sagemaker/package/package/visuals.py [8:333]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - def log_odds_to_proba(log_odds): return np.exp(log_odds) / (1 + np.exp(log_odds)) def proba_to_log_odds(proba): return np.log(proba / (1 - proba)) def summary_explanation(output, level=0, separator="__"): data = { 'shap_values': output['explanation']['shap_values'] } if 'descriptions' in output: data['feature_descriptions'] = output['descriptions'] df = pd.DataFrame(data) df.index.name = 'feature_names' df = df.reset_index() df["feature_names"] = df["feature_names"].apply( lambda e: separator.join(e.split(separator)[: level + 1]) ) df = df.groupby("feature_names")["shap_values"].sum() df = df.reset_index() if 'descriptions' in output: df["feature_descriptions"] = df["feature_names"].apply( lambda e: "All '{}' features combined.".format(e) ) explanation = { 'shap_values': df['shap_values'].tolist(), 'expected_value': output['explanation']['expected_value'], 'feature_names': df['feature_names'].tolist() } if 'descriptions' in output: explanation['feature_descriptions'] = df['feature_descriptions'].tolist() return explanation def detailed_explanation(output): data = { 'feature_values': output['features'], 'shap_values': output['explanation']['shap_values'] } if 'descriptions' in output: data['feature_descriptions'] = output['descriptions'] df = pd.DataFrame(data) df.index.name = 'feature_names' df = df.reset_index() explanation = { 'shap_values': df['shap_values'].tolist(), 'expected_value': output['explanation']['expected_value'], 'feature_names': df['feature_names'].tolist(), 'feature_values': df['feature_values'].tolist() } if 'descriptions' in output: explanation['feature_descriptions'] = df['feature_descriptions'].tolist() return explanation class WaterfallChart(): def __init__( self, baseline, shap_values, names, feature_values=None, descriptions=None, x_range_width=500, x_range_padding=0.25, x_axis_label='Score (%)', title=None, baseline_color="#808080", positive_color="#FF5733", negative_color="#69AE35", sort_order='absolute_descending', max_features=None ): self._baseline = baseline self._shap_values = np.array(shap_values) self._set_names(names, feature_values) self._descriptions = descriptions self._baseline_color = baseline_color self._positive_color = positive_color self._negative_color = negative_color self._sort_features(sort_order) self._filter_features(max_features) self._set_segments() # start, end and color of each segment self._set_data_source() self._set_tooltips() self._set_x_range(x_range_padding) self._x_range_width = x_range_width self._set_figure() self._add_baseline_span() self._add_baseline_label() self._add_prediction_span() self._add_prediction_label() self._add_segments() self.add_x_axis_label(x_axis_label) self.add_title(title) def _set_names(self, names, feature_values): if feature_values: self._names = ['{} = {}'.format(n, v) for n, v in zip(names, feature_values)] else: self._names = names def _sort_features(self, sort_order): assert sort_order in set([ 'absolute_ascending', 'absolute_descending', 'ascending', 'descending' ]) # find sorted index according to sort_order if sort_order.startswith('absolute'): values = np.abs(self._shap_values) else: values = self._shap_values if sort_order.endswith('ascending'): sorted_idxs = values.argsort().tolist() else: sorted_idxs = (-1 * values).argsort().tolist() # update features with sorted index self._shap_values = self._shap_values[sorted_idxs] self._names = [self._names[idx] for idx in sorted_idxs] if self._descriptions: self._descriptions = [self._descriptions[idx] for idx in sorted_idxs] def _filter_features(self, max_features): if max_features and max_features < len(self._shap_values): shap_values_keep = self._shap_values[:max_features] shap_values_drop = self._shap_values[max_features:] shap_values_other = shap_values_drop.sum(keepdims=True) self._shap_values = np.concatenate((shap_values_keep, shap_values_other)) self._names = self._names[:max_features] + ['other_features'] if self._descriptions: self._descriptions = self._descriptions[:max_features] + ['All other features combined.'] def _set_segments(self): self._segment_ends = self._shap_values.cumsum() + self._baseline self._segment_starts = self._segment_ends - self._shap_values segment_ends_max = self._segment_ends.max() segment_ends_min = self._segment_ends.min() segment_starts_max = self._segment_starts.max() segment_starts_min = self._segment_starts.min() self._segments_maximum = max(segment_starts_max, segment_ends_max) self._segments_minimum = min(segment_starts_min, segment_ends_min) self._segment_colors = [] for shap_value in self._shap_values: if shap_value > 0: self._segment_colors.append(self._positive_color) else: self._segment_colors.append(self._negative_color) def _set_data_source(self): data = { 'shap_value': self._shap_values, 'name': self._names, 'segment_color': self._segment_colors, 'segment_start_proba': log_odds_to_proba(self._segment_starts), 'segment_end_proba': log_odds_to_proba(self._segment_ends) } if self._descriptions: data['descriptions'] = self._descriptions self._data_source = bokeh.models.sources.ColumnDataSource(data=data) def _set_tooltips(self): self._tooltips = [ ("name", "@name"), ("description", "@description") ] def _set_x_range(self, x_range_padding): center = (self._segments_minimum + self._segments_maximum) / 2 minimum = center - ((center - self._segments_minimum) * (1 + x_range_padding)) maximum = ((self._segments_maximum - center) * (1 + x_range_padding)) + center self._x_range_minimum = log_odds_to_proba(minimum) self._x_range_maximum = log_odds_to_proba(maximum) def _set_figure(self): figure = bokeh.plotting.figure( frame_width=self._x_range_width, frame_height=25 * len(self._names), x_axis_type="log", x_axis_location="below", y_range=self._names, y_axis_location="right", tooltips=self._tooltips, toolbar_location=None, tools="" ) formatter = bokeh.models.formatters.NumeralTickFormatter(format="0.0 %") axis_label_text_font_size = "10pt" axis_line_color = "white" desired_num_ticks = 7 # set original xaxis figure.x_range = bokeh.models.Range1d( self._x_range_minimum, self._x_range_maximum ) figure.xaxis.formatter = formatter figure.xaxis.axis_label_text_font_size = axis_label_text_font_size figure.xaxis.axis_line_color = axis_line_color figure.xaxis.ticker.desired_num_ticks = desired_num_ticks # create a copy of xaxis above figure.extra_x_ranges = {"probability": figure.x_range} top_axis = bokeh.models.axes.LogAxis(x_range_name="probability") top_axis.formatter = formatter top_axis.axis_label_text_font_size = axis_label_text_font_size top_axis.axis_line_color = axis_line_color top_axis.ticker.desired_num_ticks = desired_num_ticks figure.add_layout(top_axis, 'above') # set yaxis figure.yaxis.major_label_text_font_size = axis_label_text_font_size figure.yaxis.axis_line_color = axis_line_color self._figure = figure def _add_baseline_span(self): span = bokeh.models.Span( location=log_odds_to_proba(self._baseline), dimension='height', line_color='grey', line_dash='dotted', line_width=1 ) span.level = 'underlay' self._figure.add_layout(span) def _get_x_range_offset(self, log_odd): minimum = proba_to_log_odds(self._x_range_minimum) maximum = proba_to_log_odds(self._x_range_maximum) x_range_length = maximum - minimum percentage = (log_odd - minimum) / x_range_length x_range_offset = percentage * self._x_range_width return x_range_offset def _add_baseline_label(self): baseline_proba = log_odds_to_proba(self._baseline) text = '{:.2%} baseline'.format(baseline_proba) offset = self._get_x_range_offset(self._baseline) label = bokeh.models.Title( text=text, align='left', text_font_size='10pt', offset=offset, text_font_style='italic', text_color='grey') self._figure.add_layout(label, 'below') def _get_prediction(self): return self._segment_ends[-1] def _get_prediction_color(self): prediction = self._get_prediction() if prediction > self._baseline: return self._positive_color else: return self._negative_color def _add_prediction_span(self): prediction = self._get_prediction() prediction_proba = log_odds_to_proba(prediction) prediction_color = self._get_prediction_color() span = bokeh.models.Span( location=prediction_proba, dimension='height', line_color=prediction_color, line_dash='dotted', line_width=1 ) span.level = 'underlay' self._figure.add_layout(span) def _add_prediction_label(self): prediction = self._get_prediction() prediction_proba = log_odds_to_proba(prediction) text = '{:.2%} predicted'.format(prediction_proba) offset = self._get_x_range_offset(prediction) prediction_color = self._get_prediction_color() label = bokeh.models.Title( text=text, align='left', text_font_size='10pt', offset=offset, text_font_style='italic', text_color=prediction_color ) self._figure.add_layout(label, 'above') def _add_segments(self): self._figure.segment( "segment_start_proba", "name", "segment_end_proba", "name", line_width=18, line_color="segment_color", source=self._data_source ) self._figure.diamond( y="name", x="segment_end_proba", size=16, fill_color="segment_color", line_color="segment_color", source=self._data_source ) self._figure.diamond( y="name", x="segment_start_proba", size=16, fill_color='white', line_color="white", source=self._data_source ) def add_title(self, text): if text: title = bokeh.models.Title(text=text, align='left', text_font_size='12pt') self._figure.add_layout(title, 'above') def add_x_axis_label(self, text): label = bokeh.models.Title(text=text, align='center', text_font_size='10pt') self._figure.add_layout(label, 'above') def show(self): return bokeh.plotting.show(self._figure) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - sagemaker_studio/package/package/visuals.py [8:333]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - def log_odds_to_proba(log_odds): return np.exp(log_odds) / (1 + np.exp(log_odds)) def proba_to_log_odds(proba): return np.log(proba / (1 - proba)) def summary_explanation(output, level=0, separator="__"): data = { 'shap_values': output['explanation']['shap_values'] } if 'descriptions' in output: data['feature_descriptions'] = output['descriptions'] df = pd.DataFrame(data) df.index.name = 'feature_names' df = df.reset_index() df["feature_names"] = df["feature_names"].apply( lambda e: separator.join(e.split(separator)[: level + 1]) ) df = df.groupby("feature_names")["shap_values"].sum() df = df.reset_index() if 'descriptions' in output: df["feature_descriptions"] = df["feature_names"].apply( lambda e: "All '{}' features combined.".format(e) ) explanation = { 'shap_values': df['shap_values'].tolist(), 'expected_value': output['explanation']['expected_value'], 'feature_names': df['feature_names'].tolist() } if 'descriptions' in output: explanation['feature_descriptions'] = df['feature_descriptions'].tolist() return explanation def detailed_explanation(output): data = { 'feature_values': output['features'], 'shap_values': output['explanation']['shap_values'] } if 'descriptions' in output: data['feature_descriptions'] = output['descriptions'] df = pd.DataFrame(data) df.index.name = 'feature_names' df = df.reset_index() explanation = { 'shap_values': df['shap_values'].tolist(), 'expected_value': output['explanation']['expected_value'], 'feature_names': df['feature_names'].tolist(), 'feature_values': df['feature_values'].tolist() } if 'descriptions' in output: explanation['feature_descriptions'] = df['feature_descriptions'].tolist() return explanation class WaterfallChart(): def __init__( self, baseline, shap_values, names, feature_values=None, descriptions=None, x_range_width=500, x_range_padding=0.25, x_axis_label='Score (%)', title=None, baseline_color="#808080", positive_color="#FF5733", negative_color="#69AE35", sort_order='absolute_descending', max_features=None ): self._baseline = baseline self._shap_values = np.array(shap_values) self._set_names(names, feature_values) self._descriptions = descriptions self._baseline_color = baseline_color self._positive_color = positive_color self._negative_color = negative_color self._sort_features(sort_order) self._filter_features(max_features) self._set_segments() # start, end and color of each segment self._set_data_source() self._set_tooltips() self._set_x_range(x_range_padding) self._x_range_width = x_range_width self._set_figure() self._add_baseline_span() self._add_baseline_label() self._add_prediction_span() self._add_prediction_label() self._add_segments() self.add_x_axis_label(x_axis_label) self.add_title(title) def _set_names(self, names, feature_values): if feature_values: self._names = ['{} = {}'.format(n, v) for n, v in zip(names, feature_values)] else: self._names = names def _sort_features(self, sort_order): assert sort_order in set([ 'absolute_ascending', 'absolute_descending', 'ascending', 'descending' ]) # find sorted index according to sort_order if sort_order.startswith('absolute'): values = np.abs(self._shap_values) else: values = self._shap_values if sort_order.endswith('ascending'): sorted_idxs = values.argsort().tolist() else: sorted_idxs = (-1 * values).argsort().tolist() # update features with sorted index self._shap_values = self._shap_values[sorted_idxs] self._names = [self._names[idx] for idx in sorted_idxs] if self._descriptions: self._descriptions = [self._descriptions[idx] for idx in sorted_idxs] def _filter_features(self, max_features): if max_features and max_features < len(self._shap_values): shap_values_keep = self._shap_values[:max_features] shap_values_drop = self._shap_values[max_features:] shap_values_other = shap_values_drop.sum(keepdims=True) self._shap_values = np.concatenate((shap_values_keep, shap_values_other)) self._names = self._names[:max_features] + ['other_features'] if self._descriptions: self._descriptions = self._descriptions[:max_features] + ['All other features combined.'] def _set_segments(self): self._segment_ends = self._shap_values.cumsum() + self._baseline self._segment_starts = self._segment_ends - self._shap_values segment_ends_max = self._segment_ends.max() segment_ends_min = self._segment_ends.min() segment_starts_max = self._segment_starts.max() segment_starts_min = self._segment_starts.min() self._segments_maximum = max(segment_starts_max, segment_ends_max) self._segments_minimum = min(segment_starts_min, segment_ends_min) self._segment_colors = [] for shap_value in self._shap_values: if shap_value > 0: self._segment_colors.append(self._positive_color) else: self._segment_colors.append(self._negative_color) def _set_data_source(self): data = { 'shap_value': self._shap_values, 'name': self._names, 'segment_color': self._segment_colors, 'segment_start_proba': log_odds_to_proba(self._segment_starts), 'segment_end_proba': log_odds_to_proba(self._segment_ends) } if self._descriptions: data['descriptions'] = self._descriptions self._data_source = bokeh.models.sources.ColumnDataSource(data=data) def _set_tooltips(self): self._tooltips = [ ("name", "@name"), ("description", "@description") ] def _set_x_range(self, x_range_padding): center = (self._segments_minimum + self._segments_maximum) / 2 minimum = center - ((center - self._segments_minimum) * (1 + x_range_padding)) maximum = ((self._segments_maximum - center) * (1 + x_range_padding)) + center self._x_range_minimum = log_odds_to_proba(minimum) self._x_range_maximum = log_odds_to_proba(maximum) def _set_figure(self): figure = bokeh.plotting.figure( frame_width=self._x_range_width, frame_height=25 * len(self._names), x_axis_type="log", x_axis_location="below", y_range=self._names, y_axis_location="right", tooltips=self._tooltips, toolbar_location=None, tools="" ) formatter = bokeh.models.formatters.NumeralTickFormatter(format="0.0 %") axis_label_text_font_size = "10pt" axis_line_color = "white" desired_num_ticks = 7 # set original xaxis figure.x_range = bokeh.models.Range1d( self._x_range_minimum, self._x_range_maximum ) figure.xaxis.formatter = formatter figure.xaxis.axis_label_text_font_size = axis_label_text_font_size figure.xaxis.axis_line_color = axis_line_color figure.xaxis.ticker.desired_num_ticks = desired_num_ticks # create a copy of xaxis above figure.extra_x_ranges = {"probability": figure.x_range} top_axis = bokeh.models.axes.LogAxis(x_range_name="probability") top_axis.formatter = formatter top_axis.axis_label_text_font_size = axis_label_text_font_size top_axis.axis_line_color = axis_line_color top_axis.ticker.desired_num_ticks = desired_num_ticks figure.add_layout(top_axis, 'above') # set yaxis figure.yaxis.major_label_text_font_size = axis_label_text_font_size figure.yaxis.axis_line_color = axis_line_color self._figure = figure def _add_baseline_span(self): span = bokeh.models.Span( location=log_odds_to_proba(self._baseline), dimension='height', line_color='grey', line_dash='dotted', line_width=1 ) span.level = 'underlay' self._figure.add_layout(span) def _get_x_range_offset(self, log_odd): minimum = proba_to_log_odds(self._x_range_minimum) maximum = proba_to_log_odds(self._x_range_maximum) x_range_length = maximum - minimum percentage = (log_odd - minimum) / x_range_length x_range_offset = percentage * self._x_range_width return x_range_offset def _add_baseline_label(self): baseline_proba = log_odds_to_proba(self._baseline) text = '{:.2%} baseline'.format(baseline_proba) offset = self._get_x_range_offset(self._baseline) label = bokeh.models.Title( text=text, align='left', text_font_size='10pt', offset=offset, text_font_style='italic', text_color='grey') self._figure.add_layout(label, 'below') def _get_prediction(self): return self._segment_ends[-1] def _get_prediction_color(self): prediction = self._get_prediction() if prediction > self._baseline: return self._positive_color else: return self._negative_color def _add_prediction_span(self): prediction = self._get_prediction() prediction_proba = log_odds_to_proba(prediction) prediction_color = self._get_prediction_color() span = bokeh.models.Span( location=prediction_proba, dimension='height', line_color=prediction_color, line_dash='dotted', line_width=1 ) span.level = 'underlay' self._figure.add_layout(span) def _add_prediction_label(self): prediction = self._get_prediction() prediction_proba = log_odds_to_proba(prediction) text = '{:.2%} predicted'.format(prediction_proba) offset = self._get_x_range_offset(prediction) prediction_color = self._get_prediction_color() label = bokeh.models.Title( text=text, align='left', text_font_size='10pt', offset=offset, text_font_style='italic', text_color=prediction_color ) self._figure.add_layout(label, 'above') def _add_segments(self): self._figure.segment( "segment_start_proba", "name", "segment_end_proba", "name", line_width=18, line_color="segment_color", source=self._data_source ) self._figure.diamond( y="name", x="segment_end_proba", size=16, fill_color="segment_color", line_color="segment_color", source=self._data_source ) self._figure.diamond( y="name", x="segment_start_proba", size=16, fill_color='white', line_color="white", source=self._data_source ) def add_title(self, text): if text: title = bokeh.models.Title(text=text, align='left', text_font_size='12pt') self._figure.add_layout(title, 'above') def add_x_axis_label(self, text): label = bokeh.models.Title(text=text, align='center', text_font_size='10pt') self._figure.add_layout(label, 'above') def show(self): return bokeh.plotting.show(self._figure) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -