ins-pricing 0.3.2__py3-none-any.whl → 0.3.4__py3-none-any.whl

ins_pricing/production/predict.py

@@ -23,6 +23,11 @@ from .preprocess import (
 from .scoring import batch_score
 from ..modelling.core.bayesopt.models.model_gnn import GraphNeuralNetSklearn
 from ..modelling.core.bayesopt.models.model_resn import ResNetSklearn
+from ..modelling.core.bayesopt.utils.losses import (
+    infer_loss_name_from_model_name,
+    normalize_loss_name,
+    resolve_tweedie_power,
+)
 from ins_pricing.utils import DeviceManager, get_logger
 from ins_pricing.utils.torch_compat import torch_load
 
@@ -50,6 +55,15 @@ def _default_tweedie_power(model_name: str, task_type: str) -> Optional[float]:
     return 1.5
 
 
+def _resolve_loss_name(cfg: Dict[str, Any], model_name: str, task_type: str) -> str:
+    normalized = normalize_loss_name(cfg.get("loss_name"), task_type)
+    if task_type == "classification":
+        return "logloss" if normalized == "auto" else normalized
+    if normalized == "auto":
+        return infer_loss_name_from_model_name(model_name)
+    return normalized
+
+
 def _resolve_value(
     value: Any,
     *,
@@ -182,11 +196,14 @@ def _build_resn_model(
     task_type: str,
     epochs: int,
     resn_weight_decay: float,
+    loss_name: str,
     params: Dict[str, Any],
 ) -> ResNetSklearn:
-    power = params.get("tw_power", _default_tweedie_power(model_name, task_type))
-    if power is not None:
-        power = float(power)
+    if loss_name == "tweedie":
+        power = params.get("tw_power", _default_tweedie_power(model_name, task_type))
+        power = float(power) if power is not None else None
+    else:
+        power = resolve_tweedie_power(loss_name, default=1.5)
     weight_decay = float(params.get("weight_decay", resn_weight_decay))
     return ResNetSklearn(
         model_nme=model_name,
@@ -205,6 +222,7 @@ def _build_resn_model(
         weight_decay=weight_decay,
         use_data_parallel=False,
         use_ddp=False,
+        loss_name=loss_name,
     )
 
 
@@ -215,9 +233,15 @@ def _build_gnn_model(
     task_type: str,
     epochs: int,
     cfg: Dict[str, Any],
+    loss_name: str,
     params: Dict[str, Any],
 ) -> GraphNeuralNetSklearn:
     base_tw = _default_tweedie_power(model_name, task_type)
+    if loss_name == "tweedie":
+        tw_power = params.get("tw_power", base_tw)
+        tw_power = float(tw_power) if tw_power is not None else None
+    else:
+        tw_power = resolve_tweedie_power(loss_name, default=1.5)
     return GraphNeuralNetSklearn(
         model_nme=f"{model_name}_gnn",
         input_dim=input_dim,
@@ -229,7 +253,7 @@ def _build_gnn_model(
         epochs=int(params.get("epochs", epochs)),
         patience=int(params.get("patience", 5)),
         task_type=task_type,
-        tweedie_power=float(params.get("tw_power", base_tw or 1.5)),
+        tweedie_power=tw_power,
         weight_decay=float(params.get("weight_decay", 0.0)),
         use_data_parallel=False,
         use_ddp=False,
@@ -239,6 +263,7 @@ def _build_gnn_model(
         max_gpu_knn_nodes=cfg.get("gnn_max_gpu_knn_nodes"),
         knn_gpu_mem_ratio=cfg.get("gnn_knn_gpu_mem_ratio", 0.9),
         knn_gpu_mem_overhead=cfg.get("gnn_knn_gpu_mem_overhead", 2.0),
+        loss_name=loss_name,
     )
 
 
@@ -273,6 +298,9 @@ def load_saved_model(
                 from ..modelling.core.bayesopt.models.model_ft_components import FTTransformerCore
 
                 # Reconstruct model from config
+                resolved_loss = model_config.get("loss_name")
+                if not resolved_loss:
+                    resolved_loss = _resolve_loss_name(cfg, model_name, task_type)
                 model = FTTransformerSklearn(
                     model_nme=model_config.get("model_nme", ""),
                     num_cols=model_config.get("num_cols", []),
@@ -282,6 +310,7 @@ def load_saved_model(
                     n_layers=model_config.get("n_layers", 4),
                     dropout=model_config.get("dropout", 0.1),
                     task_type=model_config.get("task_type", "regression"),
+                    loss_name=resolved_loss,
                     tweedie_power=model_config.get("tw_power", 1.5),
                     num_numeric_tokens=model_config.get("num_numeric_tokens"),
                     use_data_parallel=False,
@@ -337,12 +366,14 @@ def load_saved_model(
             params = load_best_params(output_dir, model_name, model_key)
         if params is None:
             raise RuntimeError("Best params not found for resn")
+        loss_name = _resolve_loss_name(cfg, model_name, task_type)
         model = _build_resn_model(
             model_name=model_name,
             input_dim=input_dim,
             task_type=task_type,
             epochs=int(cfg.get("epochs", 50)),
             resn_weight_decay=float(cfg.get("resn_weight_decay", 1e-4)),
+            loss_name=loss_name,
             params=params,
         )
         model.resnet.load_state_dict(state_dict)
@@ -357,12 +388,14 @@ def load_saved_model(
             raise ValueError(f"Invalid GNN checkpoint: {model_path}")
         params = payload.get("best_params") or {}
         state_dict = payload.get("state_dict")
+        loss_name = _resolve_loss_name(cfg, model_name, task_type)
         model = _build_gnn_model(
             model_name=model_name,
             input_dim=input_dim,
             task_type=task_type,
             epochs=int(cfg.get("epochs", 50)),
             cfg=cfg,
+            loss_name=loss_name,
             params=params,
         )
         model.set_params(dict(params))
@@ -628,8 +661,4 @@ def predict_from_config(
     if output_path:
         output_path = Path(output_path)
         output_path.parent.mkdir(parents=True, exist_ok=True)
-        if output_path.suffix.lower() in {".parquet", ".pq"}:
-            result.to_parquet(output_path, index=False)
-        else:
-            result.to_csv(output_path, index=False)
-    return result
+        if output_path.suffix.lower

ins_pricing/setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 
 setup(
     name="ins_pricing",
-    version="0.3.2",
+    version="0.3.4",
     description="Reusable modelling, pricing, governance, and reporting utilities.",
     author="meishi125478",
     license="Proprietary",

ins_pricing/utils/metrics.py

@@ -22,9 +22,16 @@ import numpy as np
 import pandas as pd
 
 try:
-    from sklearn.metrics import log_loss, mean_tweedie_deviance
+    from sklearn.metrics import (
+        log_loss,
+        mean_absolute_error,
+        mean_squared_error,
+        mean_tweedie_deviance,
+    )
 except ImportError:
     log_loss = None
+    mean_absolute_error = None
+    mean_squared_error = None
     mean_tweedie_deviance = None
 
 
@@ -198,6 +205,7 @@ class MetricFactory:
         self,
         task_type: str = "regression",
         tweedie_power: float = 1.5,
+        loss_name: str = "tweedie",
         clip_min: float = 1e-8,
         clip_max: float = 1 - 1e-8,
     ):
@@ -206,11 +214,13 @@ class MetricFactory:
         Args:
             task_type: Either 'regression' or 'classification'
             tweedie_power: Power parameter for Tweedie deviance (1.0-2.0)
+            loss_name: Regression loss name ('tweedie', 'poisson', 'gamma', 'mse', 'mae')
             clip_min: Minimum value for clipping predictions
             clip_max: Maximum value for clipping predictions (for classification)
         """
         self.task_type = task_type
         self.tweedie_power = tweedie_power
+        self.loss_name = loss_name
         self.clip_min = clip_min
         self.clip_max = clip_max
 
@@ -240,14 +250,28 @@ class MetricFactory:
             y_pred_clipped = np.clip(y_pred, self.clip_min, self.clip_max)
             return float(log_loss(y_true, y_pred_clipped, sample_weight=sample_weight))
 
-        # Regression: use Tweedie deviance
+        loss_name = str(self.loss_name or "tweedie").strip().lower()
+        if loss_name in {"mse", "mae"}:
+            if mean_squared_error is None or mean_absolute_error is None:
+                raise ImportError("sklearn is required for metric computation")
+            if loss_name == "mse":
+                return float(mean_squared_error(
+                    y_true, y_pred, sample_weight=sample_weight))
+            return float(mean_absolute_error(
+                y_true, y_pred, sample_weight=sample_weight))
+
         y_pred_safe = np.maximum(y_pred, self.clip_min)
+        power = self.tweedie_power
+        if loss_name == "poisson":
+            power = 1.0
+        elif loss_name == "gamma":
+            power = 2.0
         return float(
             mean_tweedie_deviance(
                 y_true,
                 y_pred_safe,
                 sample_weight=sample_weight,
-                power=self.tweedie_power,
+                power=power,
             )
         )
 

ins_pricing/utils/torch_compat.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import inspect
+import os
 from typing import Any, Optional
 
 try:
@@ -14,6 +15,7 @@ except ImportError:  # pragma: no cover - handled by callers
     torch = None
 
 _SUPPORTS_WEIGHTS_ONLY: Optional[bool] = None
+_DYNAMO_PATCHED = False
 
 
 def _supports_weights_only() -> bool:
@@ -43,3 +45,41 @@ def torch_load(
     if weights_only is not None and _supports_weights_only():
         return torch.load(path, *args, weights_only=weights_only, **kwargs)
     return torch.load(path, *args, **kwargs)
+
+
+def _env_truthy(key: str) -> bool:
+    value = os.environ.get(key)
+    if value is None:
+        return False
+    return str(value).strip().lower() in {"1", "true", "yes", "y", "on"}
+
+
+def disable_torch_dynamo_if_requested() -> None:
+    """Disable torch._dynamo wrappers when compile is explicitly disabled."""
+    global _DYNAMO_PATCHED
+    if _DYNAMO_PATCHED or not TORCH_AVAILABLE:
+        return
+
+    if not any(
+        _env_truthy(k)
+        for k in (
+            "TORCHDYNAMO_DISABLE",
+            "TORCH_DISABLE_DYNAMO",
+            "TORCH_COMPILE_DISABLE",
+            "TORCHINDUCTOR_DISABLE",
+        )
+    ):
+        return
+
+    try:
+        import torch.optim.optimizer as optim_mod
+    except Exception:
+        return
+
+    for name in ("state_dict", "load_state_dict", "zero_grad", "add_param_group"):
+        fn = getattr(optim_mod.Optimizer, name, None)
+        wrapped = getattr(fn, "__wrapped__", None)
+        if wrapped is not None:
+            setattr(optim_mod.Optimizer, name, wrapped)
+
+    _DYNAMO_PATCHED = True

{ins_pricing-0.3.2.dist-info → ins_pricing-0.3.4.dist-info}/METADATA

@@ -1,162 +1,162 @@
-Metadata-Version: 2.4
-Name: ins_pricing
-Version: 0.3.2
-Summary: Reusable modelling, pricing, governance, and reporting utilities.
-Author: meishi125478
-License: Proprietary
-Keywords: pricing,insurance,bayesopt,ml
-Classifier: Programming Language :: Python :: 3
-Classifier: Programming Language :: Python :: 3 :: Only
-Classifier: Programming Language :: Python :: 3.9
-Classifier: License :: Other/Proprietary License
-Classifier: Operating System :: OS Independent
-Classifier: Intended Audience :: Developers
-Requires-Python: >=3.9
-Description-Content-Type: text/markdown
-Requires-Dist: numpy>=1.20
-Requires-Dist: pandas>=1.4
-Provides-Extra: bayesopt
-Requires-Dist: torch>=1.13; extra == "bayesopt"
-Requires-Dist: optuna>=3.0; extra == "bayesopt"
-Requires-Dist: xgboost>=1.6; extra == "bayesopt"
-Requires-Dist: scikit-learn>=1.1; extra == "bayesopt"
-Requires-Dist: statsmodels>=0.13; extra == "bayesopt"
-Requires-Dist: joblib>=1.2; extra == "bayesopt"
-Requires-Dist: matplotlib>=3.5; extra == "bayesopt"
-Provides-Extra: plotting
-Requires-Dist: matplotlib>=3.5; extra == "plotting"
-Requires-Dist: scikit-learn>=1.1; extra == "plotting"
-Provides-Extra: explain
-Requires-Dist: torch>=1.13; extra == "explain"
-Requires-Dist: shap>=0.41; extra == "explain"
-Requires-Dist: scikit-learn>=1.1; extra == "explain"
-Provides-Extra: geo
-Requires-Dist: contextily>=1.3; extra == "geo"
-Requires-Dist: matplotlib>=3.5; extra == "geo"
-Provides-Extra: gnn
-Requires-Dist: torch>=1.13; extra == "gnn"
-Requires-Dist: pynndescent>=0.5; extra == "gnn"
-Requires-Dist: torch-geometric>=2.3; extra == "gnn"
-Provides-Extra: all
-Requires-Dist: torch>=1.13; extra == "all"
-Requires-Dist: optuna>=3.0; extra == "all"
-Requires-Dist: xgboost>=1.6; extra == "all"
-Requires-Dist: scikit-learn>=1.1; extra == "all"
-Requires-Dist: statsmodels>=0.13; extra == "all"
-Requires-Dist: joblib>=1.2; extra == "all"
-Requires-Dist: matplotlib>=3.5; extra == "all"
-Requires-Dist: shap>=0.41; extra == "all"
-Requires-Dist: contextily>=1.3; extra == "all"
-Requires-Dist: pynndescent>=0.5; extra == "all"
-Requires-Dist: torch-geometric>=2.3; extra == "all"
-
-# Insurance-Pricing
-
-A reusable toolkit for insurance modeling, pricing, governance, and reporting.
-
-## Overview
-
-Insurance-Pricing (ins_pricing) is an enterprise-grade Python library designed for machine learning model training, pricing calculations, and model governance workflows in the insurance industry.
-
-### Core Modules
-
-| Module | Description |
-|--------|-------------|
-| **modelling** | ML model training (GLM, XGBoost, ResNet, FT-Transformer, GNN) and model interpretability (SHAP, permutation importance) |
-| **pricing** | Factor table construction, numeric binning, premium calibration, exposure calculation, PSI monitoring |
-| **production** | Model prediction, batch scoring, data drift detection, production metrics monitoring |
-| **governance** | Model registry, version management, approval workflows, audit logging |
-| **reporting** | Report generation (Markdown format), report scheduling |
-| **utils** | Data validation, performance profiling, device management, logging configuration |
-
-### Quick Start
-
-```python
-# Model training with Bayesian optimization
-from ins_pricing import bayesopt as ropt
-
-model = ropt.BayesOptModel(
-    train_data, test_data,
-    model_name='my_model',
-    resp_nme='target',
-    weight_nme='weight',
-    factor_nmes=feature_list,
-    cate_list=categorical_features,
-)
-model.bayesopt_xgb(max_evals=100)      # Train XGBoost
-model.bayesopt_resnet(max_evals=50)    # Train ResNet
-model.bayesopt_ft(max_evals=50)        # Train FT-Transformer
-
-# Pricing: build factor table
-from ins_pricing.pricing import build_factor_table
-factors = build_factor_table(
-    df,
-    factor_col='age_band',
-    loss_col='claim_amount',
-    exposure_col='exposure',
-)
-
-# Production: batch scoring
-from ins_pricing.production import batch_score
-scores = batch_score(model.trainers['xgb'].predict, df)
-
-# Model governance
-from ins_pricing.governance import ModelRegistry
-registry = ModelRegistry('models.json')
-registry.register(model_name, version, metrics=metrics)
-```
-
-### Project Structure
-
-```
-ins_pricing/
-├── cli/                    # Command-line entry points
-├── modelling/
-│   ├── core/bayesopt/     # ML model training core
-│   ├── explain/           # Model interpretability
-│   └── plotting/          # Model visualization
-├── pricing/               # Insurance pricing module
-├── production/            # Production deployment module
-├── governance/            # Model governance
-├── reporting/             # Report generation
-├── utils/                 # Utilities
-└── tests/                 # Test suite
-```
-
-### Installation
-
-```bash
-# Basic installation
-pip install ins_pricing
-
-# Full installation (all optional dependencies)
-pip install ins_pricing[all]
-
-# Install specific extras
-pip install ins_pricing[bayesopt]    # Model training
-pip install ins_pricing[explain]     # Model explanation
-pip install ins_pricing[plotting]    # Visualization
-pip install ins_pricing[gnn]         # Graph neural networks
-```
-
-#### Multi-platform & GPU installation notes
-
-- **PyTorch (CPU/GPU/MPS)**: Install the correct PyTorch build for your platform/GPU first (CUDA on
-  Linux/Windows, ROCm on supported AMD platforms, or MPS on Apple Silicon). Then install the
-  optional extras you need (e.g., `bayesopt`, `explain`, or `gnn`). This avoids pip pulling a
-  mismatched wheel.
-- **Torch Geometric (GNN)**: `torch-geometric` often requires platform-specific wheels (e.g.,
-  `torch-scatter`, `torch-sparse`). Follow the official PyG installation instructions for your
-  CUDA/ROCm/CPU environment, then install `ins_pricing[gnn]`.
-- **Multi-GPU**: Training code will use CUDA when available and can enable multi-GPU via
-  `torch.distributed`/`DataParallel` where supported. On Windows, CUDA DDP is not supported and will
-  fall back to single-GPU or DataParallel where possible.
-
-### Requirements
-
-- Python >= 3.9
-- Core dependencies: numpy >= 1.20, pandas >= 1.4
-
-### License
-
-Proprietary
+Metadata-Version: 2.4
+Name: ins_pricing
+Version: 0.3.4
+Summary: Reusable modelling, pricing, governance, and reporting utilities.
+Author: meishi125478
+License: Proprietary
+Keywords: pricing,insurance,bayesopt,ml
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Programming Language :: Python :: 3.9
+Classifier: License :: Other/Proprietary License
+Classifier: Operating System :: OS Independent
+Classifier: Intended Audience :: Developers
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+Requires-Dist: numpy>=1.20
+Requires-Dist: pandas>=1.4
+Provides-Extra: bayesopt
+Requires-Dist: torch>=1.13; extra == "bayesopt"
+Requires-Dist: optuna>=3.0; extra == "bayesopt"
+Requires-Dist: xgboost>=1.6; extra == "bayesopt"
+Requires-Dist: scikit-learn>=1.1; extra == "bayesopt"
+Requires-Dist: statsmodels>=0.13; extra == "bayesopt"
+Requires-Dist: joblib>=1.2; extra == "bayesopt"
+Requires-Dist: matplotlib>=3.5; extra == "bayesopt"
+Provides-Extra: plotting
+Requires-Dist: matplotlib>=3.5; extra == "plotting"
+Requires-Dist: scikit-learn>=1.1; extra == "plotting"
+Provides-Extra: explain
+Requires-Dist: torch>=1.13; extra == "explain"
+Requires-Dist: shap>=0.41; extra == "explain"
+Requires-Dist: scikit-learn>=1.1; extra == "explain"
+Provides-Extra: geo
+Requires-Dist: contextily>=1.3; extra == "geo"
+Requires-Dist: matplotlib>=3.5; extra == "geo"
+Provides-Extra: gnn
+Requires-Dist: torch>=1.13; extra == "gnn"
+Requires-Dist: pynndescent>=0.5; extra == "gnn"
+Requires-Dist: torch-geometric>=2.3; extra == "gnn"
+Provides-Extra: all
+Requires-Dist: torch>=1.13; extra == "all"
+Requires-Dist: optuna>=3.0; extra == "all"
+Requires-Dist: xgboost>=1.6; extra == "all"
+Requires-Dist: scikit-learn>=1.1; extra == "all"
+Requires-Dist: statsmodels>=0.13; extra == "all"
+Requires-Dist: joblib>=1.2; extra == "all"
+Requires-Dist: matplotlib>=3.5; extra == "all"
+Requires-Dist: shap>=0.41; extra == "all"
+Requires-Dist: contextily>=1.3; extra == "all"
+Requires-Dist: pynndescent>=0.5; extra == "all"
+Requires-Dist: torch-geometric>=2.3; extra == "all"
+
+# Insurance-Pricing
+
+A reusable toolkit for insurance modeling, pricing, governance, and reporting.
+
+## Overview
+
+Insurance-Pricing (ins_pricing) is an enterprise-grade Python library designed for machine learning model training, pricing calculations, and model governance workflows in the insurance industry.
+
+### Core Modules
+
+| Module | Description |
+|--------|-------------|
+| **modelling** | ML model training (GLM, XGBoost, ResNet, FT-Transformer, GNN) and model interpretability (SHAP, permutation importance) |
+| **pricing** | Factor table construction, numeric binning, premium calibration, exposure calculation, PSI monitoring |
+| **production** | Model prediction, batch scoring, data drift detection, production metrics monitoring |
+| **governance** | Model registry, version management, approval workflows, audit logging |
+| **reporting** | Report generation (Markdown format), report scheduling |
+| **utils** | Data validation, performance profiling, device management, logging configuration |
+
+### Quick Start
+
+```python
+# Model training with Bayesian optimization
+from ins_pricing import bayesopt as ropt
+
+model = ropt.BayesOptModel(
+    train_data, test_data,
+    model_name='my_model',
+    resp_nme='target',
+    weight_nme='weight',
+    factor_nmes=feature_list,
+    cate_list=categorical_features,
+)
+model.bayesopt_xgb(max_evals=100)      # Train XGBoost
+model.bayesopt_resnet(max_evals=50)    # Train ResNet
+model.bayesopt_ft(max_evals=50)        # Train FT-Transformer
+
+# Pricing: build factor table
+from ins_pricing.pricing import build_factor_table
+factors = build_factor_table(
+    df,
+    factor_col='age_band',
+    loss_col='claim_amount',
+    exposure_col='exposure',
+)
+
+# Production: batch scoring
+from ins_pricing.production import batch_score
+scores = batch_score(model.trainers['xgb'].predict, df)
+
+# Model governance
+from ins_pricing.governance import ModelRegistry
+registry = ModelRegistry('models.json')
+registry.register(model_name, version, metrics=metrics)
+```
+
+### Project Structure
+
+```
+ins_pricing/
+├── cli/                    # Command-line entry points
+├── modelling/
+│   ├── core/bayesopt/     # ML model training core
+│   ├── explain/           # Model interpretability
+│   └── plotting/          # Model visualization
+├── pricing/               # Insurance pricing module
+├── production/            # Production deployment module
+├── governance/            # Model governance
+├── reporting/             # Report generation
+├── utils/                 # Utilities
+└── tests/                 # Test suite
+```
+
+### Installation
+
+```bash
+# Basic installation
+pip install ins_pricing
+
+# Full installation (all optional dependencies)
+pip install ins_pricing[all]
+
+# Install specific extras
+pip install ins_pricing[bayesopt]    # Model training
+pip install ins_pricing[explain]     # Model explanation
+pip install ins_pricing[plotting]    # Visualization
+pip install ins_pricing[gnn]         # Graph neural networks
+```
+
+#### Multi-platform & GPU installation notes
+
+- **PyTorch (CPU/GPU/MPS)**: Install the correct PyTorch build for your platform/GPU first (CUDA on
+  Linux/Windows, ROCm on supported AMD platforms, or MPS on Apple Silicon). Then install the
+  optional extras you need (e.g., `bayesopt`, `explain`, or `gnn`). This avoids pip pulling a
+  mismatched wheel.
+- **Torch Geometric (GNN)**: `torch-geometric` often requires platform-specific wheels (e.g.,
+  `torch-scatter`, `torch-sparse`). Follow the official PyG installation instructions for your
+  CUDA/ROCm/CPU environment, then install `ins_pricing[gnn]`.
+- **Multi-GPU**: Training code will use CUDA when available and can enable multi-GPU via
+  `torch.distributed`/`DataParallel` where supported. On Windows, CUDA DDP is not supported and will
+  fall back to single-GPU or DataParallel where possible.
+
+### Requirements
+
+- Python >= 3.9
+- Core dependencies: numpy >= 1.20, pandas >= 1.4
+
+### License
+
+Proprietary