modelbase2 0.1.78__py3-none-any.whl → 0.2.0__py3-none-any.whl

modelbase2/surrogates/_poly.py

@@ -0,0 +1,66 @@
+from collections.abc import Iterable
+from dataclasses import dataclass
+
+import numpy as np
+import pandas as pd
+from numpy.polynomial.polynomial import Polynomial
+
+from modelbase2.types import AbstractSurrogate, ArrayLike
+
+__all__ = ["PolySurrogate", "train_polynomial_surrogate"]
+
+
+@dataclass(kw_only=True)
+class PolySurrogate(AbstractSurrogate):
+    model: Polynomial
+
+    def predict_raw(self, y: np.ndarray) -> np.ndarray:
+        return self.model(y)
+
+
+def train_polynomial_surrogate(
+    feature: ArrayLike,
+    target: ArrayLike,
+    degrees: Iterable[int] = (1, 2, 3, 4, 5, 6, 7),
+    surrogate_args: list[str] | None = None,
+    surrogate_stoichiometries: dict[str, dict[str, float]] | None = None,
+) -> tuple[PolySurrogate, pd.DataFrame]:
+    """Train a polynomial surrogate model.
+
+    Args:
+        feature: Input data as a numpy array.
+        target: Output data as a numpy array.
+        degrees: Degrees of the polynomial to fit to the data.
+        surrogate_args: Additional arguments for the surrogate model.
+        surrogate_stoichiometries: Stoichiometries for the surrogate model.
+
+    Returns:
+        PolySurrogate: Polynomial surrogate model.
+
+    """
+    feature = np.array(feature, dtype=float)
+    target = np.array(target, dtype=float)
+
+    models = [Polynomial.fit(feature, target, degree) for degree in degrees]
+    predictions = np.array([model(feature) for model in models], dtype=float)
+    errors = np.sqrt(np.mean(np.square(predictions - target.reshape(1, -1)), axis=1))
+    log_likelihood = -0.5 * np.sum(
+        np.square(predictions - target.reshape(1, -1)), axis=1
+    )
+    score = 2 * np.array(degrees) - 2 * log_likelihood
+
+    # Choose the model with the lowest AIC
+    model = models[np.argmin(score)]
+    return (
+        PolySurrogate(
+            model=model,
+            args=surrogate_args if surrogate_args is not None else [],
+            stoichiometries=surrogate_stoichiometries
+            if surrogate_stoichiometries is not None
+            else {},
+        ),
+        pd.DataFrame(
+            {"models": models, "error": errors, "score": score},
+            index=pd.Index(np.array(degrees), name="degree"),
+        ),
+    )

modelbase2/surrogates/_torch.py

@@ -0,0 +1,249 @@
+from dataclasses import dataclass
+
+import numpy as np
+import pandas as pd
+import torch
+import tqdm
+from torch import nn
+from torch.optim.adam import Adam
+
+from modelbase2.types import AbstractSurrogate
+
+__all__ = ["DefaultDevice", "Dense", "TorchSurrogate", "train_torch_surrogate"]
+
+DefaultDevice = torch.device("cpu")
+
+
+@dataclass(kw_only=True)
+class TorchSurrogate(AbstractSurrogate):
+    """Surrogate model using PyTorch.
+
+    Attributes:
+        model: PyTorch neural network model.
+
+    Methods:
+        predict: Predict outputs based on input data using the PyTorch model.
+
+    """
+
+    model: torch.nn.Module
+
+    def predict_raw(self, y: np.ndarray) -> np.ndarray:
+        """Predict outputs based on input data using the PyTorch model.
+
+        Args:
+            y: Input data as a numpy array.
+
+        Returns:
+            dict[str, float]: Dictionary mapping output variable names to predicted values.
+
+        """
+        with torch.no_grad():
+            return self.model(
+                torch.tensor(y, dtype=torch.float32),
+            ).numpy()
+
+
+class Dense(nn.Module):
+    """Neural network approximator for surrogate modeling.
+
+    Attributes:
+        net: Sequential neural network model.
+
+    Methods:
+        forward: Forward pass through the neural network.
+
+    """
+
+    def __init__(self, n_inputs: int, n_outputs: int) -> None:
+        """Initializes the surrogate model with the given number of inputs and outputs.
+
+        Args:
+            n_inputs (int): The number of input features.
+            n_outputs (int): The number of output features.
+
+        Initializes a neural network with the following architecture:
+        - Linear layer with `n_inputs` inputs and 50 outputs
+        - ReLU activation
+        - Linear layer with 50 inputs and 50 outputs
+        - ReLU activation
+        - Linear layer with 50 inputs and `n_outputs` outputs
+
+        The weights of the linear layers are initialized with a normal distribution
+        (mean=0, std=0.1) and the biases are initialized to 0.
+
+        """
+        super().__init__()
+
+        self.net = nn.Sequential(
+            nn.Linear(n_inputs, 50),
+            nn.ReLU(),
+            nn.Linear(50, 50),
+            nn.ReLU(),
+            nn.Linear(50, n_outputs),
+        )
+
+        for m in self.net.modules():
+            if isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, mean=0, std=0.1)
+                nn.init.constant_(m.bias, val=0)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Forward pass through the neural network.
+
+        Args:
+            x: Input tensor.
+
+        Returns:
+            torch.Tensor: Output tensor.
+
+        """
+        return self.net(x)
+
+
+def _train_batched(
+    aprox: nn.Module,
+    features: pd.DataFrame,
+    targets: pd.DataFrame,
+    epochs: int,
+    optimizer: Adam,
+    device: torch.device,
+    batch_size: int,
+) -> pd.Series:
+    """Train the neural network using mini-batch gradient descent.
+
+    Args:
+        aprox: Neural network model to train.
+        features: Input features as a tensor.
+        targets: Target values as a tensor.
+        epochs: Number of training epochs.
+        optimizer: Optimizer for training.
+        device: torch device
+        batch_size: Size of mini-batches for training.
+
+    Returns:
+        pd.Series: Series containing the training loss history.
+
+    """
+    rng = np.random.default_rng()
+    losses = {}
+    for i in tqdm.trange(epochs):
+        idxs = rng.choice(features.index, size=batch_size)
+        X = torch.Tensor(features.iloc[idxs].to_numpy(), device=device)
+        Y = torch.Tensor(targets.iloc[idxs].to_numpy(), device=device)
+        optimizer.zero_grad()
+        loss = torch.mean(torch.abs(aprox(X) - Y))
+        loss.backward()
+        optimizer.step()
+        losses[i] = loss.detach().numpy()
+    return pd.Series(losses, dtype=float)
+
+
+def _train_full(
+    aprox: nn.Module,
+    features: pd.DataFrame,
+    targets: pd.DataFrame,
+    epochs: int,
+    optimizer: Adam,
+    device: torch.device,
+) -> pd.Series:
+    """Train the neural network using full-batch gradient descent.
+
+    Args:
+        aprox: Neural network model to train.
+        features: Input features as a tensor.
+        targets: Target values as a tensor.
+        epochs: Number of training epochs.
+        optimizer: Optimizer for training.
+        device: Torch device
+
+    Returns:
+        pd.Series: Series containing the training loss history.
+
+    """
+    X = torch.Tensor(features.to_numpy(), device=device)
+    Y = torch.Tensor(targets.to_numpy(), device=device)
+
+    losses = {}
+    for i in tqdm.trange(epochs):
+        optimizer.zero_grad()
+        loss = torch.mean(torch.abs(aprox(X) - Y))
+        loss.backward()
+        optimizer.step()
+        losses[i] = loss.detach().numpy()
+    return pd.Series(losses, dtype=float)
+
+
+def train_torch_surrogate(
+    features: pd.DataFrame,
+    targets: pd.DataFrame,
+    epochs: int,
+    surrogate_args: list[str] | None = None,
+    surrogate_stoichiometries: dict[str, dict[str, float]] | None = None,
+    batch_size: int | None = None,
+    approximator: nn.Module | None = None,
+    optimimzer_cls: type[Adam] = Adam,
+    device: torch.device = DefaultDevice,
+) -> tuple[TorchSurrogate, pd.Series]:
+    """Train a PyTorch surrogate model.
+
+    Examples:
+        >>> train_torch_surrogate(
+        ...     features,
+        ...     targets,
+        ...     epochs=100,
+        ...     surrogate_inputs=["x1", "x2"],
+        ...     surrogate_stoichiometries={
+        ...         "v1": {"x1": -1, "x2": 1, "ATP": -1},
+        ...     },
+        ...)
+
+    Args:
+        features: DataFrame containing the input features for training.
+        targets: DataFrame containing the target values for training.
+        epochs: Number of training epochs.
+        surrogate_args: List of input variable names for the surrogate model.
+        surrogate_stoichiometries: Dictionary mapping reaction names to stoichiometries.
+        batch_size: Size of mini-batches for training (None for full-batch).
+        approximator: Predefined neural network model (None to use default).
+        optimimzer_cls: Optimizer class to use for training (default: Adam).
+        device: Device to run the training on (default: DefaultDevice).
+
+    Returns:
+        tuple[TorchSurrogate, pd.Series]: Trained surrogate model and loss history.
+
+    """
+    if approximator is None:
+        approximator = Dense(
+            n_inputs=len(features.columns),
+            n_outputs=len(targets.columns),
+        ).to(device)
+
+    optimizer = optimimzer_cls(approximator.parameters())
+    if batch_size is None:
+        losses = _train_full(
+            aprox=approximator,
+            features=features,
+            targets=targets,
+            epochs=epochs,
+            optimizer=optimizer,
+            device=device,
+        )
+    else:
+        losses = _train_batched(
+            aprox=approximator,
+            features=features,
+            targets=targets,
+            epochs=epochs,
+            optimizer=optimizer,
+            device=device,
+            batch_size=batch_size,
+        )
+    surrogate = TorchSurrogate(
+        model=approximator,
+        args=surrogate_args if surrogate_args is not None else [],
+        stoichiometries=surrogate_stoichiometries
+        if surrogate_stoichiometries is not None
+        else {},
+    )
+    return surrogate, losses

modelbase2/surrogates.py

@@ -0,0 +1,316 @@
+"""Surrogate Models Module.
+
+This module provides classes and functions for creating and training surrogate models
+for metabolic simulations. It includes functionality for both steady-state and time-series
+data using neural networks.
+
+Classes:
+    AbstractSurrogate: Abstract base class for surrogate models.
+    TorchSurrogate: Surrogate model using PyTorch.
+    Approximator: Neural network approximator for surrogate modeling.
+
+Functions:
+    train_torch_surrogate: Train a PyTorch surrogate model.
+    train_torch_time_course_estimator: Train a PyTorch time course estimator.
+"""
+
+from __future__ import annotations
+
+from abc import abstractmethod
+from dataclasses import dataclass
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+import torch
+import tqdm
+from torch import nn
+from torch.optim.adam import Adam
+
+from modelbase2.parallel import Cache
+
+__all__ = [
+    "AbstractSurrogate",
+    "Approximator",
+    "DefaultCache",
+    "DefaultDevice",
+    "MockSurrogate",
+    "TorchSurrogate",
+    "train_torch_surrogate",
+]
+
+
+DefaultDevice = torch.device("cpu")
+DefaultCache = Cache(Path(".cache"))
+
+
+@dataclass(kw_only=True)
+class AbstractSurrogate:
+    """Abstract base class for surrogate models.
+
+    Attributes:
+        inputs: List of input variable names.
+        stoichiometries: Dictionary mapping reaction names to stoichiometries.
+
+    Methods:
+        predict: Abstract method to predict outputs based on input data.
+
+    """
+
+    inputs: list[str]
+    stoichiometries: dict[str, dict[str, float]]
+
+    @abstractmethod
+    def predict(self, y: np.ndarray) -> dict[str, float]:
+        """Predict outputs based on input data."""
+
+
+@dataclass(kw_only=True)
+class MockSurrogate(AbstractSurrogate):
+    """Mock surrogate model for testing purposes."""
+
+    def predict(
+        self,
+        y: np.ndarray,
+    ) -> dict[str, float]:
+        """Predict outputs based on input data."""
+        return dict(zip(self.stoichiometries, y, strict=True))
+
+
+@dataclass(kw_only=True)
+class TorchSurrogate(AbstractSurrogate):
+    """Surrogate model using PyTorch.
+
+    Attributes:
+        model: PyTorch neural network model.
+
+    Methods:
+        predict: Predict outputs based on input data using the PyTorch model.
+
+    """
+
+    model: torch.nn.Module
+
+    def predict(self, y: np.ndarray) -> dict[str, float]:
+        """Predict outputs based on input data using the PyTorch model.
+
+        Args:
+            y: Input data as a numpy array.
+
+        Returns:
+            dict[str, float]: Dictionary mapping output variable names to predicted values.
+
+        """
+        with torch.no_grad():
+            return dict(
+                zip(
+                    self.stoichiometries,
+                    self.model(
+                        torch.tensor(y, dtype=torch.float32),
+                    ).numpy(),
+                    strict=True,
+                )
+            )
+
+
+class Approximator(nn.Module):
+    """Neural network approximator for surrogate modeling.
+
+    Attributes:
+        net: Sequential neural network model.
+
+    Methods:
+        forward: Forward pass through the neural network.
+
+    """
+
+    def __init__(self, n_inputs: int, n_outputs: int) -> None:
+        """Initializes the surrogate model with the given number of inputs and outputs.
+
+        Args:
+            n_inputs (int): The number of input features.
+            n_outputs (int): The number of output features.
+
+        Initializes a neural network with the following architecture:
+        - Linear layer with `n_inputs` inputs and 50 outputs
+        - ReLU activation
+        - Linear layer with 50 inputs and 50 outputs
+        - ReLU activation
+        - Linear layer with 50 inputs and `n_outputs` outputs
+
+        The weights of the linear layers are initialized with a normal distribution
+        (mean=0, std=0.1) and the biases are initialized to 0.
+
+        """
+        super().__init__()
+
+        self.net = nn.Sequential(
+            nn.Linear(n_inputs, 50),
+            nn.ReLU(),
+            nn.Linear(50, 50),
+            nn.ReLU(),
+            nn.Linear(50, n_outputs),
+        )
+
+        for m in self.net.modules():
+            if isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, mean=0, std=0.1)
+                nn.init.constant_(m.bias, val=0)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Forward pass through the neural network.
+
+        Args:
+            x: Input tensor.
+
+        Returns:
+            torch.Tensor: Output tensor.
+
+        """
+        return self.net(x)
+
+
+def _train_batched(
+    aprox: nn.Module,
+    features: pd.DataFrame,
+    targets: pd.DataFrame,
+    epochs: int,
+    optimizer: Adam,
+    device: torch.device,
+    batch_size: int,
+) -> pd.Series:
+    """Train the neural network using mini-batch gradient descent.
+
+    Args:
+        aprox: Neural network model to train.
+        features: Input features as a tensor.
+        targets: Target values as a tensor.
+        epochs: Number of training epochs.
+        optimizer: Optimizer for training.
+        device: torch device
+        batch_size: Size of mini-batches for training.
+
+    Returns:
+        pd.Series: Series containing the training loss history.
+
+    """
+    rng = np.random.default_rng()
+    losses = {}
+    for i in tqdm.trange(epochs):
+        idxs = rng.choice(features.index, size=batch_size)
+        X = torch.Tensor(features.iloc[idxs].to_numpy(), device=device)
+        Y = torch.Tensor(targets.iloc[idxs].to_numpy(), device=device)
+        optimizer.zero_grad()
+        loss = torch.mean(torch.abs(aprox(X) - Y))
+        loss.backward()
+        optimizer.step()
+        losses[i] = loss.detach().numpy()
+    return pd.Series(losses, dtype=float)
+
+
+def _train_full(
+    aprox: nn.Module,
+    features: pd.DataFrame,
+    targets: pd.DataFrame,
+    epochs: int,
+    optimizer: Adam,
+    device: torch.device,
+) -> pd.Series:
+    """Train the neural network using full-batch gradient descent.
+
+    Args:
+        aprox: Neural network model to train.
+        features: Input features as a tensor.
+        targets: Target values as a tensor.
+        epochs: Number of training epochs.
+        optimizer: Optimizer for training.
+        device: Torch device
+
+    Returns:
+        pd.Series: Series containing the training loss history.
+
+    """
+    X = torch.Tensor(features.to_numpy(), device=device)
+    Y = torch.Tensor(targets.to_numpy(), device=device)
+
+    losses = {}
+    for i in tqdm.trange(epochs):
+        optimizer.zero_grad()
+        loss = torch.mean(torch.abs(aprox(X) - Y))
+        loss.backward()
+        optimizer.step()
+        losses[i] = loss.detach().numpy()
+    return pd.Series(losses, dtype=float)
+
+
+def train_torch_surrogate(
+    features: pd.DataFrame,
+    targets: pd.DataFrame,
+    epochs: int,
+    surrogate_inputs: list[str],
+    surrogate_stoichiometries: dict[str, dict[str, float]],
+    batch_size: int | None = None,
+    approximator: nn.Module | None = None,
+    optimimzer_cls: type[Adam] = Adam,
+    device: torch.device = DefaultDevice,
+) -> tuple[TorchSurrogate, pd.Series]:
+    """Train a PyTorch surrogate model.
+
+    Examples:
+        >>> train_torch_surrogate(
+        ...     features,
+        ...     targets,
+        ...     epochs=100,
+        ...     surrogate_inputs=["x1", "x2"],
+        ...     surrogate_stoichiometries={
+        ...         "v1": {"x1": -1, "x2": 1, "ATP": -1},
+        ...     },
+        ...)
+
+    Args:
+        features: DataFrame containing the input features for training.
+        targets: DataFrame containing the target values for training.
+        epochs: Number of training epochs.
+        surrogate_inputs: List of input variable names for the surrogate model.
+        surrogate_stoichiometries: Dictionary mapping reaction names to stoichiometries.
+        batch_size: Size of mini-batches for training (None for full-batch).
+        approximator: Predefined neural network model (None to use default).
+        optimimzer_cls: Optimizer class to use for training (default: Adam).
+        device: Device to run the training on (default: DefaultDevice).
+
+    Returns:
+        tuple[TorchSurrogate, pd.Series]: Trained surrogate model and loss history.
+
+    """
+    if approximator is None:
+        approximator = Approximator(
+            n_inputs=len(features.columns),
+            n_outputs=len(targets.columns),
+        ).to(device)
+
+    optimizer = optimimzer_cls(approximator.parameters())
+    if batch_size is None:
+        losses = _train_full(
+            aprox=approximator,
+            features=features,
+            targets=targets,
+            epochs=epochs,
+            optimizer=optimizer,
+            device=device,
+        )
+    else:
+        losses = _train_batched(
+            aprox=approximator,
+            features=features,
+            targets=targets,
+            epochs=epochs,
+            optimizer=optimizer,
+            device=device,
+            batch_size=batch_size,
+        )
+    surrogate = TorchSurrogate(
+        model=approximator,
+        inputs=surrogate_inputs,
+        stoichiometries=surrogate_stoichiometries,
+    )
+    return surrogate, losses