#!/usr/bin/env python3
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import math
import pandas as pd
import plotly.graph_objs as go
from ax.core.data import Data
from ax.core.experiment import Experiment
from ax.modelbridge.factory import get_empirical_bayes_thompson, get_thompson
from ax.plot.base import AxPlotConfig, AxPlotTypes, PlotMetric, Z
from ax.plot.helper import get_plot_data
from ax.plot.scatter import _error_scatter_data
COLOR_SCALE = ["#ff3333", "#ff6666", "#ffffff", "#99ff99", "#33ff33"]
[docs]def get_color(x: float, ci: float, rel: bool, reverse: bool):
"""Determine the color of the table cell."""
if not rel:
# Color coding is meant to be relative to the status quo,
# and thus doesn't make sense if rel = False
return "#ffffff"
r = min(math.floor(abs(x) / ci), 2) if ci > 0 else 2
index = int(2 + r * math.copysign(1, x))
color_scale = list(COLOR_SCALE)
if reverse:
color_scale = list(reversed(color_scale))
return color_scale[index]
[docs]def table_view_plot(
experiment: Experiment,
data: Data,
use_empirical_bayes: bool = True,
only_data_frame: bool = False,
arm_noun: str = "arm",
):
"""Table of means and confidence intervals.
Table is of the form:
+-------+------------+-----------+
| arm | metric_1 | metric_2 |
+=======+============+===========+
| 0_0 | mean +- CI | ... |
+-------+------------+-----------+
| 0_1 | ... | ... |
+-------+------------+-----------+
"""
model_func = get_empirical_bayes_thompson if use_empirical_bayes else get_thompson
model = model_func(experiment=experiment, data=data)
# We don't want to include metrics from a collection,
# or the chart will be too big to read easily.
# Example:
# experiment.metrics = {
# 'regular_metric': Metric(),
# 'collection_metric: CollectionMetric()', # collection_metric =[metric1, metric2]
# }
# model.metric_names = [regular_metric, metric1, metric2] # "exploded" out
# We want to filter model.metric_names and get rid of metric1, metric2
metric_names = [
metric_name
for metric_name in model.metric_names
if metric_name in experiment.metrics
]
metric_name_to_lower_is_better = {
metric_name: experiment.metrics[metric_name].lower_is_better
for metric_name in metric_names
}
plot_data, _, _ = get_plot_data(
model=model, generator_runs_dict={}, metric_names=metric_names
)
if plot_data.status_quo_name:
status_quo_arm = plot_data.in_sample.get(plot_data.status_quo_name)
rel = True
else:
status_quo_arm = None
rel = False
records = []
colors = []
records_with_mean = []
records_with_ci = []
for metric_name in metric_names:
arm_names, _, ys, ys_se = _error_scatter_data(
arms=list(plot_data.in_sample.values()),
y_axis_var=PlotMetric(metric_name, pred=True, rel=rel),
x_axis_var=None,
status_quo_arm=status_quo_arm,
)
results_by_arm = list(zip(arm_names, ys, ys_se))
colors.append(
[
get_color(
x=y,
ci=Z * y_se,
rel=rel,
reverse=metric_name_to_lower_is_better[metric_name],
)
for (_, y, y_se) in results_by_arm
]
)
records.append(
[
"{:.3f} ± {:.3f}".format(y, Z * y_se)
for (_, y, y_se) in results_by_arm
]
)
records_with_mean.append({arm_name: y for (arm_name, y, _) in results_by_arm})
records_with_ci.append(
{arm_name: Z * y_se for (arm_name, _, y_se) in results_by_arm}
)
if only_data_frame:
return tuple(
pd.DataFrame.from_records(records, index=metric_names)
for records in [records_with_mean, records_with_ci]
)
def transpose(m):
return [[m[j][i] for j in range(len(m))] for i in range(len(m[0]))]
records = [[name.replace(":", " : ") for name in metric_names]] + transpose(records)
colors = [["#ffffff"] * len(metric_names)] + transpose(colors)
header = [f"<b>{x}</b>" for x in [f"{arm_noun}s"] + arm_names]
column_widths = [300] + [150] * len(arm_names)
trace = go.Table(
header={"values": header, "align": ["left"]},
cells={"values": records, "align": ["left"], "fill": {"color": colors}},
columnwidth=column_widths,
)
layout = go.Layout(
width=sum(column_widths),
margin=go.layout.Margin(l=0, r=20, b=20, t=20, pad=4), # noqa E741
)
fig = go.Figure(data=[trace], layout=layout)
return AxPlotConfig(data=fig, plot_type=AxPlotTypes.GENERIC)
