#!/usr/bin/env python3
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
# pyre-strict
from logging import Logger
from typing import Any, Dict, Optional, Union
import numpy as np
import pandas as pd
import plotly.graph_objs as go
from ax.core.parameter import ChoiceParameter
from ax.exceptions.core import NoDataError
from ax.modelbridge import ModelBridge
from ax.plot.base import AxPlotConfig, AxPlotTypes
from ax.plot.helper import compose_annotation
from ax.utils.common.logger import get_logger
from plotly import subplots
logger: Logger = get_logger(__name__)
[docs]def plot_feature_importance_plotly(df: pd.DataFrame, title: str) -> go.Figure:
if df.empty:
raise NoDataError("No Data on Feature Importances found.")
df.set_index(df.columns[0], inplace=True)
data = [
go.Bar(y=df.index, x=df[column_name], name=column_name, orientation="h")
for column_name in df.columns
]
fig = subplots.make_subplots(
rows=len(df.columns),
cols=1,
subplot_titles=df.columns,
print_grid=False,
shared_xaxes=True,
)
for idx, item in enumerate(data):
fig.append_trace(item, idx + 1, 1) # pyre-ignore[16]
fig.layout.showlegend = False
fig.layout.margin = go.layout.Margin(
l=8 * min(max(len(idx) for idx in df.index), 75) # noqa E741
)
fig.layout.title = title
return fig
[docs]def plot_feature_importance(df: pd.DataFrame, title: str) -> AxPlotConfig:
"""Wrapper method to convert plot_feature_importance_plotly to
AxPlotConfig"""
return AxPlotConfig(
data=plot_feature_importance_plotly(df, title), plot_type=AxPlotTypes.GENERIC
)
[docs]def plot_feature_importance_by_metric_plotly(model: ModelBridge) -> go.Figure:
"""One plot per feature, showing importances by metric."""
importances = []
for metric_name in sorted(model.metric_names):
try:
vals: Dict[str, Any] = model.feature_importances(metric_name)
vals["index"] = metric_name
importances.append(vals)
except NotImplementedError:
logger.warning(
f"Model for {metric_name} does not support feature importances."
)
if not importances:
raise NotImplementedError(
"Feature importances could not be calculated for any metric"
)
df = pd.DataFrame(importances)
# plot_feature_importance expects index in first column
df = df.reindex(columns=(["index"] + [a for a in df.columns if a != "index"]))
plot_fi = plot_feature_importance_plotly(df, "Parameter Sensitivity by Metric")
num_subplots = len(df.columns)
num_features = len(df)
# Include per-subplot margin for subplot titles (feature names).
plot_fi["layout"]["height"] = num_subplots * (num_features + 1) * 50
return plot_fi
[docs]def plot_feature_importance_by_metric(model: ModelBridge) -> AxPlotConfig:
"""Wrapper method to convert plot_feature_importance_by_metric_plotly to
AxPlotConfig"""
return AxPlotConfig(
data=plot_feature_importance_by_metric_plotly(model),
plot_type=AxPlotTypes.GENERIC,
)
[docs]def plot_feature_importance_by_feature_plotly(
model: Optional[ModelBridge] = None,
sensitivity_values: Optional[Dict[str, Dict[str, Union[float, np.ndarray]]]] = None,
relative: bool = False,
caption: str = "",
importance_measure: str = "",
label_dict: Optional[Dict[str, str]] = None,
) -> go.Figure:
"""One plot per metric, showing importances by feature.
If sensitivity values are not all positive, the absolute value will be shown
and color will indicate positive or negative sign.
Args:
model: A model with a ``feature_importances`` method.
sensitivity_values: The sensitivity values for each metric in a dict format.
It takes the following format if only the sensitivity value is plotted:
`{"metric1":{"parameter1":value1,"parameter2":value2 ...} ...}`
It takes the following format if the sensitivity value and standard error
are plotted: `{"metric1":{"parameter1":[value1,var,se],
"parameter2":[[value2,var,se]]...}...}}`.
relative: Whether to normalize feature importances so that they add to 1.
caption: An HTML-formatted string to place at the bottom of the plot.
importance_measure: The name of the importance metric to be added to the title.
label_dict: A dictionary mapping metric names to short labels.
Returns a go.Figure of feature importances.
"""
if sensitivity_values is None:
if model is None:
raise ValueError(
"A model is required when sensitivity values are not provided"
)
try:
sensitivity_values = {
metric_name: model.feature_importances(metric_name)
for i, metric_name in enumerate(sorted(model.metric_names))
}
except NotImplementedError:
raise NotImplementedError(
"Feature importances cannot be computed by the model."
)
if label_dict is not None:
sensitivity_values = { # pyre-ignore
label_dict.get(metric_name, metric_name): v
for metric_name, v in sensitivity_values.items()
}
traces = []
dropdown = []
categorical_features = []
if model is not None:
categorical_features = [
name
for name, par in model.model_space.parameters.items()
if isinstance(par, ChoiceParameter) and not par.is_ordered
]
for i, metric_name in enumerate(sorted(sensitivity_values.keys())):
importances = sensitivity_values[metric_name]
factor_col = "Factor"
importance_col = "Importance"
sign_col = "Sign"
error_plot = np.asarray(next(iter(importances.values()))).size > 1
if error_plot:
importance_col_se = "SE"
df = pd.DataFrame(
[
{
factor_col: factor,
importance_col: np.asarray(importance)[0],
importance_col_se: np.asarray(importance)[2],
sign_col: (
0
if factor in categorical_features
else 2 * (np.asarray(importance)[0] >= 0).astype(int) - 1
),
}
for factor, importance in importances.items()
]
)
df[importance_col] = df[importance_col].abs()
df = df.sort_values(importance_col)
error_x = {"type": "data", "array": df[importance_col_se], "visible": True}
else:
df = pd.DataFrame(
[
{
factor_col: factor,
importance_col: importance,
sign_col: (
0
if factor in categorical_features
else 2 * (importance >= 0).astype(int) - 1
),
}
for factor, importance in importances.items()
]
)
df[importance_col] = df[importance_col].abs()
df = df.sort_values(importance_col)
error_x = None
if relative:
df[importance_col] = df[importance_col].div(df[importance_col].sum())
colors = {-1: "darkorange", 0: "gray", 1: "steelblue"}
names = {
-1: "Decreases metric",
0: "Affects metric (categorical choice)",
1: "Increases metric",
}
legend_counter = {-1: 0, 0: 0, 1: 0}
all_positive = all(df[sign_col] >= 0)
for _, row in df.iterrows():
traces.append(
go.Bar(
name=names[row[sign_col]],
orientation="h",
visible=i == 0,
x=np.array([row[importance_col]]),
y=np.array([row[factor_col]]),
error_x=error_x,
opacity=0.8,
marker_color=colors[row[sign_col]],
showlegend=(not all_positive)
and (legend_counter[row[sign_col]] == 0),
legendgroup=str(row[sign_col]),
)
)
legend_counter[row[sign_col]] += 1
is_visible = [False] * (len(sensitivity_values) * len(df))
for j in range(i * len(df), (i + 1) * len(df)):
is_visible[j] = True
dropdown.append(
{"args": ["visible", is_visible], "label": metric_name, "method": "restyle"}
)
if not traces:
raise NotImplementedError("No traces found for metric")
updatemenus = [
{
"x": 0,
"y": 1,
"yanchor": "top",
"xanchor": "left",
"buttons": dropdown,
"pad": {
"t": -40
}, # hack to put dropdown below title regardless of number of features
}
]
features = list(list(sensitivity_values.values())[0].keys())
title = "Normalized parameter sensitivity" if relative else "Parameter sensitivity"
if importance_measure:
title = title + " using " + importance_measure
longest_label = max(len(f) for f in features)
longest_metric = max(len(m) for m in sensitivity_values.keys())
layout = go.Layout(
height=200 + len(features) * 20,
width=10 * longest_label + max(10 * longest_metric, 400),
hovermode="closest",
margin=go.layout.Margin(
l=8 * min(max(len(idx) for idx in features), 75)
), # noqa E741
title=title,
updatemenus=updatemenus,
annotations=compose_annotation(caption=caption),
)
if relative:
layout.update({"xaxis": {"tickformat": ".0%"}})
return go.Figure(data=traces, layout=layout)
[docs]def plot_feature_importance_by_feature(
model: Optional[ModelBridge] = None,
sensitivity_values: Optional[Dict[str, Dict[str, Union[float, np.ndarray]]]] = None,
relative: bool = False,
caption: str = "",
importance_measure: str = "",
label_dict: Optional[Dict[str, str]] = None,
) -> AxPlotConfig:
"""Wrapper method to convert `plot_feature_importance_by_feature_plotly` to
AxPlotConfig"""
return AxPlotConfig(
data=plot_feature_importance_by_feature_plotly(
model=model,
sensitivity_values=sensitivity_values,
relative=relative,
caption=caption,
importance_measure=importance_measure,
label_dict=label_dict,
),
plot_type=AxPlotTypes.GENERIC,
)
[docs]def plot_relative_feature_importance_plotly(model: ModelBridge) -> go.Figure:
"""Create a stacked bar chart of feature importances per metric"""
importances = []
for metric_name in sorted(model.metric_names):
try:
vals: Dict[str, Any] = model.feature_importances(metric_name)
vals["index"] = metric_name
importances.append(vals)
except Exception:
logger.warning(
"Model for {} does not support feature importances.".format(metric_name)
)
df = pd.DataFrame(importances)
df.set_index("index", inplace=True)
df = df.div(df.sum(axis=1), axis=0)
data = [
go.Bar(y=df.index, x=df[column_name], name=column_name, orientation="h")
for column_name in df.columns
]
layout = go.Layout(
margin=go.layout.Margin(l=250), # noqa E741
barmode="group",
yaxis={"title": ""},
xaxis={"title": "Relative Parameter importance"},
showlegend=False,
title="Relative Parameter Importance per Metric",
)
return go.Figure(data=data, layout=layout)
[docs]def plot_relative_feature_importance(model: ModelBridge) -> AxPlotConfig:
"""Wrapper method to convert plot_relative_feature_importance_plotly to
AxPlotConfig"""
return AxPlotConfig(
data=plot_relative_feature_importance_plotly(model),
plot_type=AxPlotTypes.GENERIC,
)
