wizata-dsapi 2.0.0.dev24__tar.gz → 2.0.0.dev26__tar.gz

{wizata_dsapi-2.0.0.dev24/wizata_dsapi.egg-info → wizata_dsapi-2.0.0.dev26}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: wizata_dsapi
-Version: 2.0.0.dev24
+Version: 2.0.0.dev26
 Summary: Wizata Data Science Toolkit
 Author: Wizata S.A.
 Author-email: info@wizata.com

wizata_dsapi-2.0.0.dev26/wizata_dsapi/models/__init__.py

@@ -0,0 +1 @@
+from .common import linear_regression, logistic_regression, isolation_forest, gradiant_boost_classifier, setpoint_optimizer, SetpointOptimizer, random_forest_regressor, hotelling_t2_monitor, HotellingT2Monitor, ridge_regression, random_forest_classifier, autoencoder_anomaly, AutoencoderAnomaly, mahalanobis_anomaly, MahalanobisAnomaly, gmm_regime_classifier, GMMRegimeClassifier

wizata_dsapi-2.0.0.dev26/wizata_dsapi/models/common.py

@@ -0,0 +1,681 @@
+import wizata_dsapi
+
+import pandas
+import numpy
+
+import sklearn
+import sklearn.linear_model
+import sklearn.ensemble
+import sklearn.mixture
+import sklearn.neighbors
+import sklearn.neural_network
+import sklearn.pipeline
+import sklearn.preprocessing
+import sklearn.decomposition
+import sklearn.covariance
+
+
+def extract_target_feat(context: wizata_dsapi.Context, single: bool = True):
+    """
+    return a list of target_feat columns names if not single value or the single value target feat name
+    raise an error if configuration mismatch
+    """
+
+    if "target_feat" not in context.properties:
+        raise ValueError(f"training script requires a proper target_feat")
+
+    target_feat = context.properties["target_feat"]
+    if isinstance(target_feat, str):
+        if single:
+            return target_feat
+        else:
+            return [target_feat]
+    elif isinstance(target_feat, list):
+        if single:
+            if len(target_feat) == 1:
+                return target_feat[0]
+            else:
+                raise ValueError(f"expecting only one target_feat but found {len(target_feat)}")
+        else:
+            return [target_feat]
+    else:
+        raise TypeError(f'target_feat must be a str or a list of str but found {target_feat.__class__.__name__}')
+
+
+def linear_regression(context: wizata_dsapi.Context):
+    """Train a linear regression model on all features to predict a single target column."""
+    df = context.dataframe
+
+    model_config = context.get_model_config()
+    if not model_config.has_target_feat():
+        raise ValueError(f'linear_regression requires a target feat')
+    target_feat_name = context.properties["target_feat"]
+
+    x = df.drop(columns=[target_feat_name])
+    y = df[target_feat_name]
+
+    model = sklearn.linear_model.LinearRegression()
+    model.fit(x, y)
+
+    context.set_model(model, features=x.columns)
+
+
+def logistic_regression(context: wizata_dsapi.Context):
+    """Train a logistic regression classifier on all features to predict a binary target column."""
+    df = context.dataframe
+
+    model_config = context.get_model_config()
+    if not model_config.has_target_feat():
+        raise ValueError(f'logistic_regression requires a target feat')
+    target_feat_name = context.properties["target_feat"]
+
+    x = df.drop(columns=[target_feat_name])
+    y = df[target_feat_name]
+
+    model = sklearn.linear_model.LogisticRegression()
+    model.fit(x, y.astype(int))
+
+    context.set_model(model, features=x.columns)
+
+
+def isolation_forest(context: wizata_dsapi.Context):
+    """Train an Isolation Forest for unsupervised anomaly detection using a sensitivity level (1-5)."""
+
+    model_config = context.get_model_config()
+    if model_config.has_target_feat():
+        raise ValueError(f'isolation_forest does not requires a target feat')
+
+    try:
+        if context.properties['sensitivity'] is None:
+            raise KeyError("sensitivity is none")
+        sensitivity = int(context.properties['sensitivity'])
+        sensitivities = [0.05, 0.15, 0.25, 0.35, 0.4]
+        contamination = sensitivities[sensitivity - 1]
+    except Exception as e:
+        raise ValueError(f'cannot extract sensitivity integer from 0 to 4 due to {e}')
+
+    df = context.dataframe.copy()
+    model = sklearn.ensemble.IsolationForest(contamination=contamination)
+    df['isolation_forest_predict'] = model.fit_predict(df)
+    context.set_model(model, features=df.columns)
+    return df
+
+
+def gradiant_boost_classifier(context: wizata_dsapi.Context):
+    """Train a Gradient Boosting classifier on all features to predict a target column."""
+    df = context.dataframe
+
+    model_config = context.get_model_config()
+    if not model_config.has_target_feat():
+        raise ValueError(f'gradiant_boost_classifier requires a target feat')
+    target_feat_name = context.properties["target_feat"]
+
+    x = df.drop(columns=[target_feat_name])
+    y = df[target_feat_name]
+
+    model = sklearn.ensemble.GradientBoostingClassifier(random_state=0).fit(x, y)
+    context.set_model(model, features=df.columns)
+
+
+class SetpointOptimizer:
+    """
+    Wraps a KNN quality forecaster (StandardScaler + KNeighborsRegressor) with a grid-search
+    setpoint recommendation method.
+
+    At training time, the underlying pipeline learns quality = f(telemetry + setpoints) from
+    historical data. At inference, recommend(X) keeps each row's telemetry fixed and grid-searches
+    over stored setpoint bounds (5th-95th percentile of train data) to find the combination that
+    minimizes (or maximizes) predicted quality.
+
+    :ivar pipeline: fitted sklearn Pipeline (StandardScaler -> KNeighborsRegressor).
+    :ivar setpoint_cols: ordered list of setpoint column names auto-detected at train time.
+    :ivar feature_cols: ordered list of all feature columns used at training (telemetry + setpoints).
+    :ivar bounds: dict mapping each setpoint column name to a (low, high) tuple.
+    :ivar direction: 'minimize' or 'maximize' the target quality.
+    :ivar grid_size: number of points per setpoint axis in the grid search.
+    """
+
+    def __init__(self, pipeline, setpoint_cols, feature_cols, bounds, direction, grid_size):
+        self.pipeline = pipeline
+        self.setpoint_cols = list(setpoint_cols)
+        self.feature_cols = list(feature_cols)
+        self.bounds = dict(bounds)
+        self.direction = direction
+        self.grid_size = int(grid_size)
+
+    def predict(self, X):
+        """Return predicted quality for each row — dual-use for validation or regular predict-mode pipelines."""
+        return self.pipeline.predict(X)
+
+    def recommend(self, X):
+        """
+        For each row of X, return the grid-search best setpoint combination.
+        :param X: features matrix (DataFrame or ndarray) with columns matching feature_cols.
+        :return: ndarray of shape [n_rows, n_setpoints] in the order of self.setpoint_cols.
+        """
+        if not isinstance(X, pandas.DataFrame):
+            X = pandas.DataFrame(X, columns=self.feature_cols)
+
+        grids = [
+            numpy.linspace(self.bounds[sp][0], self.bounds[sp][1], self.grid_size)
+            for sp in self.setpoint_cols
+        ]
+        mesh = numpy.array(numpy.meshgrid(*grids)).reshape(len(self.setpoint_cols), -1).T
+
+        recs = numpy.zeros((len(X), len(self.setpoint_cols)))
+        for i, (_, row) in enumerate(X.iterrows()):
+            candidates = pandas.DataFrame(
+                numpy.tile(row.values, (len(mesh), 1)),
+                columns=self.feature_cols
+            )
+            for j, sp in enumerate(self.setpoint_cols):
+                candidates[sp] = mesh[:, j]
+            preds = self.pipeline.predict(candidates)
+            if self.direction == "maximize":
+                best_idx = int(numpy.argmax(preds))
+            else:
+                best_idx = int(numpy.argmin(preds))
+            recs[i] = mesh[best_idx]
+
+        return recs
+
+    def output_names(self, suffix: str = "_recommended"):
+        """Suggest MLModelConfig.output_columns_names matching the setpoint order (e.g. for UI prefill)."""
+        return [f"{sp}{suffix}" for sp in self.setpoint_cols]
+
+    def get_inference_contract(self):
+        """Self-describing inference contract — the platform reads this to override MLModelConfig at runtime.
+        Returns the function to invoke (.recommend) and the output column names that will be produced, in order."""
+        return {
+            "function": "recommend",
+            "output_columns_names": self.output_names(),
+        }
+
+
+def setpoint_optimizer(context: wizata_dsapi.Context):
+    """Train a KNN-based setpoint optimizer that learns quality = f(telemetry + setpoints) and at
+    inference recommends optimal setpoint values. Setpoint columns are auto-detected via
+    BusinessType.SET_POINTS on context.datapoints; bounds are the 5th-95th percentile of train data.
+
+    Required MLModelConfig:
+      - train_script = 'wizata.models.setpoint_optimizer'
+      - target_feat  = '<quality column name>'
+      - function     = 'recommend'  (or 'predict' for quality forecasting only)
+      - output_columns_names = ['<sp1>_recommended', '<sp2>_recommended', ...] in the order setpoints appear.
+        The trained model exposes `.output_names()` to suggest a matching default.
+
+    Properties:
+      - k: KNN neighbors (default 5)
+      - grid_size: points per setpoint axis (default 10 — total cost is grid_size^n_setpoints per row)
+      - direction: 'minimize' (default) or 'maximize' the target
+    """
+    df = context.dataframe
+
+    model_config = context.get_model_config()
+    if not model_config.has_target_feat():
+        raise ValueError(f'setpoint_optimizer requires a target_feat (the quality column to optimize)')
+
+    target = context.properties["target_feat"]
+    if isinstance(target, list):
+        if len(target) != 1:
+            raise ValueError(f'setpoint_optimizer requires exactly one target_feat column')
+        target = target[0]
+
+    if target not in df.columns:
+        raise ValueError(f"target_feat '{target}' not found in dataframe columns")
+
+    datapoints = context.datapoints or {}
+    setpoint_cols = [
+        col for col, dp in datapoints.items()
+        if col in df.columns
+        and col != target
+        and dp.business_type == wizata_dsapi.BusinessType.SET_POINTS
+    ]
+
+    if not setpoint_cols:
+        raise ValueError(
+            "no setpoint datapoints found in context (BusinessType.SET_POINTS) — "
+            "the optimizer needs at least one setpoint column to optimize"
+        )
+
+    k = int(context.properties.get("k", 5))
+    grid_size = int(context.properties.get("grid_size", 10))
+    direction = context.properties.get("direction", "minimize")
+    if direction not in ("minimize", "maximize"):
+        raise ValueError(f"direction must be 'minimize' or 'maximize', got '{direction}'")
+
+    x = df.drop(columns=[target])
+    y = df[target]
+
+    pipeline = sklearn.pipeline.Pipeline([
+        ("scaler", sklearn.preprocessing.StandardScaler()),
+        ("knn", sklearn.neighbors.KNeighborsRegressor(n_neighbors=k)),
+    ])
+    pipeline.fit(x, y)
+
+    bounds = {
+        sp: (float(x[sp].quantile(0.05)), float(x[sp].quantile(0.95)))
+        for sp in setpoint_cols
+    }
+
+    optimizer = SetpointOptimizer(
+        pipeline=pipeline,
+        setpoint_cols=setpoint_cols,
+        feature_cols=list(x.columns),
+        bounds=bounds,
+        direction=direction,
+        grid_size=grid_size,
+    )
+
+    context.set_model(optimizer, features=x.columns)
+
+
+def random_forest_regressor(context: wizata_dsapi.Context):
+    """Train a Random Forest Regressor on all features to predict a single target column — non-linear,
+    robust to outliers, no scaling required. Exposes feature_importances_ on the trained model
+    for interpretability (root-cause hints).
+
+    Properties:
+      - n_estimators: number of trees (default 100)
+      - max_depth: max depth per tree (default None = unbounded)
+      - random_state: RNG seed for reproducibility (default 0)
+    """
+    df = context.dataframe
+
+    model_config = context.get_model_config()
+    if not model_config.has_target_feat():
+        raise ValueError(f'random_forest_regressor requires a target_feat')
+    target_feat_name = context.properties["target_feat"]
+
+    x = df.drop(columns=[target_feat_name])
+    y = df[target_feat_name]
+
+    n_estimators = int(context.properties.get("n_estimators", 100))
+    max_depth_prop = context.properties.get("max_depth")
+    max_depth = int(max_depth_prop) if max_depth_prop not in (None, "", "None") else None
+    random_state = int(context.properties.get("random_state", 0))
+
+    model = sklearn.ensemble.RandomForestRegressor(
+        n_estimators=n_estimators,
+        max_depth=max_depth,
+        random_state=random_state,
+    )
+    model.fit(x, y)
+
+    context.set_model(model, features=x.columns)
+
+
+class HotellingT2Monitor:
+    """
+    Multivariate statistical process control (SPC) anomaly monitor.
+
+    Fits a StandardScaler + PCA on training data (assumed to represent normal operation), then at
+    inference computes two classic industrial anomaly statistics per observation:
+
+      - T² (Hotelling's T²): sum of squared PC scores divided by eigenvalues — measures anomaly
+        *within* the principal subspace (drift of the operating regime while still explained by
+        the learned model).
+      - SPE / Q statistic: squared reconstruction error — measures anomaly *orthogonal* to the
+        PC space (novel behaviour not captured by the learned model).
+
+    Both statistics together enable decomposition of what is abnormal and why.
+
+    :ivar scaler: fitted sklearn StandardScaler.
+    :ivar pca: fitted sklearn PCA.
+    :ivar feature_cols: ordered list of feature columns used at training.
+    """
+
+    def __init__(self, scaler, pca, feature_cols):
+        self.scaler = scaler
+        self.pca = pca
+        self.feature_cols = list(feature_cols)
+
+    def _compute(self, X):
+        if not isinstance(X, pandas.DataFrame):
+            X = pandas.DataFrame(X, columns=self.feature_cols)
+        X_scaled = self.scaler.transform(X)
+        scores = self.pca.transform(X_scaled)
+        eigenvalues = self.pca.explained_variance_
+        # T² = Σ (tᵢ² / λᵢ) per row
+        t2 = (scores ** 2 / eigenvalues).sum(axis=1)
+        # SPE = ||x − x̂||² per row
+        reconstructed = self.pca.inverse_transform(scores)
+        spe = ((X_scaled - reconstructed) ** 2).sum(axis=1)
+        return t2, spe
+
+    def predict(self, X):
+        """Return the T² statistic per row (1-D) — dual-use with the default predict interface."""
+        t2, _ = self._compute(X)
+        return t2
+
+    def monitor(self, X):
+        """Return [n_rows, 2] array with T² and SPE per row."""
+        t2, spe = self._compute(X)
+        return numpy.column_stack([t2, spe])
+
+    def get_inference_contract(self):
+        """Self-describing contract — platform reads this to override MLModelConfig at runtime."""
+        return {
+            "function": "monitor",
+            "output_columns_names": ["T2", "SPE"],
+        }
+
+
+def hotelling_t2_monitor(context: wizata_dsapi.Context):
+    """Train a PCA-based multivariate SPC monitor on training data representative of *normal*
+    operation. At inference, produces two columns per row: T² (Hotelling's statistic — anomaly
+    within the principal subspace) and SPE (squared prediction error — novelty orthogonal to the
+    PC space). The classical industrial monitoring pattern — far more interpretable than a black-box
+    outlier score because T² and SPE can be decomposed back to per-sensor contributions.
+
+    Unsupervised: do **not** set target_feat. NaN values must be handled upstream.
+
+    Property:
+      - n_components: int (exact number of components) or float in (0, 1] (minimum explained
+        variance ratio, default 0.95).
+    """
+    df = context.dataframe
+
+    model_config = context.get_model_config()
+    if model_config.has_target_feat():
+        raise ValueError(f'hotelling_t2_monitor is unsupervised — do not set target_feat')
+
+    if df.isna().any().any():
+        raise ValueError(f'hotelling_t2_monitor cannot handle NaN values — run interpolate or fillna upstream')
+
+    n_components_prop = context.properties.get("n_components", 0.95)
+    try:
+        n_components = float(n_components_prop)
+        if n_components > 1:
+            n_components = int(n_components)
+    except (TypeError, ValueError):
+        raise ValueError(f"n_components must be an int (# components) or a float in (0, 1] (variance ratio)")
+
+    scaler = sklearn.preprocessing.StandardScaler()
+    x_scaled = scaler.fit_transform(df)
+
+    pca = sklearn.decomposition.PCA(n_components=n_components)
+    pca.fit(x_scaled)
+
+    monitor = HotellingT2Monitor(
+        scaler=scaler,
+        pca=pca,
+        feature_cols=list(df.columns),
+    )
+
+    context.set_model(monitor, features=df.columns)
+
+
+def ridge_regression(context: wizata_dsapi.Context):
+    """Train a Ridge (L2-regularized) linear regression to predict a single target column.
+    A drop-in replacement for linear_regression when features are highly correlated (common with
+    multi-sensor industrial data) — produces more stable coefficients.
+
+    Property:
+      - alpha: L2 regularization strength (default 1.0 — larger means more shrinkage)
+    """
+    df = context.dataframe
+
+    model_config = context.get_model_config()
+    if not model_config.has_target_feat():
+        raise ValueError(f'ridge_regression requires a target_feat')
+    target_feat_name = context.properties["target_feat"]
+
+    x = df.drop(columns=[target_feat_name])
+    y = df[target_feat_name]
+
+    alpha = float(context.properties.get("alpha", 1.0))
+
+    model = sklearn.linear_model.Ridge(alpha=alpha)
+    model.fit(x, y)
+
+    context.set_model(model, features=x.columns)
+
+
+def random_forest_classifier(context: wizata_dsapi.Context):
+    """Train a Random Forest Classifier on all features to predict a target class column — robust,
+    handles multi-class, no scaling required, exposes feature_importances_ for interpretability.
+    Good default for process-regime or fault-type classification.
+
+    Properties:
+      - n_estimators: number of trees (default 100)
+      - max_depth: max depth per tree (default None = unbounded)
+      - random_state: RNG seed for reproducibility (default 0)
+    """
+    df = context.dataframe
+
+    model_config = context.get_model_config()
+    if not model_config.has_target_feat():
+        raise ValueError(f'random_forest_classifier requires a target_feat')
+    target_feat_name = context.properties["target_feat"]
+
+    x = df.drop(columns=[target_feat_name])
+    y = df[target_feat_name]
+
+    n_estimators = int(context.properties.get("n_estimators", 100))
+    max_depth_prop = context.properties.get("max_depth")
+    max_depth = int(max_depth_prop) if max_depth_prop not in (None, "", "None") else None
+    random_state = int(context.properties.get("random_state", 0))
+
+    model = sklearn.ensemble.RandomForestClassifier(
+        n_estimators=n_estimators,
+        max_depth=max_depth,
+        random_state=random_state,
+    )
+    model.fit(x, y)
+
+    context.set_model(model, features=x.columns)
+
+
+class AutoencoderAnomaly:
+    """
+    MLP-based bottleneck autoencoder — trained to reconstruct its input; at inference returns the
+    per-row reconstruction error as a single anomaly score. Complements hotelling_t2_monitor when
+    the normal-operation manifold is non-linear and PCA misses meaningful structure.
+
+    Implementation detail: built on sklearn's MLPRegressor with fit(X, X) — keeps dependencies
+    minimal (no torch) while preserving the autoencoder concept via symmetric hidden_layer_sizes
+    like (16, 8, 16) which yields a bottleneck of 8.
+
+    :ivar scaler: fitted sklearn StandardScaler.
+    :ivar mlp: fitted sklearn MLPRegressor used as encoder/decoder.
+    :ivar feature_cols: ordered list of feature columns used at training.
+    """
+
+    def __init__(self, scaler, mlp, feature_cols):
+        self.scaler = scaler
+        self.mlp = mlp
+        self.feature_cols = list(feature_cols)
+
+    def predict(self, X):
+        """Return the per-row reconstruction error (anomaly score) as a 1-D array."""
+        if not isinstance(X, pandas.DataFrame):
+            X = pandas.DataFrame(X, columns=self.feature_cols)
+        x_scaled = self.scaler.transform(X)
+        reconstructed = self.mlp.predict(x_scaled)
+        return ((x_scaled - reconstructed) ** 2).sum(axis=1)
+
+    def get_inference_contract(self):
+        """Self-describing contract — platform reads this to override MLModelConfig at runtime."""
+        return {
+            "function": "predict",
+            "output_columns_names": ["anomaly_score"],
+        }
+
+
+def autoencoder_anomaly(context: wizata_dsapi.Context):
+    """Train an MLP bottleneck autoencoder on normal-operation data. At inference, returns a single
+    'anomaly_score' column per row (squared reconstruction error on scaled features). Pair with
+    hotelling_t2_monitor when you need non-linear anomaly detection — PCA captures linear variance,
+    this captures non-linear structure.
+
+    Unsupervised: do **not** set target_feat. NaN values must be handled upstream.
+
+    Properties:
+      - hidden_layer_sizes: tuple/list of hidden layer widths (default (16, 8, 16) = bottleneck 8).
+        Pass as a comma-separated string from the UI (e.g. '16,8,16').
+      - max_iter: training iterations (default 500)
+      - random_state: RNG seed for reproducibility (default 0)
+    """
+    df = context.dataframe
+
+    model_config = context.get_model_config()
+    if model_config.has_target_feat():
+        raise ValueError(f'autoencoder_anomaly is unsupervised — do not set target_feat')
+
+    if df.isna().any().any():
+        raise ValueError(f'autoencoder_anomaly cannot handle NaN values — run interpolate or fillna upstream')
+
+    hidden_prop = context.properties.get("hidden_layer_sizes", (16, 8, 16))
+    if isinstance(hidden_prop, str):
+        hidden = tuple(int(x.strip()) for x in hidden_prop.split(",") if x.strip())
+    elif isinstance(hidden_prop, (list, tuple)):
+        hidden = tuple(int(x) for x in hidden_prop)
+    else:
+        raise ValueError(f"hidden_layer_sizes must be a tuple, list, or comma-separated string")
+
+    max_iter = int(context.properties.get("max_iter", 500))
+    random_state = int(context.properties.get("random_state", 0))
+
+    scaler = sklearn.preprocessing.StandardScaler()
+    x_scaled = scaler.fit_transform(df)
+
+    mlp = sklearn.neural_network.MLPRegressor(
+        hidden_layer_sizes=hidden,
+        max_iter=max_iter,
+        random_state=random_state,
+    )
+    mlp.fit(x_scaled, x_scaled)
+
+    ae = AutoencoderAnomaly(scaler=scaler, mlp=mlp, feature_cols=list(df.columns))
+
+    context.set_model(ae, features=df.columns)
+
+
+class MahalanobisAnomaly:
+    """
+    Lightweight multivariate anomaly detector based on Mahalanobis distance from training mean.
+    Uses a robust inverse covariance matrix from sklearn's EmpiricalCovariance. Complements
+    hotelling_t2_monitor — same underlying idea but no PCA reduction, giving a single interpretable
+    distance score per row.
+
+    :ivar covariance: fitted sklearn EmpiricalCovariance.
+    :ivar feature_cols: ordered list of feature columns used at training.
+    """
+
+    def __init__(self, covariance, feature_cols):
+        self.covariance = covariance
+        self.feature_cols = list(feature_cols)
+
+    def predict(self, X):
+        """Return Mahalanobis distance per row (1-D array)."""
+        if not isinstance(X, pandas.DataFrame):
+            X = pandas.DataFrame(X, columns=self.feature_cols)
+        return numpy.sqrt(self.covariance.mahalanobis(X.values))
+
+    def get_inference_contract(self):
+        """Self-describing contract — platform reads this to override MLModelConfig at runtime."""
+        return {
+            "function": "predict",
+            "output_columns_names": ["mahalanobis_distance"],
+        }
+
+
+def mahalanobis_anomaly(context: wizata_dsapi.Context):
+    """Train a Mahalanobis-distance anomaly detector on normal-operation data. At inference,
+    returns a single 'mahalanobis_distance' column per row — larger distance = more anomalous,
+    interpretable because each feature contributes proportional to its deviation weighted by
+    the inverse covariance. No dimensionality reduction (unlike hotelling_t2_monitor), so useful
+    when you have few-to-medium sensors and want a single calibrated score.
+
+    Unsupervised: do **not** set target_feat. NaN values must be handled upstream.
+
+    Properties: (none — uses sklearn's EmpiricalCovariance defaults)
+    """
+    df = context.dataframe
+
+    model_config = context.get_model_config()
+    if model_config.has_target_feat():
+        raise ValueError(f'mahalanobis_anomaly is unsupervised — do not set target_feat')
+
+    if df.isna().any().any():
+        raise ValueError(f'mahalanobis_anomaly cannot handle NaN values — run interpolate or fillna upstream')
+
+    cov = sklearn.covariance.EmpiricalCovariance()
+    cov.fit(df.values)
+
+    ma = MahalanobisAnomaly(covariance=cov, feature_cols=list(df.columns))
+
+    context.set_model(ma, features=df.columns)
+
+
+class GMMRegimeClassifier:
+    """
+    Gaussian Mixture Model for soft process-regime detection. At inference, assigns each row to
+    the most likely regime (cluster index 0..n_components-1).
+
+    :ivar gmm: fitted sklearn GaussianMixture.
+    :ivar scaler: fitted sklearn StandardScaler (so clusters are scale-invariant).
+    :ivar feature_cols: ordered list of feature columns used at training.
+    """
+
+    def __init__(self, gmm, scaler, feature_cols):
+        self.gmm = gmm
+        self.scaler = scaler
+        self.feature_cols = list(feature_cols)
+
+    def predict(self, X):
+        """Return the most likely regime index (0..n_components-1) per row as a 1-D array."""
+        if not isinstance(X, pandas.DataFrame):
+            X = pandas.DataFrame(X, columns=self.feature_cols)
+        x_scaled = self.scaler.transform(X)
+        return self.gmm.predict(x_scaled).astype(float)
+
+    def get_inference_contract(self):
+        """Self-describing contract — platform reads this to override MLModelConfig at runtime."""
+        return {
+            "function": "predict",
+            "output_columns_names": ["regime"],
+        }
+
+
+def gmm_regime_classifier(context: wizata_dsapi.Context):
+    """Train a Gaussian Mixture Model to identify process regimes from multi-sensor data. At
+    inference, returns a single 'regime' column per row (integer cluster index, 0..n_components-1).
+    Good for separating operating modes (e.g. startup / steady / shutdown) without labels.
+
+    Unsupervised: do **not** set target_feat. NaN values must be handled upstream.
+
+    Properties:
+      - n_components: number of regimes / clusters (default 3)
+      - covariance_type: 'full' (default), 'tied', 'diag', or 'spherical'
+      - random_state: RNG seed for reproducibility (default 0)
+    """
+    df = context.dataframe
+
+    model_config = context.get_model_config()
+    if model_config.has_target_feat():
+        raise ValueError(f'gmm_regime_classifier is unsupervised — do not set target_feat')
+
+    if df.isna().any().any():
+        raise ValueError(f'gmm_regime_classifier cannot handle NaN values — run interpolate or fillna upstream')
+
+    n_components = int(context.properties.get("n_components", 3))
+    covariance_type = context.properties.get("covariance_type", "full")
+    if covariance_type not in ("full", "tied", "diag", "spherical"):
+        raise ValueError(f"covariance_type must be 'full', 'tied', 'diag', or 'spherical', got '{covariance_type}'")
+    random_state = int(context.properties.get("random_state", 0))
+
+    scaler = sklearn.preprocessing.StandardScaler()
+    x_scaled = scaler.fit_transform(df)
+
+    gmm = sklearn.mixture.GaussianMixture(
+        n_components=n_components,
+        covariance_type=covariance_type,
+        random_state=random_state,
+    )
+    gmm.fit(x_scaled)
+
+    classifier = GMMRegimeClassifier(gmm=gmm, scaler=scaler, feature_cols=list(df.columns))
+
+    context.set_model(classifier, features=df.columns)

wizata_dsapi-2.0.0.dev26/wizata_dsapi/version.py

@@ -0,0 +1 @@
+__version__ = "2.0.0.dev26"

{wizata_dsapi-2.0.0.dev24 → wizata_dsapi-2.0.0.dev26/wizata_dsapi.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: wizata_dsapi
-Version: 2.0.0.dev24
+Version: 2.0.0.dev26
 Summary: Wizata Data Science Toolkit
 Author: Wizata S.A.
 Author-email: info@wizata.com

wizata_dsapi-2.0.0.dev24/wizata_dsapi/models/__init__.py

@@ -1 +0,0 @@
-from .common import linear_regression, logistic_regression, isolation_forest, gradiant_boost_classifier

wizata_dsapi-2.0.0.dev24/wizata_dsapi/models/common.py

@@ -1,116 +0,0 @@
-import wizata_dsapi
-
-import pandas
-import numpy
-
-import sklearn
-import sklearn.linear_model
-import sklearn.ensemble
-
-
-def extract_target_feat(context: wizata_dsapi.Context, single: bool = True):
-    """
-    return a list of target_feat columns names if not single value or the single value target feat name
-    raise an error if configuration mismatch
-    """
-
-    if "target_feat" not in context.properties:
-        raise ValueError(f"training script requires a proper target_feat")
-
-    target_feat = context.properties["target_feat"]
-    if isinstance(target_feat, str):
-        if single:
-            return target_feat
-        else:
-            return [target_feat]
-    elif isinstance(target_feat, list):
-        if single:
-            if len(target_feat) == 1:
-                return target_feat[0]
-            else:
-                raise ValueError(f"expecting only one target_feat but found {len(target_feat)}")
-        else:
-            return [target_feat]
-    else:
-        raise TypeError(f'target_feat must be a str or a list of str but found {target_feat.__class__.__name__}')
-
-
-def linear_regression(context: wizata_dsapi.Context):
-    """Train a linear regression model on all features to predict a single target column."""
-    df = context.dataframe
-
-    model_config = context.get_model_config()
-    if not model_config.has_target_feat():
-        raise ValueError(f'linear_regression requires a target feat')
-    target_feat_name = context.properties["target_feat"]
-
-    x = df.drop(columns=[target_feat_name])
-    y = df[target_feat_name]
-
-    model = sklearn.linear_model.LinearRegression()
-    model.fit(x, y)
-
-    context.set_model(model, features=x.columns)
-
-
-def logistic_regression(context: wizata_dsapi.Context):
-    """Train a logistic regression classifier on all features to predict a binary target column."""
-    df = context.dataframe
-
-    model_config = context.get_model_config()
-    if not model_config.has_target_feat():
-        raise ValueError(f'logistic_regression requires a target feat')
-    target_feat_name = context.properties["target_feat"]
-
-    x = df.drop(columns=[target_feat_name])
-    y = df[target_feat_name]
-
-    model = sklearn.linear_model.LogisticRegression()
-    model.fit(x, y.astype(int))
-
-    context.set_model(model, features=x.columns)
-
-
-def isolation_forest(context: wizata_dsapi.Context):
-    """Train an Isolation Forest for unsupervised anomaly detection using a sensitivity level (1-5)."""
-
-    model_config = context.get_model_config()
-    if model_config.has_target_feat():
-        raise ValueError(f'isolation_forest does not requires a target feat')
-
-    try:
-        if context.properties['sensitivity'] is None:
-            raise KeyError("sensitivity is none")
-        sensitivity = int(context.properties['sensitivity'])
-        sensitivities = [0.05, 0.15, 0.25, 0.35, 0.4]
-        contamination = sensitivities[sensitivity - 1]
-    except Exception as e:
-        raise ValueError(f'cannot extract sensitivity integer from 0 to 4 due to {e}')
-
-    df = context.dataframe.copy()
-    model = sklearn.ensemble.IsolationForest(contamination=contamination)
-    df['isolation_forest_predict'] = model.fit_predict(df)
-    context.set_model(model, features=df.columns)
-    return df
-
-
-def gradiant_boost_classifier(context: wizata_dsapi.Context):
-    """Train a Gradient Boosting classifier on all features to predict a target column."""
-    df = context.dataframe
-
-    model_config = context.get_model_config()
-    if not model_config.has_target_feat():
-        raise ValueError(f'gradiant_boost_classifier requires a target feat')
-    target_feat_name = context.properties["target_feat"]
-
-    x = df.drop(columns=[target_feat_name])
-    y = df[target_feat_name]
-
-    model = sklearn.ensemble.GradientBoostingClassifier(random_state=0).fit(x, y)
-    context.set_model(model, features=df.columns)
-
-
-
-
-
-

wizata_dsapi-2.0.0.dev24/wizata_dsapi/version.py

@@ -1 +0,0 @@
-__version__ = "2.0.0.dev24"