ins-pricing 0.4.3__py3-none-any.whl → 0.4.5__py3-none-any.whl

ins_pricing/frontend/config_builder.py

@@ -32,6 +32,7 @@ class ConfigBuilder:
             "save_preprocess": False,
             "preprocess_artifact_path": None,
             "bo_sample_limit": None,
+            "build_oht": True,
             "cache_predictions": False,
             "prediction_cache_dir": None,
             "prediction_cache_format": "parquet",

ins_pricing/modelling/README.md

@@ -0,0 +1,67 @@
+# Modelling
+
+This directory contains reusable training tooling and frameworks centered on BayesOpt.
+
+## Key locations
+
+- `core/bayesopt/` - core training/tuning package
+- `explain/` - explainability helpers
+- `plotting/` - plotting utilities
+- `ins_pricing/cli/` - CLI entry points
+- `examples/` - example configs and notebooks (repo only)
+
+## Common usage
+
+- CLI training: `python ins_pricing/cli/BayesOpt_entry.py --config-json config_template.json`
+- Notebook API: `from ins_pricing.modelling import BayesOptModel`
+
+## Explainability
+
+- CLI: `python ins_pricing/cli/Explain_entry.py --config-json config_explain_template.json`
+- Notebook: `examples/04 Explain_Run.ipynb`
+
+## Loss functions
+
+Configure the regression/classification loss with `loss_name` in the BayesOpt config.
+
+Supported `loss_name` values:
+- `auto` (default): legacy behavior based on model name
+- `tweedie`: Tweedie deviance
+- `poisson`: Poisson deviance
+- `gamma`: Gamma deviance
+- `mse`: mean squared error
+- `mae`: mean absolute error
+
+Mapping summary:
+- Tweedie deviance -> `tweedie`
+- Poisson deviance -> `poisson`
+- Gamma deviance -> `gamma`
+- Mean squared error -> `mse`
+- Mean absolute error -> `mae`
+- Classification log loss -> `logloss` (classification only)
+- Classification BCE -> `bce` (classification only)
+
+Classification tasks:
+- `loss_name` can be `auto`, `logloss`, or `bce`.
+- Training uses `BCEWithLogits` for torch models; evaluation uses log loss.
+
+Where to set `loss_name`:
+
+```json
+{
+  "task_type": "regression",
+  "loss_name": "mse"
+}
+```
+
+Behavior notes:
+- When `loss_name` is `mse` or `mae`, tuning does not sample Tweedie power.
+- When `loss_name` is `poisson` or `gamma`, power is fixed (1.0 / 2.0).
+- When `loss_name` is `tweedie`, power is sampled as usual.
+- XGBoost objective is selected from the loss name.
+
+## Notes
+
+- Models load from `output_dir/model` by default (override with `explain.model_dir`).
+- Training outputs are written to `plot/`, `Results/`, and `model/` under `output_dir`.
+- Keep large data and secrets outside the repo; use environment variables or `.env` files.

ins_pricing/modelling/core/bayesopt/README.md

@@ -0,0 +1,59 @@
+# BayesOpt
+
+BayesOpt is the training/tuning core for GLM, XGBoost, ResNet, FT-Transformer, and GNN workflows.
+It supports JSON-driven CLI runs and a Python API for notebooks/scripts.
+
+## Recommended API (config-based)
+
+```python
+from ins_pricing.modelling.core.bayesopt import BayesOptConfig
+from ins_pricing.modelling import BayesOptModel
+
+config = BayesOptConfig(
+    model_nme="my_model",
+    resp_nme="target",
+    weight_nme="weight",
+    factor_nmes=["f1", "f2"],
+    cate_list=["f2"],
+    task_type="regression",
+    epochs=50,
+    output_dir="./Results",
+)
+
+model = BayesOptModel(train_data, test_data, config=config)
+model.optimize_model("xgb", max_evals=50)
+```
+
+## Load config from file
+
+```python
+from ins_pricing.modelling.core.bayesopt import BayesOptConfig
+from ins_pricing.modelling import BayesOptModel
+
+config = BayesOptConfig.from_file("config.json")
+model = BayesOptModel(train_data, test_data, config=config)
+```
+
+## CLI entry
+
+```bash
+python ins_pricing/cli/BayesOpt_entry.py --config-json config_template.json
+```
+
+## FT roles
+
+- `model`: FT is a prediction model (writes `pred_ft`).
+- `embedding`: FT trains with labels but exports embeddings (`pred_<prefix>_*`).
+- `unsupervised_embedding`: FT trains without labels and exports embeddings.
+
+## Output layout
+
+`output_dir/` contains:
+- `plot/` plots and diagnostics
+- `Results/` metrics, params, and snapshots
+- `model/` saved models
+
+## Notes
+
+- Relative paths in config are resolved from the config file directory.
+- For multi-GPU, use `torchrun` and set `runner.nproc_per_node` in config.

ins_pricing/modelling/core/bayesopt/config_preprocess.py

@@ -97,6 +97,7 @@ class BayesOptConfig:
         use_gnn_ddp: Use DDP for GNN
         ft_role: FT-Transformer role ('model', 'embedding', 'unsupervised_embedding')
         cv_strategy: CV strategy ('random', 'group', 'time', 'stratified')
+        build_oht: Whether to build one-hot encoded features (default True)
 
     Example:
         >>> config = BayesOptConfig(
@@ -192,6 +193,7 @@ class BayesOptConfig:
     preprocess_artifact_path: Optional[str] = None
     plot_path_style: str = "nested"
     bo_sample_limit: Optional[int] = None
+    build_oht: bool = True
     cache_predictions: bool = False
     prediction_cache_dir: Optional[str] = None
     prediction_cache_format: str = "parquet"
@@ -465,6 +467,16 @@ class DatasetPreprocessor:
         self.num_features = [
             nme for nme in cfg.factor_nmes if nme not in cate_list]
 
+        build_oht = bool(getattr(cfg, "build_oht", True))
+        if not build_oht:
+            print("[Preprocess] build_oht=False; skip one-hot features.", flush=True)
+            self.train_oht_data = None
+            self.test_oht_data = None
+            self.train_oht_scl_data = None
+            self.test_oht_scl_data = None
+            self.var_nmes = list(cfg.factor_nmes)
+            return self
+
         # Memory optimization: Single copy + in-place operations
         train_oht = self.train_data[cfg.factor_nmes +
                                     [cfg.weight_nme] + [cfg.resp_nme]].copy()

ins_pricing/modelling/core/bayesopt/core.py

@@ -201,6 +201,8 @@ class BayesOptModel(BayesOptPlottingMixin, BayesOptExplainMixin):
                 raise ValueError("weight_nme is required when not using config parameter")
 
             # Infer categorical features if needed
+            # Only use user-specified categorical list for one-hot; do not auto-infer.
+            user_cate_list = [] if cate_list is None else list(cate_list)
             inferred_factors, inferred_cats = infer_factor_and_cate_list(
                 train_df=train_data,
                 test_df=test_data,
@@ -208,7 +210,7 @@ class BayesOptModel(BayesOptPlottingMixin, BayesOptExplainMixin):
                 weight_nme=weight_nme,
                 binary_resp_nme=binary_resp_nme,
                 factor_nmes=factor_nmes,
-                cate_list=cate_list,
+                cate_list=user_cate_list,
                 infer_categorical_max_unique=int(infer_categorical_max_unique),
                 infer_categorical_max_ratio=float(infer_categorical_max_ratio),
             )