smelt-ml 0.1.0__tar.gz

smelt_ml-0.1.0/.dockerignore

@@ -0,0 +1,4 @@
+target/
+.git/
+docs/WOS/
+*.profraw

smelt_ml-0.1.0/.gitignore

@@ -0,0 +1,3 @@
+/target
+docs/WOS/*.bib
+catboost_info/

smelt_ml-0.1.0/CLAUDE.md

@@ -0,0 +1,118 @@
+# Smelt — Machine Learning Framework for Rust
+
+## Overview
+
+Smelt is an ML framework inspired by [mlr3](https://mlr3.mlr-org.com/) (R), designed for Rust's performance and safety guarantees. The name refers to smelting — refining raw data into useful models.
+
+## Architecture
+
+```
+Task → Learner → TrainedModel → Prediction → Measure
+                      ↑
+                  Resampling (CV, Holdout)
+                  Tuning (Grid, Random, Bayesian)
+                  Preprocessing (Scale, Encode, Impute)
+```
+
+### Core Abstractions (mlr3 mapping)
+
+| Smelt | mlr3 | Purpose |
+|-------|------|---------|
+| `Task` | `Task` | Data container with target |
+| `ClassificationTask` | `TaskClassif` | Discrete target |
+| `RegressionTask` | `TaskRegr` | Continuous target |
+| `Learner` | `Learner` | Algorithm that trains |
+| `TrainedModel` | trained Learner | Fitted model that predicts |
+| `Prediction` | `Prediction` | Output with optional truth |
+| `Measure` | `Measure` | Evaluation metric |
+| `Resample` | `Resampling` | Train/test splitting strategy |
+
+### Module Structure
+
+```
+src/
+├── lib.rs          # Public API + prelude
+├── error.rs        # SmeltError enum (thiserror)
+├── task/mod.rs     # Task, ClassificationTask, RegressionTask
+├── learner/mod.rs  # Learner trait, TrainedModel trait
+├── prediction/     # Prediction enum (Classification/Regression)
+├── measure/        # Accuracy, RMSE, MAE (+ trait Measure)
+├── resample/       # CrossValidation, Holdout (+ trait Resample)
+├── preprocess/     # TODO: StandardScaler, MinMaxScaler, OneHotEncoder
+└── tuning/         # TODO: GridSearch, RandomSearch, BayesianOpt
+```
+
+## Build & Test
+
+```bash
+cargo check          # Type check
+cargo test           # Run tests
+cargo bench          # Run benchmarks (criterion)
+cargo doc --open     # Generate docs
+```
+
+## Design Principles
+
+1. **Type safety** — Classification and Regression are separate types, not runtime tags
+2. **Trait-based extensibility** — Implement `Learner` to add new algorithms
+3. **Zero-copy where possible** — ndarray views, references over clones
+4. **Parallel by default** — rayon for data parallelism (CV folds, ensemble training)
+5. **Composable pipeline** — Task → Learner → Prediction → Measure is always the flow
+
+## Implementation Roadmap
+
+### Phase 1 — Core (current)
+- [x] Task system (Classification + Regression)
+- [x] Learner + TrainedModel traits
+- [x] Prediction enum
+- [x] Measures: Accuracy, RMSE, MAE
+- [x] Resampling: CrossValidation, Holdout
+
+### Phase 2 — First Learners
+- [ ] Decision Tree (CART)
+- [ ] K-Nearest Neighbors
+- [ ] Logistic Regression
+- [ ] Linear Regression (OLS)
+- [ ] Benchmark pipeline (resample + measure loop)
+
+### Phase 3 — Ensembles
+- [ ] Random Forest
+- [ ] Gradient Boosting (XGBoost-style)
+- [ ] Bagging
+
+### Phase 4 — Preprocessing
+- [ ] StandardScaler, MinMaxScaler
+- [ ] OneHotEncoder, LabelEncoder
+- [ ] Missing value imputation
+- [ ] Pipeline chaining (preprocess → learner)
+
+### Phase 5 — Tuning
+- [ ] GridSearch
+- [ ] RandomSearch
+- [ ] Bayesian Optimization
+
+### Phase 6 — Advanced
+- [ ] Feature importance (permutation, SHAP-like)
+- [ ] Spatial cross-validation (for geo applications)
+- [ ] CSV/Parquet data loading
+- [ ] Model serialization (serde)
+- [ ] Python bindings (PyO3) — expose as `smelt-py`
+
+## Dependencies
+
+- `ndarray` — N-dimensional arrays (feature matrices)
+- `rand` — Random number generation (resampling, stochastic algorithms)
+- `rayon` — Data parallelism
+- `thiserror` — Error types
+- `serde` — Serialization
+- `criterion` — Benchmarks (dev)
+
+## Author
+
+Francisco Parra — francisco.parra.o@usach.cl
+
+## Inspiration
+
+- [mlr3](https://mlr3.mlr-org.com/) (R) — Task/Learner/Measure architecture
+- [scikit-learn](https://scikit-learn.org/) (Python) — fit/predict API
+- [linfa](https://github.com/rust-ml/linfa) (Rust) — Existing Rust ML, but different design philosophy