libinephany 0.16.3__tar.gz → 0.16.5__tar.gz

libinephany-0.16.5/CODE_VERSION.cfg

@@ -0,0 +1 @@
+0.16.5

{libinephany-0.16.3/libinephany.egg-info → libinephany-0.16.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: libinephany
-Version: 0.16.3
+Version: 0.16.5
 Summary: Inephany library containing code commonly used by multiple subpackages.
 Author-email: Inephany <info@inephany.com>
 License: Apache 2.0

{libinephany-0.16.3 → libinephany-0.16.5}/libinephany/observations/observation_utils.py

@@ -64,20 +64,8 @@ def get_exponential_weighted_average(values: list[int | float]) -> float:
     :param values: List of values to average via EWA.
     :return: EWA of the given values.
     """
-
-    # Check for NaN and infinite values in input
-    valid_values = [float(val) for val in values if not math.isnan(float(val)) and not math.isinf(float(val))]
-
-    if not valid_values:
-        raise ValueError("Cannot compute exponential weighted average on empty list")
-
-    if len(valid_values) == 1:
-        return valid_values[0]
-
-    exp_weighted_average = pd.Series(valid_values).ewm(alpha=0.1).mean().iloc[-1]
+    exp_weighted_average = pd.Series(values).ewm(alpha=0.1).mean().iloc[-1]
     assert isinstance(exp_weighted_average, float)
-    assert not math.isnan(exp_weighted_average)
-    assert not math.isinf(exp_weighted_average)
     return exp_weighted_average
 
 
@@ -395,7 +383,7 @@ def _weighted_interval_expectation(
     interval_gradient = (end_value - start_value) / (end_time_point - start_time_point)
     start_exp_time = math.exp(log_decay_factor * start_time_point)
     end_exp_time = math.exp(log_decay_factor * end_time_point)
-    return (1 / log_decay_factor) * (end_value * end_exp_time - start_value * start_exp_time) + (
+    return (1 / log_decay_factor) * (end_value * end_exp_time - start_value * start_exp_time) - (
         1 / log_decay_factor**2
     ) * interval_gradient * (end_exp_time - start_exp_time)
 

libinephany-0.16.5/libinephany/observations/observers/global_observers/__init__.py

@@ -0,0 +1,69 @@
+# ======================================================================================================================
+#
+# GLOBAL OBSERVERS PACKAGE
+#
+# This package contains all global observer classes used for collecting observations
+# across the entire training process, not specific to individual agents.
+#
+# ======================================================================================================================
+
+
+from .gradient_observers import GlobalFirstOrderGradients, GlobalSecondOrderGradients, LHOPTGradientVarianceFraction
+from .hyperparameter_observers import (
+    InitialHyperparameters,
+    LHOPTHyperparameterRatio,
+    ModelFamilyOneHot,
+    OptimizerTypeOneHot,
+)
+from .loss_observers import (
+    LHOPTLossRatio,
+    LHOPTTrainingLoss,
+    LHOPTValidationLoss,
+    LossRatio,
+    PercentileOfLossAtEachCheckpoint,
+    TrainingLoss,
+    TrainingScore,
+    ValidationLoss,
+    ValidationScore,
+)
+from .model_observers import (
+    GlobalActivations,
+    GlobalLAMBTrustRatio,
+    GlobalParameters,
+    GlobalParameterUpdates,
+    LogRatioOfPreviousAndCurrentParamNormEnvStepObserver,
+    LogRatioOfUpdateAndPreviousParamNormEnvStepObserver,
+    NumberOfLayers,
+    NumberOfParameters,
+)
+from .progress_observers import EpochsCompleted, ProgressAtEachCheckpoint, TrainingProgress
+
+__all__ = [
+    InitialHyperparameters.__name__,
+    LHOPTHyperparameterRatio.__name__,
+    OptimizerTypeOneHot.__name__,
+    ModelFamilyOneHot.__name__,
+    TrainingLoss.__name__,
+    ValidationLoss.__name__,
+    LossRatio.__name__,
+    TrainingScore.__name__,
+    ValidationScore.__name__,
+    GlobalFirstOrderGradients.__name__,
+    GlobalSecondOrderGradients.__name__,
+    LHOPTGradientVarianceFraction.__name__,
+    GlobalActivations.__name__,
+    GlobalParameterUpdates.__name__,
+    GlobalParameters.__name__,
+    GlobalLAMBTrustRatio.__name__,
+    NumberOfParameters.__name__,
+    NumberOfLayers.__name__,
+    LogRatioOfPreviousAndCurrentParamNormEnvStepObserver.__name__,
+    LogRatioOfUpdateAndPreviousParamNormEnvStepObserver.__name__,
+    TrainingProgress.__name__,
+    EpochsCompleted.__name__,
+    ProgressAtEachCheckpoint.__name__,
+    LHOPTTrainingLoss.__name__,
+    LHOPTValidationLoss.__name__,
+    LHOPTLossRatio.__name__,
+    PercentileOfLossAtEachCheckpoint.__name__,
+]

libinephany-0.16.5/libinephany/observations/observers/global_observers/base_classes.py

@@ -0,0 +1,224 @@
+# ======================================================================================================================
+#
+# BASE CLASSES
+#
+# ======================================================================================================================
+
+import math
+from abc import ABC, abstractmethod
+from typing import Any
+
+from libinephany.observations.observation_utils import StatisticStorageTypes, compute_cdf_feature
+from libinephany.observations.observers.base_observers import GlobalObserver
+from libinephany.observations.observers.global_observers.constants import LHOPT_CONSTANTS
+from libinephany.pydantic_models.schemas.observation_models import ObservationInputs
+from libinephany.pydantic_models.schemas.tensor_statistics import TensorStatistics
+from libinephany.pydantic_models.states.hyperparameter_states import HyperparameterStates
+
+
+class LHOPTBaseObserver(GlobalObserver, ABC):
+    """
+    Base class for LHOPT outer step observers to eliminate duplicate code.
+    """
+
+    def __init__(
+        self,
+        decay_factor: float = LHOPT_CONSTANTS["DEFAULT_DECAY_FACTOR"],
+        time_window: int = LHOPT_CONSTANTS["DEFAULT_TIME_WINDOW"],
+        **kwargs,
+    ) -> None:
+        """
+        :param decay_factor: Decay factor for CDF calculation in [1, 2.5, 5, 10, 20]
+        :param time_window: Number of time steps to consider for CDF calculation
+        :param kwargs: Other observation keyword arguments.
+        """
+        super().__init__(**kwargs)
+        self.decay_factor = max(0.0, decay_factor)
+        self.time_window = max(1, time_window)
+
+        # Store time series data for CDF calculation
+        self._time_series: list[tuple[float, float]] = []  # (time, value) pairs
+        self._current_time: float = 0.0
+
+    @property
+    def can_standardize(self) -> bool:
+        """
+        This observer has its own CDF calculation, no need to standardize.
+        :return: Whether the observation can be standardized.
+        """
+        return False
+
+    def _get_observation_format(self) -> StatisticStorageTypes:
+        """
+        :return: Format the observation returns data in. Must be one of the StatisticStorageTypes
+        enumeration class.
+        """
+        return StatisticStorageTypes.VECTOR
+
+    def _compute_cdf_feature(self, value: float) -> float:
+        """
+        Compute CDF feature for the given value.
+        training loss will be added to the time series after this call.
+        :param value: The value to compute CDF feature for
+        :return: CDF feature value
+        """
+        return compute_cdf_feature(value, self._time_series, self.decay_factor, self._current_time, self.time_window)
+
+    def _update_time(self) -> None:
+        """Update the current time counter."""
+        self._current_time += 1.0
+
+    def get_required_trackers(self) -> dict[str, dict[str, Any] | None]:
+        """
+        :return: Dictionary mapping statistic tracker class names to kwargs for the class or None if no kwargs are
+        needed.
+        """
+        return {}
+
+    def reset(self) -> None:
+        """Reset the observer by clearing the time series."""
+        self._time_series = []
+        self._current_time = 0.0
+
+    @abstractmethod
+    def _observe(
+        self,
+        observation_inputs: ObservationInputs,
+        hyperparameter_states: HyperparameterStates,
+        tracked_statistics: dict[str, dict[str, float | TensorStatistics]],
+        action_taken: float | int | None,
+    ) -> float | int | list[int | float] | TensorStatistics:
+        """
+        :param observation_inputs: Observation input metrics not calculated with statistic trackers.
+        :param hyperparameter_states: HyperparameterStates that manages the hyperparameters.
+        :param tracked_statistics: Dictionary mapping statistic tracker class names to dictionaries mapping module
+        :param action_taken: Action taken by the agent this class instance is assigned to.
+        """
+        raise NotImplementedError
+
+    def _compute_log_ratio(self, numerator: float, denominator: float) -> float:
+        """
+        Compute the log ratio.
+
+        :param numerator: Numerator value
+        :param denominator: Denominator value
+        :return: Log ratio value
+        """
+        # Calculate the ratio of numerator to denominator
+
+        if denominator <= LHOPT_CONSTANTS["ZERO_DIVISION_TOLERANCE"]:
+            return 0.0
+
+        ratio = numerator / denominator
+
+        if ratio <= 0:
+            return 0.0
+
+        return math.log(ratio)
+
+
+class LHOPTCheckpointBaseObserver(GlobalObserver, ABC):
+    """
+    Base class for checkpoint-based observers to eliminate duplicate code.
+    """
+
+    def __init__(self, checkpoint_interval: int = LHOPT_CONSTANTS["DEFAULT_CHECKPOINT_INTERVAL"], **kwargs) -> None:
+        """
+        :param checkpoint_interval: How often to create checkpoints (in outer model steps).
+        :param kwargs: Miscellaneous keyword arguments.
+        """
+        super().__init__(**kwargs)
+        self.checkpoint_interval = checkpoint_interval
+        self._history: list[float] = []
+        self.last_value: float | None = None
+
+    @property
+    def can_standardize(self) -> bool:
+        """
+        This observer has its own CDF calculation, no need to standardize.
+        :return: Whether the observation can be standardized.
+        """
+        return False
+
+    def _get_observation_format(self) -> StatisticStorageTypes:
+        """
+        :return: Format the observation returns data in.
+        """
+        return StatisticStorageTypes.FLOAT
+
+    def _update_history(self, value: float) -> None:
+        """
+        Update the history with a new value and maintain sliding window.
+
+        :param value: The new value to add to history
+        """
+        self._history.append(value)
+
+        # Keep only the last checkpoint_interval values for sliding window
+        if len(self._history) > self.checkpoint_interval:
+            self._history = self._history[-self.checkpoint_interval :]
+
+    def _should_create_checkpoint(self) -> bool:
+        """
+        Check if we should create a checkpoint.
+
+        :return: True if checkpoint should be created, False otherwise
+        """
+        return len(self._history) >= self.checkpoint_interval
+
+    def _cold_start(self, value: float) -> None:
+        """
+        Handle cold start by setting the last value if not already set.
+
+        :param value: The value to set as last value if cold start
+        """
+        if self.last_value is None:
+            self.last_value = value
+
+    def get_required_trackers(self) -> dict[str, dict[str, Any] | None]:
+        """
+        :return: Dictionary mapping statistic tracker class names to kwargs for the class or None if no kwargs are
+        needed.
+        """
+        return {}
+
+    def reset(self) -> None:
+        """Reset the observer by clearing history."""
+        self._history = []
+        self.last_value = None
+
+    @abstractmethod
+    def _observe(
+        self,
+        observation_inputs: ObservationInputs,
+        hyperparameter_states: HyperparameterStates,
+        tracked_statistics: dict[str, dict[str, float | TensorStatistics]],
+        action_taken: float | int | None,
+    ) -> float | int | list[int | float] | TensorStatistics:
+        """
+        :param observation_inputs: Observation input metrics not calculated with statistic trackers.
+        :param hyperparameter_states: HyperparameterStates that manages the hyperparameters.
+        :param tracked_statistics: Dictionary mapping statistic tracker class names to dictionaries mapping module
+        :param action_taken: Action taken by the agent this class instance is assigned to.
+        """
+        raise NotImplementedError
+
+    def _compute_log_ratio(self, numerator: float, denominator: float) -> float:
+        """
+        Compute the log ratio.
+
+        :param numerator: Numerator value
+        :param denominator: Denominator value
+        :return: Log ratio value
+        """
+        # Calculate the ratio of numerator to denominator
+
+        if denominator <= LHOPT_CONSTANTS["ZERO_DIVISION_TOLERANCE"]:
+            return 0.0
+
+        ratio = numerator / denominator
+
+        if ratio <= 0:
+            return 0.0
+
+        return math.log(ratio)

libinephany-0.16.5/libinephany/observations/observers/global_observers/constants.py

@@ -0,0 +1,39 @@
+# ======================================================================================================================
+#
+# CONSTANTS
+#
+# ======================================================================================================================
+
+from typing import TypedDict
+
+
+class LHOPTConstants(TypedDict):
+    IS_NAN: float
+    NOT_NAN: float
+    IS_INF: float
+    NOT_INF: float
+    TANH_BOUND: float
+    DEFAULT_DECAY_FACTOR: float
+    DEFAULT_TIME_WINDOW: int
+    DEFAULT_CHECKPOINT_INTERVAL: int
+    DEFAULT_PERCENTILE: float
+    ZERO_DIVISION_TOLERANCE: float
+    DEFAULT_SAMPLE_FREQUENCY: int
+    DEFAULT_VARIANCE_THRESHOLD: float
+
+
+# Create the constants instance
+LHOPT_CONSTANTS: LHOPTConstants = LHOPTConstants(
+    IS_NAN=1.0,
+    NOT_NAN=0.0,
+    IS_INF=1.0,
+    NOT_INF=0.0,
+    TANH_BOUND=10.0,
+    DEFAULT_DECAY_FACTOR=1.25,
+    DEFAULT_TIME_WINDOW=32,
+    DEFAULT_CHECKPOINT_INTERVAL=100,
+    DEFAULT_PERCENTILE=0.6,
+    ZERO_DIVISION_TOLERANCE=1e-8,
+    DEFAULT_SAMPLE_FREQUENCY=4,
+    DEFAULT_VARIANCE_THRESHOLD=1e-6,
+)

libinephany-0.16.5/libinephany/observations/observers/global_observers/gradient_observers.py

@@ -0,0 +1,193 @@
+# ======================================================================================================================
+#
+# GRADIENT OBSERVERS
+#
+# ======================================================================================================================
+
+from typing import Any
+
+from libinephany.observations import observation_utils, statistic_trackers
+from libinephany.observations.observation_utils import StatisticStorageTypes
+from libinephany.observations.observers.base_observers import GlobalObserver
+from libinephany.observations.observers.global_observers.base_classes import LHOPTBaseObserver
+from libinephany.observations.observers.global_observers.constants import LHOPT_CONSTANTS
+from libinephany.pydantic_models.schemas.observation_models import ObservationInputs
+from libinephany.pydantic_models.schemas.tensor_statistics import TensorStatistics
+from libinephany.pydantic_models.states.hyperparameter_states import HyperparameterStates
+
+
+class GlobalFirstOrderGradients(GlobalObserver):
+
+    def _get_observation_format(self) -> StatisticStorageTypes:
+        """
+        :return: Format the observation returns data in. Must be one of the enum attributes in the StatisticStorageTypes
+        enumeration class.
+        """
+
+        return StatisticStorageTypes.TENSOR_STATISTICS
+
+    def _observe(
+        self,
+        observation_inputs: ObservationInputs,
+        hyperparameter_states: HyperparameterStates,
+        tracked_statistics: dict[str, dict[str, float | TensorStatistics]],
+        action_taken: float | int | None,
+    ) -> float | int | list[int | float] | TensorStatistics:
+        """
+        :param observation_inputs: Observation input metrics not calculated with statistic trackers.
+        :param hyperparameter_states: HyperparameterStates that manages the hyperparameters.
+        :param tracked_statistics: Dictionary mapping statistic tracker class names to dictionaries mapping module
+        names to floats or TensorStatistic models.
+        :param action_taken: Action taken by the agent this class instance is assigned to.
+        :return: Single float/int, list of floats/ints or TensorStatistics model to add to the observation vector.
+        """
+
+        statistics = tracked_statistics[statistic_trackers.FirstOrderGradients.__name__]
+
+        return observation_utils.average_tensor_statistics(tensor_statistics=list(statistics.values()))  # type: ignore
+
+    def get_required_trackers(self) -> dict[str, dict[str, Any] | None]:
+        """
+        :return: Dictionary mapping statistic tracker class names to kwargs for the class or None if no kwargs are
+        needed.
+        """
+
+        return {statistic_trackers.FirstOrderGradients.__name__: dict(skip_statistics=self.skip_statistics)}
+
+
+class GlobalSecondOrderGradients(GlobalObserver):
+
+    def __init__(
+        self,
+        *,
+        compute_hessian_diagonal: bool = False,
+        **kwargs,
+    ) -> None:
+        """
+        :param compute_hessian_diagonal: Whether to compute the Hessian diagonal to determine second order gradients
+        or use the squared first order gradients as approximations in the same way Adam does.
+        :param kwargs: Miscellaneous keyword arguments.
+        """
+
+        super().__init__(**kwargs)
+
+        self.compute_hessian_diagonal = compute_hessian_diagonal
+
+    def _get_observation_format(self) -> StatisticStorageTypes:
+        """
+        :return: Format the observation returns data in. Must be one of the enum attributes in the StatisticStorageTypes
+        enumeration class.
+        """
+
+        return StatisticStorageTypes.TENSOR_STATISTICS
+
+    def _observe(
+        self,
+        observation_inputs: ObservationInputs,
+        hyperparameter_states: HyperparameterStates,
+        tracked_statistics: dict[str, dict[str, float | TensorStatistics]],
+        action_taken: float | int | None,
+    ) -> float | int | list[int | float] | TensorStatistics:
+        """
+        :param observation_inputs: Observation input metrics not calculated with statistic trackers.
+        :param hyperparameter_states: HyperparameterStates that manages the hyperparameters.
+        :param tracked_statistics: Dictionary mapping statistic tracker class names to dictionaries mapping module
+        names to floats or TensorStatistic models.
+        :param action_taken: Action taken by the agent this class instance is assigned to.
+        :return: Single float/int, list of floats/ints or TensorStatistics model to add to the observation vector.
+        """
+
+        statistics = tracked_statistics[statistic_trackers.SecondOrderGradients.__name__]
+
+        return observation_utils.average_tensor_statistics(tensor_statistics=list(statistics.values()))  # type: ignore
+
+    def get_required_trackers(self) -> dict[str, dict[str, Any] | None]:
+        """
+        :return: Dictionary mapping statistic tracker class names to kwargs for the class or None if no kwargs are
+        needed.
+        """
+
+        return {
+            statistic_trackers.SecondOrderGradients.__name__: dict(
+                skip_statistics=self.skip_statistics, compute_hessian_diagonal=self.compute_hessian_diagonal
+            )
+        }
+
+
+class LHOPTGradientVarianceFraction(LHOPTBaseObserver):
+    """
+    This is a global observer from the OpenAI paper "Learning to Optimize with Reinforcement Learning"
+    https://arxiv.org/abs/2305.18291.
+
+    It returns two-dimensional observations: [raw_value, cdf_feature] for gradient variance fraction values.
+    """
+
+    def __init__(
+        self,
+        *,
+        variance_threshold: float = LHOPT_CONSTANTS["DEFAULT_VARIANCE_THRESHOLD"],
+        **kwargs,
+    ) -> None:
+        """
+        :param variance_threshold: Threshold for variance comparison in gradient variance fraction calculation
+        :param kwargs: Other observation keyword arguments.
+        """
+        super().__init__(**kwargs)
+        self.variance_threshold = variance_threshold
+
+    @property
+    def can_standardize(self) -> bool:
+        """
+        This observer has its own CDF calculation, no need to standardize.
+        :return: Whether the observation can be standardized.
+        """
+        return False
+
+    def _get_observation_format(self) -> StatisticStorageTypes:
+        """
+        :return: Format the observation returns data in. Must be one of the enum attributes in the StatisticStorageTypes
+        enumeration class.
+        """
+        return StatisticStorageTypes.VECTOR
+
+    @property
+    def vector_length(self) -> int:
+        """
+        :return: Length of the vector returned by this observation if it returns a vector.
+        """
+        return 2  # [raw_value, cdf_feature]
+
+    def _observe(
+        self,
+        observation_inputs: ObservationInputs,
+        hyperparameter_states: HyperparameterStates,
+        tracked_statistics: dict[str, dict[str, float | TensorStatistics]],
+        action_taken: float | int | None,
+    ) -> float | int | list[int | float] | TensorStatistics:
+        """
+        :param observation_inputs: Observation input metrics not calculated with statistic trackers.
+        :param hyperparameter_states: HyperparameterStates that manages the hyperparameters.
+        :param tracked_statistics: Dictionary mapping statistic tracker class names to dictionaries mapping module
+        names to floats or TensorStatistic models.
+        :param action_taken: Action taken by the agent this class instance is assigned to.
+        :return: Single float/int, list of floats/ints or TensorStatistics model to add to the observation vector.
+        """
+        if statistic_trackers.GradientVarianceFraction.__name__ not in tracked_statistics:
+            return [0.0, 0.0]
+
+        raw_value = list(tracked_statistics[statistic_trackers.GradientVarianceFraction.__name__].values())[0]  # type: ignore[list-item]
+
+        cdf_feature = self._compute_cdf_feature(raw_value)  # type: ignore[arg-type]
+        self._update_time()
+
+        return [raw_value, cdf_feature]  # type: ignore[list-item]
+
+    def get_required_trackers(self) -> dict[str, dict[str, Any] | None]:
+        """
+        :return: Dictionary mapping statistic tracker class names to kwargs for the class or None if no kwargs are
+        needed.
+        """
+
+        return {
+            statistic_trackers.GradientVarianceFraction.__name__: dict(variance_threshold=self.variance_threshold),
+        }