dataset_ml/

covtype.rs

//! Forest Cover Type dataset.
//!
//! The Forest CoverType dataset for multi-class classification, identical to the
//! one scikit-learn exposes through `fetch_covtype`. Each of the 581,012 samples
//! describes a 30×30 metre cell of wilderness in the Roosevelt National Forest of
//! northern Colorado, and the task is to predict which of seven forest cover types
//! the cell belongs to from 54 cartographic features.
//!
//! **Features (54):** these encode 12 logical attributes (10 numeric + 2
//! categorical), with the two categorical attributes already **one-hot expanded**.
//! By 0-based column index:
//! - cols `0..=9` — 10 distinct quantitative variables: `Elevation`, `Aspect`,
//!   `Slope`, `Horizontal_Distance_To_Hydrology`, `Vertical_Distance_To_Hydrology`,
//!   `Horizontal_Distance_To_Roadways`, `Hillshade_9am`, `Hillshade_Noon`,
//!   `Hillshade_3pm`, `Horizontal_Distance_To_Fire_Points`
//! - cols `10..=13` — `Wilderness_Area`: **one** categorical attribute (4 areas)
//!   one-hot encoded, so exactly one of these columns is `1` and the rest are `0`
//! - cols `14..=53` — `Soil_Type`: **one** categorical attribute (40 soil types)
//!   one-hot encoded, so exactly one of these columns is `1` and the rest are `0`
//!
//! All 54 are stored as `f64` (the one-hot columns hold `0.0`/`1.0`), so each
//! one-hot block sums to `1` per row. See the struct docs for a per-column table.
//!
//! **Target:** `cover_type` - the forest cover type, one of `1`–`7` (stored as
//! `u8`): `1` = Spruce/Fir, `2` = Lodgepole Pine, `3` = Ponderosa Pine,
//! `4` = Cottonwood/Willow, `5` = Aspen, `6` = Douglas-fir, `7` = Krummholz.
//!
//! **Samples:** 581,012 total
//! **Application:** Multi-class classification / forest cover type prediction
//!
//! **Source:** UCI Machine Learning Repository, via the gzip-compressed
//! `covtype.data.gz` mirror that scikit-learn's `fetch_covtype` downloads.
//! <https://archive.ics.uci.edu/dataset/31/covertype>

use csv::ReaderBuilder;
use dataset_core::{Dataset, DatasetError, acquire_dataset, download_to, gunzip};
use ndarray::{Array1, Array2};
use std::fs::File;

/// The URL for the Forest Cover Type dataset.
///
/// This is the gzip-compressed `covtype.data.gz` mirror used by scikit-learn's
/// `fetch_covtype`. The URL has no filename segment, so the download is saved under
/// an explicit name ([`COVTYPE_GZ_FILENAME`]).
///
/// # Citation
///
/// J. A. Blackard and D. J. Dean. "Covertype," UCI Machine Learning Repository,
/// \[Online\]. Available: <https://doi.org/10.24432/C50K5N>
const COVTYPE_DATA_URL: &str = "https://ndownloader.figshare.com/files/5976039";

/// The name the downloaded gzip archive is saved under inside the temp directory.
const COVTYPE_GZ_FILENAME: &str = "covtype.data.gz";

/// The name of the final cached (decompressed) Cover Type dataset file.
const COVTYPE_FILENAME: &str = "covtype.csv";

/// The SHA256 hash of the **decompressed** Cover Type dataset file (`covtype.csv`).
///
/// Note this is the hash of the uncompressed comma-separated data, not of the
/// downloaded `covtype.data.gz`, because the cached file is the decompressed one.
const COVTYPE_SHA256: &str = "0a9371cef7c964b5475d6053cc3e0894a5aa6f65ad1ed3ecb01c45aa96217945";

/// The name of the dataset
const COVTYPE_DATASET_NAME: &str = "covtype";

/// The number of cartographic features per sample.
const N_FEATURES: usize = 54;

/// The number of columns per CSV record (54 features + 1 cover-type label).
const N_COLUMNS: usize = N_FEATURES + 1;

/// The expected number of samples, used only to pre-allocate the parse buffers.
const N_SAMPLES: usize = 581_012;

/// Type alias for the Cover Type dataset: (features, labels).
type CovtypeData = (Array2<f64>, Array1<u8>);

/// A struct representing the Forest Cover Type dataset with lazy loading.
///
/// The dataset is not loaded until you call one of the data accessor methods.
/// Once loaded, the data is cached for subsequent accesses.
///
/// # About Dataset
///
/// The Forest CoverType dataset contains 581,012 cartographic samples, each
/// describing a 30×30 metre cell of the Roosevelt National Forest in northern
/// Colorado. The 54 features combine 10 quantitative measurements (elevation,
/// slope, distances to hydrology/roadways/fire points, hillshade indices) with two
/// one-hot blocks: 4 `Wilderness_Area` columns and 40 `Soil_Type` columns. The
/// target is the forest cover type (`1`–`7`).
///
/// This is the same data scikit-learn exposes through `fetch_covtype`.
///
/// # Feature columns
///
/// The 54 feature columns are **not** 54 independent variables: they encode 12
/// logical attributes (10 numeric + 2 categorical), where the two categorical
/// attributes are already **one-hot expanded** into many binary indicator columns.
/// By 0-based column index in the feature matrix:
///
/// | Columns   | Attribute(s)                                  | Encoding                                   |
/// |-----------|-----------------------------------------------|--------------------------------------------|
/// | `0`       | `Elevation`                                   | quantitative (metres)                      |
/// | `1`       | `Aspect`                                      | quantitative (azimuth degrees)             |
/// | `2`       | `Slope`                                       | quantitative (degrees)                     |
/// | `3`       | `Horizontal_Distance_To_Hydrology`            | quantitative                               |
/// | `4`       | `Vertical_Distance_To_Hydrology`              | quantitative (may be negative)             |
/// | `5`       | `Horizontal_Distance_To_Roadways`             | quantitative                               |
/// | `6`       | `Hillshade_9am`                               | quantitative (`0..=255`)                   |
/// | `7`       | `Hillshade_Noon`                              | quantitative (`0..=255`)                   |
/// | `8`       | `Hillshade_3pm`                               | quantitative (`0..=255`)                   |
/// | `9`       | `Horizontal_Distance_To_Fire_Points`          | quantitative                               |
/// | `10..=13` | `Wilderness_Area` (one attribute, 4 areas)    | one-hot: exactly one column is `1`, rest `0` |
/// | `14..=53` | `Soil_Type` (one attribute, 40 soil types)    | one-hot: exactly one column is `1`, rest `0` |
///
/// So columns `0..=9` are ten distinct numeric features, but columns `10..=13`
/// jointly answer "which of 4 wilderness areas" and columns `14..=53` jointly
/// answer "which of 40 soil types" — each block is a single categorical variable,
/// with `1` marking the active category and `0` everywhere else (the block sums to
/// `1`). All 54 columns are stored as `f64` (the one-hot columns hold `0.0`/`1.0`),
/// matching scikit-learn's dense `fetch_covtype` matrix.
///
/// # Labels
///
/// - cover type (in `u8`): `1` = Spruce/Fir, `2` = Lodgepole Pine,
///   `3` = Ponderosa Pine, `4` = Cottonwood/Willow, `5` = Aspen,
///   `6` = Douglas-fir, `7` = Krummholz
///
/// See more information at
/// <https://archive.ics.uci.edu/dataset/31/covertype>
///
/// # Citation
///
/// J. A. Blackard and D. J. Dean. "Covertype," UCI Machine Learning Repository,
/// \[Online\]. Available: <https://doi.org/10.24432/C50K5N>
///
/// # Thread Safety
///
/// This struct automatically implements `Send` and `Sync` (All fields implement them), making it safe to share across threads.
/// The internal [`Dataset`] ensures thread-safe lazy initialization.
///
/// # Example
/// ```no_run
/// use dataset_ml::covtype::Covtype;
///
/// let download_dir = "./covtype"; // the code will create the directory if it doesn't exist
///
/// let mut dataset = Covtype::new(download_dir);
/// let features = dataset.features().unwrap();
/// let labels = dataset.labels().unwrap();
///
/// let (features, labels) = dataset.data().unwrap(); // this is also a way to get features and labels
/// assert_eq!(features.shape(), &[581012, 54]);
/// assert_eq!(labels.len(), 581012);
///
/// // `get_data()` borrows the cached arrays without reloading; `get_data_mut()`
/// // edits them in place — no clone, no reload, the change stays cached. Prefer
/// // this over cloning with `.to_owned()` when you only need to tweak values.
/// if let Some((features, labels)) = dataset.get_data_mut() {
///     features[[0, 0]] = 2596.0;
///     labels[0] = 5;
/// }
/// assert!(dataset.get_data().is_some());
///
/// // `take_data()` moves owned arrays out (no `to_owned()` clone) and leaves the
/// // instance reusable — the next access reloads from the cached file.
/// let (owned_features, owned_labels) = dataset.take_data().unwrap();
/// assert_eq!(owned_features.shape(), &[581012, 54]);
/// assert_eq!(owned_labels.len(), 581012);
///
/// // `into_data()` also returns owned arrays with no clone, but consumes the
/// // instance (use it when you are done with the dataset).
/// let (owned_features, owned_labels) = dataset.into_data().unwrap();
/// assert_eq!(owned_features.shape(), &[581012, 54]);
/// assert_eq!(owned_labels.len(), 581012);
/// ```
#[derive(Debug)]
pub struct Covtype {
    dataset: Dataset<CovtypeData, DatasetError>,
}

impl Covtype {
    /// Create a new Covtype instance without loading data.
    ///
    /// The dataset will be loaded lazily when you first call any data accessor method.
    /// This is a lightweight operation that only stores the storage directory.
    ///
    /// # Parameters
    ///
    /// - `storage_dir` - Directory where the dataset will be stored.
    ///
    /// # Returns
    ///
    /// - `Self` - `Covtype` instance ready for lazy loading.
    pub fn new(storage_dir: &str) -> Self {
        Covtype {
            dataset: Dataset::new(storage_dir, Self::load_data),
        }
    }

    /// Acquire and parse the Forest Cover Type dataset.
    fn load_data(dir: &str) -> Result<CovtypeData, DatasetError> {
        // Prepare the dataset file: download the gzip-compressed `covtype.data.gz`
        // and decompress it into the plain comma-separated `covtype.csv`.
        let file_path = acquire_dataset(
            dir,
            COVTYPE_FILENAME,
            COVTYPE_DATASET_NAME,
            Some(COVTYPE_SHA256),
            |temp_path| {
                download_to(COVTYPE_DATA_URL, temp_path, Some(COVTYPE_GZ_FILENAME))?;
                let gz_path = temp_path.join(COVTYPE_GZ_FILENAME);
                let csv_path = temp_path.join(COVTYPE_FILENAME);
                gunzip(&gz_path, &csv_path)?;
                Ok(csv_path)
            },
        )?;

        // `covtype.data` is a headerless comma-separated file: every line is a
        // record of 54 cartographic features followed by the cover-type label.
        let file = File::open(&file_path)?;
        let mut rdr = ReaderBuilder::new().has_headers(false).from_reader(file);

        // Pre-allocate for the known sample count to avoid repeatedly growing a
        // ~250 MB feature buffer; parsing still works for any actual row count.
        let mut features = Vec::with_capacity(N_SAMPLES * N_FEATURES);
        let mut labels = Vec::with_capacity(N_SAMPLES);

        for (idx, result) in rdr.records().enumerate() {
            let record =
                result.map_err(|e| DatasetError::csv_read_error(COVTYPE_DATASET_NAME, e))?;
            let line_num = idx + 1; // headerless file, lines are 1-indexed

            if record.len() != N_COLUMNS {
                return Err(DatasetError::invalid_column_count(
                    COVTYPE_DATASET_NAME,
                    N_COLUMNS,
                    record.len(),
                    line_num,
                ));
            }

            for (col, field) in record.iter().take(N_FEATURES).enumerate() {
                let value: f64 = field.trim().parse().map_err(|e| {
                    DatasetError::parse_failed(
                        COVTYPE_DATASET_NAME,
                        &format!("feature_{}", col),
                        line_num,
                        e,
                    )
                })?;
                features.push(value);
            }

            let raw_label = record[N_FEATURES].trim();
            let label: u8 = raw_label.parse().map_err(|e| {
                DatasetError::parse_failed(COVTYPE_DATASET_NAME, "cover_type", line_num, e)
            })?;
            if !(1..=7).contains(&label) {
                return Err(DatasetError::invalid_value(
                    COVTYPE_DATASET_NAME,
                    "cover_type",
                    raw_label,
                    line_num,
                ));
            }
            labels.push(label);
        }

        let n_samples = labels.len();
        if n_samples == 0 {
            return Err(DatasetError::empty_dataset(COVTYPE_DATASET_NAME));
        }

        // Cover Type has a fixed schema of 54 numeric features per sample.
        let features_array = Array2::from_shape_vec((n_samples, N_FEATURES), features)
            .map_err(|e| DatasetError::array_shape_error(COVTYPE_DATASET_NAME, "features", e))?;
        let labels_array = Array1::from_vec(labels);

        Ok((features_array, labels_array))
    }

    /// Get a reference to the feature matrix.
    ///
    /// This method triggers lazy loading on first call. Subsequent calls return
    /// the cached data instantly.
    ///
    /// # Returns
    ///
    /// - `&Array2<f64>` - Reference to feature matrix with shape `(581012, 54)`
    ///   containing the 10 quantitative variables followed by the 4 one-hot
    ///   `Wilderness_Area` and 40 one-hot `Soil_Type` columns.
    ///
    /// # Errors
    ///
    /// Returns `DatasetError` if:
    /// - Download fails due to network issues
    /// - File decompression or I/O operations fail
    /// - Data format is invalid (wrong number of columns, unparseable values, or invalid labels)
    /// - Dataset size doesn't match expected dimensions (581012 samples, 54 features)
    pub fn features(&self) -> Result<&Array2<f64>, DatasetError> {
        Ok(&self.dataset.load()?.0)
    }

    /// Get a reference to the labels vector.
    ///
    /// This method triggers lazy loading on first call. Subsequent calls return
    /// the cached data instantly.
    ///
    /// # Returns
    ///
    /// - `&Array1<u8>` - Reference to labels vector with shape `(581012,)` containing the cover-type classes (`1`–`7`).
    ///
    /// # Errors
    ///
    /// Returns `DatasetError` if:
    /// - Download fails due to network issues
    /// - File decompression or I/O operations fail
    /// - Data format is invalid (wrong number of columns, unparseable values, or invalid labels)
    /// - Dataset size doesn't match expected dimensions (581012 samples)
    pub fn labels(&self) -> Result<&Array1<u8>, DatasetError> {
        Ok(&self.dataset.load()?.1)
    }

    /// Get both features and labels as references.
    ///
    /// This method triggers lazy loading on first call. Subsequent calls return
    /// the cached data instantly.
    ///
    /// # Returns
    ///
    /// - `&CovtypeData` - reference to the cached `(features, labels)` tuple: the
    ///   feature matrix has shape `(581012, 54)` and the label vector has shape
    ///   `(581012,)` containing the cover-type classes (`1`–`7`).
    ///
    /// # Errors
    ///
    /// Returns `DatasetError` if:
    /// - Download fails due to network issues
    /// - File decompression or I/O operations fail
    /// - Data format is invalid (wrong number of columns, unparseable values, or invalid labels)
    /// - Dataset size doesn't match expected dimensions (581012 samples, 54 features)
    pub fn data(&self) -> Result<&CovtypeData, DatasetError> {
        self.dataset.load()
    }

    /// Get both features and labels as references **without** triggering loading.
    ///
    /// Unlike [`Covtype::data`], which loads the dataset on first call, this never
    /// runs the loader: if the data has not been loaded yet, it returns `None`
    /// instead of downloading and parsing. Use it when you only want the data if
    /// it is already cached and want to avoid paying the download/parse cost
    /// otherwise.
    ///
    /// # Returns
    ///
    /// - `Some(&CovtypeData)` - reference to the cached `(features, labels)` tuple
    ///   (feature matrix `(581012, 54)`, label vector `(581012,)`), if loaded.
    /// - `None` - if the dataset has not been loaded yet.
    pub fn get_data(&self) -> Option<&CovtypeData> {
        self.dataset.get()
    }

    /// Get mutable references to features and labels for **in-place** editing.
    ///
    /// This lets you modify the cached arrays directly (e.g. normalize features,
    /// replace label values) with no `to_owned()` clone and without removing them
    /// from the cache: the changes persist, so later [`Covtype::features`],
    /// [`Covtype::data`], or [`Covtype::get_data`] calls observe them.
    ///
    /// Like [`Covtype::get_data`], this does **not** trigger loading: it returns
    /// `None` if the dataset has not been loaded. Call a loading accessor (e.g.
    /// [`Covtype::data`]) first if you need to ensure the data is present.
    ///
    /// # Returns
    ///
    /// - `Some(&mut CovtypeData)` - mutable reference to the cached
    ///   `(features, labels)` tuple (feature matrix `(581012, 54)`, label vector
    ///   `(581012,)`), if loaded.
    /// - `None` - if the dataset has not been loaded yet.
    pub fn get_data_mut(&mut self) -> Option<&mut CovtypeData> {
        self.dataset.get_mut()
    }

    /// Consume the dataset and return **owned** features and labels.
    ///
    /// Unlike [`Covtype::data`], which borrows the cached data, this moves it out and
    /// returns owned arrays directly — no `to_owned()` clone needed. The dataset is
    /// loaded on first access if it has not been loaded yet.
    ///
    /// This **consumes** `self`, so the instance cannot be used afterwards. If you
    /// want owned data but need to keep using the instance, use [`Covtype::take_data`]
    /// instead — it takes `&mut self` and leaves the instance reusable.
    ///
    /// # Returns
    ///
    /// - `(Array2<f64>, Array1<u8>)` - owned feature matrix with shape
    ///   `(581012, 54)` and owned label vector with shape `(581012,)`.
    ///
    /// # Errors
    ///
    /// Returns `DatasetError` if loading fails (network, file I/O, parsing, invalid
    /// labels, or a dimension mismatch).
    pub fn into_data(self) -> Result<CovtypeData, DatasetError> {
        self.dataset.load()?;
        Ok(self
            .dataset
            .into_inner()
            .expect("data is present after a successful load"))
    }

    /// Take **owned** features and labels out of the dataset, leaving it reusable.
    ///
    /// Like [`Covtype::into_data`], this returns owned arrays with no `to_owned()`
    /// clone. But instead of consuming the instance, it takes `&mut self` and moves
    /// the cached data out, resetting the instance to its unloaded state: the next
    /// accessor call (e.g. [`Covtype::features`] or [`Covtype::data`]) loads the
    /// dataset again.
    ///
    /// Use [`Covtype::into_data`] instead if you are done with the instance.
    ///
    /// # Returns
    ///
    /// - `(Array2<f64>, Array1<u8>)` - owned feature matrix with shape
    ///   `(581012, 54)` and owned label vector with shape `(581012,)`.
    ///
    /// # Errors
    ///
    /// Returns `DatasetError` if loading fails (network, file I/O, parsing, invalid
    /// labels, or a dimension mismatch).
    pub fn take_data(&mut self) -> Result<CovtypeData, DatasetError> {
        self.dataset.load()?;
        Ok(self
            .dataset
            .take()
            .expect("data is present after a successful load"))
    }
}