Add by_wallclock_time option to ProgressionPlot (facebook#3383)

Summary:
Pull Request resolved: facebook#3383

Allows a user to specify that they would like to plot with wallclock time on the x axis. Starts a t=0 as the earliest start time

Reviewed By: saitcakmak

Differential Revision: D69800761

fbshipit-source-id: abf2eb647ad22c4f2cbad2932b0841cd3d5330e1

ax/analysis/plotly/progression.py

@@ -4,6 +4,9 @@
 # LICENSE file in the root directory of this source tree.
 
 # pyre-strict
+from logging import Logger
+
+import numpy as np
 import plotly.express as px
 from ax.analysis.analysis import AnalysisCardLevel
 
@@ -14,7 +17,11 @@
 from ax.core.trial_status import TrialStatus
 from ax.exceptions.core import UserInputError
 from ax.generation_strategy.generation_strategy import GenerationStrategy
+from ax.utils.common.logger import get_logger
 from plotly import graph_objects as go
+from pyre_extensions import assert_is_instance
+
+logger: Logger = get_logger(__name__)
 
 
 class ProgressionPlot(PlotlyAnalysis):
@@ -28,17 +35,24 @@ class ProgressionPlot(PlotlyAnalysis):
         - arm_name: The name of the arm
         - METRIC_NAME: The observed mean of the metric specified
         - progression: The progression at which the metric was observed
+        - wallclock_time: The wallclock time at which the metric was observed, in
+            seconds and starting at 0 from the first trial's start time.
     """
 
-    def __init__(self, metric_name: str | None = None) -> None:
+    def __init__(
+        self, metric_name: str | None = None, by_wallclock_time: bool = False
+    ) -> None:
         """
         Args:
             metric_name: The name of the metric to plot. If not specified the objective
                 will be used. Note that the metric cannot be inferred for
                 multi-objective or scalarized-objective experiments.
+            wallclock_time: If True, plot the relative wallclock time instead of the
+                progression on the x-axis.
         """
 
         self._metric_name = metric_name
+        self._by_wallclock_time = by_wallclock_time
 
     def compute(
         self,
@@ -74,28 +88,111 @@ def compute(
                     for trial in experiment.trials_by_status[TrialStatus.EARLY_STOPPED]
                 ]
             ),
-            ["mean", map_key],
+            ["trial_index", "mean", map_key],
         ].rename(columns={map_key: "progression", "mean": metric_name})
 
+        # Add the wallclock time column
+        wallclock_series = _calculate_wallclock_timeseries(
+            experiment=experiment, metric_name=metric_name
+        )
+
+        df["wallclock_time"] = df.apply(
+            lambda row: wallclock_series[row["trial_index"]][row["progression"]],
+            axis=1,
+        )
+        if len(terminal_points) > 0:
+            terminal_points["wallclock_time"] = terminal_points.apply(
+                lambda row: wallclock_series[row["trial_index"]][row["progression"]],
+                axis=1,
+            )
+
         # Plot the progression lines with one curve for each arm.
-        fig = px.line(df, x="progression", y=metric_name, color="arm_name")
+        if self._by_wallclock_time:
+            x_axis_name = "wallclock_time"
+        else:
+            x_axis_name = "progression"
+
+        fig = px.line(df, x=x_axis_name, y=metric_name, color="arm_name")
 
         # Add a marker for each terminal point on early stopped trials.
-        fig.add_trace(
-            go.Scatter(
-                x=terminal_points["progression"],
-                y=terminal_points[metric_name],
-                mode="markers",
-                showlegend=False,
-                line_color="red",
-                hoverinfo="none",
+        if len(terminal_points) > 0:
+            fig.add_trace(
+                go.Scatter(
+                    x=terminal_points[x_axis_name],
+                    y=terminal_points[metric_name],
+                    mode="markers",
+                    showlegend=False,
+                    line_color="red",
+                    hoverinfo="none",
+                )
             )
-        )
 
         return self._create_plotly_analysis_card(
-            title=f"{metric_name} by progression",
+            title=f"{metric_name} by {x_axis_name.replace('_', ' ')}",
             subtitle="Observe how the metric changes as each trial progresses",
             level=AnalysisCardLevel.MID,
             df=df,
             fig=fig,
         )
+
+
+def _calculate_wallclock_timeseries(
+    experiment: Experiment,
+    metric_name: str,
+) -> dict[int, dict[float, float]]:
+    """
+    Calculate a mapping from each trial index and progression to the time since the
+    first trial started, in seconds. Assume that the first trial started at t=0, and
+    that progressions are linearly spaced between the start and completion times of
+    each trial.
+
+    If a trial does not have either a start or completion time the wallclock time
+    cannot be calculated and the value will be nan (which will not be plotted).
+
+    Returns:
+        trial_index => (progression => timestamp)
+    """
+    # Find the earliest start time.
+    start_time = min(
+        trial.time_run_started.timestamp()
+        for trial in experiment.trials.values()
+        if trial.time_run_started is not None
+    )
+    # Calculate all start and completion times relative to the earliest start time.
+    # Give nan for trials that don't have a start or completion time.
+    relative_timestamps = {
+        idx: (
+            trial.time_run_started.timestamp() - start_time
+            if trial.time_run_started is not None
+            else np.nan,
+            trial.time_completed.timestamp() - start_time
+            if trial.time_completed is not None
+            else np.nan,
+        )
+        for idx, trial in experiment.trials.items()
+    }
+
+    data = assert_is_instance(experiment.lookup_data(), MapData)
+    df = data.map_df[data.map_df["metric_name"] == metric_name]
+    map_key = data.map_key_infos[0].key
+
+    return {
+        trial_index: dict(
+            zip(
+                df[df["trial_index"] == trial_index][map_key].to_numpy(),
+                # Map the progressions to linspace if the start and completion times
+                # are both available, otherwise map to nans
+                np.linspace(
+                    relative_timestamps[trial_index][0],
+                    relative_timestamps[trial_index][1],
+                    len(df[df["trial_index"] == trial_index]),
+                )
+                if (
+                    relative_timestamps[trial_index][0] is not None
+                    and relative_timestamps[trial_index][1] is not None
+                )
+                else np.full(len(df[df["trial_index"] == trial_index]), np.nan),
+            )
+        )
+        for trial_index in experiment.trials.keys()
+    }

ax/analysis/plotly/tests/test_progression.py

@@ -5,8 +5,12 @@
 
 # pyre-strict
 
+import pandas as pd
 from ax.analysis.analysis import AnalysisCardLevel
-from ax.analysis.plotly.progression import ProgressionPlot
+from ax.analysis.plotly.progression import (
+    _calculate_wallclock_timeseries,
+    ProgressionPlot,
+)
 from ax.exceptions.core import UserInputError
 from ax.utils.common.testutils import TestCase
 from ax.utils.testing.core_stubs import get_test_map_data_experiment
@@ -33,8 +37,25 @@ def test_compute(self) -> None:
         )
         self.assertEqual(card.level, AnalysisCardLevel.MID)
         self.assertEqual(
-            {*card.df.columns}, {"trial_index", "arm_name", "branin_map", "progression"}
+            {*card.df.columns},
+            {"trial_index", "arm_name", "branin_map", "progression", "wallclock_time"},
         )
 
         self.assertIsNotNone(card.blob)
         self.assertEqual(card.blob_annotation, "plotly")
+
+    def test_calculate_wallclock_timeseries(self) -> None:
+        experiment = get_test_map_data_experiment(
+            num_trials=2, num_fetches=5, num_complete=2
+        )
+        wallclock_timeseries = _calculate_wallclock_timeseries(
+            experiment=experiment, metric_name="branin_map"
+        )
+
+        self.assertEqual(len(wallclock_timeseries), 2)
+        self.assertTrue(
+            all(len(timeseries) == 5 for timeseries in wallclock_timeseries.values())
+        )
+
+        for timeseries in wallclock_timeseries.values():
+            self.assertTrue(pd.Series(timeseries).is_monotonic_increasing)