feat: Cost tracker for CostTriggers

modyn/config/schema/pipeline/trigger/__init__.py

@@ -4,6 +4,11 @@
 
 from pydantic import Field
 
+from .cost import *  # noqa
+from .cost import (
+    AvoidableMisclassificationCostTriggerConfig,
+    DataIncorporationLatencyCostTriggerConfig,
+)
 from .drift import *  # noqa
 from .drift import DataDriftTriggerConfig
 from .ensemble import *  # noqa
@@ -14,6 +19,11 @@
 from .simple import SimpleTriggerConfig
 
 TriggerConfig = Annotated[
-    SimpleTriggerConfig | DataDriftTriggerConfig | EnsembleTriggerConfig | PerformanceTriggerConfig,
+    AvoidableMisclassificationCostTriggerConfig
+    | DataIncorporationLatencyCostTriggerConfig
+    | DataDriftTriggerConfig
+    | EnsembleTriggerConfig
+    | PerformanceTriggerConfig
+    | SimpleTriggerConfig,
     Field(discriminator="id"),
 ]

modyn/config/schema/pipeline/trigger/cost/__init__.py

@@ -0,0 +1,9 @@
+from .cost import (
+    AvoidableMisclassificationCostTriggerConfig,
+    DataIncorporationLatencyCostTriggerConfig,
+)
+
+__all__ = [
+    "AvoidableMisclassificationCostTriggerConfig",
+    "DataIncorporationLatencyCostTriggerConfig",
+]

modyn/config/schema/pipeline/trigger/cost/cost.py

@@ -0,0 +1,74 @@
+"""Cost aware triggers are evaluate the trade-off between the cost of a trigger
+(e.g. measured by training machine time) and the benefit of the trigger (e.g.
+measured by a regret metric like data incorporation latency or avoidable
+misclassification incorporation latency)."""
+
+from typing import Literal
+
+from pydantic import Field
+
+from modyn.config.schema.base_model import ModynBaseModel
+
+from ..performance.performance import _InternalPerformanceTriggerConfig
+
+
+class _CostTriggerConfig(ModynBaseModel):
+    """Base class for cost aware trigger policies."""
+
+    cost_tracking_window_size: int = Field(
+        description="How many trigger into the past should be considered for the linear cost w.r.t. data amount model."
+    )
+
+
+class DataIncorporationLatencyCostTriggerConfig(_CostTriggerConfig):
+    """Cost aware trigger policy configuration that uses the data incorporation
+    latency as a regret metric.
+
+    While no trigger occurs the un-integrated data is cumulated and the sum of all seconds of every unincorporated data
+    point is added up.
+
+    `incorporation_delay_per_training_second` servers as a conversion rate between budget (training time) and regret
+    metric (incorporation latency).
+
+    When the regret metric exceeds the training-time based budget metric, a trigger is fired.
+    """
+
+    id: Literal["DataIncorporationLatencyCostTrigger"] = Field("DataIncorporationLatencyCostTrigger")
+
+    # Conversion rate between budget (training time) and regret metric (misclassifications)
+    incorporation_delay_per_training_second: float = Field(
+        description=(
+            "How many seconds of samples having spent in the incorporation queue are we willing to accept "
+            "per second of training time saved."
+        )
+    )
+
+
+class AvoidableMisclassificationCostTriggerConfig(_CostTriggerConfig, _InternalPerformanceTriggerConfig):
+    """Cost aware trigger policy configuration that using the number of
+    avoidable misclassifications integration latency as a regret metric.
+
+    `avoidable_misclassification_latency_per_training_second` servers as a conversion rate between budget (training time)
+    and regret metric (misclassifications).
+
+    When a the regret metric exceeds the budget, a trigger is fired.
+
+    This policy is an extension of performance aware triggers.
+
+    Offers the same set of `decision_criteria` as `PerformanceTriggerConfig`
+    but implicitly adds a cost criterion to the list.
+
+    Not only evaluates data density and model performance but also
+    consider the cost of a trigger both in terms of wall clock time and
+    number of triggers.
+
+    We use the `_InternalPerformanceTriggerConfig` base class as we cannot override the `id` field of
+    `PerformanceTriggerConfig`.
+    """
+
+    id: Literal["AvoidableMisclassificationCostTrigger"] = Field("AvoidableMisclassificationCost")
+
+    # Conversion rate between budget (training time) and regret metric (misclassifications)
+    avoidable_misclassification_latency_per_training_second: float = Field(
+        description="How many seconds of unaddressed avoidable misclassifications are we willing to accept per second of training time saved."
+    )

modyn/supervisor/internal/triggers/cost/cost_tracker.py

@@ -0,0 +1,44 @@
+from collections import deque
+
+import numpy as np
+from sklearn import linear_model
+
+
+class CostTracker:
+    """Observes a stream trigger costs (wall clack time measurements) and
+    maintains a linear model assuming a linear relationship between the number
+    of samples and the time it takes to process them."""
+
+    def __init__(self, window_size: int = 1000) -> None:
+        """
+        Args:
+            window_size: How many trigger into the past should be considered for the linear model.
+        """
+        self.measurements: deque[tuple[int, float]] = deque(maxlen=window_size)
+        """List of measurements of number of samples and the resulted training
+        time for the last `window_size` triggers."""
+
+        self._linear_model = linear_model.Ridge()
+
+    def inform_trigger(self, number_samples: int, elapsed_time: float) -> None:
+        """Informs the tracker about new data batch."""
+        self.measurements.append((number_samples, elapsed_time))
+
+        samples = np.array([x[0] for x in self.measurements]).reshape(-1, 1)
+        observations = [x[1] for x in self.measurements]
+
+        self._linear_model.fit(samples, observations)
+
+    def needs_calibration(self) -> bool:
+        """Checks if the tracker has enough data for a forecast.
+
+        After one trigger inform, the tracker forecasts with a constant
+        model. With the second trigger the model can learn the
+        y-intercept and start making meaningful forecasts.
+        """
+        return len(self.measurements) == 0 or not hasattr(self._linear_model, "coef_")
+
+    def forecast_training_time(self, number_samples: int) -> float:
+        """Forecasts the training time for a given number of samples."""
+        assert not self.needs_calibration(), "The tracker needs more data to make a forecast."
+        return self._linear_model.predict(np.array([[number_samples]]))[0]

modyn/tests/supervisor/internal/triggers/cost/test_cost_tracker.py

@@ -0,0 +1,50 @@
+from collections import deque
+
+import numpy as np
+import pytest
+
+from modyn.supervisor.internal.triggers.cost.cost_tracker import CostTracker
+
+
+@pytest.fixture
+def tracker() -> CostTracker:
+    return CostTracker(window_size=3)
+
+
+def test_initial_state_and_needs_calibration(tracker: CostTracker) -> None:
+    assert len(tracker.measurements) == 0
+    assert tracker.needs_calibration()
+
+    tracker.inform_trigger(10, 0.5)
+    assert not tracker.needs_calibration()
+
+
+def test_inform_trigger_updates_measurements(tracker: CostTracker) -> None:
+    tracker.inform_trigger(10, 0.5)
+    tracker.inform_trigger(20, 1.0)
+    assert tracker.measurements == deque([(10, 0.5), (20, 1.0)], maxlen=3)
+
+    tracker.inform_trigger(30, 1.5)
+    assert tracker.measurements == deque([(10, 0.5), (20, 1.0), (30, 1.5)], maxlen=3)
+
+    tracker.inform_trigger(40, 2.0)
+    assert tracker.measurements == deque([(20, 1.0), (30, 1.5), (40, 2.0)], maxlen=3)
+
+
+def test_forecast_training_time_and_model_retraining(
+    tracker: CostTracker,
+) -> None:
+    # Inform with data points
+    with pytest.raises(AssertionError, match="The tracker needs more data to make a forecast."):
+        tracker.forecast_training_time(20)
+
+    tracker.inform_trigger(10, 0.5)
+    assert tracker.forecast_training_time(20) == 0.5  # first inform will only configure a constant model
+    assert tracker.forecast_training_time(0) == 0.5
+
+    tracker.inform_trigger(20, 1.0)
+    initial_coef = tracker._linear_model.coef_.copy()
+
+    tracker.inform_trigger(30, 1.5)
+    updated_coef = tracker._linear_model.coef_.copy()
+    assert not np.array_equal(initial_coef, updated_coef)