jon.recoil.org
A fork of mtelver's day10 project
tessera-linalg: polish docs and generate opam file
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
+87
+61
tessera-linalg/lib/linalg.mli
+61
tessera-linalg/lib/linalg.mli
···
1
type vec = (float, Bigarray.float32_elt, Bigarray.c_layout) Bigarray.Array1.t
2
3
type mat = {
4
data : vec;
5
rows : int;
6
cols : int;
7
}
8
9
val create_mat : rows:int -> cols:int -> mat
10
val mat_get : mat -> int -> int -> float
11
val mat_set : mat -> int -> int -> float -> unit
12
13
type pca_model
14
15
val pca_fit : ?max_samples:int -> mat -> n_components:int -> pca_model
16
val pca_transform : pca_model -> mat -> mat
17
18
type knn_model
19
20
val knn_fit : embeddings:mat -> labels:int array -> knn_model
21
22
type knn_result = {
23
predictions : int array;
24
confidences : float array;
25
}
26
27
val knn_predict : knn_model -> k:int -> mat -> knn_result
···
1
+
(** PCA and kNN for float32 Bigarray data.
2
+
3
+
Portable OCaml implementations of PCA dimensionality reduction and
4
+
k-nearest-neighbors classification. Operates on flat float32 Bigarray
5
+
data with explicit shapes.
6
+
7
+
{2 Example: embedding classification}
8
+
9
+
{[
10
+
(* PCA: reduce 128-dim embeddings to 3 components *)
11
+
let pca = Linalg.pca_fit embeddings ~n_components:3 in
12
+
let projected = Linalg.pca_transform pca embeddings in
13
+
14
+
(* kNN: classify using labeled training points *)
15
+
let model = Linalg.knn_fit ~embeddings:train ~labels in
16
+
let result = Linalg.knn_predict model ~k:5 projected in
17
+
(* result.predictions.(i) is the class, result.confidences.(i) in [0,1] *)
18
+
]} *)
19
+
20
+
(** {1 Data types} *)
21
+
22
type vec = (float, Bigarray.float32_elt, Bigarray.c_layout) Bigarray.Array1.t
23
+
(** A flat float32 vector backed by Bigarray. *)
24
25
type mat = {
26
data : vec;
27
rows : int;
28
cols : int;
29
}
30
+
(** A 2D float32 matrix in row-major order. Element [(i, j)] is at flat
31
+
index [i * cols + j] in {!data}. *)
32
+
33
+
(** {1 Matrix operations} *)
34
35
val create_mat : rows:int -> cols:int -> mat
36
+
(** Create a zero-initialized matrix. *)
37
+
38
val mat_get : mat -> int -> int -> float
39
+
(** [mat_get m row col] returns element at [(row, col)]. *)
40
+
41
val mat_set : mat -> int -> int -> float -> unit
42
+
(** [mat_set m row col v] sets element at [(row, col)] to [v]. *)
43
+
44
+
(** {1 PCA} *)
45
46
type pca_model
47
+
(** A fitted PCA model storing the mean vector and principal components. *)
48
49
val pca_fit : ?max_samples:int -> mat -> n_components:int -> pca_model
50
+
(** Fit PCA on a data matrix of shape [(n_samples, n_features)].
51
+
52
+
@param max_samples Limit rows used for covariance estimation
53
+
(default: 100,000). Rows are sampled evenly if the matrix is larger.
54
+
@param n_components Number of principal components to compute.
55
+
56
+
Uses power iteration on the covariance matrix, which is efficient for
57
+
the typical case of 128 features (128x128 covariance). *)
58
+
59
val pca_transform : pca_model -> mat -> mat
60
+
(** Project data through a fitted PCA model.
61
+
62
+
Input: [(n_samples, n_features)]. Output: [(n_samples, n_components)]. *)
63
+
64
+
(** {1 kNN} *)
65
66
type knn_model
67
+
(** A fitted kNN model storing training embeddings and labels. *)
68
69
val knn_fit : embeddings:mat -> labels:int array -> knn_model
70
+
(** Create a kNN model from labeled training data.
71
+
72
+
@param embeddings Training data of shape [(n_training, n_features)].
73
+
@param labels Integer class labels, length [n_training]. *)
74
75
type knn_result = {
76
predictions : int array;
77
confidences : float array;
78
}
79
+
(** Classification output. [predictions.(i)] is the predicted class for
80
+
sample [i]. [confidences.(i)] is in [\[0, 1\]] — the fraction of
81
+
distance-weighted votes for the winning class. *)
82
83
val knn_predict : knn_model -> k:int -> mat -> knn_result
84
+
(** Classify samples using k-nearest neighbors with distance weighting.
85
+
86
+
Input: [(n_samples, n_features)]. [k] is capped to the number of
87
+
training points. Weights are [1 / (distance + epsilon)] to handle
88
+
exact matches gracefully. *)
+26
tessera-linalg/tessera-linalg.opam
+26
tessera-linalg/tessera-linalg.opam
···
···
1
+
# This file is generated by dune, edit dune-project instead
2
+
opam-version: "2.0"
3
+
synopsis: "PCA and kNN for float32 Bigarray data"
4
+
description:
5
+
"Portable OCaml implementations of PCA dimensionality reduction and k-nearest-neighbors classification. Operates on flat float32 Bigarray data with explicit shapes."
6
+
license: "ISC"
7
+
depends: [
8
+
"dune" {>= "3.17"}
9
+
"ocaml" {>= "5.2"}
10
+
"alcotest" {with-test & >= "0.8"}
11
+
"odoc" {with-doc}
12
+
]
13
+
build: [
14
+
["dune" "subst"] {dev}
15
+
[
16
+
"dune"
17
+
"build"
18
+
"-p"
19
+
name
20
+
"-j"
21
+
jobs
22
+
"@install"
23
+
"@runtest" {with-test}
24
+
"@doc" {with-doc}
25
+
]
26
+
]