lecrapaud 0.18.7__py3-none-any.whl → 0.22.6__py3-none-any.whl

lecrapaud/db/models/experiment.py

@@ -1,5 +1,7 @@
 from itertools import chain
 import joblib
+import pandas as pd
+import os
 
 from sqlalchemy import (
     Column,
@@ -14,18 +16,18 @@ from sqlalchemy import (
     TIMESTAMP,
     UniqueConstraint,
     func,
+    event,
 )
-from sqlalchemy.orm import relationship, aliased
+from sqlalchemy.orm import relationship, aliased, mapper
 from sqlalchemy.ext.hybrid import hybrid_property
 from sqlalchemy import func
 from statistics import fmean as mean
 from lecrapaud.db.models.model_selection import ModelSelection
-from lecrapaud.db.models.model_training import ModelTraining
-from lecrapaud.db.models.score import Score
+from lecrapaud.db.models.model_selection_score import ModelSelectionScore
 
 from lecrapaud.db.models.base import Base, with_db
 from lecrapaud.db.models.utils import create_association_table
-from lecrapaud.utils import logger, contains_best
+from lecrapaud.utils import logger, contains_best, strip_timestamp_suffix
 
 # jointures
 lecrapaud_experiment_target_association = create_association_table(
@@ -49,107 +51,15 @@ class Experiment(Base):
         onupdate=func.now(),
         nullable=False,
     )
-    name = Column(String(50), nullable=False)
+    name = Column(String(255), nullable=False)
     path = Column(String(255))  # we do not have this at creation time
-    type = Column(String(50), nullable=False)
     size = Column(Integer, nullable=False)
     train_size = Column(Integer)
     val_size = Column(Integer)
-
-    # Relationships
-    model_selections = relationship(
-        "ModelSelection",
-        back_populates="experiment",
-        cascade="all, delete-orphan",
-        lazy="selectin",
-    )
-
-    @hybrid_property
-    def best_rmse(self):
-        """Best RMSE score across all model selections and trainings."""
-        # Get the minimum RMSE for each model selection
-        min_scores = [
-            min(
-                score.rmse
-                for mt in ms.model_trainings
-                for score in mt.score
-                if score.rmse is not None
-            )
-            for ms in self.model_selections
-            if any(
-                score.rmse is not None
-                for mt in ms.model_trainings
-                for score in mt.score
-            )
-        ]
-        return min(min_scores) if min_scores else None
-
-    @hybrid_property
-    def best_logloss(self):
-        """Best LogLoss score across all model selections and trainings."""
-        # Get the minimum LogLoss for each model selection
-        min_scores = [
-            min(
-                score.logloss
-                for mt in ms.model_trainings
-                for score in mt.score
-                if score.logloss is not None
-            )
-            for ms in self.model_selections
-            if any(
-                score.logloss is not None
-                for mt in ms.model_trainings
-                for score in mt.score
-            )
-        ]
-        return min(min_scores) if min_scores else None
-
-    @hybrid_property
-    def avg_rmse(self):
-        """Average RMSE score across all model selections and trainings."""
-        # Get the minimum RMSE for each model selection
-        min_scores = [
-            min(
-                score.rmse
-                for mt in ms.model_trainings
-                for score in mt.score
-                if score.rmse is not None
-            )
-            for ms in self.model_selections
-            if any(
-                score.rmse is not None
-                for mt in ms.model_trainings
-                for score in mt.score
-            )
-        ]
-        return mean(min_scores) if min_scores else None
-
-    @hybrid_property
-    def avg_logloss(self):
-        """Average LogLoss score across all model selections and trainings."""
-        # Get the minimum LogLoss for each model selection
-        min_scores = [
-            min(
-                score.logloss
-                for mt in ms.model_trainings
-                for score in mt.score
-                if score.logloss is not None
-            )
-            for ms in self.model_selections
-            if any(
-                score.logloss is not None
-                for mt in ms.model_trainings
-                for score in mt.score
-            )
-        ]
-        return mean(min_scores) if min_scores else None
-
     test_size = Column(Integer)
-    corr_threshold = Column(Float, nullable=False)
-    max_features = Column(Integer, nullable=False)
-    percentile = Column(Float, nullable=False)
     number_of_groups = Column(Integer)
     list_of_groups = Column(JSON)
+    number_of_targets = Column(Integer)
     start_date = Column(DateTime)
     end_date = Column(DateTime)
     train_start_date = Column(DateTime)
@@ -181,6 +91,142 @@ class Experiment(Base):
         ),
     )
 
+    # Relationships
+    model_selections = relationship(
+        "ModelSelection",
+        back_populates="experiment",
+        cascade="all, delete-orphan",
+        lazy="selectin",
+    )
+
+    # Hooks
+    # @event.listens_to(Experiment, "after_commit")
+    # def set_score(mapper, connection, target):
+    #     target.score = target.score
+
+    # Properties
+    @hybrid_property
+    def rmse_scores(self):
+        """best RMSE scores across all model selections, for each targets."""
+        # Get the minimum RMSE for each model selection
+        min_scores = [
+            min(mss.rmse for mss in ms.model_selection_scores if mss.rmse is not None)
+            for ms in self.model_selections
+            if any(mss.rmse is not None for mss in ms.model_selection_scores)
+        ]
+
+        if not min_scores:
+            # fallback to path if no model_selection_scores found
+            for target in self.targets:
+                path = f"{self.path}/{target.name}/scores_tracking.csv"
+                if not os.path.exists(path):
+                    continue
+                score = pd.read_csv(path)
+                if "RMSE" not in score.columns:
+                    continue
+                min_scores.append(min(score["RMSE"]))
+
+        return min_scores
+
+    @hybrid_property
+    def logloss_scores(self):
+        """best LogLoss scores across all model selections, for each targets."""
+        # Get the minimum LogLoss for each model selection
+        min_scores = [
+            min(
+                mss.logloss
+                for mss in ms.model_selection_scores
+                if mss.logloss is not None
+            )
+            for ms in self.model_selections
+            if any(mss.logloss is not None for mss in ms.model_selection_scores)
+        ]
+
+        if not min_scores:
+            # fallback to path if no model_selection_scores found
+            for target in self.targets:
+                path = f"{self.path}/{target.name}/scores_tracking.csv"
+                if not os.path.exists(path):
+                    continue
+                score = pd.read_csv(path)
+                if "LOGLOSS" not in score.columns:
+                    continue
+                min_scores.append(min(score["LOGLOSS"]))
+
+        return min_scores
+
+    @hybrid_property
+    def best_rmse(self):
+        """Best RMSE score within targets, across all model selections."""
+        return min(self.rmse_scores) if self.rmse_scores else None
+
+    @hybrid_property
+    def best_logloss(self):
+        """Best LogLoss score within targets, across all model selections."""
+        return min(self.logloss_scores) if self.logloss_scores else None
+
+    @hybrid_property
+    def score(self):
+        # Calculate a combined score: average of normalized best RMSE and LogLoss per targets
+        # This ensures we're comparing apples to apples by normalizing the scores
+
+        if not self.rmse_scores and not self.logloss_scores:
+            logger.error("No experiments found with RMSE or LogLoss scores.")
+            return None
+
+        # Normalize scores (subtract min and divide by range)
+        # Guard against division by zero when only one observation or all equal
+        # Let's gather all the data from similar experiments to calculate the range
+
+        similar_experiments = Experiment.get_all_by_name(name=self.name)
+        if not similar_experiments:
+            similar_experiments = [self]
+        rmse_scores = [
+            score for exp in similar_experiments for score in exp.rmse_scores or []
+        ]
+        logloss_scores = [
+            score for exp in similar_experiments for score in exp.logloss_scores or []
+        ]
+
+        min_rmse = min(rmse_scores)
+        max_rmse = max(rmse_scores)
+        range_rmse = max_rmse - min_rmse
+        min_logloss = min(logloss_scores)
+        max_logloss = max(logloss_scores)
+        range_logloss = max_logloss - min_logloss
+
+        # Calculate combined score for each experiment
+        normed_scores = []
+        for rmse_score in self.rmse_scores:
+
+            # Normalize both scores (safe when range == 0)
+            norm_rmse = 0.0 if range_rmse == 0 else (rmse_score - min_rmse) / range_rmse
+            normed_scores.append(norm_rmse)
+
+        for logloss_score in self.logloss_scores:
+            norm_logloss = (
+                0.0
+                if range_logloss == 0
+                else (logloss_score - min_logloss) / range_logloss
+            )
+            normed_scores.append(norm_logloss)
+
+        # Calculate score (average of normalized scores)
+        score = sum(normed_scores) / len(normed_scores)
+
+        return score
+
+    @hybrid_property
+    def avg_rmse(self):
+        """Average within targets of best RMSE score across all model selections ."""
+        return mean(self.rmse_scores) if self.rmse_scores else None
+
+    @hybrid_property
+    def avg_logloss(self):
+        """Average within targets of best LogLoss score across all model selections ."""
+        return mean(self.logloss_scores) if self.logloss_scores else None
+
+    # Class methods
     @classmethod
     @with_db
     def get_all_by_name(cls, name: str | None = None, limit: int = 1000, db=None):
@@ -194,10 +240,11 @@ class Experiment(Base):
         Returns:
             Experiment or None: The most recent matching experiment or None if not found
         """
+        base_name = strip_timestamp_suffix(name)
         if name is not None:
             return (
                 db.query(cls)
-                .filter(cls.name.ilike(f"%{name}%"))
+                .filter(cls.name.ilike(f"%{base_name}%"))
                 .order_by(cls.created_at.desc())
                 .limit(limit)
                 .all()
@@ -217,27 +264,24 @@ class Experiment(Base):
         Returns:
             Experiment or None: The most recent matching experiment or None if not found
         """
+        base_name = strip_timestamp_suffix(name)
         return (
             db.query(cls)
-            .filter(cls.name.ilike(f"%{name}%"))
+            .filter(cls.name.ilike(f"%{base_name}%"))
             .order_by(cls.created_at.desc())
             .first()
         )
 
     @classmethod
     @with_db
-    def get_best_by_score(cls, name: str, metric="both", db=None):
+    def get_best_by_score(cls, name: str, db=None):
         """
-        Find the experiment with the best score based on average RMSE, LogLoss, or both.
-
-        Args:
-            metric (str): 'rmse', 'logloss', or 'both' to determine which score to optimize
-            db: SQLAlchemy session
+        Find the experiment with the best normalized score across RMSE and LogLoss.
 
         Returns:
             Experiment or None: The experiment with the best score or None if not found
         """
-        experiments = db.query(cls).filter(cls.name.ilike(f"%{name}%")).all()
+        experiments = Experiment.get_all_by_name(name=name)
         if not experiments:
             logger.error(f"No experiments found with the given name: {name}")
             return None
@@ -255,66 +299,22 @@ class Experiment(Base):
             )
             return None
 
-        if metric == "both":
-            # Calculate a combined score: average of normalized RMSE and LogLoss
-            # This ensures we're comparing apples to apples by normalizing the scores
-
-            # Get all scores
-            rmse_scores = [e.avg_rmse for e in experiments if e.avg_rmse is not None]
-            logloss_scores = [
-                e.avg_logloss for e in experiments if e.avg_logloss is not None
-            ]
-
-            if not rmse_scores or not logloss_scores:
-                logger.error(
-                    "No experiments found with both RMSE and LogLoss scores. Maybe try with only one metric."
-                )
-                return None
-
-            # Normalize scores (subtract min and divide by range)
-            min_rmse = min(rmse_scores)
-            range_rmse = max(rmse_scores) - min_rmse
-            min_logloss = min(logloss_scores)
-            range_logloss = max(logloss_scores) - min_logloss
-
-            # Calculate combined score for each experiment
-            experiment_scores = []
-            for experiment in experiments:
-                if experiment.avg_rmse is None or experiment.avg_logloss is None:
-                    continue
-
-                # Normalize both scores
-                norm_rmse = (experiment.avg_rmse - min_rmse) / range_rmse
-                norm_logloss = (experiment.avg_logloss - min_logloss) / range_logloss
+        scored_experiments = []
+        for experiment in experiments:
+            score = experiment.score
+            if score is not None:
+                scored_experiments.append((experiment, score))
 
-                # Calculate combined score (average of normalized scores)
-                combined_score = (norm_rmse + norm_logloss) / 2
-                experiment_scores.append((experiment, combined_score))
-
-            # Sort by combined score (ascending since lower is better)
-            experiment_scores.sort(key=lambda x: x[1])
-
-            return experiment_scores[0][0] if experiment_scores else None
-
-        elif metric == "rmse" or metric == "logloss":
-            # For single metric case (rmse or logloss)
-
-            # Filter out experiments without scores and sort by the selected metric
-            filtered_experiments = []
-            for exp in experiments:
-                score = exp.avg_rmse if metric == "rmse" else exp.avg_logloss
-                if score is not None:
-                    filtered_experiments.append((exp, score))
+        if not scored_experiments:
+            logger.error(
+                f"No experiments with calculable scores found with the given name: {name}"
+            )
+            return None
 
-            if not filtered_experiments:
-                return None
-
-            # Sort by score (ascending since lower is better)
-            filtered_experiments.sort(key=lambda x: x[1])
-            return filtered_experiments[0][0]
-        else:
-            raise ValueError("Invalid metric. Must be 'rmse', 'logloss', or 'both'.")
+        scored_experiments.sort(key=lambda x: x[1])
+        return scored_experiments[0][0]
 
+    # Instance methods
     def best_score(self, target_number: int) -> dict:
         """
         Returns the scores for the best model of the specified target.
@@ -342,7 +342,7 @@ class Experiment(Base):
             (ms for ms in self.model_selections if ms.target_id == target.id), None
         )
 
-        if not best_model_selection or not best_model_selection.model_trainings:
+        if not best_model_selection or not best_model_selection.model_selection_scores:
             return {
                 "experiment_name": self.name,
                 "target_number": target_number,
@@ -350,22 +350,31 @@ class Experiment(Base):
                 "scores": {},
             }
 
-        # Get the best model training (assuming the first one is the best)
-        best_training = best_model_selection.model_trainings[0]
-
-        # Get the validation score for this training
-        validation_scores = [s for s in best_training.score if s.type == "validation"]
+        # Get the best model score based on lowest logloss or rmse
+        model_scores = best_model_selection.model_selection_scores
 
-        if not validation_scores:
+        # Determine if we should use logloss or rmse based on what's available
+        if any(ms.logloss is not None for ms in model_scores):
+            # Classification: find lowest logloss
+            best_score = min(
+                (ms for ms in model_scores if ms.logloss is not None),
+                key=lambda x: x.logloss,
+            )
+        elif any(ms.rmse is not None for ms in model_scores):
+            # Regression: find lowest rmse
+            best_score = min(
+                (ms for ms in model_scores if ms.rmse is not None), key=lambda x: x.rmse
+            )
+        else:
             return {
                 "experiment_name": self.name,
                 "target_number": target_number,
-                "error": "No validation scores found for the best model",
+                "error": "No scores found for the best model",
                 "scores": {},
             }
 
-        # Get all available metrics from the first validation score
-        score = validation_scores[0]
+        # Use the best score found
+        score = best_score
         available_metrics = [
             "rmse",
             "mae",
@@ -386,13 +395,9 @@ class Experiment(Base):
 
         # Get the model info
         model_info = {
-            "model_type": (
-                best_training.model.model_type if best_training.model else "unknown"
-            ),
-            "model_name": (
-                best_training.model.name if best_training.model else "unknown"
-            ),
-            "training_time_seconds": best_training.training_time,
+            "model_type": (score.model.model_type if score.model else "unknown"),
+            "model_name": (score.model.name if score.model else "unknown"),
+            "training_time_seconds": score.training_time,
         }
 
         return {
@@ -402,7 +407,10 @@ class Experiment(Base):
             "scores": scores,
         }
 
-    def get_features(self, target_number: int):
+    @with_db
+    def get_features(self, target_number: int, db=None):
+        # Ensure we have a fresh instance attached to the session
+        self = db.merge(self)
         targets = [t for t in self.targets if t.name == f"TARGET_{target_number}"]
         if targets:
             target_id = targets[0].id
@@ -418,7 +426,12 @@ class Experiment(Base):
         features = joblib.load(f"{self.path}/TARGET_{target_number}/features.pkl")
         return features
 
-    def get_all_features(self, date_column: str = None, group_column: str = None):
+    @with_db
+    def get_all_features(
+        self, date_column: str = None, group_column: str = None, db=None
+    ):
+        # Ensure we have a fresh instance attached to the session
+        self = db.merge(self)
         target_idx = [target.id for target in self.targets]
         _all_features = chain.from_iterable(
             [f.name for f in fs.features]

lecrapaud/db/models/feature_selection.py

@@ -115,7 +115,4 @@ class FeatureSelection(Base):
             if feature not in self.features:
                 self.features.append(feature)
 
-        db.flush()
-        db.refresh(self)
-        print(self.features)
         return self

lecrapaud/db/models/feature_selection_rank.py

@@ -65,21 +65,3 @@ class FeatureSelectionRank(Base):
             name="uq_feature_selection_rank_composite",
         ),
     )
-
-    @classmethod
-    @with_db
-    def bulk_upsert(cls, rows, db=None):
-        stmt = insert(cls).values(rows)
-
-        update_fields = {
-            key: stmt.inserted[key]
-            for key in rows[0]
-            if key not in ("feature_selection_id", "feature_id", "method")
-        }
-
-        stmt = stmt.on_duplicate_key_update(**update_fields)
-
-        db.execute(stmt)
-        db.commit()
-
-        return len(rows)

lecrapaud/db/models/model_selection.py

@@ -54,8 +54,8 @@ class ModelSelection(Base):
     )
 
     best_model = relationship("Model", lazy="selectin")
-    model_trainings = relationship(
-        "ModelTraining",
+    model_selection_scores = relationship(
+        "ModelSelectionScore",
         back_populates="model_selection",
         cascade="all, delete-orphan",
         lazy="selectin",

lecrapaud/db/models/{score.py → model_selection_score.py}

@@ -3,10 +3,11 @@ from sqlalchemy import (
     Integer,
     String,
     Float,
+    JSON,
     ForeignKey,
     BigInteger,
     TIMESTAMP,
-    JSON,
+    UniqueConstraint,
 )
 from sqlalchemy import func
 from sqlalchemy.orm import relationship
@@ -14,7 +15,9 @@ from lecrapaud.db.models.base import Base
 from lecrapaud.config import LECRAPAUD_TABLE_PREFIX
 
 
-class Score(Base):
+class ModelSelectionScore(Base):
+    __tablename__ = f"{LECRAPAUD_TABLE_PREFIX}_model_selection_scores"
+    
     id = Column(BigInteger, primary_key=True, index=True, autoincrement=True)
     created_at = Column(
         TIMESTAMP(timezone=True), server_default=func.now(), nullable=False
@@ -25,10 +28,21 @@ class Score(Base):
         onupdate=func.now(),
         nullable=False,
     )
-    type = Column(
-        String(50), nullable=False
-    )  # either hyperopts or validation or crossval
+    
+    # From ModelTraining
+    best_params = Column(JSON)
+    model_path = Column(String(255))
     training_time = Column(Integer)
+    model_id = Column(
+        BigInteger, ForeignKey(f"{LECRAPAUD_TABLE_PREFIX}_models.id"), nullable=False
+    )
+    model_selection_id = Column(
+        BigInteger,
+        ForeignKey(f"{LECRAPAUD_TABLE_PREFIX}_model_selections.id", ondelete="CASCADE"),
+        nullable=False,
+    )
+    
+    # From Score (excluding type and training_time which is already in ModelTraining)
     eval_data_std = Column(Float)
     rmse = Column(Float)
     rmse_std_ratio = Column(Float)
@@ -45,16 +59,20 @@ class Score(Base):
     recall = Column(Float)
     f1 = Column(Float)
     roc_auc = Column(Float)
+    avg_precision = Column(Float)
     thresholds = Column(JSON)
     precision_at_threshold = Column(Float)
     recall_at_threshold = Column(Float)
     f1_at_threshold = Column(Float)
-    model_training_id = Column(
-        BigInteger,
-        ForeignKey(f"{LECRAPAUD_TABLE_PREFIX}_model_trainings.id", ondelete="CASCADE"),
-        nullable=False,
-    )
 
-    model_trainings = relationship(
-        "ModelTraining", back_populates="score", lazy="selectin"
+    # Relationships
+    model = relationship("Model", lazy="selectin")
+    model_selection = relationship(
+        "ModelSelection", back_populates="model_selection_scores", lazy="selectin"
     )
+
+    __table_args__ = (
+        UniqueConstraint(
+            "model_id", "model_selection_id", name="uq_model_selection_score_composite"
+        ),
+    )

lecrapaud/db/session.py

@@ -12,6 +12,7 @@ from lecrapaud.config import DB_USER, DB_PASSWORD, DB_HOST, DB_PORT, DB_NAME, DB
 
 _engine = None
 _SessionLocal = None
+
 if DB_URI:
     if "mysql://" in DB_URI and "pymysql://" not in DB_URI:
         DB_URI = DB_URI.replace("mysql://", "mysql+pymysql://")
@@ -26,6 +27,9 @@ else:
 
 def init_db(uri: str = None):
     global _engine, _SessionLocal, DATABASE_URL, DB_URI
+    if _SessionLocal is not None:
+        return
+
     if uri:
         if "mysql://" in uri and "pymysql://" not in uri:
             uri = uri.replace("mysql://", "mysql+pymysql://")
@@ -53,11 +57,27 @@ def init_db(uri: str = None):
         conn.execute(text(f"CREATE DATABASE IF NOT EXISTS `{db_name}`"))
         conn.commit()
 
-    # Step 3: Connect to the newly created database
-    _engine = create_engine(DATABASE_URL, echo=False)
+    # Step 3: Configure main engine with connection pooling and reconnection settings
+    _engine = create_engine(
+        DATABASE_URL,
+        echo=False,
+        pool_pre_ping=True,  # Check connection health before using
+        pool_recycle=3600,  # Recycle connections after 1 hour
+        pool_timeout=30,  # Wait 30 seconds for a connection from the pool
+        pool_size=10,  # Maintain up to 10 connections
+        max_overflow=10,  # Allow up to 10 more connections during spikes
+        connect_args={
+            "connect_timeout": 10,  # 10 second connection timeout
+        },
+    )
 
-    # Step 4: Create session factory
-    _SessionLocal = sessionmaker(autocommit=False, autoflush=False, bind=_engine)
+    # Step 4: Configure session factory
+    _SessionLocal = sessionmaker(
+        autocommit=False,
+        autoflush=False,
+        bind=_engine,
+        expire_on_commit=False,  # Prevent detached instance errors
+    )
 
     # Step 5: Apply Alembic migrations programmatically
     current_dir = os.path.dirname(__file__)  # → lecrapaud/db
@@ -76,8 +96,17 @@ def init_db(uri: str = None):
 # Dependency to get a session instance
 @contextmanager
 def get_db():
+    """Get a database session with automatic connection management.
+
+    Yields:
+        Session: A SQLAlchemy session instance
+
+    The session is automatically committed if no exceptions occur, and rolled back otherwise.
+    The connection is automatically returned to the pool when the session is closed.
+    """
     if _SessionLocal is None:
         init_db()
+
     db = _SessionLocal()
     try:
         yield db