l0-python 0.1.0__py3-none-any.whl → 0.3.0__py3-none-any.whl

l0/__init__.py

@@ -26,6 +26,7 @@ from .penalties import (
     get_sparsity_stats,
     update_temperatures,
 )
+from .sparse import SparseL0Linear
 
 __all__ = [
     # Distributions
@@ -36,6 +37,8 @@ __all__ = [
     "L0DepthwiseConv2d",
     "SparseMLP",
     "prune_model",
+    # Sparse
+    "SparseL0Linear",
     # Gates
     "L0Gate",
     "SampleGate",

l0/calibration.py

@@ -0,0 +1,436 @@
+"""
+L0-regularized calibration weights for survey data.
+
+Implements sparse calibration weights that are constrained to be non-negative,
+suitable for survey weighting and importance sampling applications.
+"""
+
+# ruff: noqa: N803 N806
+
+import torch
+import torch.nn as nn
+import numpy as np
+from scipy import sparse as sp
+
+
+class SparseCalibrationWeights(nn.Module):
+    """
+    L0-regularized calibration weights with positivity constraint.
+
+    Designed for survey calibration where weights must be non-negative
+    and sparsity is desired to select a subset of households/units.
+
+    Parameters
+    ----------
+    n_features : int
+        Number of calibration weights (e.g., number of households)
+    beta : float
+        Temperature parameter for Hard Concrete distribution
+    gamma : float
+        Lower bound of stretched concrete distribution
+    zeta : float
+        Upper bound of stretched concrete distribution
+    init_keep_prob : float
+        Initial probability of keeping each weight active
+    init_weight_scale : float
+        Initial scale for log weights (controls initial weight magnitude)
+    device : str or torch.device
+        Device to run computations on ('cpu' or 'cuda')
+    """
+
+    def __init__(
+        self,
+        n_features: int,
+        beta: float = 2 / 3,
+        gamma: float = -0.1,
+        zeta: float = 1.1,
+        init_keep_prob: float = 0.5,
+        init_weight_scale: float = 1.0,
+        device: str | torch.device = "cpu",
+    ):
+        super().__init__()
+        self.n_features = n_features
+        self.beta = beta
+        self.gamma = gamma
+        self.zeta = zeta
+        self.device = torch.device(device)
+
+        # Log weights to ensure positivity via exp transformation
+        self.log_weight = nn.Parameter(
+            torch.normal(
+                mean=0.0,
+                std=init_weight_scale,
+                size=(n_features,),
+                device=self.device,
+            )
+        )
+
+        # L0 gate parameters
+        mu = torch.log(torch.tensor(init_keep_prob / (1 - init_keep_prob)))
+        self.log_alpha = nn.Parameter(
+            torch.normal(
+                mu.item(), 0.01, size=(n_features,), device=self.device
+            )
+        )
+
+        # Cache for sparse tensor conversion
+        self._cached_M_torch: torch.sparse.Tensor | None = None
+        self._cached_M_shape: tuple[int, int] | None = None
+
+    def _convert_sparse_to_torch(
+        self, M_sparse: sp.spmatrix
+    ) -> torch.sparse.Tensor:
+        """
+        Convert scipy sparse matrix to torch sparse tensor.
+
+        Caches the result if the shape matches to avoid redundant conversions.
+        """
+        if (
+            self._cached_M_torch is not None
+            and self._cached_M_shape == M_sparse.shape
+            and M_sparse.nnz == self._cached_M_torch._nnz()
+        ):
+            return self._cached_M_torch
+
+        M_coo = M_sparse.tocoo()
+        indices = torch.LongTensor(np.vstack([M_coo.row, M_coo.col])).to(
+            self.device
+        )
+        values = torch.FloatTensor(M_coo.data).to(self.device)
+        M_torch = torch.sparse_coo_tensor(
+            indices,
+            values,
+            M_sparse.shape,
+            dtype=torch.float32,
+            device=self.device,
+        )
+
+        # Cache for future use
+        self._cached_M_torch = M_torch
+        self._cached_M_shape = M_sparse.shape
+
+        return M_torch
+
+    def _sample_gates(self) -> torch.Tensor:
+        """Sample gates using Hard Concrete distribution."""
+        eps = 1e-6
+        u = torch.rand_like(self.log_alpha).clamp(eps, 1 - eps)
+        s = (torch.log(u) - torch.log(1 - u) + self.log_alpha) / self.beta
+        s = torch.sigmoid(s)
+        s_bar = s * (self.zeta - self.gamma) + self.gamma
+        return s_bar.clamp(0, 1)
+
+    def get_deterministic_gates(self) -> torch.Tensor:
+        """Get deterministic gate values (for inference)."""
+        s = torch.sigmoid(self.log_alpha / self.beta)
+        s_bar = s * (self.zeta - self.gamma) + self.gamma
+        return s_bar.clamp(0, 1)
+
+    def get_weights(self, deterministic: bool = False) -> torch.Tensor:
+        """
+        Get effective calibration weights.
+
+        Parameters
+        ----------
+        deterministic : bool
+            Whether to use deterministic gates (for inference)
+
+        Returns
+        -------
+        torch.Tensor
+            Positive calibration weights with L0 sparsity applied
+        """
+        # Sample or get deterministic gates
+        if deterministic:
+            gates = self.get_deterministic_gates()
+        else:
+            gates = self._sample_gates()
+
+        # Apply exp transformation and gates together
+        # This ensures gradient flow matches successful implementations
+        return torch.exp(self.log_weight) * gates
+
+    def forward(
+        self,
+        M: sp.spmatrix,
+        deterministic: bool = False,
+    ) -> torch.Tensor:
+        """
+        Apply calibration weights to metric matrix.
+
+        Parameters
+        ----------
+        M : scipy.sparse matrix
+            Metric matrix of shape (n_targets, n_features)
+        deterministic : bool
+            Whether to use deterministic gates (for inference)
+
+        Returns
+        -------
+        torch.Tensor
+            Weighted sums of shape (n_targets,)
+        """
+        # Convert sparse matrix to torch
+        M_torch = self._convert_sparse_to_torch(M)
+
+        # Get effective weights
+        weights = self.get_weights(deterministic=deterministic)
+
+        # Sparse matrix multiplication
+        y = torch.sparse.mm(M_torch, weights.unsqueeze(1)).squeeze(1)
+
+        return y
+
+    def get_l0_penalty(self) -> torch.Tensor:
+        """
+        Compute L0 complexity penalty.
+
+        Returns expected number of active weights.
+        """
+        c = -self.beta * torch.log(
+            torch.tensor(-self.gamma / self.zeta, device=self.device)
+        )
+        pi = torch.sigmoid(self.log_alpha + c)
+        return pi.sum()
+
+    def get_l2_penalty(self) -> torch.Tensor:
+        """
+        Compute L2 penalty on positive weights.
+
+        Helps prevent weight explosion when gates are partially open.
+
+        Returns
+        -------
+        torch.Tensor
+            L2 norm of positive weights
+        """
+        positive_weights = torch.exp(self.log_weight)
+        return (positive_weights**2).sum()
+
+    def get_sparsity(self) -> float:
+        """
+        Get current sparsity level.
+
+        Returns
+        -------
+        float
+            Fraction of weights that are effectively zero
+        """
+        with torch.no_grad():
+            gates = self.get_deterministic_gates()
+            return (gates == 0).float().mean().item()
+
+    def get_active_weights(self) -> dict:
+        """
+        Get indices and values of active (non-zero) weights.
+
+        Returns
+        -------
+        dict
+            Dictionary with 'indices' and 'values' of active weights
+        """
+        with torch.no_grad():
+            weights = self.get_weights(deterministic=True)
+            active_mask = weights > 0
+
+            return {
+                "indices": torch.where(active_mask)[0],
+                "values": weights[active_mask],
+                "count": active_mask.sum().item(),
+            }
+
+    def fit(
+        self,
+        M: sp.spmatrix,
+        y: np.ndarray,
+        lambda_l0: float = 0.01,
+        lambda_l2: float = 0.0,
+        lr: float = 0.01,
+        epochs: int = 1000,
+        loss_type: str = "mse",
+        verbose: bool = False,
+        verbose_freq: int = 100,
+        target_groups: np.ndarray | None = None,
+    ) -> "SparseCalibrationWeights":
+        """
+        Fit calibration weights using gradient descent.
+
+        Parameters
+        ----------
+        M : scipy.sparse matrix
+            Metric matrix of shape (n_targets, n_features)
+        y : numpy.ndarray
+            Target values to match
+        lambda_l0 : float
+            L0 regularization strength
+        lambda_l2 : float
+            L2 regularization strength on positive weights
+        lr : float
+            Learning rate
+        epochs : int
+            Number of training epochs
+        loss_type : str
+            Type of loss function: 'mse' or 'relative'
+        verbose : bool
+            Whether to print progress
+        verbose_freq : int
+            How often to print progress
+        target_groups : numpy.ndarray, optional
+            Array of group IDs for each target. Targets in the same group
+            will be averaged together so each group contributes equally to loss.
+            If None, all targets are treated independently.
+
+        Returns
+        -------
+        self
+            Fitted model
+        """
+        # Convert y to tensor
+        y = torch.tensor(y, dtype=torch.float32, device=self.device)
+
+        # Convert M to torch sparse (will be cached)
+        M_torch = self._convert_sparse_to_torch(M)
+
+        # Compute group weights for loss averaging
+        if target_groups is not None:
+            # Convert to tensor
+            target_groups = torch.tensor(
+                target_groups, dtype=torch.long, device=self.device
+            )
+
+            # Calculate group weights: 1 / group_size for each target
+            unique_groups = torch.unique(target_groups)
+            group_weights = torch.zeros_like(y)
+
+            for group_id in unique_groups:
+                group_mask = target_groups == group_id
+                group_size = group_mask.sum().item()
+                # Each target in the group gets weight 1/group_size
+                # so the group's total contribution is 1
+                group_weights[group_mask] = 1.0 / group_size
+        else:
+            # No grouping - all targets weighted equally
+            group_weights = torch.ones_like(y)
+
+        # Initialize weights
+        nn.init.normal_(self.log_weight, 0, 0.5)
+
+        # Setup optimizer
+        optimizer = torch.optim.Adam([self.log_weight, self.log_alpha], lr=lr)
+
+        # Training loop
+        for epoch in range(epochs):
+            # Forward pass
+            y_pred = self.forward(M, deterministic=False)
+
+            # Compute loss with group weighting
+            if loss_type == "relative":
+                # Relative error: (y - y_pred)^2 / (y + 1)^2
+                # Adding 1 to avoid division by zero
+                relative_errors = (y - y_pred) / (y + 1)
+                # Apply group weights and then average
+                weighted_squared_errors = (
+                    relative_errors.pow(2) * group_weights
+                )
+                data_loss = (
+                    weighted_squared_errors.sum()
+                )  # Sum because weights already normalize
+            else:
+                # Standard MSE with group weighting
+                squared_errors = (y - y_pred).pow(2)
+                weighted_squared_errors = squared_errors * group_weights
+                data_loss = (
+                    weighted_squared_errors.sum()
+                )  # Sum because weights already normalize
+
+            l0_loss = self.get_l0_penalty()
+            loss = data_loss + lambda_l0 * l0_loss
+
+            # Add L2 penalty if specified
+            if lambda_l2 > 0:
+                l2_loss = self.get_l2_penalty()
+                loss = loss + lambda_l2 * l2_loss
+
+            # Backward pass
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+
+            # Print progress
+            if verbose and (epoch + 1) % verbose_freq == 0:
+                with torch.no_grad():
+                    active_info = self.get_active_weights()
+                    weights = self.get_weights(deterministic=True)
+                    active_weights = weights[weights > 0]
+
+                    # Compute relative errors for meaningful output
+                    y_det = self.forward(M, deterministic=True)
+                    if loss_type == "relative":
+                        rel_errors = torch.abs((y - y_det) / (y + 1))
+                    else:
+                        # For MSE, show relative errors anyway for interpretability
+                        rel_errors = torch.abs((y - y_det) / (y + 1))
+
+                    # For reporting, we can show both overall and group-averaged errors
+                    mean_rel_err = rel_errors.mean().item()
+                    max_rel_err = rel_errors.max().item()
+
+                    # Compute mean group loss if groups are used
+                    if target_groups is not None:
+                        # Calculate mean loss per group
+                        group_losses = []
+                        for group_id in torch.unique(target_groups):
+                            group_mask = target_groups == group_id
+                            group_mean_err = (
+                                rel_errors[group_mask].mean().item()
+                            )
+                            group_losses.append(group_mean_err)
+                        mean_group_loss = np.mean(group_losses)
+                    else:
+                        mean_group_loss = mean_rel_err
+
+                    # Calculate sparsity percentage
+                    sparsity_pct = 100 * (
+                        1 - active_info["count"] / self.n_features
+                    )
+
+                    # Calculate components of the actual loss being minimized
+                    actual_data_loss = data_loss.item()
+                    actual_l0_loss = l0_loss.item()
+                    actual_total_loss = loss.item()
+
+                    if target_groups is not None:
+                        print(
+                            f"Epoch {epoch+1:4d}: "
+                            f"mean_group_loss={mean_group_loss:.1%}, "
+                            f"max_error={max_rel_err:.1%}, "
+                            f"total_loss={actual_total_loss:.3f}, "
+                            f"active={active_info['count']:4d}/{self.n_features} ({sparsity_pct:.1f}% sparse)"
+                        )
+                    else:
+                        print(
+                            f"Epoch {epoch+1:4d}: "
+                            f"mean_error={mean_rel_err:.1%}, "
+                            f"max_error={max_rel_err:.1%}, "
+                            f"total_loss={actual_total_loss:.3f}, "
+                            f"active={active_info['count']:4d}/{self.n_features} ({sparsity_pct:.1f}% sparse)"
+                        )
+
+        return self
+
+    def predict(self, M: sp.spmatrix) -> torch.Tensor:
+        """
+        Apply calibration weights to new data.
+
+        Parameters
+        ----------
+        M : scipy.sparse matrix
+            Metric matrix of shape (n_targets, n_features)
+
+        Returns
+        -------
+        torch.Tensor
+            Calibrated predictions
+        """
+        with torch.no_grad():
+            return self.forward(M, deterministic=True)

l0/sparse.py

@@ -0,0 +1,321 @@
+"""
+L0 regularized linear models for sparse data.
+
+Implements L0-regularized linear regression that efficiently handles
+sparse input matrices without converting them to dense format.
+"""
+
+# ruff: noqa: N803 N806
+
+
+import numpy as np
+import torch
+import torch.nn as nn
+from scipy import sparse as sp
+
+
+class SparseL0Linear(nn.Module):
+    """
+    L0-regularized linear model for sparse input data.
+
+    Efficiently handles sparse matrices (scipy.sparse) without
+    converting to dense format during computation.
+
+    Parameters
+    ----------
+    n_features : int
+        Number of input features
+    fit_intercept : bool
+        Whether to fit an intercept term
+    beta : float
+        Temperature parameter for Hard Concrete distribution
+    gamma : float
+        Lower bound of stretched concrete distribution
+    zeta : float
+        Upper bound of stretched concrete distribution
+    init_keep_prob : float
+        Initial probability of keeping each feature
+    device : str or torch.device
+        Device to run computations on ('cpu' or 'cuda')
+    """
+
+    def __init__(
+        self,
+        n_features: int,
+        fit_intercept: bool = True,
+        beta: float = 2 / 3,
+        gamma: float = -0.1,
+        zeta: float = 1.1,
+        init_keep_prob: float = 0.5,
+        device: str | torch.device = "cpu",
+    ):
+        super().__init__()
+        self.n_features = n_features
+        self.fit_intercept = fit_intercept
+        self.beta = beta
+        self.gamma = gamma
+        self.zeta = zeta
+        self.device = torch.device(device)
+
+        # Model parameters
+        self.weight = nn.Parameter(torch.zeros(n_features, device=self.device))
+        if fit_intercept:
+            self.bias = nn.Parameter(torch.zeros(1, device=self.device))
+        else:
+            self.register_parameter("bias", None)
+
+        # L0 gate parameters
+        mu = torch.log(torch.tensor(init_keep_prob / (1 - init_keep_prob)))
+        self.log_alpha = nn.Parameter(
+            torch.normal(
+                mu.item(), 0.01, size=(n_features,), device=self.device
+            )
+        )
+
+        # Cache for sparse tensor conversion
+        self._cached_X_torch: torch.sparse.Tensor | None = None
+        self._cached_X_shape: tuple[int, int] | None = None
+
+    def _convert_sparse_to_torch(
+        self, X_sparse: sp.spmatrix
+    ) -> torch.sparse.Tensor:
+        """
+        Convert scipy sparse matrix to torch sparse tensor.
+
+        Caches the result if the shape matches to avoid redundant conversions.
+        """
+        if (
+            self._cached_X_torch is not None
+            and self._cached_X_shape == X_sparse.shape
+            and X_sparse.nnz == self._cached_X_torch._nnz()
+        ):
+            return self._cached_X_torch
+
+        X_coo = X_sparse.tocoo()
+        indices = torch.LongTensor(np.vstack([X_coo.row, X_coo.col])).to(
+            self.device
+        )
+        values = torch.FloatTensor(X_coo.data).to(self.device)
+        X_torch = torch.sparse_coo_tensor(
+            indices,
+            values,
+            X_sparse.shape,
+            dtype=torch.float32,
+            device=self.device,
+        )
+
+        # Cache for future use
+        self._cached_X_torch = X_torch
+        self._cached_X_shape = X_sparse.shape
+
+        return X_torch
+
+    def _sample_gates(self) -> torch.Tensor:
+        """Sample gates using Hard Concrete distribution."""
+        eps = 1e-6
+        u = torch.rand_like(self.log_alpha).clamp(eps, 1 - eps)
+        X = (torch.log(u) - torch.log(1 - u) + self.log_alpha) / self.beta
+        s = torch.sigmoid(X)
+        s_bar = s * (self.zeta - self.gamma) + self.gamma
+        return s_bar.clamp(0, 1)
+
+    def get_deterministic_gates(self) -> torch.Tensor:
+        """Get deterministic gate values (for inference)."""
+        X = self.log_alpha / self.beta
+        s = torch.sigmoid(X)
+        s_bar = s * (self.zeta - self.gamma) + self.gamma
+        return s_bar.clamp(0, 1)
+
+    def forward(
+        self,
+        X: torch.sparse.Tensor | sp.spmatrix,
+        deterministic: bool = False,
+    ) -> torch.Tensor:
+        """
+        Forward pass maintaining sparse computation.
+
+        Parameters
+        ----------
+        X : torch.sparse.Tensor or scipy.sparse matrix
+            Sparse input tensor of shape (n_samples, n_features)
+        deterministic : bool
+            Whether to use deterministic gates (for inference)
+
+        Returns
+        -------
+        torch.Tensor
+            Predictions of shape (n_samples,)
+        """
+        # Convert scipy sparse to torch if needed
+        if isinstance(X, sp.spmatrix):
+            X = self._convert_sparse_to_torch(X)
+
+        # Sample or get deterministic gates
+        if deterministic:
+            gates = self.get_deterministic_gates()
+        else:
+            gates = self._sample_gates()
+
+        # Apply gates to weights
+        gated_weight = self.weight * gates
+
+        # Sparse matrix multiplication
+        # X is (n_samples, n_features), gated_weight is (n_features,)
+        y = torch.sparse.mm(X, gated_weight.unsqueeze(1)).squeeze(1)
+
+        # Add bias if needed
+        if self.fit_intercept and self.bias is not None:
+            y = y + self.bias
+
+        return y  # type: ignore[no-any-return]
+
+    def get_l0_penalty(self) -> torch.Tensor:
+        """
+        Compute L0 complexity penalty.
+
+        Returns expected number of active parameters.
+        """
+        c = -self.beta * torch.log(
+            torch.tensor(-self.gamma / self.zeta, device=self.device)
+        )
+        pi = torch.sigmoid(self.log_alpha + c)
+        return pi.sum()
+
+    def get_sparsity(self) -> float:
+        """
+        Get current sparsity level.
+
+        Returns
+        -------
+        float
+            Fraction of weights that are effectively zero
+        """
+        with torch.no_grad():
+            gates = self.get_deterministic_gates()
+            return (gates < 0.01).float().mean().item()
+
+    def get_selected_features(self, threshold: float = 0.01) -> torch.Tensor:
+        """
+        Get indices of selected (non-zero) features.
+
+        Parameters
+        ----------
+        threshold : float
+            Gate values below this are considered zero
+
+        Returns
+        -------
+        torch.Tensor
+            Indices of selected features
+        """
+        with torch.no_grad():
+            gates = self.get_deterministic_gates()
+            return torch.where(gates > threshold)[0]
+
+    def fit(
+        self,
+        X_sparse: sp.spmatrix,
+        y: torch.Tensor | np.ndarray,
+        lambda_reg: float = 0.01,
+        lr: float = 0.01,
+        epochs: int = 1000,
+        verbose: bool = False,
+        verbose_freq: int = 100,
+    ) -> "SparseL0Linear":
+        """
+        Fit the model using gradient descent.
+
+        Parameters
+        ----------
+        X_sparse : scipy.sparse matrix
+            Sparse input features
+        y : torch.Tensor or numpy.ndarray
+            Target values
+        lambda_reg : float
+            L0 regularization strength
+        lr : float
+            Learning rate
+        epochs : int
+            Number of training epochs
+        verbose : bool
+            Whether to print progress
+        verbose_freq : int
+            How often to print progress
+
+        Returns
+        -------
+        self
+            Fitted model
+        """
+        import numpy as np
+
+        # Convert y to tensor if needed
+        if isinstance(y, np.ndarray):
+            y = torch.tensor(y, dtype=torch.float32, device=self.device)
+        elif not isinstance(y, torch.Tensor):
+            y = torch.tensor(y, dtype=torch.float32, device=self.device)
+        else:
+            y = y.to(self.device)
+
+        # Convert sparse matrix to torch sparse tensor (will be cached)
+        X_torch = self._convert_sparse_to_torch(X_sparse)
+
+        # Initialize weights with small random values
+        nn.init.normal_(self.weight, 0, 0.1)
+        if self.fit_intercept and self.bias is not None:
+            self.bias.data.fill_(y.mean())
+
+        # Setup optimizer
+        params = [self.weight, self.log_alpha]
+        if self.fit_intercept and self.bias is not None:
+            params.append(self.bias)
+        optimizer = torch.optim.Adam(params, lr=lr)
+
+        # Training loop
+        for epoch in range(epochs):
+            # Forward pass
+            y_pred = self.forward(X_torch, deterministic=False)
+
+            # Compute loss
+            mse_loss = (y - y_pred).pow(2).mean()
+            l0_loss = self.get_l0_penalty()
+            loss = mse_loss + lambda_reg * l0_loss
+
+            # Backward pass
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+
+            # Print progress
+            if verbose and (epoch + 1) % verbose_freq == 0:
+                with torch.no_grad():
+                    n_active = (self.get_deterministic_gates() > 0.01).sum()
+                    print(
+                        f"Epoch {epoch+1:4d}: "
+                        f"loss={loss.item():.4f}, "
+                        f"mse={mse_loss.item():.4f}, "
+                        f"l0={l0_loss.item():.2f}, "
+                        f"active={n_active.item()}"
+                    )
+
+        return self
+
+    def predict(
+        self, X_sparse: sp.spmatrix | torch.sparse.Tensor
+    ) -> torch.Tensor:
+        """
+        Make predictions on new data.
+
+        Parameters
+        ----------
+        X_sparse : scipy.sparse matrix or torch.sparse.Tensor
+            Sparse input features
+
+        Returns
+        -------
+        torch.Tensor
+            Predictions
+        """
+        # Make deterministic predictions
+        with torch.no_grad():
+            return self.forward(X_sparse, deterministic=True)

{l0_python-0.1.0.dist-info → l0_python-0.3.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: l0-python
-Version: 0.1.0
+Version: 0.3.0
 Summary: L0 regularization for sparse neural networks and intelligent sampling
 Author-email: PolicyEngine <hello@policyengine.org>
 License: MIT License
@@ -41,14 +41,15 @@ Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: torch>=2.0.0
 Requires-Dist: numpy>=1.20.0
+Requires-Dist: scipy>=1.7.0
 Provides-Extra: dev
 Requires-Dist: pytest>=7.0; extra == "dev"
 Requires-Dist: pytest-cov>=3.0; extra == "dev"
 Requires-Dist: black>=22.0; extra == "dev"
-Requires-Dist: mypy>=0.900; extra == "dev"
-Requires-Dist: ruff>=0.0.280; extra == "dev"
-Requires-Dist: pre-commit>=3.0; extra == "dev"
 Requires-Dist: build>=1.3; extra == "dev"
+Requires-Dist: scikit-learn>=1.0; extra == "dev"
+Requires-Dist: yaml-changelog>=0.3.0; extra == "dev"
+Requires-Dist: twine>=4.0.0; extra == "dev"
 Provides-Extra: docs
 Requires-Dist: mystmd>=1.3; extra == "docs"
 Requires-Dist: jupyter>=1.0; extra == "docs"

l0_python-0.3.0.dist-info/RECORD

@@ -0,0 +1,12 @@
+l0/__init__.py,sha256=ezGVLGbdU_jAElAJSdLKjSWCnZRILAZi8qMBPjcHwQc,1192
+l0/calibration.py,sha256=H-5bo0SOCQXNBFuAQCzfBxw9-j4Ta0DcmJ66Hv-3Qtc,14860
+l0/distributions.py,sha256=GvVoDasdxoDlPPdzt33lFzkuJgQsfkEn9e-abxru3nM,6325
+l0/gates.py,sha256=54OORMg9qSsdbIyzYIL9t9CCHWsKbumSKl_d4-YFuNU,11443
+l0/layers.py,sha256=j_DPEFdfUc37ojtoK2ufGzVYzZ_DwusqTcJXCVxZS68,13785
+l0/penalties.py,sha256=okqXgX2eGUmg5prcDclicTQALPPPF9FA5eQP09V4GTE,9569
+l0/sparse.py,sha256=UVpkrrJBnQ9iSNDoWK5F_sr43tQareWXzYHyMq_cKss,9726
+l0_python-0.3.0.dist-info/licenses/LICENSE,sha256=LoyFJ8qmpwWPp5RmxFGHhdBxsemvp_BqGyiCgNAZHlk,1069
+l0_python-0.3.0.dist-info/METADATA,sha256=sBDt2EahzEyNWSBsufl0geYWY5mwq9JMkSTVGppjPkg,6833
+l0_python-0.3.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+l0_python-0.3.0.dist-info/top_level.txt,sha256=brP2SoElgmGhNJk-qRFEQZ5UE-d1e87Tr60hSSh8n34,3
+l0_python-0.3.0.dist-info/RECORD,,

l0_python-0.1.0.dist-info/RECORD

@@ -1,10 +0,0 @@
-l0/__init__.py,sha256=meL6jJ1CxO6pO373nY-nwGXnpzWZGC1Ts6q2TLyAeLM,1122
-l0/distributions.py,sha256=GvVoDasdxoDlPPdzt33lFzkuJgQsfkEn9e-abxru3nM,6325
-l0/gates.py,sha256=54OORMg9qSsdbIyzYIL9t9CCHWsKbumSKl_d4-YFuNU,11443
-l0/layers.py,sha256=j_DPEFdfUc37ojtoK2ufGzVYzZ_DwusqTcJXCVxZS68,13785
-l0/penalties.py,sha256=okqXgX2eGUmg5prcDclicTQALPPPF9FA5eQP09V4GTE,9569
-l0_python-0.1.0.dist-info/licenses/LICENSE,sha256=LoyFJ8qmpwWPp5RmxFGHhdBxsemvp_BqGyiCgNAZHlk,1069
-l0_python-0.1.0.dist-info/METADATA,sha256=hrAugZ-F7oKmleyJ8WjVTykRixBmuTiQ2M4oae1Ceu4,6794
-l0_python-0.1.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-l0_python-0.1.0.dist-info/top_level.txt,sha256=brP2SoElgmGhNJk-qRFEQZ5UE-d1e87Tr60hSSh8n34,3
-l0_python-0.1.0.dist-info/RECORD,,