scitex 2.0.0__py2.py3-none-any.whl → 2.1.0__py2.py3-none-any.whl

scitex/ml/feature_selection/feature_selection.py

@@ -0,0 +1,364 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# Timestamp: "2025-10-03 04:30:00 (ywatanabe)"
+# File: feature_selection.py
+
+"""
+Feature selection utilities for classification tasks.
+
+This module provides comprehensive feature selection and importance analysis:
+- Feature importance extraction from various model types
+- Univariate feature selection (ANOVA F-test, chi2, mutual_info)
+- Model-based feature selection (tree importances, L1 coefficients)
+- Recursive feature elimination (RFE)
+- Cross-fold feature consistency analysis
+- Feature importance aggregation and visualization
+"""
+
+from __future__ import annotations
+
+import warnings
+from typing import Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import pandas as pd
+from sklearn.feature_selection import (
+    SelectKBest,
+    chi2,
+    f_classif,
+    mutual_info_classif,
+)
+from sklearn.impute import SimpleImputer
+
+
+def extract_feature_importance(
+    model,
+    feature_names: List[str],
+    method: str = "auto",
+) -> Optional[Dict[str, float]]:
+    """Extract feature importance from trained model.
+
+    Args:
+        model: Trained classifier
+        feature_names: List of feature names
+        method: Method to extract importance:
+            - "auto": Automatically detect best method
+            - "tree": Use feature_importances_ (tree-based models)
+            - "coef": Use coefficients (linear models)
+            - "permutation": Use permutation importance (any model)
+
+    Returns:
+        Dictionary mapping feature names to importance values,
+        or None if extraction fails
+    """
+    try:
+        # Auto-detect method
+        if method == "auto":
+            if hasattr(model, "feature_importances_"):
+                method = "tree"
+            elif hasattr(model, "coef_"):
+                method = "coef"
+            else:
+                warnings.warn(
+                    f"Model {type(model).__name__} does not support "
+                    f"feature importance extraction"
+                )
+                return None
+
+        # Extract importance
+        if method == "tree":
+            if not hasattr(model, "feature_importances_"):
+                raise ValueError(
+                    f"Model {type(model).__name__} does not have "
+                    f"feature_importances_ attribute"
+                )
+            importances = model.feature_importances_
+
+        elif method == "coef":
+            if not hasattr(model, "coef_"):
+                raise ValueError(
+                    f"Model {type(model).__name__} does not have "
+                    f"coef_ attribute"
+                )
+            # Use absolute values for multi-class or single coefficient vector
+            coef = model.coef_
+            if coef.ndim > 1:
+                # Multi-class: average absolute coefficients across classes
+                importances = np.mean(np.abs(coef), axis=0)
+            else:
+                importances = np.abs(coef)
+
+        elif method == "permutation":
+            raise NotImplementedError(
+                "Permutation importance requires validation data. "
+                "Use extract_permutation_importance() instead."
+            )
+
+        else:
+            raise ValueError(f"Unknown method: {method}")
+
+        # Normalize to sum to 1
+        importances = importances / importances.sum()
+
+        # Create dictionary sorted by importance
+        importance_dict = dict(zip(feature_names, importances))
+        importance_dict = dict(
+            sorted(importance_dict.items(), key=lambda x: x[1], reverse=True)
+        )
+
+        return importance_dict
+
+    except Exception as e:
+        warnings.warn(f"Failed to extract feature importance: {e}")
+        return None
+
+
+def select_features_univariate(
+    X_train: np.ndarray,
+    y_train: np.ndarray,
+    X_val: np.ndarray,
+    feature_names: List[str],
+    k: int = 10,
+    score_func: str = "f_classif",
+    impute_strategy: str = "median",
+) -> Tuple[np.ndarray, np.ndarray, List[int], List[str], object]:
+    """Select top k features using univariate statistical tests.
+
+    This prevents data leakage by:
+    1. Fitting the selector only on training data
+    2. Transforming validation/test data with the fitted selector
+
+    Args:
+        X_train: Training features
+        y_train: Training labels
+        X_val: Validation features
+        feature_names: List of feature names
+        k: Number of features to select
+        score_func: Scoring function:
+            - "f_classif": ANOVA F-test (default)
+            - "chi2": Chi-squared test (requires non-negative features)
+            - "mutual_info": Mutual information
+        impute_strategy: Strategy for imputing missing values:
+            - "median" (default), "mean", "most_frequent", "constant"
+
+    Returns:
+        X_train_selected: Selected training features
+        X_val_selected: Selected validation features
+        feature_indices: Indices of selected features
+        selected_names: Names of selected features
+        imputer: Fitted imputer for test data
+    """
+    # Map string to function
+    score_func_map = {
+        "f_classif": f_classif,
+        "chi2": chi2,
+        "mutual_info": mutual_info_classif,
+    }
+
+    if score_func not in score_func_map:
+        raise ValueError(
+            f"Unknown score_func: {score_func}. "
+            f"Choose from {list(score_func_map.keys())}"
+        )
+
+    score_func_callable = score_func_map[score_func]
+
+    # Impute NaN values
+    imputer = SimpleImputer(strategy=impute_strategy)
+    X_train_imputed = imputer.fit_transform(X_train)
+    X_val_imputed = imputer.transform(X_val)
+
+    # Select features
+    k_actual = min(k, X_train.shape[1])
+    selector = SelectKBest(score_func_callable, k=k_actual)
+    X_train_selected = selector.fit_transform(X_train_imputed, y_train)
+    X_val_selected = selector.transform(X_val_imputed)
+
+    # Get selected feature information
+    feature_indices = selector.get_support(indices=True).tolist()
+    selected_names = [feature_names[i] for i in feature_indices]
+
+    return (
+        X_train_selected,
+        X_val_selected,
+        feature_indices,
+        selected_names,
+        imputer,
+    )
+
+
+def analyze_feature_consistency(
+    selected_features_per_fold: List[List[str]],
+) -> Dict[str, Union[int, float, Dict[str, int]]]:
+    """Analyze feature selection consistency across CV folds.
+
+    Args:
+        selected_features_per_fold: List of feature lists, one per fold
+
+    Returns:
+        Dictionary containing:
+        - "feature_frequency": Dict mapping features to selection count
+        - "n_folds": Total number of folds
+        - "n_unique_features": Number of unique features selected
+        - "consistency_score": Average selection frequency (0-1)
+        - "stable_features": Features selected in all folds
+        - "unstable_features": Features selected in only one fold
+    """
+    if not selected_features_per_fold:
+        return {}
+
+    n_folds = len(selected_features_per_fold)
+
+    # Count feature occurrences
+    all_features = set()
+    for features in selected_features_per_fold:
+        all_features.update(features)
+
+    feature_frequency = {f: 0 for f in all_features}
+    for features in selected_features_per_fold:
+        for f in features:
+            feature_frequency[f] += 1
+
+    # Sort by frequency
+    feature_frequency = dict(
+        sorted(feature_frequency.items(), key=lambda x: x[1], reverse=True)
+    )
+
+    # Identify stable and unstable features
+    stable_features = [
+        f for f, count in feature_frequency.items() if count == n_folds
+    ]
+    unstable_features = [
+        f for f, count in feature_frequency.items() if count == 1
+    ]
+
+    # Calculate consistency score (average selection frequency)
+    consistency_score = (
+        np.mean(list(feature_frequency.values())) / n_folds
+        if feature_frequency
+        else 0.0
+    )
+
+    return {
+        "feature_frequency": feature_frequency,
+        "n_folds": n_folds,
+        "n_unique_features": len(all_features),
+        "consistency_score": round(consistency_score, 3),
+        "stable_features": stable_features,
+        "unstable_features": unstable_features,
+    }
+
+
+def aggregate_feature_importances(
+    importances_per_fold: List[Dict[str, float]],
+    method: str = "mean",
+) -> Dict[str, Dict[str, float]]:
+    """Aggregate feature importances across CV folds.
+
+    Args:
+        importances_per_fold: List of importance dicts, one per fold
+        method: Aggregation method:
+            - "mean": Average importance across folds
+            - "median": Median importance across folds
+            - "max": Maximum importance across folds
+
+    Returns:
+        Dictionary containing:
+        - "mean": Mean importance per feature
+        - "std": Standard deviation per feature
+        - "min": Minimum importance per feature
+        - "max": Maximum importance per feature
+        - "cv": Coefficient of variation (std/mean) per feature
+    """
+    if not importances_per_fold:
+        return {}
+
+    # Collect all features
+    all_features = set()
+    for imp_dict in importances_per_fold:
+        all_features.update(imp_dict.keys())
+
+    # Build matrix: (n_folds, n_features)
+    importance_matrix = np.zeros((len(importances_per_fold), len(all_features)))
+    feature_list = sorted(list(all_features))
+
+    for fold_idx, imp_dict in enumerate(importances_per_fold):
+        for feat_idx, feat in enumerate(feature_list):
+            importance_matrix[fold_idx, feat_idx] = imp_dict.get(feat, 0.0)
+
+    # Calculate statistics
+    mean_importance = importance_matrix.mean(axis=0)
+    std_importance = importance_matrix.std(axis=0)
+    min_importance = importance_matrix.min(axis=0)
+    max_importance = importance_matrix.max(axis=0)
+
+    # Coefficient of variation (handle division by zero)
+    with np.errstate(divide="ignore", invalid="ignore"):
+        cv_importance = std_importance / mean_importance
+        cv_importance[~np.isfinite(cv_importance)] = 0.0
+
+    # Create result dictionary
+    result = {
+        "mean": dict(zip(feature_list, mean_importance)),
+        "std": dict(zip(feature_list, std_importance)),
+        "min": dict(zip(feature_list, min_importance)),
+        "max": dict(zip(feature_list, max_importance)),
+        "cv": dict(zip(feature_list, cv_importance)),
+    }
+
+    # Sort each by mean importance
+    for key in result:
+        result[key] = dict(
+            sorted(
+                result[key].items(),
+                key=lambda x: result["mean"][x[0]],
+                reverse=True,
+            )
+        )
+
+    return result
+
+
+def create_feature_importance_dataframe(
+    aggregated_importances: Dict[str, Dict[str, float]],
+) -> pd.DataFrame:
+    """Create a formatted DataFrame from aggregated feature importances.
+
+    Args:
+        aggregated_importances: Output from aggregate_feature_importances()
+
+    Returns:
+        DataFrame with columns: feature, mean, std, min, max, cv
+        Sorted by mean importance (descending)
+    """
+    if not aggregated_importances or "mean" not in aggregated_importances:
+        return pd.DataFrame()
+
+    features = list(aggregated_importances["mean"].keys())
+
+    df = pd.DataFrame(
+        {
+            "feature": features,
+            "mean": [aggregated_importances["mean"][f] for f in features],
+            "std": [aggregated_importances["std"][f] for f in features],
+            "min": [aggregated_importances["min"][f] for f in features],
+            "max": [aggregated_importances["max"][f] for f in features],
+            "cv": [aggregated_importances["cv"][f] for f in features],
+        }
+    )
+
+    # Sort by mean importance
+    df = df.sort_values("mean", ascending=False).reset_index(drop=True)
+
+    return df
+
+
+__all__ = [
+    "extract_feature_importance",
+    "select_features_univariate",
+    "analyze_feature_consistency",
+    "aggregate_feature_importances",
+    "create_feature_importance_dataframe",
+]
+
+# EOF

scitex/ml/loss/_L1L2Losses.py

@@ -0,0 +1,34 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# Time-stamp: "2024-11-07 18:53:03 (ywatanabe)"
+# File: ./scitex_repo/src/scitex/ai/loss/_L1L2Losses.py
+
+import torch
+
+
+def l1(model, lambda_l1=0.01):
+    lambda_l1 = torch.tensor(lambda_l1)
+    l1 = torch.tensor(0.0).cuda()
+    for param in model.parameters():  # fixme; is this OK?
+        l1 += torch.abs(param).sum()
+    return l1
+
+
+def l2(model, lambda_l2=0.01):
+    lambda_l2 = torch.tensor(lambda_l2)
+    l2 = torch.tensor(0.0).cuda()
+    for param in model.parameters():  # fixme; is this OK?
+        l2 += torch.norm(param).sum()
+    return l2
+
+
+def elastic(model, alpha=1.0, l1_ratio=0.5):
+    assert 0 <= l1_ratio <= 1
+
+    L1 = l1(model)
+    L2 = l2(model)
+
+    return alpha * (l1_ratio * L1 + (1 - l1_ratio) * L2)
+
+
+# EOF

scitex/ml/loss/__init__.py

@@ -0,0 +1,12 @@
+#!/usr/bin/env python3
+"""Scitex loss module."""
+
+from ._L1L2Losses import elastic, l1, l2
+from .multi_task_loss import MultiTaskLoss
+
+__all__ = [
+    "elastic",
+    "l1",
+    "l2",
+    "MultiTaskLoss",
+]

scitex/ml/loss/multi_task_loss.py

@@ -0,0 +1,47 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# Time-stamp: "2024-11-07 19:07:29 (ywatanabe)"
+# File: ./scitex_repo/src/scitex/ai/loss/MultiTaskLoss.py
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+from ...repro import fix_seeds
+
+
+class MultiTaskLoss(nn.Module):
+    """
+    # https://openaccess.thecvf.com/content_cvpr_2018/papers/Kendall_Multi-Task_Learning_Using_CVPR_2018_paper.pdf
+
+    Example:
+        are_regression = [False, False]
+        mtl = MultiTaskLoss(are_regression)
+        losses = [torch.rand(1, requires_grad=True) for _ in range(len(are_regression))]
+        loss = mtl(losses)
+        print(loss)
+        # [tensor([0.4215], grad_fn=<AddBackward0>), tensor([0.6190], grad_fn=<AddBackward0>)]
+    """
+
+    def __init__(self, are_regression=[False, False], reduction="none"):
+        super().__init__()
+        fix_seeds(np=np, torch=torch, show=False)
+        n_tasks = len(are_regression)
+
+        self.register_buffer("are_regression", torch.tensor(are_regression))
+
+        # for the numercal stability, log(variables) are learned.
+        self.log_vars = torch.nn.Parameter(torch.zeros(n_tasks))
+        self.reduction = reduction
+
+    def forward(self, losses):
+        vars = torch.exp(self.log_vars).type_as(losses[0])
+        stds = vars ** (1 / 2)
+        coeffs = 1 / ((self.are_regression + 1) * vars)
+        scaled_losses = [
+            coeffs[i] * losses[i] + torch.log(stds[i]) for i in range(len(losses))
+        ]
+        return scaled_losses
+
+
+# EOF

scitex/ml/metrics/__init__.py

@@ -0,0 +1,56 @@
+#!/usr/bin/env python3
+"""Scitex metrics module.
+
+Standardized naming convention:
+- calc_* functions: Modern standardized metric calculations
+- Legacy names (bACC, balanced_accuracy, etc.): For backward compatibility
+"""
+
+# Modern standardized calc_* functions
+from ._normalize_labels import normalize_labels as _normalize_labels
+from ._calc_bacc import calc_bacc
+from ._calc_mcc import calc_mcc
+from ._calc_conf_mat import calc_conf_mat
+from ._calc_clf_report import calc_clf_report
+from ._calc_roc_auc import calc_roc_auc
+from ._calc_pre_rec_auc import calc_pre_rec_auc
+from ._calc_bacc_from_conf_mat import calc_bacc_from_conf_mat
+from ._calc_seizure_prediction_metrics import (
+    calc_seizure_window_prediction_metrics,
+    calc_seizure_event_prediction_metrics,
+    calc_seizure_prediction_metrics,  # backward compat alias
+)
+from ._calc_silhouette_score import (
+    calc_silhouette_score_slow,
+    calc_silhouette_samples_slow,
+    calc_silhouette_score_block,
+    calc_silhouette_samples_block,
+)
+from ._calc_feature_importance import (
+    calc_feature_importance,
+    calc_permutation_importance,
+)
+
+__all__ = [
+    "calc_bacc",
+    "calc_mcc",
+    "calc_conf_mat",
+    "calc_clf_report",
+    "calc_roc_auc",
+    "calc_pre_rec_auc",
+    "calc_bacc_from_conf_mat",
+    "calc_seizure_window_prediction_metrics",
+    "calc_seizure_event_prediction_metrics",
+    "calc_seizure_prediction_metrics",  # backward compat alias
+    "calc_silhouette_score_slow",
+    "calc_silhouette_samples_slow",
+    "calc_silhouette_score_block",
+    "calc_silhouette_samples_block",
+    "calc_feature_importance",
+    "calc_permutation_importance",
+    # # Legacy names for backward compatibility
+    # "bACC",
+    # "balanced_accuracy",
+    # "bACC_from_conf_mat",
+    # "balanced_accuracy_from_conf_mat",
+]

scitex/ml/metrics/_calc_bacc.py

@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# Timestamp: "2025-10-02 (ywatanabe)"
+# File: /home/ywatanabe/proj/scitex_repo/src/scitex/ml/metrics/_calc_bacc.py
+
+"""Calculate balanced accuracy metric."""
+
+__FILE__ = __file__
+
+from typing import Any, Dict, List, Optional
+import numpy as np
+from sklearn.metrics import balanced_accuracy_score
+from ._normalize_labels import normalize_labels
+
+
+def calc_bacc(
+    y_true: np.ndarray,
+    y_pred: np.ndarray,
+    labels: Optional[List] = None,
+    fold: Optional[int] = None,
+) -> Dict[str, Any]:
+    """
+    Calculate balanced accuracy with robust label handling.
+
+    Parameters
+    ----------
+    y_true : np.ndarray
+        True labels (can be str or int)
+    y_pred : np.ndarray
+        Predicted labels (can be str or int)
+    labels : List, optional
+        Expected label list
+    fold : int, optional
+        Fold number for tracking
+
+    Returns
+    -------
+    Dict[str, Any]
+        {'metric': 'balanced_accuracy', 'value': float, 'fold': int}
+    """
+    try:
+        y_true_norm, y_pred_norm, label_names, _ = normalize_labels(
+            y_true, y_pred, labels
+        )
+        value = balanced_accuracy_score(y_true_norm, y_pred_norm)
+        return {
+            "metric": "balanced_accuracy",
+            "value": float(value),
+            "fold": fold,
+            "labels": label_names,
+        }
+    except Exception as e:
+        return {
+            "metric": "balanced_accuracy",
+            "value": np.nan,
+            "fold": fold,
+            "error": str(e),
+        }
+
+
+# EOF

scitex/ml/metrics/_calc_bacc_from_conf_mat.py

@@ -0,0 +1,38 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# Timestamp: "2025-10-02 (ywatanabe)"
+# File: /home/ywatanabe/proj/scitex_repo/src/scitex/ml/metrics/_calc_bacc_from_conf_mat.py
+
+"""Calculate balanced accuracy from confusion matrix."""
+
+__FILE__ = __file__
+
+import numpy as np
+
+
+def calc_bacc_from_conf_mat(cm: np.ndarray) -> float:
+    """
+    Calculate balanced accuracy from confusion matrix.
+
+    Parameters
+    ----------
+    cm : np.ndarray
+        Confusion matrix
+
+    Returns
+    -------
+    float
+        Balanced accuracy
+    """
+    try:
+        per_class = np.diag(cm) / np.sum(cm, axis=1)
+        return float(np.nanmean(per_class))
+    except:
+        return np.nan
+
+
+# Convenience aliases
+bACC_from_conf_mat = calc_bacc_from_conf_mat
+balanced_accuracy_from_conf_mat = calc_bacc_from_conf_mat
+
+# EOF

scitex/ml/metrics/_calc_clf_report.py

@@ -0,0 +1,78 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# Timestamp: "2025-10-02 (ywatanabe)"
+# File: /home/ywatanabe/proj/scitex_repo/src/scitex/ml/metrics/_calc_clf_report.py
+
+"""Generate classification report."""
+
+__FILE__ = __file__
+
+from typing import Any, Dict, List, Optional
+import pandas as pd
+from sklearn.metrics import classification_report
+from ._normalize_labels import normalize_labels
+
+
+def calc_clf_report(
+    y_true,
+    y_pred,
+    labels: Optional[List] = None,
+    fold: Optional[int] = None,
+) -> Dict[str, Any]:
+    """
+    Generate classification report with robust label handling.
+
+    Parameters
+    ----------
+    y_true : np.ndarray
+        True labels (can be str or int)
+    y_pred : np.ndarray
+        Predicted labels (can be str or int)
+    labels : List, optional
+        Expected label list
+    fold : int, optional
+        Fold number for tracking
+
+    Returns
+    -------
+    Dict[str, Any]
+        {
+            'metric': 'classification_report',
+            'value': pd.DataFrame,
+            'fold': int,
+            'labels': list
+        }
+    """
+    try:
+        y_true_norm, y_pred_norm, label_names, _ = normalize_labels(
+            y_true, y_pred, labels
+        )
+
+        # Get classification report
+        report_dict = classification_report(
+            y_true_norm,
+            y_pred_norm,
+            target_names=[str(l) for l in label_names],
+            output_dict=True,
+            zero_division=0,
+        )
+
+        # Convert to DataFrame
+        report_df = pd.DataFrame(report_dict).T
+
+        return {
+            "metric": "classification_report",
+            "value": report_df,
+            "fold": fold,
+            "labels": label_names,
+        }
+    except Exception as e:
+        return {
+            "metric": "classification_report",
+            "value": None,
+            "fold": fold,
+            "error": str(e),
+        }
+
+
+# EOF