line().pointAt(0.90)
if left_quad:
# Anchor the text under the axis spine
anchor = (0.5, -0.1)
else:
# Anchor the text over the axis spine
anchor = (0.5, 1.1)
self._label.setPos(label_pos)
self._label.setAnchor(pg.Point(*anchor))
self._label.setRotation(-angle if left_quad else 180 - angle)
self._arrow.setPos(self._spine.line().p2())
self._arrow.setRotation(180 - angle)
def changeEvent(self, event):
if event.type() == QEvent.PaletteChange:
self._label.setColor(self.palette().color(QPalette.Text))
super().changeEvent(event)
def _offseted_line(ax, ay):
r = QGraphicsLineItem(x + ax, y + ay, x + (ax or w),
y + (ay or h))
self.canvas.addItem(r)
r.setPen(pen)
class StripeItem(QGraphicsWidget):
WIDTH = 70
def __init__(self, parent):
super().__init__(parent)
self.__range = None # type: Tuple[float]
self.__value_range = None # type: Tuple[float]
self.__model_output = None # type: float
self.__base_value = None # type: float
self.__group = QGraphicsItemGroup(self)
low_color, high_color = QColor(*RGB_LOW), QColor(*RGB_HIGH)
self.__low_item = QGraphicsRectItem()
self.__low_item.setPen(QPen(low_color))
self.__low_item.setBrush(QBrush(low_color))
self.__high_item = QGraphicsRectItem()
self.__high_item.setPen(QPen(high_color))
self.__high_item.setBrush(QBrush(high_color))
self.__low_cover_item = LowCoverItem()
self.__high_cover_item = HighCoverItem()
pen = QPen(IndicatorItem.COLOR)
pen.setStyle(Qt.DashLine)
pen.setWidth(1)
self.__model_output_line = QGraphicsLineItem()
self.__model_output_line.setPen(pen)
self.__base_value_line = QGraphicsLineItem()
self.__base_value_line.setPen(pen)
self.__model_output_ind = IndicatorItem("Model prediction: {}")
self.__base_value_ind = IndicatorItem("Base value: {}\nThe average "
"prediction for selected class.")
self.__group.addToGroup(self.__low_item)
self.__group.addToGroup(self.__high_item)
self.__group.addToGroup(self.__low_cover_item)
self.__group.addToGroup(self.__high_cover_item)
self.__group.addToGroup(self.__model_output_line)
self.__group.addToGroup(self.__base_value_line)
self.__group.addToGroup(self.__model_output_ind)
self.__group.addToGroup(self.__base_value_ind)
self.__low_parts = [] # type: List[LowPartItem]
self.__high_parts = [] # type: List[HighPartItem]
@property
def total_width(self):
widths = [
part.label_item.boundingRect().width()
for part in self.__low_parts + self.__high_parts
] + [0]
return self.WIDTH + StripePlot.SPACING + max(widths)
@property
def model_output_ind(self) -> IndicatorItem:
return self.__model_output_ind
@property
def base_value_ind(self) -> IndicatorItem:
return self.__base_value_ind
def set_data(self, data: PlotData, y_range: Tuple[float, float],
height: float):
self.__range = y_range
self.__value_range = data.value_range
self.__model_output = data.model_output
self.__base_value = data.base_value
r = self.__range[1] - self.__range[0]
self.__model_output_ind.set_text(self.__model_output, r)
self.__base_value_ind.set_text(self.__base_value, r)
for value, label in zip(data.low_values, data.low_labels):
self.__add_part(value, label, value / sum(data.low_values),
self.__low_parts, LowPartItem)
for value, label in zip(data.high_values, data.high_labels):
self.__add_part(value, label, value / sum(data.high_values),
self.__high_parts, HighPartItem)
if self.__low_parts:
self.__low_parts[-1].setVisible(False)
else:
self.__low_item.setVisible(False)
self.__low_cover_item.setVisible(False)
if self.__high_parts:
self.__high_parts[-1].setVisible(False)
else:
self.__high_item.setVisible(False)
self.__high_cover_item.setVisible(False)
self.set_z_values()
self.set_height(height)
def __add_part(self, value: float, label: Tuple[str, str], norm_val: float,
list_: List[PartItem], cls_: Type[PartItem]):
item = cls_(value, label, norm_val)
list_.append(item)
self.__group.addToGroup(item)
self.__group.addToGroup(item.value_item)
self.__group.addToGroup(item.label_item)
def set_z_values(self):
if len(self.__high_parts) < len(self.__low_parts):
self.__high_cover_item.setZValue(-2)
self.__high_item.setZValue(-3)
for i, item in enumerate(self.__high_parts):
item.setZValue(-1)
else:
self.__low_cover_item.setZValue(-2)
self.__low_item.setZValue(-3)
for i, item in enumerate(self.__low_parts):
item.setZValue(-1)
def set_height(self, height: float):
if self.__range[1] == self.__range[0]:
return
height = height / (self.__range[1] - self.__range[0])
y_top = height * (self.__range[1] - self.__value_range[1])
h_top = height * (self.__value_range[1] - self.__model_output)
y_bot = height * (self.__range[1] - self.__model_output)
h_bot = height * (self.__model_output - self.__value_range[0])
self.__low_item.setRect(QRectF(0, y_top, self.WIDTH, h_top))
self.__high_item.setRect(QRectF(0, y_bot, self.WIDTH, h_bot))
self.__low_cover_item.setY(y_top)
self.__high_cover_item.setY(y_bot + h_bot)
self._set_indicators_pos(height)
def adjust_y_text_low(i):
# adjust label y according to TIP_LEN
# adjust for 0.8 * TIP_LEN, because label is not 1 pixel wide
k = 0.4 if i == len(self.__low_parts) - 1 else 0.8
return PartItem.TIP_LEN * k
def adjust_y_text_high(i):
k = 0.4 if i == 0 else 0.8
return -PartItem.TIP_LEN * k
self._set_parts_pos(height, y_top, h_top / height, self.__low_parts,
adjust_y_text_low)
self._set_parts_pos(height, y_bot, h_bot / height, self.__high_parts,
adjust_y_text_high)
def _set_parts_pos(self, height: float, y: float, diff: float,
parts: List[PartItem], adjust_y: Callable):
for i, item in enumerate(parts):
y_delta = height * item.norm_value * diff
y_text = y + y_delta / 2 - item.value_height / 2
visible = y_delta > item.value_height + 8
y_test_adj = y_text + adjust_y(i)
y_mid = height * (self.__range[1] - self.__model_output)
collides = _collides(
y_mid, y_mid, y_test_adj,
y_test_adj + item.value_item.boundingRect().height())
item.value_item.setVisible(visible and not collides)
item.value_item.setY(y_test_adj)
item.label_item.setVisible(visible)
item.label_item.setY(y_text)
y = y + y_delta
item.setY(y)
def _set_indicators_pos(self, height: float):
mo_y = height * (self.__range[1] - self.__model_output)
mo_h = self.__model_output_ind.boundingRect().height()
self.__model_output_ind.setY(mo_y - mo_h / 2)
self.__model_output_line.setLine(
0, mo_y, -StripePlot.SPACING - IndicatorItem.MARGIN, mo_y)
bv_y = height * (self.__range[1] - self.__base_value)
bv_h = self.__base_value_ind.boundingRect().height()
self.__base_value_ind.setY(bv_y - bv_h / 2)
self.__base_value_line.setLine(
-StripePlot.SPACING, bv_y,
-StripePlot.SPACING - IndicatorItem.MARGIN, bv_y)
collides = _collides(mo_y, mo_y + mo_h, bv_y, bv_y + bv_h, d=6)
self.__base_value_ind.setVisible(not collides)
class OWNomogram(OWWidget):
name = "列线图"
description = "朴素贝叶斯可视化的诺莫图和逻辑回归分类器"
icon = "icons/Nomogram.svg"
priority = 2000
keywords = []
class Inputs:
classifier = Input("分类器", Model)
data = Input("数据", Table)
MAX_N_ATTRS = 1000
POINT_SCALE = 0
ALIGN_LEFT = 0
ALIGN_ZERO = 1
ACCEPTABLE = (NaiveBayesModel, LogisticRegressionClassifier)
settingsHandler = ClassValuesContextHandler()
target_class_index = ContextSetting(0)
normalize_probabilities = Setting(False)
scale = Setting(1)
display_index = Setting(1)
n_attributes = Setting(10)
sort_index = Setting(SortBy.ABSOLUTE)
cont_feature_dim_index = Setting(0)
graph_name = "场景"
class Error(OWWidget.Error):
invalid_classifier = Msg("诺莫图只接受朴素贝叶斯和逻辑回归分类器")
def __init__(self):
super().__init__()
self.instances = None
self.domain = None
self.data = None
self.classifier = None
self.align = OWNomogram.ALIGN_ZERO
self.log_odds_ratios = []
self.log_reg_coeffs = []
self.log_reg_coeffs_orig = []
self.log_reg_cont_data_extremes = []
self.p = None
self.b0 = None
self.points = []
self.feature_items = {}
self.feature_marker_values = []
self.scale_marker_values = lambda x: x
self.nomogram_main = None
self.vertical_line = None
self.hidden_vertical_line = None
self.old_target_class_index = self.target_class_index
self.repaint = False
# GUI
box = gui.vBox(self.controlArea, "目标类")
self.class_combo = gui.comboBox(
box, self, "target_class_index", callback=self._class_combo_changed,
contentsLength=12)
self.norm_check = gui.checkBox(
box, self, "normalize_probabilities", "Normalize probabilities",
hidden=True, callback=self.update_scene,
tooltip="For multiclass data 1 vs. all probabilities do not"
" sum to 1 and therefore could be normalized.")
self.scale_radio = gui.radioButtons(
self.controlArea, self, "scale", ["点量表", "记录比值比"],
box="规模", callback=self.update_scene)
box = gui.vBox(self.controlArea, "显示特征")
grid = QGridLayout()
radio_group = gui.radioButtonsInBox(
box, self, "display_index", [], orientation=grid,
callback=self.update_scene)
radio_all = gui.appendRadioButton(
radio_group, "所有", addToLayout=False)
radio_best = gui.appendRadioButton(
radio_group, "最佳排名:", addToLayout=False)
spin_box = gui.hBox(None, margin=0)
self.n_spin = gui.spin(
spin_box, self, "n_attributes", 1, self.MAX_N_ATTRS, label=" ",
controlWidth=60, callback=self._n_spin_changed)
grid.addWidget(radio_all, 1, 1)
grid.addWidget(radio_best, 2, 1)
grid.addWidget(spin_box, 2, 2)
self.sort_combo = gui.comboBox(
box, self, "sort_index", label="排名靠前:", items=SortBy.items(),
orientation=Qt.Horizontal, callback=self.update_scene)
self.cont_feature_dim_combo = gui.comboBox(
box, self, "cont_feature_dim_index", label="数字特征: ",
items=["1维投影", "二维曲线"], orientation=Qt.Horizontal,
callback=self.update_scene)
gui.rubber(self.controlArea)
class _GraphicsView(QGraphicsView):
def __init__(self, scene, parent, **kwargs):
for k, v in dict(verticalScrollBarPolicy=Qt.ScrollBarAlwaysOff,
horizontalScrollBarPolicy=Qt.ScrollBarAlwaysOff,
viewportUpdateMode=QGraphicsView.BoundingRectViewportUpdate,
renderHints=(QPainter.Antialiasing |
QPainter.TextAntialiasing |
QPainter.SmoothPixmapTransform),
alignment=(Qt.AlignTop |
Qt.AlignLeft),
sizePolicy=QSizePolicy(QSizePolicy.MinimumExpanding,
QSizePolicy.MinimumExpanding)).items():
kwargs.setdefault(k, v)
super().__init__(scene, parent, **kwargs)
class GraphicsView(_GraphicsView):
def __init__(self, scene, parent):
super().__init__(scene, parent,
verticalScrollBarPolicy=Qt.ScrollBarAlwaysOn,
styleSheet='QGraphicsView {background: white}')
self.viewport().setMinimumWidth(300) # XXX: This prevents some tests failing
self._is_resizing = False
w = self
def resizeEvent(self, resizeEvent):
# Recompute main scene on window width change
if resizeEvent.size().width() != resizeEvent.oldSize().width():
self._is_resizing = True
self.w.update_scene()
self._is_resizing = False
return super().resizeEvent(resizeEvent)
def is_resizing(self):
return self._is_resizing
def sizeHint(self):
return QSize(400, 200)
class FixedSizeGraphicsView(_GraphicsView):
def __init__(self, scene, parent):
super().__init__(scene, parent,
sizePolicy=QSizePolicy(QSizePolicy.MinimumExpanding,
QSizePolicy.Minimum))
def sizeHint(self):
return QSize(400, 85)
scene = self.scene = QGraphicsScene(self)
top_view = self.top_view = FixedSizeGraphicsView(scene, self)
mid_view = self.view = GraphicsView(scene, self)
bottom_view = self.bottom_view = FixedSizeGraphicsView(scene, self)
for view in (top_view, mid_view, bottom_view):
self.mainArea.layout().addWidget(view)
def _class_combo_changed(self):
with np.errstate(invalid='ignore'):
coeffs = [np.nan_to_num(p[self.target_class_index] /
p[self.old_target_class_index])
for p in self.points]
points = [p[self.old_target_class_index] for p in self.points]
self.feature_marker_values = [
self.get_points_from_coeffs(v, c, p) for (v, c, p) in
zip(self.feature_marker_values, coeffs, points)]
self.feature_marker_values = np.asarray(self.feature_marker_values)
self.update_scene()
self.old_target_class_index = self.target_class_index
def _n_spin_changed(self):
self.display_index = 1
self.update_scene()
def update_controls(self):
self.class_combo.clear()
self.norm_check.setHidden(True)
self.cont_feature_dim_combo.setEnabled(True)
if self.domain is not None:
self.class_combo.addItems(self.domain.class_vars[0].values)
if len(self.domain.attributes) > self.MAX_N_ATTRS:
self.display_index = 1
if len(self.domain.class_vars[0].values) > 2:
self.norm_check.setHidden(False)
if not self.domain.has_continuous_attributes():
self.cont_feature_dim_combo.setEnabled(False)
self.cont_feature_dim_index = 0
model = self.sort_combo.model()
item = model.item(SortBy.POSITIVE)
item.setFlags(item.flags() | Qt.ItemIsEnabled)
item = model.item(SortBy.NEGATIVE)
item.setFlags(item.flags() | Qt.ItemIsEnabled)
self.align = OWNomogram.ALIGN_ZERO
if self.classifier and isinstance(self.classifier,
LogisticRegressionClassifier):
self.align = OWNomogram.ALIGN_LEFT
@Inputs.data
def set_data(self, data):
self.instances = data
self.feature_marker_values = []
self.set_feature_marker_values()
self.update_scene()
@Inputs.classifier
def set_classifier(self, classifier):
self.closeContext()
self.classifier = classifier
self.Error.clear()
if self.classifier and not isinstance(self.classifier, self.ACCEPTABLE):
self.Error.invalid_classifier()
self.classifier = None
self.domain = self.classifier.domain if self.classifier else None
self.data = None
self.calculate_log_odds_ratios()
self.calculate_log_reg_coefficients()
self.update_controls()
self.target_class_index = 0
self.openContext(self.domain.class_var if self.domain is not None
else None)
self.points = self.log_odds_ratios or self.log_reg_coeffs
self.feature_marker_values = []
self.old_target_class_index = self.target_class_index
self.update_scene()
def calculate_log_odds_ratios(self):
self.log_odds_ratios = []
self.p = None
if self.classifier is None or self.domain is None:
return
if not isinstance(self.classifier, NaiveBayesModel):
return
log_cont_prob = self.classifier.log_cont_prob
class_prob = self.classifier.class_prob
for i in range(len(self.domain.attributes)):
ca = np.exp(log_cont_prob[i]) * class_prob[:, None]
_or = (ca / (1 - ca)) / (class_prob / (1 - class_prob))[:, None]
self.log_odds_ratios.append(np.log(_or))
self.p = class_prob
def calculate_log_reg_coefficients(self):
self.log_reg_coeffs = []
self.log_reg_cont_data_extremes = []
self.b0 = None
if self.classifier is None or self.domain is None:
return
if not isinstance(self.classifier, LogisticRegressionClassifier):
return
self.domain = self.reconstruct_domain(self.classifier.original_domain,
self.domain)
self.data = self.classifier.original_data.transform(self.domain)
attrs, ranges, start = self.domain.attributes, [], 0
for attr in attrs:
stop = start + len(attr.values) if attr.is_discrete else start + 1
ranges.append(slice(start, stop))
start = stop
self.b0 = self.classifier.intercept
coeffs = self.classifier.coefficients
if len(self.domain.class_var.values) == 2:
self.b0 = np.hstack((self.b0 * (-1), self.b0))
coeffs = np.vstack((coeffs * (-1), coeffs))
self.log_reg_coeffs = [coeffs[:, ranges[i]] for i in range(len(attrs))]
self.log_reg_coeffs_orig = self.log_reg_coeffs.copy()
min_values = nanmin(self.data.X, axis=0)
max_values = nanmax(self.data.X, axis=0)
for i, min_t, max_t in zip(range(len(self.log_reg_coeffs)),
min_values, max_values):
if self.log_reg_coeffs[i].shape[1] == 1:
coef = self.log_reg_coeffs[i]
self.log_reg_coeffs[i] = np.hstack((coef * min_t, coef * max_t))
self.log_reg_cont_data_extremes.append(
[sorted([min_t, max_t], reverse=(c < 0)) for c in coef])
else:
self.log_reg_cont_data_extremes.append([None])
def update_scene(self):
self.clear_scene()
if self.domain is None or not len(self.points[0]):
return
n_attrs = self.n_attributes if self.display_index else int(1e10)
attr_inds, attributes = zip(*self.get_ordered_attributes()[:n_attrs])
name_items = [QGraphicsTextItem(attr.name) for attr in attributes]
point_text = QGraphicsTextItem("Points")
probs_text = QGraphicsTextItem("Probabilities (%)")
all_items = name_items + [point_text, probs_text]
name_offset = -max(t.boundingRect().width() for t in all_items) - 10
w = self.view.viewport().rect().width()
max_width = w + name_offset - 30
points = [self.points[i][self.target_class_index]
for i in attr_inds]
if self.align == OWNomogram.ALIGN_LEFT:
points = [p - p.min() for p in points]
max_ = np.nan_to_num(max(max(abs(p)) for p in points))
d = 100 / max_ if max_ else 1
minimums = [p[self.target_class_index].min() for p in self.points]
if self.scale == OWNomogram.POINT_SCALE:
points = [p * d for p in points]
if self.align == OWNomogram.ALIGN_LEFT:
self.scale_marker_values = lambda x: (x - minimums) * d
else:
self.scale_marker_values = lambda x: x * d
else:
if self.align == OWNomogram.ALIGN_LEFT:
self.scale_marker_values = lambda x: x - minimums
else:
self.scale_marker_values = lambda x: x
point_item, nomogram_head = self.create_main_nomogram(
attributes, attr_inds,
name_items, points, max_width, point_text, name_offset)
probs_item, nomogram_foot = self.create_footer_nomogram(
probs_text, d, minimums, max_width, name_offset)
for item in self.feature_items.values():
item.dot.point_dot = point_item.dot
item.dot.probs_dot = probs_item.dot
item.dot.vertical_line = self.hidden_vertical_line
self.nomogram = nomogram = NomogramItem()
nomogram.add_items([nomogram_head, self.nomogram_main, nomogram_foot])
self.scene.addItem(nomogram)
self.set_feature_marker_values()
rect = QRectF(self.scene.itemsBoundingRect().x(),
self.scene.itemsBoundingRect().y(),
self.scene.itemsBoundingRect().width(),
self.nomogram.preferredSize().height()).adjusted(10, 0, 20, 0)
self.scene.setSceneRect(rect)
# Clip top and bottom (60 and 150) parts from the main view
self.view.setSceneRect(rect.x(), rect.y() + 80, rect.width() - 10, rect.height() - 160)
self.view.viewport().setMaximumHeight(rect.height() - 160)
# Clip main part from top/bottom views
# below point values are imprecise (less/more than required) but this
# is not a problem due to clipped scene content still being drawn
self.top_view.setSceneRect(rect.x(), rect.y() + 3, rect.width() - 10, 20)
self.bottom_view.setSceneRect(rect.x(), rect.height() - 110, rect.width() - 10, 30)
def create_main_nomogram(self, attributes, attr_inds, name_items, points,
max_width, point_text, name_offset):
cls_index = self.target_class_index
min_p = min(p.min() for p in points)
max_p = max(p.max() for p in points)
values = self.get_ruler_values(min_p, max_p, max_width)
min_p, max_p = min(values), max(values)
diff_ = np.nan_to_num(max_p - min_p)
scale_x = max_width / diff_ if diff_ else max_width
nomogram_header = NomogramItem()
point_item = RulerItem(point_text, values, scale_x, name_offset,
- scale_x * min_p)
point_item.setPreferredSize(point_item.preferredWidth(), 35)
nomogram_header.add_items([point_item])
self.nomogram_main = NomogramItem()
cont_feature_item_class = ContinuousFeature2DItem if \
self.cont_feature_dim_index else ContinuousFeatureItem
feature_items = [
DiscreteFeatureItem(
name_item, attr.values, point,
scale_x, name_offset, - scale_x * min_p)
if attr.is_discrete else
cont_feature_item_class(
name_item, self.log_reg_cont_data_extremes[i][cls_index],
self.get_ruler_values(
point.min(), point.max(),
scale_x * point.ptp(), False),
scale_x, name_offset, - scale_x * min_p)
for i, attr, name_item, point in zip(attr_inds, attributes, name_items, points)]
self.nomogram_main.add_items(feature_items)
self.feature_items = OrderedDict(sorted(zip(attr_inds, feature_items)))
x = - scale_x * min_p
y = self.nomogram_main.layout().preferredHeight() + 10
self.vertical_line = QGraphicsLineItem(x, -6, x, y)
self.vertical_line.setPen(QPen(Qt.DotLine))
self.vertical_line.setParentItem(point_item)
self.hidden_vertical_line = QGraphicsLineItem(x, -6, x, y)
pen = QPen(Qt.DashLine)
pen.setBrush(QColor(Qt.red))
self.hidden_vertical_line.setPen(pen)
self.hidden_vertical_line.setParentItem(point_item)
return point_item, nomogram_header
def get_ordered_attributes(self):
"""Return (in_domain_index, attr) pairs, ordered by method in SortBy combo"""
if self.domain is None or not self.domain.attributes:
return []
attrs = self.domain.attributes
sort_by = self.sort_index
class_value = self.target_class_index
if sort_by == SortBy.NO_SORTING:
return list(enumerate(attrs))
elif sort_by == SortBy.NAME:
def key(x):
_, attr = x
return attr.name.lower()
elif sort_by == SortBy.ABSOLUTE:
def key(x):
i, attr = x
if attr.is_discrete:
ptp = self.points[i][class_value].ptp()
else:
coef = np.abs(self.log_reg_coeffs_orig[i][class_value]).mean()
ptp = coef * np.ptp(self.log_reg_cont_data_extremes[i][class_value])
return -ptp
elif sort_by == SortBy.POSITIVE:
def key(x):
i, attr = x
max_value = (self.points[i][class_value].max()
if attr.is_discrete else
np.mean(self.log_reg_cont_data_extremes[i][class_value]))
return -max_value
elif sort_by == SortBy.NEGATIVE:
def key(x):
i, attr = x
min_value = (self.points[i][class_value].min()
if attr.is_discrete else
np.mean(self.log_reg_cont_data_extremes[i][class_value]))
return min_value
return sorted(enumerate(attrs), key=key)
def create_footer_nomogram(self, probs_text, d, minimums,
max_width, name_offset):
# pylint: disable=invalid-unary-operand-type
eps, d_ = 0.05, 1
k = - np.log(self.p / (1 - self.p)) if self.p is not None else - self.b0
min_sum = k[self.target_class_index] - np.log((1 - eps) / eps)
max_sum = k[self.target_class_index] - np.log(eps / (1 - eps))
if self.align == OWNomogram.ALIGN_LEFT:
max_sum = max_sum - sum(minimums)
min_sum = min_sum - sum(minimums)
for i in range(len(k)): # pylint: disable=consider-using-enumerate
k[i] = k[i] - sum([min(q) for q in [p[i] for p in self.points]])
if self.scale == OWNomogram.POINT_SCALE:
min_sum *= d
max_sum *= d
d_ = d
values = self.get_ruler_values(min_sum, max_sum, max_width)
min_sum, max_sum = min(values), max(values)
diff_ = np.nan_to_num(max_sum - min_sum)
scale_x = max_width / diff_ if diff_ else max_width
cls_var, cls_index = self.domain.class_var, self.target_class_index
nomogram_footer = NomogramItem()
def get_normalized_probabilities(val):
if not self.normalize_probabilities:
return 1 / (1 + np.exp(k[cls_index] - val / d_))
totals = self.__get_totals_for_class_values(minimums)
p_sum = np.sum(1 / (1 + np.exp(k - totals / d_)))
return 1 / (1 + np.exp(k[cls_index] - val / d_)) / p_sum
def get_points(prob):
if not self.normalize_probabilities:
return (k[cls_index] - np.log(1 / prob - 1)) * d_
totals = self.__get_totals_for_class_values(minimums)
p_sum = np.sum(1 / (1 + np.exp(k - totals / d_)))
return (k[cls_index] - np.log(1 / (prob * p_sum) - 1)) * d_
probs_item = ProbabilitiesRulerItem(
probs_text, values, scale_x, name_offset, - scale_x * min_sum,
get_points=get_points,
title="{}='{}'".format(cls_var.name, cls_var.values[cls_index]),
get_probabilities=get_normalized_probabilities)
nomogram_footer.add_items([probs_item])
return probs_item, nomogram_footer
def __get_totals_for_class_values(self, minimums):
cls_index = self.target_class_index
marker_values = self.scale_marker_values(self.feature_marker_values)
totals = np.full(len(self.domain.class_var.values), np.nan)
totals[cls_index] = marker_values.sum()
for i in range(len(self.domain.class_var.values)):
if i == cls_index:
continue
coeffs = [np.nan_to_num(p[i] / p[cls_index]) for p in self.points]
points = [p[cls_index] for p in self.points]
total = sum([self.get_points_from_coeffs(v, c, p) for (v, c, p)
in zip(self.feature_marker_values, coeffs, points)])
if self.align == OWNomogram.ALIGN_LEFT:
points = [p - m for m, p in zip(minimums, points)]
total -= sum([min(p) for p in [p[i] for p in self.points]])
d = 100 / max(max(abs(p)) for p in points)
if self.scale == OWNomogram.POINT_SCALE:
total *= d
totals[i] = total
assert not np.any(np.isnan(totals))
return totals
def set_feature_marker_values(self):
if not (len(self.points) and len(self.feature_items)):
return
if not len(self.feature_marker_values):
self._init_feature_marker_values()
marker_values = self.scale_marker_values(self.feature_marker_values)
invisible_sum = 0
for i, marker in enumerate(marker_values):
try:
item = self.feature_items[i]
except KeyError:
invisible_sum += marker
else:
item.dot.move_to_val(marker)
item.dot.probs_dot.move_to_sum(invisible_sum)
def _init_feature_marker_values(self):
self.feature_marker_values = []
cls_index = self.target_class_index
instances = Table(self.domain, self.instances) \
if self.instances else None
values = []
for i, attr in enumerate(self.domain.attributes):
value, feature_val = 0, None
if len(self.log_reg_coeffs):
if attr.is_discrete:
ind, n = unique(self.data.X[:, i], return_counts=True)
feature_val = np.nan_to_num(ind[np.argmax(n)])
else:
feature_val = nanmean(self.data.X[:, i])
# If data is provided on a separate signal, use the first data
# instance to position the points instead of the mean
inst_in_dom = instances and attr in instances.domain
if inst_in_dom and not np.isnan(instances[0][attr]):
feature_val = instances[0][attr]
if feature_val is not None:
value = (self.points[i][cls_index][int(feature_val)]
if attr.is_discrete else
self.log_reg_coeffs_orig[i][cls_index][0] * feature_val)
values.append(value)
self.feature_marker_values = np.asarray(values)
def clear_scene(self):
self.feature_items = {}
self.scale_marker_values = lambda x: x
self.nomogram = None
self.nomogram_main = None
self.vertical_line = None
self.hidden_vertical_line = None
self.scene.clear()
def send_report(self):
self.report_plot()
@staticmethod
def reconstruct_domain(original, preprocessed):
# abuse dict to make "in" comparisons faster
attrs = OrderedDict()
for attr in preprocessed.attributes:
cv = attr._compute_value.variable._compute_value
var = cv.variable if cv else original[attr.name]
if var in attrs: # the reason for OrderedDict
continue
attrs[var] = None # we only need keys
attrs = list(attrs.keys())
return Domain(attrs, original.class_var, original.metas)
@staticmethod
def get_ruler_values(start, stop, max_width, round_to_nearest=True):
if max_width == 0:
return [0]
diff = np.nan_to_num((stop - start) / max_width)
if diff <= 0:
return [0]
decimals = int(np.floor(np.log10(diff)))
if diff > 4 * pow(10, decimals):
step = 5 * pow(10, decimals + 2)
elif diff > 2 * pow(10, decimals):
step = 2 * pow(10, decimals + 2)
elif diff > 1 * pow(10, decimals):
step = 1 * pow(10, decimals + 2)
else:
step = 5 * pow(10, decimals + 1)
round_by = int(- np.floor(np.log10(step)))
r = start % step
if not round_to_nearest:
_range = np.arange(start + step, stop + r, step) - r
start, stop = np.floor(start * 100) / 100, np.ceil(stop * 100) / 100
return np.round(np.hstack((start, _range, stop)), 2)
return np.round(np.arange(start, stop + r + step, step) - r, round_by)
@staticmethod
def get_points_from_coeffs(current_value, coefficients, possible_values):
if np.isnan(possible_values).any():
return 0
# pylint: disable=undefined-loop-variable
indices = np.argsort(possible_values)
sorted_values = possible_values[indices]
sorted_coefficients = coefficients[indices]
for i, val in enumerate(sorted_values):
if current_value < val:
break
diff = sorted_values[i] - sorted_values[i - 1]
k = 0 if diff < 1e-6 else (sorted_values[i] - current_value) / \
(sorted_values[i] - sorted_values[i - 1])
return sorted_coefficients[i - 1] * sorted_values[i - 1] * k + \
sorted_coefficients[i] * sorted_values[i] * (1 - k)
