jon.recoil.org
this repo has no description
tessera-linalg: implement kNN with distance-weighted voting
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
+1
-3
tessera-linalg/lib/dune
+1
-3
tessera-linalg/lib/dune
+42
-4
tessera-linalg/lib/linalg.ml
+42
-4
tessera-linalg/lib/linalg.ml
···
123
123
done;
124
124
result
125
125
126
-
(* ---- kNN (stub) ---- *)
126
+
(* ---- kNN ---- *)
127
127
128
128
type knn_model = {
129
129
embeddings : mat;
···
138
138
confidences : float array;
139
139
}
140
140
141
-
let knn_predict _model ~k:_ query =
142
-
{ predictions = Array.make query.rows 0;
143
-
confidences = Array.make query.rows 0.0 }
141
+
let knn_predict model ~k query =
142
+
let n_train = model.embeddings.rows in
143
+
let n_features = model.embeddings.cols in
144
+
let k = min k n_train in
145
+
let predictions = Array.make query.rows 0 in
146
+
let confidences = Array.make query.rows 0.0 in
147
+
for qi = 0 to query.rows - 1 do
148
+
(* Compute squared distances *)
149
+
let dists = Array.init n_train (fun ti ->
150
+
let d = ref 0.0 in
151
+
for j = 0 to n_features - 1 do
152
+
let diff = mat_get query qi j -. mat_get model.embeddings ti j in
153
+
d := !d +. diff *. diff
154
+
done;
155
+
(!d, model.labels.(ti))
156
+
) in
157
+
(* Sort by distance *)
158
+
Array.sort (fun (d1, _) (d2, _) -> Float.compare d1 d2) dists;
159
+
(* Distance-weighted vote for k nearest *)
160
+
let class_weights = Hashtbl.create 16 in
161
+
let total_weight = ref 0.0 in
162
+
for i = 0 to k - 1 do
163
+
let (dist, label) = dists.(i) in
164
+
let w = 1.0 /. (dist +. 1e-10) in
165
+
total_weight := !total_weight +. w;
166
+
let prev = try Hashtbl.find class_weights label with Not_found -> 0.0 in
167
+
Hashtbl.replace class_weights label (prev +. w)
168
+
done;
169
+
(* Find class with highest weight *)
170
+
let best_class = ref 0 in
171
+
let best_weight = ref Float.neg_infinity in
172
+
Hashtbl.iter (fun cls w ->
173
+
if w > !best_weight then begin
174
+
best_class := cls;
175
+
best_weight := w
176
+
end
177
+
) class_weights;
178
+
predictions.(qi) <- !best_class;
179
+
confidences.(qi) <- !best_weight /. !total_weight
180
+
done;
181
+
{ predictions; confidences }
+50
-5
tessera-linalg/test/test_linalg.ml
+50
-5
tessera-linalg/test/test_linalg.ml
···
38
38
(* ---- PCA tests ---- *)
39
39
40
40
let test_pca_x_axis_points () =
41
-
(* 4 points along x-axis in 3D: (1,0,0), (2,0,0), (3,0,0), (4,0,0)
42
-
PCA to 1D should produce evenly spaced output *)
43
41
let data = mat_of_arrays [|
44
42
[| 1.0; 0.0; 0.0 |];
45
43
[| 2.0; 0.0; 0.0 |];
···
50
48
let result = pca_transform model data in
51
49
Alcotest.(check int) "result rows" 4 result.rows;
52
50
Alcotest.(check int) "result cols" 1 result.cols;
53
-
(* Should be evenly spaced: differences between consecutive should be equal *)
54
51
let v0 = mat_get result 0 0 in
55
52
let v1 = mat_get result 1 0 in
56
53
let v2 = mat_get result 2 0 in
···
62
59
Alcotest.(check (float eps)) "spacing d12~d23" d12 d23
63
60
64
61
let test_pca_diagonal_monotonic () =
65
-
(* 100 points along y=x in 2D -> PCA to 1D -> monotonic output *)
66
62
let rows_data = Array.init 100 (fun i ->
67
63
let v = Float.of_int i in
68
64
[| v; v |]
···
71
67
let model = pca_fit data ~n_components:1 in
72
68
let result = pca_transform model data in
73
69
Alcotest.(check int) "result rows" 100 result.rows;
74
-
(* Check monotonic (either all increasing or all decreasing) *)
75
70
let increasing = ref true in
76
71
let decreasing = ref true in
77
72
for i = 1 to 99 do
···
93
88
Alcotest.(check int) "rows" 3 result.rows;
94
89
Alcotest.(check int) "cols" 2 result.cols
95
90
91
+
(* ---- kNN tests ---- *)
92
+
93
+
let test_knn_two_clusters () =
94
+
let train = mat_of_arrays [|
95
+
[| 0.0; 0.0 |];
96
+
[| 0.1; 0.1 |];
97
+
[| 0.2; 0.0 |];
98
+
[| 100.0; 100.0 |];
99
+
[| 100.1; 100.1 |];
100
+
[| 99.9; 100.0 |];
101
+
|] in
102
+
let labels = [| 0; 0; 0; 1; 1; 1 |] in
103
+
let model = knn_fit ~embeddings:train ~labels in
104
+
let test_pts = mat_of_arrays [|
105
+
[| 0.05; 0.05 |];
106
+
[| 99.95; 100.05 |];
107
+
|] in
108
+
let res = knn_predict model ~k:3 test_pts in
109
+
Alcotest.(check int) "pred 0" 0 res.predictions.(0);
110
+
Alcotest.(check int) "pred 1" 1 res.predictions.(1)
111
+
112
+
let test_knn_distance_weighting () =
113
+
let train = mat_of_arrays [|
114
+
[| 0.0; 0.0 |];
115
+
[| 100.0; 0.0 |];
116
+
[| 100.0; 1.0 |];
117
+
|] in
118
+
let labels = [| 0; 1; 1 |] in
119
+
let model = knn_fit ~embeddings:train ~labels in
120
+
let query = mat_of_arrays [| [| 0.01; 0.01 |] |] in
121
+
let res = knn_predict model ~k:3 query in
122
+
Alcotest.(check int) "close wins" 0 res.predictions.(0)
123
+
124
+
let test_knn_k1_confidence () =
125
+
let train = mat_of_arrays [|
126
+
[| 0.0; 0.0 |];
127
+
[| 10.0; 10.0 |];
128
+
|] in
129
+
let labels = [| 0; 1 |] in
130
+
let model = knn_fit ~embeddings:train ~labels in
131
+
let query = mat_of_arrays [| [| 0.1; 0.1 |] |] in
132
+
let res = knn_predict model ~k:1 query in
133
+
Alcotest.(check int) "k=1 pred" 0 res.predictions.(0);
134
+
Alcotest.(check (float 1e-6)) "k=1 confidence" 1.0 res.confidences.(0)
135
+
96
136
(* ---- Test runner ---- *)
97
137
98
138
let () =
···
106
146
Alcotest.test_case "x-axis points evenly spaced" `Quick test_pca_x_axis_points;
107
147
Alcotest.test_case "diagonal monotonic" `Quick test_pca_diagonal_monotonic;
108
148
Alcotest.test_case "output dimensions" `Quick test_pca_output_dims;
149
+
];
150
+
"knn", [
151
+
Alcotest.test_case "two clusters" `Quick test_knn_two_clusters;
152
+
Alcotest.test_case "distance weighting" `Quick test_knn_distance_weighting;
153
+
Alcotest.test_case "k=1 confidence" `Quick test_knn_k1_confidence;
109
154
];
110
155
]