Source code for ax.plot.benchmark

# 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.

from typing import Iterable, List, Optional

from ax.benchmark.benchmark_result import AggregatedBenchmarkResult
from ax.plot.base import AxPlotConfig, AxPlotTypes
from ax.plot.color import COLORS, DISCRETE_COLOR_SCALE, rgba
from ax.plot.helper import rgb
from plotly import graph_objs as go


[docs]def plot_modeling_times( aggregated_results: Iterable[AggregatedBenchmarkResult], ) -> AxPlotConfig: """Plots wall times of each method's fit and gen calls as a stack bar chart.""" data = [ go.Bar( name="fit", x=[result.name for result in aggregated_results], y=[result.fit_time[0] for result in aggregated_results], text=["fit" for _ in aggregated_results], error_y={ "type": "data", "array": [result.fit_time[1] for result in aggregated_results], "visible": True, }, opacity=0.6, ), go.Bar( name="gen", x=[result.name for result in aggregated_results], y=[result.gen_time[0] for result in aggregated_results], text=["gen" for _ in aggregated_results], error_y={ "type": "data", "array": [agg.gen_time[1] for agg in aggregated_results], "visible": True, }, opacity=0.9, ), ] layout = go.Layout( title="Modeling Times", showlegend=False, yaxis={"title": "Time (s)"}, xaxis={"title": "Method"}, barmode="stack", ) return AxPlotConfig( data=go.Figure(layout=layout, data=data), plot_type=AxPlotTypes.GENERIC )
[docs]def plot_optimization_trace( aggregated_results: List[AggregatedBenchmarkResult], optimum: Optional[float] = None, by_progression: bool = False, final_progression_only: bool = False, ) -> AxPlotConfig: """Plots optimization trace for each aggregated result with mean and SEM. When `by_progression` is True, the results are plotted with progressions on the x-axis. In that case, if `final_progression_only` is True, then the value of a trial is taken to be the value of its final progression. If an optimum is provided (can represent either an optimal value or maximum hypervolume in the case of multi-objective problems) it will be plotted as an orange dashed line as well. """ x_axes = [] dfs = [] for agg_res in aggregated_results: if not by_progression: x_axes.append([*range(len(agg_res.optimization_trace))]) dfs.append(agg_res.optimization_trace) else: optim_trace_by_prog_res = agg_res.optimization_trace_by_progression( final_progression_only=final_progression_only ) x_axes.append(optim_trace_by_prog_res["progression"]) dfs.append(optim_trace_by_prog_res) mean_sem_scatters = [ [ go.Scatter( x=x_axis, y=df["mean"], line={ "color": rgba(DISCRETE_COLOR_SCALE[i % len(DISCRETE_COLOR_SCALE)]) }, mode="lines", name=r.name, customdata=df["sem"], hovertemplate="<br><b>Mean:</b> %{y}<br><b>SEM</b>: %{customdata}", ), go.Scatter( x=x_axis, y=df["mean"] + df["sem"], line={"width": 0}, mode="lines", fillcolor=rgba( DISCRETE_COLOR_SCALE[i % len(DISCRETE_COLOR_SCALE)], 0.3 ), fill="tonexty", showlegend=False, hoverinfo="skip", ), go.Scatter( x=x_axis, y=df["mean"] - df["sem"], line={"width": 0}, mode="lines", fillcolor=rgba( DISCRETE_COLOR_SCALE[i % len(DISCRETE_COLOR_SCALE)], 0.3 ), fill="tonexty", showlegend=False, hoverinfo="skip", ), ] for i, (x_axis, df, r) in enumerate(zip(x_axes, dfs, aggregated_results)) ] optimum_scatter = ( [ go.Scatter( x=x_axes[0], y=[optimum] * len(x_axes[0]), mode="lines", line={"dash": "dash", "color": rgb(COLORS.ORANGE.value)}, name="Optimum", hovertemplate="Optimum: %{y}", ) ] if optimum is not None else [] ) layout = go.Layout( title="Optimization Traces", yaxis={"title": "Best Found"}, xaxis={"title": "Iteration"}, hovermode="x unified", ) return AxPlotConfig( data=go.Figure( layout=layout, data=[scatter for sublist in mean_sem_scatters for scatter in sublist] + optimum_scatter, ), plot_type=AxPlotTypes.GENERIC, )