ins-pricing 0.4.4__py3-none-any.whl → 0.4.5__py3-none-any.whl

ins_pricing/CHANGELOG.md

@@ -1,272 +0,0 @@
-# Changelog
-
-All notable changes to the ins_pricing project will be documented in this file.
-
-The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
-and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
-
-## [0.2.11] - 2026-01-15
-
-### Changed
-
-#### Refactoring Phase 3: Utils Module Consolidation
-- **Eliminated code duplication** - Consolidated duplicated utility classes:
-  - `DeviceManager` and `GPUMemoryManager` now imported from `ins_pricing.utils`
-  - Removed 181 lines of duplicate code from `bayesopt/utils/metrics_and_devices.py`
-  - File size reduced from 721 to 540 lines (25% reduction)
-  - **Benefit**: Single source of truth for device management utilities
-  - **Impact**: Bug fixes now propagate automatically, no risk of code drift
-  - **Compatibility**: 100% backward compatible - all import patterns continue working
-
-**Technical Details**:
-- Package-level `ins_pricing/utils/device.py` is now the canonical implementation
-- BayesOpt utils automatically re-export these classes for backward compatibility
-- No breaking changes required in existing code
-
-## [0.2.10] - 2026-01-15
-
-### Added
-
-#### Refactoring Phase 2: Simplified BayesOptModel API
-- **BayesOptModel config-based initialization** - New recommended API using configuration objects:
-  - Added `config` parameter accepting `BayesOptConfig` instances
-  - **Before**: 56 individual parameters required
-  - **After**: Single config object parameter
-  - **Benefits**: Improved code clarity, reusability, type safety, and testability
-
-### Changed
-
-#### API Improvements
-- **BayesOptModel initialization** - Enhanced parameter handling:
-  - New API: `BayesOptModel(train_df, test_df, config=BayesOptConfig(...))`
-  - Old API still supported with deprecation warning
-  - Made `model_nme`, `resp_nme`, `weight_nme` optional (validated when config=None)
-  - Added type validation for config parameter
-  - Added helpful error messages for missing required parameters
-
-### Deprecated
-
-- **BayesOptModel individual parameters** - Passing 56 individual parameters to `__init__`:
-  - Use `config=BayesOptConfig(...)` instead
-  - Old API will be removed in v0.4.0
-  - Migration guide: See `modelling/core/bayesopt/PHASE2_REFACTORING_SUMMARY.md`
-
-### Fixed
-
-- **Type hints** - Improved type safety in BayesOptModel initialization
-- **Documentation** - Added comprehensive examples of both old and new APIs
-
-## [0.2.9] - 2026-01-15
-
-### Added
-
-#### Refactoring Phase 1: Utils Module Split
-- **Modular utils package** - Split monolithic 1,503-line utils.py into focused modules:
-  - `utils/constants.py` (183 lines) - Core constants and simple helpers
-  - `utils/io_utils.py` (110 lines) - File I/O and parameter loading
-  - `utils/distributed_utils.py` (163 lines) - DDP and CUDA management
-  - `utils/torch_trainer_mixin.py` (587 lines) - PyTorch training infrastructure
-  - `utils/metrics_and_devices.py` (721 lines) - Metrics, GPU, device, CV, plotting
-  - `utils/__init__.py` (86 lines) - Backward compatibility re-exports
-
-- **Upload automation** - Cross-platform PyPI upload scripts:
-  - `upload_to_pypi.sh` - Shell script for Linux/macOS with auto-version extraction
-  - `upload_to_pypi.bat` - Updated Windows batch script with auto-version extraction
-  - `Makefile` - Cross-platform build automation (build, check, upload, clean)
-  - `README_UPLOAD.md` - Comprehensive upload documentation in English
-  - `UPLOAD_QUICK_START.md` - Quick start guide for package publishing
-
-### Changed
-
-#### Code Organization
-- **utils module structure** - Improved maintainability and testability:
-  - Average file size reduced from 1,503 to 351 lines per module
-  - Each module has single responsibility
-  - Independent testing now possible for each component
-  - **Impact**: 100% backward compatibility maintained via re-exports
-
-### Deprecated
-
-- **utils.py single file import** - Direct import from `bayesopt/utils.py`:
-  - Use `from .utils import ...` instead (package import)
-  - Old single-file import shows deprecation warning
-  - File will be removed in v0.4.0
-  - **Note**: All imports continue to work identically
-
-### Removed
-
-- **verify_core_decoupling.py** - Obsolete test script for unimplemented refactoring
-  - Cleanup logged in `.cleanup_log.md`
-
-## [0.2.8] - 2026-01-14
-
-### Added
-
-#### New Utility Modules
-- **utils/validation.py** - Comprehensive data validation toolkit with 8 validation functions:
-  - `validate_required_columns()` - Validate required DataFrame columns
-  - `validate_column_types()` - Validate and optionally coerce column types
-  - `validate_value_range()` - Validate numeric value ranges
-  - `validate_no_nulls()` - Check for null values
-  - `validate_categorical_values()` - Validate categorical values against allowed set
-  - `validate_positive()` - Ensure positive numeric values
-  - `validate_dataframe_not_empty()` - Check DataFrame is not empty
-  - `validate_date_range()` - Validate date ranges
-
-- **utils/profiling.py** - Performance profiling and memory monitoring utilities:
-  - `profile_section()` - Context manager for execution time and memory tracking
-  - `get_memory_info()` - Get current memory usage statistics
-  - `log_memory_usage()` - Log memory usage with custom prefix
-  - `check_memory_threshold()` - Check if memory exceeds threshold
-  - `cleanup_memory()` - Force memory cleanup for CPU and GPU
-  - `MemoryMonitor` - Context manager with automatic cleanup
-  - `profile_training_epoch()` - Periodic memory profiling during training
-
-- **pricing/factors.py** - LRU caching for binning operations:
-  - `_compute_bins_cached()` - Cached bin edge computation (maxsize=128)
-  - `clear_binning_cache()` - Clear binning cache
-  - `get_cache_info()` - Get cache statistics (hits, misses, size)
-  - Enhanced `bin_numeric()` with `use_cache` parameter
-
-#### Test Coverage Expansion
-- **tests/production/** - Complete production module test suite (4 files, 247 test scenarios):
-  - `test_predict.py` - Prediction and model loading tests (87 scenarios)
-  - `test_scoring.py` - Scoring metrics validation (60 scenarios)
-  - `test_monitoring.py` - Drift detection and monitoring (55 scenarios)
-  - `test_preprocess.py` - Preprocessing pipeline tests (45 scenarios)
-
-- **tests/pricing/** - Pricing module test suite (4 files):
-  - `test_factors.py` - Factor table construction and binning
-  - `test_exposure.py` - Exposure calculation tests
-  - `test_calibration.py` - Calibration factor fitting tests
-  - `test_rate_table.py` - Rate table generation tests
-
-- **tests/governance/** - Governance workflow test suite (3 files):
-  - `test_registry.py` - Model registry operations
-  - `test_release.py` - Release management and rollback
-  - `test_audit.py` - Audit logging and trail verification
-
-### Enhanced
-
-#### SHAP Computation Parallelization
-- **modelling/explain/shap_utils.py** - Added parallel SHAP computation:
-  - `_compute_shap_parallel()` - Parallel SHAP value computation using joblib
-  - All SHAP functions now support `use_parallel` and `n_jobs` parameters:
-    - `compute_shap_glm()` - GLM model SHAP with parallelization
-    - `compute_shap_xgb()` - XGBoost model SHAP with parallelization
-    - `compute_shap_resn()` - ResNet model SHAP with parallelization
-    - `compute_shap_ft()` - FT-Transformer model SHAP with parallelization
-  - Automatic batch size optimization based on CPU cores
-  - **Performance**: 3-6x speedup on multi-core systems (n_samples > 100)
-  - Graceful fallback to sequential computation if joblib unavailable
-
-#### Documentation Improvements
-- **production/preprocess.py** - Complete documentation overhaul:
-  - Module-level docstring with workflow explanation and examples
-  - `load_preprocess_artifacts()` - Full parameter and return value documentation
-  - `prepare_raw_features()` - Detailed data preparation steps and examples
-  - `apply_preprocess_artifacts()` - Complete preprocessing pipeline documentation
-
-- **pricing/calibration.py** - Comprehensive documentation:
-  - Module-level docstring with business context and use cases
-  - `fit_calibration_factor()` - Mathematical formulas, multiple examples, business guidance
-  - `apply_calibration()` - Usage examples showing ratio preservation
-
-#### Configuration Validation
-- **modelling/core/bayesopt/config_preprocess.py** - BayesOptConfig validation already comprehensive:
-  - Task type validation
-  - Parameter range validation
-  - Distributed training conflict detection
-  - Cross-validation strategy validation
-  - GNN memory settings validation
-
-### Performance Improvements
-
-- **Memory optimization** - DatasetPreprocessor reduces unnecessary DataFrame copies:
-  - Conditional copying only when scaling needed
-  - Direct reference assignment where safe
-  - **Impact**: 30-40% reduction in memory usage during preprocessing
-
-- **Binning cache** - LRU cache for factor table binning operations:
-  - Cache size: 128 entries
-  - **Impact**: 5-10x speedup for repeated binning of same columns
-
-- **SHAP parallelization** - Multi-core SHAP value computation:
-  - **Impact**: 3-6x speedup depending on CPU cores and sample size
-  - Automatic batch size tuning
-  - Memory-efficient batch processing
-
-### Fixed
-
-- **Distributed training** - State dict key mismatch issues already resolved in previous versions:
-  - model_ft_trainer.py: Lines 409, 738
-  - model_resn.py: Line 405
-  - utils.py: Line 796
-
-### Technical Debt
-
-- Custom exception hierarchy fully implemented in `exceptions.py`:
-  - `InsPricingError` - Base exception
-  - `ConfigurationError` - Invalid configuration
-  - `DataValidationError` - Data validation failures
-  - `ModelLoadError` - Model loading failures
-  - `DistributedTrainingError` - DDP/DataParallel errors
-  - `PreprocessingError` - Preprocessing failures
-  - `PredictionError` - Prediction failures
-  - `GovernanceError` - Governance workflow errors
-
-### Testing
-
-- **Test coverage increase**: From 35% → 60%+ (estimated)
-  - 250+ new test scenarios across 11 test files
-  - Coverage for previously untested modules: production, pricing, governance
-  - Integration tests for end-to-end workflows
-
-### Documentation
-
-- **Docstring coverage**: 0% → 95% for improved modules
-  - 150+ lines of new documentation
-  - 8+ complete code examples
-  - Business context and use case explanations
-  - Parameter constraints and edge case documentation
-
----
-
-## [0.2.7] - Previous Release
-
-(Previous changelog entries would go here)
-
----
-
-## Release Notes for 0.2.8
-
-This release focuses on **code quality, performance optimization, and documentation** improvements. Major highlights:
-
-### 🚀 Performance
-- **3-6x faster SHAP computation** with parallel processing
-- **30-40% memory reduction** in preprocessing
-- **5-10x faster binning** with LRU cache
-
-### 📚 Documentation
-- **Complete module documentation** for production and pricing modules
-- **150+ lines of new documentation** with practical examples
-- **Business context** explanations for insurance domain
-
-### 🧪 Testing
-- **250+ new test scenarios** across 11 test files
-- **60%+ test coverage** (up from 35%)
-- **Complete coverage** for production, pricing, governance modules
-
-### 🛠️ Developer Experience
-- **Comprehensive validation toolkit** for data quality checks
-- **Performance profiling utilities** for optimization
-- **Enhanced error messages** with clear troubleshooting guidance
-
-### Migration Notes
-- All changes are **backward compatible**
-- New features are **opt-in** (e.g., `use_parallel=True`)
-- No breaking changes to existing APIs
-
-### Dependencies
-- Optional: `joblib>=1.2` for parallel SHAP computation
-- Optional: `psutil` for memory profiling utilities

ins_pricing/RELEASE_NOTES_0.2.8.md

@@ -1,344 +0,0 @@
-# Release Notes: ins_pricing v0.2.8
-
-**Release Date:** January 14, 2026
-**Type:** Minor Release (Quality & Performance Improvements)
-
----
-
-## 🎯 Overview
-
-Version 0.2.8 is a significant quality and performance improvement release that focuses on:
-- **Code quality and maintainability**
-- **Performance optimization** (3-6x faster SHAP, 30-40% memory reduction)
-- **Comprehensive documentation**
-- **Extensive test coverage** (35% → 60%+)
-
-**All changes are backward compatible.** No breaking changes.
-
----
-
-## ⭐ Highlights
-
-### 1. 🚀 Performance Optimizations
-
-#### SHAP Parallelization (3-6x Speedup)
-```python
-# Before (slow - serial processing)
-result = compute_shap_xgb(ctx, n_samples=200)  # ~10 minutes
-
-# After (fast - parallel processing)
-result = compute_shap_xgb(ctx, n_samples=200, use_parallel=True)  # ~2 minutes
-```
-**Impact:** 3-6x faster on multi-core systems for n_samples > 100
-
-#### Memory Optimization (30-40% Reduction)
-- DatasetPreprocessor reduces unnecessary DataFrame copies
-- Conditional copying only when needed
-- Direct reference assignment where safe
-
-#### Binning Cache (5-10x Speedup)
-```python
-from ins_pricing.pricing.factors import get_cache_info, clear_binning_cache
-
-# Automatic caching for repeated binning
-factor_table = build_factor_table(df, factor_col='age', n_bins=10)  # Cached!
-
-# Check cache performance
-info = get_cache_info()
-print(f"Cache hit rate: {info['hits'] / (info['hits'] + info['misses']):.1%}")
-```
-
----
-
-### 2. 🛠️ New Utility Modules
-
-#### Data Validation Toolkit
-```python
-from ins_pricing.utils.validation import (
-    validate_required_columns,
-    validate_column_types,
-    validate_value_range,
-    validate_no_nulls,
-    validate_positive
-)
-
-# Validate DataFrame structure
-validate_required_columns(df, ['age', 'premium', 'exposure'], df_name='policy_data')
-
-# Validate data types
-df = validate_column_types(df, {'age': 'int64', 'premium': 'float64'}, coerce=True)
-
-# Validate value ranges
-validate_value_range(df, 'age', min_val=0, max_val=120)
-validate_positive(df, ['premium', 'exposure'], allow_zero=False)
-```
-
-#### Performance Profiling
-```python
-from ins_pricing.utils.profiling import profile_section, MemoryMonitor
-
-# Simple profiling
-with profile_section("Data Processing", logger):
-    process_large_dataset()
-# Output: [Profile] Data Processing: 5.23s, RAM: +1250.3MB, GPU peak: 2048.5MB
-
-# Memory monitoring with auto-cleanup
-with MemoryMonitor("Training", threshold_gb=16.0, logger=logger):
-    train_model()
-```
-
----
-
-### 3. 📚 Documentation Overhaul
-
-#### Complete Module Documentation
-- **production/preprocess.py**: Module + 3 functions fully documented
-- **pricing/calibration.py**: Module + 2 functions with business context
-- All docs include practical examples and business rationale
-
-#### Example Quality
-```python
-def fit_calibration_factor(pred, actual, *, weight=None, target_lr=None):
-    """Fit a scalar calibration factor to align predictions with actuals.
-
-    This function computes a multiplicative calibration factor...
-
-    Args:
-        pred: Model predictions (premiums or pure premiums)
-        actual: Actual observed values (claims or losses)
-        weight: Optional weights (e.g., exposure, earned premium)
-        target_lr: Target loss ratio to achieve (0 < target_lr < 1)
-
-    Returns:
-        Calibration factor (scalar multiplier)
-
-    Example:
-        >>> # Calibrate to achieve 70% loss ratio
-        >>> pred_premium = np.array([100, 150, 200])
-        >>> actual_claims = np.array([75, 100, 130])
-        >>> factor = fit_calibration_factor(pred_premium, actual_claims, target_lr=0.70)
-        >>> print(f"{factor:.3f}")
-        1.143  # Adjust premiums to achieve 70% loss ratio
-
-    Note:
-        - target_lr typically in range [0.5, 0.9] for insurance pricing
-    """
-```
-
----
-
-### 4. 🧪 Test Coverage Expansion
-
-#### New Test Suites
-- **tests/production/** (247 scenarios)
-  - Prediction, scoring, monitoring, preprocessing
-- **tests/pricing/** (60+ scenarios)
-  - Factors, exposure, calibration, rate tables
-- **tests/governance/** (40+ scenarios)
-  - Registry, release, audit workflows
-
-#### Coverage Increase
-- **Before:** 35% overall coverage
-- **After:** 60%+ overall coverage
-- **Impact:** Better reliability, fewer production bugs
-
----
-
-## 📦 What's New
-
-### Added
-
-#### Core Utilities
-- `utils/validation.py` - 8 validation functions for data quality
-- `utils/profiling.py` - Performance and memory monitoring tools
-- `pricing/factors.py` - LRU caching for binning operations
-
-#### Test Coverage
-- 11 new test files with 250+ test scenarios
-- Complete coverage for production, pricing, governance modules
-
-#### Documentation
-- Module-level docstrings with business context
-- 150+ lines of comprehensive documentation
-- 8+ complete working examples
-
-### Enhanced
-
-#### SHAP Computation
-- Parallel processing support via joblib
-- Automatic batch size optimization
-- Graceful fallback if joblib unavailable
-- All SHAP functions support `use_parallel=True`
-
-#### Configuration Validation
-- BayesOptConfig with comprehensive `__post_init__` validation
-- Clear error messages for configuration issues
-- Validation of distributed training settings
-
-### Performance
-
-| Feature | Before | After | Improvement |
-|---------|--------|-------|-------------|
-| SHAP (200 samples) | 10 min | 2-3 min | **3-6x faster** |
-| Preprocessing memory | 2.5 GB | 1.5 GB | **40% reduction** |
-| Repeated binning | 5.2s | 0.5s | **10x faster** |
-
----
-
-## 🔄 Migration Guide
-
-### No Breaking Changes
-
-All changes are **backward compatible**. Existing code will continue to work without modifications.
-
-### Opt-in Features
-
-New features are opt-in and don't affect existing behavior:
-
-```python
-# SHAP parallelization - opt-in
-result = compute_shap_xgb(ctx, use_parallel=True)  # New parameter
-
-# Binning cache - automatic, but can be disabled
-binned = bin_numeric(series, bins=10, use_cache=False)  # Opt-out if needed
-```
-
-### Recommended Updates
-
-While not required, consider adopting these improvements:
-
-#### 1. Enable Parallel SHAP (if using SHAP)
-```python
-# Before
-shap_result = compute_shap_xgb(ctx, n_samples=200)
-
-# After (recommended for n_samples > 100)
-shap_result = compute_shap_xgb(ctx, n_samples=200, use_parallel=True, n_jobs=-1)
-```
-
-#### 2. Add Data Validation (for production code)
-```python
-from ins_pricing.utils.validation import validate_required_columns, validate_positive
-
-def score_policies(df):
-    # Add validation at entry points
-    validate_required_columns(df, ['age', 'premium', 'exposure'], df_name='input_data')
-    validate_positive(df, ['premium', 'exposure'])
-
-    # Your existing code...
-```
-
-#### 3. Use Profiling (for optimization)
-```python
-from ins_pricing.utils.profiling import profile_section
-
-def expensive_operation():
-    with profile_section("Data Processing"):
-        # Your code...
-```
-
----
-
-## 📋 Installation
-
-### Standard Installation
-```bash
-pip install ins_pricing==0.2.8
-```
-
-### With Optional Dependencies
-```bash
-# For parallel SHAP computation
-pip install "ins_pricing[explain]==0.2.8"
-
-# For memory profiling
-pip install psutil
-
-# All features
-pip install "ins_pricing[all]==0.2.8" psutil
-```
-
----
-
-## 🔧 Dependencies
-
-### New Optional Dependencies
-- `joblib>=1.2` - For parallel SHAP computation (optional)
-- `psutil` - For memory profiling utilities (optional)
-
-### Unchanged Core Dependencies
-- `numpy>=1.20`
-- `pandas>=1.4`
-- All existing optional dependencies remain the same
-
----
-
-## 🐛 Known Issues
-
-None identified in this release.
-
----
-
-## 🔮 What's Next (v0.2.9)
-
-Planned improvements for the next release:
-
-1. **Governance Module Documentation** - Complete docs for registry, approval, release modules
-2. **Plotting Module Documentation** - Enhanced visualization guidance
-3. **CI/CD Pipeline** - Automated testing and code quality checks
-4. **Additional Performance Optimizations** - Vectorized operations in pricing modules
-
----
-
-## 📊 Metrics Summary
-
-| Metric | Before | After | Change |
-|--------|--------|-------|--------|
-| **Test Coverage** | 35% | 60%+ | +25% ✅ |
-| **Documentation Coverage** | ~40% | ~70% | +30% ✅ |
-| **SHAP Performance** | 1x | 3-6x | +3-6x ✅ |
-| **Memory Usage** | 100% | 60-70% | -30-40% ✅ |
-| **Binning Performance** | 1x | 5-10x | +5-10x ✅ |
-
----
-
-## 🙏 Acknowledgments
-
-This release includes comprehensive code review findings and implements best practices for:
-- Performance optimization
-- Memory management
-- Code documentation
-- Test coverage
-- Developer experience
-
----
-
-## 📞 Support
-
-For issues or questions about this release:
-1. Check the [CHANGELOG.md](CHANGELOG.md) for detailed changes
-2. Review module documentation in updated files
-3. Check test files for usage examples
-
----
-
-## ✅ Upgrade Checklist
-
-Before upgrading to 0.2.8:
-
-- [ ] Review [CHANGELOG.md](CHANGELOG.md) for all changes
-- [ ] No breaking changes - safe to upgrade
-- [ ] Consider enabling parallel SHAP if using SHAP
-- [ ] Consider adding data validation for production workflows
-- [ ] Install optional dependencies if needed: `pip install joblib psutil`
-
-After upgrading:
-
-- [ ] Verify existing functionality still works
-- [ ] Consider adopting new validation utilities
-- [ ] Consider adding performance profiling
-- [ ] Review new test examples for your use cases
-
----
-
-**Happy modeling! 🎉**

ins_pricing/docs/LOSS_FUNCTIONS.md

@@ -1,78 +0,0 @@
-LOSS FUNCTIONS
-
-Overview
-This document describes the loss-function changes in ins_pricing. The training
-stack now supports multiple regression losses (not just Tweedie deviance) and
-propagates the selected loss into tuning, training, and inference.
-
-Supported loss_name values
-- auto (default): keep legacy behavior based on model name
-- tweedie: Tweedie deviance (uses tw_power / tweedie_variance_power when tuning)
-- poisson: Poisson deviance (power=1)
-- gamma: Gamma deviance (power=2)
-- mse: mean squared error
-- mae: mean absolute error
-
-Loss name mapping (all options)
-- 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 continues to use BCEWithLogits for torch models; evaluation uses logloss
-
-Where to set loss_name
-Add to any BayesOpt config JSON:
-
-{
-  "task_type": "regression",
-  "loss_name": "mse"
-}
-
-Behavior changes
-1) Tuning and metrics
-   - When loss_name is mse/mae, tuning does not sample Tweedie power.
-   - When loss_name is poisson/gamma, power is fixed (1.0/2.0).
-   - When loss_name is tweedie, power is sampled as before.
-
-2) Torch training (ResNet/FT/GNN)
-   - Loss computation is routed by loss_name.
-   - For tweedie/poisson/gamma, predictions are clamped positive.
-   - For mse/mae, no Tweedie power is used.
-
-3) XGBoost objective
-   - loss_name controls XGB objective:
-     - tweedie -> reg:tweedie
-     - poisson -> count:poisson
-     - gamma -> reg:gamma
-     - mse -> reg:squarederror
-     - mae -> reg:absoluteerror
-
-4) Inference
-   - ResNet/GNN constructors now receive loss_name.
-   - When loss_name is not tweedie, tw_power is not applied at inference.
-
-Legacy defaults (auto)
-- If loss_name is omitted, behavior is unchanged:
-  - model name contains "f" -> poisson
-  - model name contains "s" -> gamma
-  - otherwise -> tweedie
-
-Examples
-- ResNet direct training (MSE):
-  "loss_name": "mse"
-
-- FT embed -> ResNet (MSE):
-  "loss_name": "mse"
-
-- XGB direct training (unchanged):
-  omit loss_name or set "loss_name": "auto"
-
-Notes
-- loss_name is global per config. If you need different losses for different
-  models, split into separate configs and run them independently.