Clump

https://docs.rs/clump
Admin
Clump is a Rust library providing SIMD-accelerated dense clustering algorithms and evaluation...
Tokens:434,662
Snippets:3,124
Trust Score:9.1
Update:3 weeks ago
Show doc for...
Tokens:
Raw
### Test macro_clustering_finds_groups setup

Source: https://docs.rs/clump/latest/src/clump/cluster/denstream.rs.html?search=Option%3CT%3E%2C+%28T+-%3E+U%29+-%3E+Option%3CU%3E

Starts a test for macro clustering with a DenStream configured with macro_epsilon 3.0 and a long pruning period.

```rust
#[test]
fn macro_clustering_finds_groups() {
    let mut ds = DenStream::new(1.0, 2)
        .with_beta(0.2)
        .with_lambda(0.0001)
        .with_mu(1.0)
        .with_macro_epsilon(3.0)
        .with_pruning_period(10_000);
```

--------------------------------

### Example for Result::is_ok

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Example demonstrating `is_ok()` which returns `true` if the result is `Ok`.

```rust
let x: Result<i32, &str> = Ok(-3);
assert_eq!(x.is_ok(), true);

let x: Result<i32, &str> = Err("Some error message");
assert_eq!(x.is_ok(), false);
```

--------------------------------

### Property-based test module setup and arbitrary data strategy

Source: https://docs.rs/clump/latest/src/clump/cluster/kmeans.rs.html?search=Option%3CT%3E%2C+%28T+-%3E+U%29+-%3E+Option%3CU%3E

Starts the proptest module, defining an arbitrary data strategy that generates vectors of vectors of f32 with values in [-100, 100] and a fixed dimension.

```rust
#[cfg(test)]
mod proptests {
    use super::*;
    use proptest::prelude::*;

    fn arb_data(max_n: usize, d: usize) -> impl Strategy<Value = Vec<Vec<f32>>> {
        proptest::collection::vec(
            proptest::collection::vec(-100.0f32..100.0, d..=d),
            3..=max_n,
        )
    }

    proptest! {
        /// All labels must be in [0, k).
        #[test]
```

--------------------------------

### Example for Result::is_ok_and

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Example demonstrating `is_ok_and()` which returns `true` if the result is `Ok` and the value matches a predicate.

```rust
let x: Result<u32, &str> = Ok(2);
assert_eq!(x.is_ok_and(|x| x > 1), true);

let x: Result<u32, &str> = Ok(0);
assert_eq!(x.is_ok_and(|x| x > 1), false);

let x: Result<u32, &str> = Err("hey");
assert_eq!(x.is_ok_and(|x| x > 1), false);

let x: Result<String, &str> = Ok("ownership".to_string());
assert_eq!(x.as_ref().is_ok_and(|x| x.len() > 1), true);
println!("still alive {:?}", x);
```

--------------------------------

### iter example

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Shows how to use iter to get an iterator over the possibly contained value, yielding one value if Ok and none if Err.

```rust
let x: Result<u32, &str> = Ok(7);
assert_eq!(x.iter().next(), Some(&7));

let x: Result<u32, &str> = Err("nothing!");
assert_eq!(x.iter().next(), None);
```

--------------------------------

### COP-Kmeans usage example with constraints

Source: https://docs.rs/clump/latest/clump/cluster/constrained/struct.CopKmeans.html?search=u32+-%3E+bool

Example showing how to create a CopKmeans clusterer, define must-link and cannot-link constraints, and fit the model. The example asserts that points 0 and 1 are in the same cluster and points 0 and 2 are in different clusters.

```rust
use clump::cluster::constrained::{CopKmeans, Constraint};

let data = vec![
    vec![0.0f32, 0.0],
    vec![0.1, 0.1],
    vec![10.0, 10.0],
    vec![10.1, 10.1],
];

// Force points 0 and 1 together, points 0 and 2 apart.
let constraints = vec![
    Constraint::MustLink(0, 1),
    Constraint::CannotLink(0, 2),
];

let labels = CopKmeans::new(2)
    .with_seed(42)
    .fit_predict_constrained(&data, &constraints)
    .unwrap();

assert_eq!(labels[0], labels[1]);
assert_ne!(labels[0], labels[2]);
```

--------------------------------

### MiniBatchKmeans usage example

Source: https://docs.rs/clump/latest/clump/cluster/streaming/struct.MiniBatchKmeans.html?search=u32+-%3E+bool

Example showing creation, seeding, batch updates, point updates, and cluster count retrieval.

```rust
use clump::MiniBatchKmeans;

let mut mbk = MiniBatchKmeans::new(2).with_seed(42);

// First batch seeds centroids via k-means++
let batch1 = vec![
    vec![0.0f32, 0.0], vec![0.1, 0.1],
    vec![10.0, 10.0], vec![10.1, 10.1],
];
let labels = mbk.update_batch(&batch1).unwrap();
assert_eq!(labels.len(), 4);

// Subsequent updates refine centroids
let label = mbk.update(&[0.05, 0.05]).unwrap();
assert_eq!(mbk.n_clusters(), 2);
```

--------------------------------

### K-means fit: validation, initialization, and buffer setup

Source: https://docs.rs/clump/latest/src/clump/cluster/kmeans.rs.html

Validates k, dimension, cluster count, and uniform dimensionality; normalizes tolerance by data variance; initializes RNG, centroids (warm-start or k-means++), and working buffers including Hamerly bounds. Precomputes squared norms for expanded-form assignment, optionally sets up Metal GPU assigner (feature gpu, macOS) and BLAS GEMM (feature blas) for large problems, then starts the iteration loop.

```rust
273            return Err(Error::InvalidParameter {
274                name: "k",
275                message: "must be at least 1",
276            });
277        }
278
279        let n = data.n();
280        let d = data.d();
281        if d == 0 {
282            return Err(Error::InvalidParameter {
283                name: "dimension",
284                message: "must be at least 1",
285            });
286        }
287
288        if self.k > n {
289            return Err(Error::InvalidClusterCount {
290                requested: self.k,
291                n_items: n,
292            });
293        }
294
295        // Validate uniform dimensionality.
296        for i in 0..n {
297            if data.row(i).len() != d {
298                return Err(Error::DimensionMismatch {
299                    expected: d,
300                    found: data.row(i).len(),
301                });
302            }
303        }
304
305        util::validate_finite(data)?;
306
307        // Normalize tolerance by data variance so it scales with data magnitude.
308        // Without this, the raw tol is compared against the sum of k*d squared
309        // centroid shifts, which becomes meaninglessly tight for high-dimensional
310        // or large-scale data. (Matches scikit-learn's _tolerance approach.)
311        let effective_tol = (self.tol * util::mean_variance(data) * self.k as f64) as f32;
312
313        // Initialize RNG.
314        let mut rng = match self.seed {
315            Some(s) => StdRng::seed_from_u64(s),
316            None => StdRng::from_os_rng(),
317        };
318
319        // Initialize centroids: use provided centroids (warm-start) or k-means++.
320        let mut centroids = if let Some(ref init) = self.init_centroids {
321            assert_eq!(
322                init.len(),
323                self.k,
324                "init_centroids length ({}) must equal k ({})",
325                init.len(),
326                self.k
327            );
328            init.clone()
329        } else {
330            util::kmeanspp_init(data, self.k, &self.metric, self.seeding_alpha, &mut rng)
331        };
332        let mut labels = vec![0usize; n];
333
334        // Pre-allocate working buffers outside the iteration loop to avoid
335        // per-iteration allocation overhead (AMD ROCm pattern).
336        let mut new_centroids = vec![vec![0.0f32; d]; self.k];
337        let mut counts = vec![0usize; self.k];
338
        // Hamerly bounds: per-point upper (dist to assigned centroid) and
        // lower (dist to second-nearest centroid). When upper <= lower,
        // the assignment cannot change and we skip distance computation.
        // O(n) extra memory (Hamerly, SDM 2010).
        let mut upper_bounds = vec![f32::MAX; n];
        let mut lower_bounds = vec![0.0f32; n];
        let mut centroid_shifts = vec![0.0f32; self.k];
        let mut sums_f64 = vec![vec![0.0f64; d]; self.k];
        let mut flat_buf: Vec<f32> = Vec::with_capacity(self.k * d);
        let mut inertia_trace: Vec<f32> = Vec::with_capacity(self.max_iter);

        // Precompute squared norms for expanded-form first-iteration assignment.
        // Only valid for SquaredEuclidean (||x-c||^2 = ||x||^2 + ||c||^2 - 2*x.c).
        let use_expanded = self.metric.supports_expanded_form();
        let data_norms: Vec<f32> = if use_expanded {
            util::squared_norms(data)
        } else {
            Vec::new()
        };

        // GPU acceleration: initialize Metal compute pipeline for assignment
        // when using SquaredEuclidean and problem is large enough to amortize
        // GPU setup overhead. Buffers for data, labels, and params are allocated
        // once here and reused across iterations.
        #[cfg(all(feature = "gpu", target_os = "macos"))]
        let gpu_assigner = if self.metric.supports_expanded_form() && n * self.k >= 500_000 {
            let data_flat = super::gpu::flatten(data);
            super::gpu::GpuAssigner::new(&data_flat, n, self.k, d)
        } else {
            None
        };

        // BLAS GEMM for first-iteration assignment (when blas feature enabled).
        // matrixmultiply's optimized SGEMM beats per-point distance loops
        // for large n*k due to micro-kernel SIMD and cache optimization.
        #[cfg(feature = "blas")]
        let use_blas = n * self.k >= 100_000 && self.metric.supports_expanded_form();
        #[cfg(feature = "blas")]
        let blas_data = if use_blas {
            let fd = super::flat::FlatMatrix::from_data(data);
            let xn = fd.row_norms_sq();
            Some((fd, xn))
        } else {
            None
        };

        let mut iters = 0usize;
        for iter in 0..self.max_iter {
            iters = iter + 1;

            // Zero the accumulators for this iteration.
```

--------------------------------

### expect example

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Shows how to use expect to return the contained Ok value, panicking with a message if the value is Err. The first example demonstrates a panic, and the second shows a recommended message style for environment variables.

```rust
let x: Result<u32, &str> = Err("emergency failure");
x.expect("Testing expect"); // panics with `Testing expect: emergency failure`
```

```rust
let path = std::env::var("IMPORTANT_PATH")
    .expect("env variable `IMPORTANT_PATH` should be set by `wrapper_script.sh`");
```

--------------------------------

### Example for Result::copied on &mut T

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Example demonstrating `copied()` on `Result<&mut T, E>` to copy the `Ok` value.

```rust
let mut val = 12;
let x: Result<&mut i32, i32> = Ok(&mut val);
assert_eq!(x, Ok(&mut 12));
let copied = x.copied();
assert_eq!(copied, Ok(12));
```

--------------------------------

### Example for Result::flatten

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Example demonstrating `flatten()` which converts `Result<Result<T, E>, E>` to `Result<T, E>`.

```rust
let x: Result<Result<&'static str, u32>, u32> = Ok(Ok("hello"));
assert_eq!(Ok("hello"), x.flatten());

let x: Result<Result<&'static str, u32>, u32> = Ok(Err(6));
assert_eq!(Err(6), x.flatten());

let x: Result<Result<&'static str, u32>, u32> = Err(6);
assert_eq!(Err(6), x.flatten());
```

--------------------------------

### Example for Result::transpose

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Example demonstrating `transpose()` which maps `Result<Option<T>, E>` to `Option<Result<T, E>>`.

```rust
#[derive(Debug, Eq, PartialEq)]
struct SomeErr;

let x: Result<Option<i32>, SomeErr> = Ok(Some(5));
let y: Option<Result<i32, SomeErr>> = Some(Ok(5));
assert_eq!(x.transpose(), y);
```

--------------------------------

### Example for Result::copied on &T

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Example demonstrating `copied()` on `Result<&T, E>` to copy the `Ok` value.

```rust
let val = 12;
let x: Result<&i32, i32> = Ok(&val);
assert_eq!(x, Ok(&12));
let copied = x.copied();
assert_eq!(copied, Ok(12));
```

--------------------------------

### Example for Result::cloned on &T

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Example demonstrating `cloned()` on `Result<&T, E>` to clone the `Ok` value.

```rust
let val = 12;
let x: Result<&i32, i32> = Ok(&val);
assert_eq!(x, Ok(&12));
let cloned = x.cloned();
assert_eq!(cloned, Ok(12));
```

--------------------------------

### and_then example function

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=Option%3CT%3E%2C+%28T+-%3E+U%29+-%3E+Option%3CU%3E

Example function for and_then, showing control flow based on Result values.

```rust
fn sq_then_to_string(x: u32) -> Result<String, &'static str> {
    x.checked_mul(x).map(|sq| sq.to_string()).ok_or("overflowed")
}
```

--------------------------------

### Example for Result::cloned on &mut T

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Example demonstrating `cloned()` on `Result<&mut T, E>` to clone the `Ok` value.

```rust
let mut val = 12;
let x: Result<&mut i32, i32> = Ok(&mut val);
assert_eq!(x, Ok(&mut 12));
let cloned = x.cloned();
assert_eq!(cloned, Ok(12));
```

--------------------------------

### as_mut example

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Demonstrates using Result::as_mut to get mutable references to the contained values, allowing mutation of the Ok or Err value.

```rust
fn mutate(r: &mut Result<i32, i32>) {
    match r.as_mut() {
        Ok(v) => *v = 42,
        Err(e) => *e = 0,
    }
}

let mut x: Result<i32, i32> = Ok(2);
mutate(&mut x);
assert_eq!(x.unwrap(), 42);

let mut x: Result<i32, i32> = Err(13);
mutate(&mut x);
assert_eq!(x.unwrap_err(), 0);
```

--------------------------------

### Start main iteration loop

Source: https://docs.rs/clump/latest/src/clump/cluster/kmeans.rs.html?search=

Starts the main iteration loop, tracking the iteration count and zeroing accumulators at the beginning of each iteration.

```rust
let mut iters = 0usize;
for iter in 0..self.max_iter {
    iters = iter + 1;

    // Zero the accumulators for this iteration.
```

--------------------------------

### flat_to_vecs usage example

Source: https://docs.rs/clump/latest/clump/cluster/adapt/fn.flat_to_vecs.html?search=Option%3CT%3E%2C+%28T+-%3E+U%29+-%3E+Option%3CU%3E

Example usage of flat_to_vecs, converting a flat vector of six f32 values into three vectors of dimension two. The example asserts the resulting length and the content of the second vector.

```rust
use clump::cluster::adapt::flat_to_vecs;

let flat = vec![1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0];
let vecs = flat_to_vecs(&flat, 3, 2);
assert_eq!(vecs.len(), 3);
assert_eq!(vecs[1], vec![3.0, 4.0]);
```

--------------------------------

### HDBSCAN clustering example with Hdbscan::new()

Source: https://docs.rs/clump/latest/clump/cluster/struct.Hdbscan.html

Example usage of Hdbscan to cluster two well-separated clusters. Uses with_min_samples(3) and with_min_cluster_size(5), then fit_predict. Asserts 30 labels and at least 2 non-noise clusters.

```rust
use clump::{Hdbscan, NOISE};

// Two tight clusters, well-separated
let mut data: Vec<Vec<f32>> = (0..15).map(|i| {
    vec![(i % 3) as f32 * 0.1, (i / 3) as f32 * 0.1]
}).collect();
data.extend((0..15).map(|i| {
    vec![50.0 + (i % 3) as f32 * 0.1, 50.0 + (i / 3) as f32 * 0.1]
}));

let labels = Hdbscan::new()
    .with_min_samples(3)
    .with_min_cluster_size(5)
    .fit_predict(&data)
    .unwrap();

assert_eq!(labels.len(), 30);
// At least 2 distinct non-noise clusters found
let clusters: std::collections::HashSet<_> =
    labels.iter().copied().filter(|&l| l != NOISE).collect();
assert!(clusters.len() >= 2);
```

--------------------------------

### map example

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Shows how to use Result::map to apply a function to the Ok value, leaving Err untouched. The example parses each line of a string and prints the doubled number.

```rust
let line = "1\n2\n3\n4\n";

for num in line.lines() {
    match num.parse::<i32>().map(|i| i * 2) {
        Ok(n) => println!("{n}"),
        Err(..) => {}
    }
}
```

--------------------------------

### inspect example

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Shows how to use inspect to call a function with a reference to the contained value if Ok, returning the original result. The example parses a string to u8, inspects the value, and then maps it to its cube.

```rust
let x: u8 = "4"
    .parse::<u8>()
    .inspect(|x| println!("original: {x}"))
    .map(|x| x.pow(3))
    .expect("literal `4` should parse as a `u8`");
```

--------------------------------

### as_deref_mut example

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Shows how to use as_deref_mut to convert a Result<T, E> to Result<&mut <T as DerefMut>::Target, &mut E>, coercing the Ok variant via DerefMut. The example demonstrates with String and u32 types, including mutation.

```rust
let mut s = "HELLO".to_string();
let mut x: Result<String, u32> = Ok("hello".to_string());
let y: Result<&mut str, &mut u32> = Ok(&mut s);
assert_eq!(x.as_deref_mut().map(|x| { x.make_ascii_uppercase(); x }), y);

let mut i = 42;
let mut x: Result<String, u32> = Err(42);
let y: Result<&mut str, &mut u32> = Err(&mut i);
assert_eq!(x.as_deref_mut().map(|x| { x.make_ascii_uppercase(); x }), y);
```

--------------------------------

### Kmeans fit example

Source: https://docs.rs/clump/latest/clump/cluster/struct.Kmeans.html

Fits k-means and returns a KmeansFit containing centroids, labels, and iteration count. The example asserts the expected lengths and that iterations occurred.

```rust
use clump::Kmeans;

let data = vec![vec![0.0f32, 0.0], vec![0.1, 0.1], vec![5.0, 5.0], vec![5.1, 5.1]];
let fit = Kmeans::new(2).with_seed(42).fit(&data).unwrap();

assert_eq!(fit.centroids.len(), 2);
assert_eq!(fit.labels.len(), 4);
assert!(fit.iters > 0);
```

--------------------------------

### as_deref example

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Shows how to use as_deref to convert a Result<T, E> to Result<&<T as Deref>::Target, &E>, coercing the Ok variant via Deref. The example demonstrates with String and u32 types.

```rust
let x: Result<String, u32> = Ok("hello".to_string());
let y: Result<&str, &u32> = Ok("hello");
assert_eq!(x.as_deref(), y);

let x: Result<String, u32> = Err(42);
let y: Result<&str, &u32> = Err(&42);
assert_eq!(x.as_deref(), y);
```

--------------------------------

### DenStream usage example

Source: https://docs.rs/clump/latest/clump/cluster/denstream/struct.DenStream.html

Shows how to create a DenStream with epsilon 1.0 and min_pts 3, configure beta, lambda, and mu, feed points from two clusters, and assert that at least two clusters are found.

```rust
use clump::DenStream;

let mut ds = DenStream::new(1.0, 3)
    .with_beta(0.5)
    .with_lambda(0.01)
    .with_mu(1.0);

// Feed points from two clusters.
for i in 0..20 {
    let offset = i as f32 * 0.1;
    ds.update(&[offset, offset]).unwrap();
}
for i in 0..20 {
    let offset = 50.0 + i as f32 * 0.1;
    ds.update(&[offset, offset]).unwrap();
}

assert!(ds.n_clusters() >= 2);
```

--------------------------------

### iter_mut example

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Shows how to use iter_mut to get a mutable iterator over the possibly contained value, allowing modification of the Ok value.

```rust
let mut x: Result<u32, &str> = Ok(7);
match x.iter_mut().next() {
    Some(v) => *v = 40,
    None => {},
}
assert_eq!(x, Ok(40));

let mut x: Result<u32, &str> = Err("nothing!");
assert_eq!(x.iter_mut().next(), None);
```

--------------------------------

### Usage example: K-means, DBSCAN, and EVoC clustering

Source: https://docs.rs/clump/latest/clump/cluster/index.html

Shows how to use `Kmeans`, `Dbscan`, and `EVoC` from `clump::cluster` on a small dataset. The example creates a 4x2 data matrix, fits K-means with 2 clusters, DBSCAN with epsilon 0.5 and min points 2, and EVoC with custom parameters. It asserts label lengths and relationships to verify clustering behavior.

```rust
use clump::cluster::{Dbscan, EVoC, EVoCParams, Kmeans};

let data = vec![
    vec![0.0, 0.0],
    vec![0.1, 0.1],
    vec![10.0, 10.0],
    vec![10.1, 10.1],
];

// Hard clustering with K-means
let labels = Kmeans::new(2).fit_predict(&data).unwrap();
assert_eq!(labels[0], labels[1]);  // First two together
assert_ne!(labels[0], labels[2]);  // Separate from last two

// Density-based clustering with DBSCAN
let labels = Dbscan::new(0.5, 2).fit_predict(&data).unwrap();
assert_eq!(labels.len(), data.len());

// Hierarchical clustering with EVōC (noise as `None`)
let mut evoc = EVoC::new(EVoCParams {
    intermediate_dim: 1,
    min_cluster_size: 2,
    seed: Some(42),
    ..Default::default()
});
let labels = evoc.fit_predict(&data).unwrap();
assert_eq!(labels.len(), data.len());
```

--------------------------------

### and_then: chaining fallible operations

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Shows how to chain operations that may return Err. The first example uses a helper function sq_then_to_string (not shown) to convert a number to a string or return an error on overflow. The second example uses metadata().and_then() to get file modification times, handling errors with ErrorKind::NotFound.

```rust
assert_eq!(Ok(2).and_then(sq_then_to_string), Ok(4.to_string()));
assert_eq!(Ok(1_000_000).and_then(sq_then_to_string), Err("overflowed"));
assert_eq!(Err("not a number").and_then(sq_then_to_string), Err("not a number"));
```

```rust
use std::{io::ErrorKind, path::Path};

// Note: on Windows "/" maps to "C:\"
let root_modified_time = Path::new("/").metadata().and_then(|md| md.modified());
assert!(root_modified_time.is_ok());

let should_fail = Path::new("/bad/path").metadata().and_then(|md| md.modified());
assert!(should_fail.is_err());
assert_eq!(should_fail.unwrap_err().kind(), ErrorKind::NotFound);
```

--------------------------------

### Example for Result::is_err

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Example demonstrating `is_err()` which returns `true` if the result is `Err`.

```rust
let x: Result<i32, &str> = Ok(-3);
assert_eq!(x.is_err(), false);

let x: Result<i32, &str> = Err("Some error message");
assert_eq!(x.is_err(), true);
```

--------------------------------

### ok example

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Demonstrates converting a Result to an Option using Result::ok, which returns Some(value) for Ok and None for Err.

```rust
let x: Result<u32, &str> = Ok(2);
assert_eq!(x.ok(), Some(2));

let x: Result<u32, &str> = Err("Nothing here");
assert_eq!(x.ok(), None);
```

--------------------------------

### CompositeDistance usage example

Source: https://docs.rs/clump/latest/clump/cluster/distance/struct.CompositeDistance.html

Example showing how to create a CompositeDistance with SquaredEuclidean and Euclidean metrics, and compute a distance. The result is 0.5 * 25.0 + 0.5 * 5.0 = 15.0. The example includes an attribute to allow deprecated code.

```rust
#![allow(deprecated)]
use clump::cluster::distance::{CompositeDistance, SquaredEuclidean, Euclidean, DistanceMetric};

let metric = CompositeDistance::new(SquaredEuclidean, Euclidean, 0.5, 0.5);
let d = metric.distance(&[0.0, 0.0], &[3.0, 4.0]);
// 0.5 * 25.0 + 0.5 * 5.0 = 15.0
assert!((d - 15.0).abs() < 1e-6);
```

--------------------------------

### CorrelationClustering::edges_from_distances

Source: https://docs.rs/clump/latest/clump/cluster/correlation/struct.CorrelationClustering.html?search=std%3A%3Avec

Creates signed edges from a distance matrix and a threshold. Pairs closer than the threshold get positive weight, pairs farther get negative weight, and pairs at exactly the threshold get weight 0.

```APIDOC
## CorrelationClustering::edges_from_distances

### Description
Creates signed edges from a distance matrix and a threshold. Pairs closer than the threshold get positive weight, pairs farther get negative weight, and pairs at exactly the threshold get weight 0.

### Method
edges_from_distances

### Signature
```rust
pub fn edges_from_distances<D: DistanceMetric>(
    data: &[Vec<f32>],
    metric: &D,
    threshold: f32,
) -> Vec<SignedEdge>
```

### Parameters
- **data** (&[Vec<f32>]) - Required - The distance matrix or data points.
- **metric** (&D) - Required - The distance metric used to compute distances.
- **threshold** (f32) - Required - The threshold for determining positive or negative edges.

### Returns
- **Vec<SignedEdge>** - A vector of signed edges.

### Example
```rust
// Example usage with a distance metric
let data = vec![vec![0.0, 1.0], vec![1.0, 0.0]];
let metric = EuclideanDistance;
let edges = CorrelationClustering::edges_from_distances(&data, &metric, 0.5);
```
```

--------------------------------

### as_ref example

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=std%3A%3Avec

Shows how to use Result::as_ref to convert a Result to a Result of references, leaving the original intact.

```rust
let x: Result<u32, &str> = Ok(2);
assert_eq!(x.as_ref(), Ok(&2));

let x: Result<u32, &str> = Err("Error");
assert_eq!(x.as_ref(), Err(&"Error"));
```

--------------------------------

### CorrelationClustering::edges_from_distances

Source: https://docs.rs/clump/latest/clump/cluster/correlation/struct.CorrelationClustering.html?search=Option%3CT%3E%2C+%28T+-%3E+U%29+-%3E+Option%3CU%3E

Creates signed edges from a distance matrix and a threshold. Pairs closer than the threshold get positive weight, pairs farther get negative weight, and pairs at exactly the threshold get weight 0.

```APIDOC
## CorrelationClustering::edges_from_distances

### Description
Create signed edges from a distance matrix and a threshold. Pairs closer than `threshold` get positive weight `= threshold - distance`. Pairs farther than `threshold` get negative weight `= threshold - distance` (which is negative). Pairs at exactly `threshold` get weight 0 and are included but contribute no cost.

### Method
`pub fn edges_from_distances<D: DistanceMetric>(data: &[Vec<f32>], metric: &D, threshold: f32) -> Vec<SignedEdge>`

### Parameters
- **data** (&[Vec<f32>]) - Required - The distance matrix (or feature vectors) used to compute distances.
- **metric** (&D) - Required - The distance metric to use.
- **threshold** (f32) - Required - The threshold for determining positive/negative edges.

### Returns
A `Vec<SignedEdge>` containing the signed edges.

### Example
```rust
// Example usage with a distance metric
let data = vec![vec![0.0, 1.0], vec![1.0, 0.0]];
let metric = EuclideanDistance;
let edges = CorrelationClustering::edges_from_distances(&data, &metric, 0.5);
```
```

--------------------------------

### expect example: panic on Err

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=

Shows how to use expect to return the contained Ok value, panicking with a message if the result is Err.

```rust
let x: Result<u32, &str> = Err("emergency failure");
x.expect("Testing expect"); // panics with `Testing expect: emergency failure`
```

--------------------------------

### CorrelationClustering::edges_from_distances

Source: https://docs.rs/clump/latest/clump/cluster/correlation/struct.CorrelationClustering.html?search=

Creates signed edges from a distance matrix and a threshold. Pairs closer than the threshold get positive weight, farther pairs get negative weight, and pairs at exactly the threshold get weight 0.

```APIDOC
## CorrelationClustering::edges_from_distances

### Description
Creates signed edges from a distance matrix and a threshold. Pairs closer than `threshold` get positive weight `= threshold - distance`. Pairs farther than `threshold` get negative weight `= threshold - distance` (which is negative). Pairs at exactly `threshold` get weight 0 and are included but contribute no cost.

### Method
`pub fn edges_from_distances<D: DistanceMetric>(data: &[Vec<f32>], metric: &D, threshold: f32) -> Vec<SignedEdge>`

### Parameters
- **data** (&[Vec<f32>]) - Required - The distance matrix.
- **metric** (&D) - Required - The distance metric used to compute distances.
- **threshold** (f32) - Required - The threshold for determining positive/negative edges.

### Returns
- `Vec<SignedEdge>` - A vector of signed edges.

### Example
```rust
// Assuming a DistanceMetric implementation exists
let edges = CorrelationClustering::edges_from_distances(&data, &metric, 0.5);
```
```

--------------------------------

### Fit K-means and inspect centroids, labels, and iterations

Source: https://docs.rs/clump/latest/src/clump/cluster/kmeans.rs.html?search=Option%3CT%3E%2C+%28T+-%3E+U%29+-%3E+Option%3CU%3E

Demonstrates fitting a Kmeans clusterer with 2 clusters on a small dataset, then verifying the number of centroids, labels, and iteration count.

```rust
use clump::Kmeans;

let data = vec![vec![0.0f32, 0.0], vec![0.1, 0.1], vec![5.0, 5.0], vec![5.1, 5.1]];
let fit = Kmeans::new(2).with_seed(42).fit(&data).unwrap();

assert_eq!(fit.centroids.len(), 2);
assert_eq!(fit.labels.len(), 4);
assert!(fit.iters > 0);
```

--------------------------------

### KMeans fit() method: validation and initialization

Source: https://docs.rs/clump/latest/src/clump/cluster/kmeans.rs.html?search=Option%3CT%3E%2C+%28T+-%3E+U%29+-%3E+Option%3CU%3E

This is the start of the KMeans fit() method. It validates the k parameter, data dimension, cluster count, uniform dimensionality, and finite data. It then normalizes tolerance, initializes the RNG, centroids (warm-start or k-means++), and pre-allocates working buffers including Hamerly bounds, expanded-form norms, and optional GPU/BLAS acceleration.

```rust
273            return Err(Error::InvalidParameter {
274                name: "k",
275                message: "must be at least 1",
276            });
277        }
278
279        let n = data.n();
280        let d = data.d();
281        if d == 0 {
282            return Err(Error::InvalidParameter {
283                name: "dimension",
284                message: "must be at least 1",
285            });
286        }
287
288        if self.k > n {
289            return Err(Error::InvalidClusterCount {
290                requested: self.k,
291                n_items: n,
292            });
293        }
294
295        // Validate uniform dimensionality.
296        for i in 0..n {
297            if data.row(i).len() != d {
298                return Err(Error::DimensionMismatch {
299                    expected: d,
300                    found: data.row(i).len(),
301                });
302            }
303        }
304
305        util::validate_finite(data)?;
306
307        // Normalize tolerance by data variance so it scales with data magnitude.
308        // Without this, the raw tol is compared against the sum of k*d squared
309        // centroid shifts, which becomes meaninglessly tight for high-dimensional
310        // or large-scale data. (Matches scikit-learn's _tolerance approach.)
311        let effective_tol = (self.tol * util::mean_variance(data) * self.k as f64) as f32;
312
313        // Initialize RNG.
314        let mut rng = match self.seed {
315            Some(s) => StdRng::seed_from_u64(s),
316            None => StdRng::from_os_rng(),
317        };
318
319        // Initialize centroids: use provided centroids (warm-start) or k-means++.
320        let mut centroids = if let Some(ref init) = self.init_centroids {
321            assert_eq!(
322                init.len(),
323                self.k,
324                "init_centroids length ({}) must equal k ({})",
325                init.len(),
326                self.k
327            );
328            init.clone()
329        } else {
330            util::kmeanspp_init(data, self.k, &self.metric, self.seeding_alpha, &mut rng)
331        };
332        let mut labels = vec![0usize; n];
333
334        // Pre-allocate working buffers outside the iteration loop to avoid
335        // per-iteration allocation overhead (AMD ROCm pattern).
336        let mut new_centroids = vec![vec![0.0f32; d]; self.k];
337        let mut counts = vec![0usize; self.k];
338
339        // Hamerly bounds: per-point upper (dist to assigned centroid) and
340        // lower (dist to second-nearest centroid). When upper <= lower,
341        // the assignment cannot change and we skip distance computation.
342        // O(n) extra memory (Hamerly, SDM 2010).
343        let mut upper_bounds = vec![f32::MAX; n];
344        let mut lower_bounds = vec![0.0f32; n];
345        let mut centroid_shifts = vec![0.0f32; self.k];
346        let mut sums_f64 = vec![vec![0.0f64; d]; self.k];
347        let mut flat_buf: Vec<f32> = Vec::with_capacity(self.k * d);
348        let mut inertia_trace: Vec<f32> = Vec::with_capacity(self.max_iter);
349
350        // Precompute squared norms for expanded-form first-iteration assignment.
351        // Only valid for SquaredEuclidean (||x-c||^2 = ||x||^2 + ||c||^2 - 2*x.c).
352        let use_expanded = self.metric.supports_expanded_form();
353        let data_norms: Vec<f32> = if use_expanded {
354            util::squared_norms(data)
355        } else {
356            Vec::new()
357        };
358
359        // GPU acceleration: initialize Metal compute pipeline for assignment
360        // when using SquaredEuclidean and problem is large enough to amortize
361        // GPU setup overhead. Buffers for data, labels, and params are allocated
362        // once here and reused across iterations.
363        #[cfg(all(feature = "gpu", target_os = "macos"))]
364        let gpu_assigner = if self.metric.supports_expanded_form() && n * self.k >= 500_000 {
365            let data_flat = super::gpu::flatten(data);
366            super::gpu::GpuAssigner::new(&data_flat, n, self.k, d)
367        } else {
368            None
369        };
370
371        // BLAS GEMM for first-iteration assignment (when blas feature enabled).
372        // matrixmultiply's optimized SGEMM beats per-point distance loops
373        // for large n*k due to micro-kernel SIMD and cache optimization.
374        #[cfg(feature = "blas")]
375        let use_blas = n * self.k >= 100_000 && self.metric.supports_expanded_form();
376        #[cfg(feature = "blas")]
377        let blas_data = if use_blas {
378            let fd = super::flat::FlatMatrix::from_data(data);
379            let xn = fd.row_norms_sq();
380            Some((fd, xn))
381        } else {
382            None
383        };
384
385        let mut iters = 0usize;
386        for iter in 0..self.max_iter {
387            iters = iter + 1;
388
389            // Zero the accumulators for this iteration.
390        }
```

--------------------------------

### is_ok method

Source: https://docs.rs/clump/latest/clump/error/type.Result.html?search=Option%3CT%3E%2C+%28T+-%3E+U%29+-%3E+Option%3CU%3E

Example for the `is_ok` method, showing it returns `true` for `Ok` and `false` for `Err`.

```rust
let x: Result<i32, &str> = Ok(-3);
assert_eq!(x.is_ok(), true);

let x: Result<i32, &str> = Err("Some error message");
assert_eq!(x.is_ok(), false);
```

--------------------------------

### Create signed edges from distance matrix and threshold

Source: https://docs.rs/clump/latest/src/clump/cluster/correlation.rs.html?search=

Builds a vector of SignedEdge for all pairs of items. Pairs closer than the threshold get positive weight (threshold - distance), farther pairs get negative weight, and pairs at exactly the threshold get weight 0 and are included but contribute no cost.

```rust
pub fn edges_from_distances<D: DistanceMetric>(
    data: &[Vec<f32>],
    metric: &D,
    threshold: f32,
) -> Vec<SignedEdge> {
    let n = data.len();
    let mut edges = Vec::with_capacity(n * (n - 1) / 2);
    for i in 0..n {
        for j in (i + 1)..n {
            let dist = metric.distance(&data[i], &data[j]);
            let weight = threshold - dist;
            edges.push(SignedEdge { i, j, weight });
        }
    }
    edges
}
```

--------------------------------

### fit() method: full clustering pipeline

Source: https://docs.rs/clump/latest/src/clump/cluster/correlation.rs.html

Main entry point for clustering. Validates edge indices, runs preclustering, PIVOT, merge, local search, and iterated flip on a contracted graph, then expands back to original vertices. Returns Error::InvalidParameter if any edge index exceeds n_items. Handles n_items == 0 by returning an empty result.

```rust
    pub fn fit(&self, n_items: usize, edges: &[SignedEdge]) -> Result<CorrelationResult> {
        if n_items == 0 {
            return Ok(CorrelationResult {
                labels: Vec::new(),
                n_clusters: 0,
                cost: 0.0,
            });
        }

        // Validate edge indices.
        for edge in edges {
            if edge.i >= n_items || edge.j >= n_items {
                return Err(Error::InvalidParameter {
                    name: "edge index",
                    message: "exceeds n_items",
                });
            }
        }

        // Build adjacency list: for each item, list of (neighbor, weight).
        let adj = Self::build_adj(n_items, edges);

        // Phase 0: Preclustering -- merge vertices with high positive-
        // neighborhood overlap.
        let mapping = precluster(n_items, &adj, 0.5);
        let (n_contracted, contracted_edges) = contract_graph(n_items, edges, &mapping);
        let contracted_adj = Self::build_adj(n_contracted, &contracted_edges);

        // Phase 1: PIVOT + merge + local search on contracted graph.
        let mut labels1 = self.pivot(n_contracted, &contracted_adj);
        if self.max_iter > 0 {
            Self::merge_pass(n_contracted, &contracted_adj, &mut labels1);
            self.local_search(n_contracted, &contracted_adj, &mut labels1);
        }
        let cost1 = compute_cost(&labels1, &contracted_edges);

        // Phase 2: Iterated flip (Cohen-Addad et al. STOC 2024 warm-up).
        // Double weights of inter-cluster edges from phase 1, then re-run.
        let flipped_edges: Vec<SignedEdge> = contracted_edges
            .iter()
            .map(|e| {
                if labels1[e.i] != labels1[e.j] {
                    SignedEdge {
                        weight: e.weight * 2.0,
                        ..*e
                    }
                } else {
                    *e
                }
            })
            .collect();
        let adj2 = Self::build_adj(n_contracted, &flipped_edges);

        let mut labels2 = self.pivot(n_contracted, &adj2);
        if self.max_iter > 0 {
            Self::merge_pass(n_contracted, &adj2, &mut labels2);
            self.local_search(n_contracted, &adj2, &mut labels2);
        }
        // Evaluate cost on contracted (not reweighted) edges.
        let cost2 = compute_cost(&labels2, &contracted_edges);

        // Take better solution on the contracted graph.
        let contracted_labels = if cost2 < cost1 { labels2 } else { labels1 };

        // Expand back to original vertices.
        let mut labels = vec![0usize; n_items];
        for i in 0..n_items {
            labels[i] = contracted_labels[mapping[i]];
        }

        // Evaluate cost on ORIGINAL edges.
        let cost = compute_cost(&labels, edges);
        let (labels, n_clusters) = relabel_contiguous(&labels);

        Ok(CorrelationResult {
            labels,
            n_clusters,
            cost,
        })
    }
```

--------------------------------

### DBSCAN clustering usage example with noise detection

Source: https://docs.rs/clump/latest/clump/cluster/struct.Dbscan.html?search=

Demonstrates creating a Dbscan with epsilon 0.5 and min_pts 2 to cluster data, marking an outlier as noise via the NOISE sentinel. Shows that core cluster points share a label and the outlier is labeled as noise.

```rust
use clump::{Dbscan, NOISE};

let data = vec![
    vec![0.0f32, 0.0], vec![0.1, 0.0], vec![0.0, 0.1], // cluster
    vec![100.0, 100.0], // outlier
];

let labels = Dbscan::new(0.5, 2).fit_predict(&data).unwrap();
assert_eq!(labels[0], labels[1]); // same cluster
assert_eq!(labels[3], NOISE);     // outlier is noise
```
Tokens:
Raw