forked from
anil.recoil.org/monopam-myspace
My aggregated monorepo of OCaml code, automaintained
Add tessera-tfjs: TF.js SVD-based PCA for GeoTessera notebook
New tessera-tfjs library provides minimal js_of_ocaml bindings to
TensorFlow.js for PCA computation. Replaces the slow power-iteration
PCA in tessera-linalg with SVD on the covariance matrix via tf.linalg.svd.
- Tfjs.pca: single-call API, converts Linalg.mat to tf.tensor2d,
mean-centers, computes covariance SVD, projects, converts back
- All intermediate tensors disposed to prevent memory leaks
- TF.js loaded via CDN importScripts in the notebook setup cell
- Notebook updated to use Tfjs.pca instead of Linalg.pca_fit/transform
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
+531
-4
+1
-1
build-site.sh
+1
-1
build-site.sh
···
70
70
dune exec -- jtw opam astring base brr note mime_printer fpath rresult \
71
71
opam-format bos odoc.model tyxml yojson uri jsonm \
72
72
js_top_worker-widget-leaflet \
73
-
tessera-geotessera-jsoo tessera-viz-jsoo \
73
+
tessera-geotessera-jsoo tessera-viz-jsoo tessera-tfjs \
74
74
onnxrt -o "$SITE/_opam"
75
75
echo " universe built → $SITE/_opam/"
76
76
fi
+378
docs/superpowers/plans/2026-03-18-tessera-tfjs-pca.md
+378
docs/superpowers/plans/2026-03-18-tessera-tfjs-pca.md
···
1
+
# tessera-tfjs: TF.js PCA Bindings Implementation Plan
2
+
3
+
> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
4
+
5
+
**Goal:** Replace the slow power-iteration PCA in tessera-linalg with GPU-accelerated SVD-based PCA via TensorFlow.js bindings, improving both speed and accuracy for the GeoTessera notebook.
6
+
7
+
**Architecture:** A new `tessera-tfjs` library provides minimal js_of_ocaml bindings to `@tensorflow/tfjs`, exposing a single `Tfjs.pca` function. The bindings use `Js.Unsafe` to call TF.js APIs loaded via CDN `importScripts` in the web worker. The notebook's setup cell loads TF.js and `#require`s the new library.
8
+
9
+
**Tech Stack:** OCaml, js_of_ocaml, js_of_ocaml-ppx, TensorFlow.js (CDN), dune
10
+
11
+
---
12
+
13
+
## File Structure
14
+
15
+
| File | Responsibility |
16
+
|------|---------------|
17
+
| `tessera-tfjs/dune-project` | Package metadata |
18
+
| `tessera-tfjs/lib/dune` | Library build config |
19
+
| `tessera-tfjs/lib/tfjs.ml` | TF.js bindings and PCA implementation |
20
+
| `tessera-tfjs/lib/tfjs.mli` | Public API |
21
+
| `tessera-tfjs/tessera-tfjs.opam` | Generated by dune |
22
+
| `site/notebooks/interactive_map.mld` | Notebook: load TF.js CDN, use `Tfjs.pca` |
23
+
| `build-site.sh` | Add `tessera-tfjs` to jtw universe package list |
24
+
25
+
---
26
+
27
+
### Task 1: Project scaffold
28
+
29
+
**Files:**
30
+
- Create: `tessera-tfjs/dune-project`
31
+
- Create: `tessera-tfjs/lib/dune`
32
+
- Create: `tessera-tfjs/lib/tfjs.mli`
33
+
- Create: `tessera-tfjs/lib/tfjs.ml`
34
+
35
+
- [ ] **Step 1: Create dune-project**
36
+
37
+
```
38
+
tessera-tfjs/dune-project
39
+
```
40
+
41
+
```
42
+
(lang dune 3.17)
43
+
(name tessera-tfjs)
44
+
(generate_opam_files true)
45
+
(license ISC)
46
+
(package
47
+
(name tessera-tfjs)
48
+
(synopsis "TensorFlow.js PCA bindings for OCaml")
49
+
(description "Minimal js_of_ocaml bindings to TensorFlow.js for SVD-based PCA. Operates on tessera-linalg mat types.")
50
+
(depends
51
+
(ocaml (>= 5.2))
52
+
(tessera-linalg (>= 0.1))
53
+
(js_of_ocaml (>= 5.0))
54
+
(js_of_ocaml-ppx (>= 5.0))))
55
+
```
56
+
57
+
- [ ] **Step 2: Create lib/dune**
58
+
59
+
```
60
+
tessera-tfjs/lib/dune
61
+
```
62
+
63
+
```
64
+
(library
65
+
(name tfjs)
66
+
(public_name tessera-tfjs)
67
+
(libraries tessera-linalg js_of_ocaml)
68
+
(preprocess (pps js_of_ocaml-ppx)))
69
+
```
70
+
71
+
- [ ] **Step 3: Create stub tfjs.mli**
72
+
73
+
```
74
+
tessera-tfjs/lib/tfjs.mli
75
+
```
76
+
77
+
```ocaml
78
+
(** TensorFlow.js PCA via SVD.
79
+
80
+
Requires TensorFlow.js to be loaded in the JavaScript environment
81
+
(e.g., via [importScripts] in a web worker). *)
82
+
83
+
val pca : Linalg.mat -> n_components:int -> Linalg.mat
84
+
(** [pca mat ~n_components] computes PCA on a matrix of shape
85
+
[(n_samples, n_features)] and returns the projected data of shape
86
+
[(n_samples, n_components)].
87
+
88
+
Uses TF.js SVD for accurate principal component extraction.
89
+
All intermediate tensors are disposed to avoid memory leaks.
90
+
91
+
@raise Failure if TensorFlow.js is not loaded. *)
92
+
```
93
+
94
+
- [ ] **Step 4: Create stub tfjs.ml**
95
+
96
+
```
97
+
tessera-tfjs/lib/tfjs.ml
98
+
```
99
+
100
+
```ocaml
101
+
let pca _mat ~n_components:_ =
102
+
failwith "TODO: implement TF.js PCA"
103
+
```
104
+
105
+
- [ ] **Step 5: Build to verify scaffold compiles**
106
+
107
+
Run: `opam exec -- dune build tessera-tfjs`
108
+
Expected: success, no errors
109
+
110
+
- [ ] **Step 6: Commit**
111
+
112
+
```bash
113
+
git add tessera-tfjs/
114
+
git commit -m "tessera-tfjs: scaffold project with stub PCA API"
115
+
```
116
+
117
+
---
118
+
119
+
### Task 2: Implement TF.js bindings and PCA
120
+
121
+
**Files:**
122
+
- Modify: `tessera-tfjs/lib/tfjs.ml`
123
+
124
+
The implementation converts a `Linalg.mat` (float32 bigarray) to a TF.js tensor,
125
+
computes mean-centered SVD, projects onto the top-k components, and converts back.
126
+
All TF.js calls go through `Js.Unsafe` since we're binding to a globally-loaded
127
+
library, not a compiled dependency.
128
+
129
+
- [ ] **Step 1: Implement the TF.js PCA**
130
+
131
+
```
132
+
tessera-tfjs/lib/tfjs.ml
133
+
```
134
+
135
+
```ocaml
136
+
open Js_of_ocaml
137
+
138
+
(* Access the global tf object loaded via importScripts/CDN *)
139
+
let tf () : Js.Unsafe.any =
140
+
let v = Js.Unsafe.global##.tf in
141
+
if Js.Optdef.test v then v
142
+
else failwith "TensorFlow.js not loaded. Load it via importScripts before calling Tfjs.pca."
143
+
144
+
(* Helper: call a method on a JS object *)
145
+
let call obj meth args =
146
+
Js.Unsafe.meth_call obj meth (Array.of_list args)
147
+
148
+
(* Helper: get a nested property *)
149
+
let get obj prop =
150
+
Js.Unsafe.get obj (Js.string prop)
151
+
152
+
(* Convert Linalg.mat (float32 bigarray, row-major) to tf.tensor2d *)
153
+
let mat_to_tensor (mat : Linalg.mat) : Js.Unsafe.any =
154
+
let tf = tf () in
155
+
(* Create a Float32Array view of the bigarray data *)
156
+
let len = mat.rows * mat.cols in
157
+
let js_arr = new%js Typed_array.float32Array len in
158
+
for i = 0 to len - 1 do
159
+
Typed_array.set js_arr i (Bigarray.Array1.get mat.data i)
160
+
done;
161
+
Js.Unsafe.meth_call tf "tensor2d"
162
+
[| Js.Unsafe.inject js_arr;
163
+
Js.Unsafe.inject (Js.array [| mat.rows; mat.cols |]) |]
164
+
165
+
(* Convert tf.tensor2d back to Linalg.mat *)
166
+
let tensor_to_mat (t : Js.Unsafe.any) : Linalg.mat =
167
+
let shape = Js.to_array (get t "shape") in
168
+
let rows = shape.(0) in
169
+
let cols = shape.(1) in
170
+
let result = Linalg.create_mat ~rows ~cols in
171
+
(* dataSync returns a Float32Array *)
172
+
let data : Typed_array.float32Array Js.t = call t "dataSync" [] in
173
+
let len = rows * cols in
174
+
for i = 0 to len - 1 do
175
+
Bigarray.Array1.set result.data i (Typed_array.get data i)
176
+
done;
177
+
result
178
+
179
+
(* Dispose a tensor to free GPU/CPU memory *)
180
+
let dispose t = ignore (call t "dispose" [])
181
+
182
+
let pca (mat : Linalg.mat) ~n_components : Linalg.mat =
183
+
let tf = tf () in
184
+
(* 1. Convert input to tensor *)
185
+
let x = mat_to_tensor mat in
186
+
(* 2. Compute mean along axis 0 *)
187
+
let mean = call x "mean" [Js.Unsafe.inject 0] in
188
+
(* 3. Center the data: x_centered = x - mean *)
189
+
let x_centered = call x "sub" [Js.Unsafe.inject mean] in
190
+
(* 4. Compute covariance-like matrix: X^T X / (n-1)
191
+
For SVD-based PCA, we compute SVD of the centered data directly.
192
+
For n_samples >> n_features (typical: 40k x 128), computing
193
+
the covariance matrix (128x128) and then eigendecomposing is faster
194
+
than SVD of the full matrix. *)
195
+
let n_samples = Float.of_int mat.rows in
196
+
let xt = call x_centered "transpose" [] in
197
+
(* cov = X^T X / (n-1), shape (n_features, n_features) *)
198
+
let xtx = call (get tf "linalg") "matMul" [Js.Unsafe.inject xt; Js.Unsafe.inject x_centered] in
199
+
let scale = Js.Unsafe.meth_call tf "scalar" [| Js.Unsafe.inject (n_samples -. 1.0) |] in
200
+
let cov = call xtx "div" [Js.Unsafe.inject scale] in
201
+
(* 5. Eigendecompose the symmetric covariance matrix.
202
+
tf.linalg.eig is not available, but we can use SVD on the symmetric
203
+
matrix: for symmetric A, SVD gives A = U S V^T where U = V and S = eigenvalues. *)
204
+
let svd_result = Js.Unsafe.meth_call (get tf "linalg") "svd"
205
+
[| Js.Unsafe.inject cov; Js.Unsafe.inject Js._true |] in
206
+
(* svd returns {s, u, v} or [s, u, v] — TF.js returns an object *)
207
+
let v_full = get svd_result "v" in
208
+
(* 6. Take top n_components columns of V *)
209
+
let v_top = Js.Unsafe.meth_call v_full "slice"
210
+
[| Js.Unsafe.inject (Js.array [| 0; 0 |]);
211
+
Js.Unsafe.inject (Js.array [| mat.cols; n_components |]) |] in
212
+
(* 7. Project: result = x_centered @ v_top, shape (n_samples, n_components) *)
213
+
let projected = call (get tf "linalg") "matMul"
214
+
[Js.Unsafe.inject x_centered; Js.Unsafe.inject v_top] in
215
+
(* 8. Convert back to Linalg.mat *)
216
+
let result = tensor_to_mat projected in
217
+
(* 9. Clean up all tensors *)
218
+
List.iter dispose [x; mean; x_centered; xt; xtx; scale; cov;
219
+
get svd_result "s"; get svd_result "u"; v_full;
220
+
v_top; projected];
221
+
result
222
+
```
223
+
224
+
- [ ] **Step 2: Build to verify it compiles**
225
+
226
+
Run: `opam exec -- dune build tessera-tfjs`
227
+
Expected: success, no errors
228
+
229
+
- [ ] **Step 3: Commit**
230
+
231
+
```bash
232
+
git add tessera-tfjs/lib/tfjs.ml
233
+
git commit -m "tessera-tfjs: implement PCA via TF.js SVD on covariance matrix"
234
+
```
235
+
236
+
---
237
+
238
+
### Task 3: Integrate into the notebook
239
+
240
+
**Files:**
241
+
- Modify: `site/notebooks/interactive_map.mld`
242
+
- Modify: `build-site.sh`
243
+
244
+
- [ ] **Step 1: Add tessera-tfjs to the jtw universe in build-site.sh**
245
+
246
+
In `build-site.sh`, add `tessera-tfjs` to the jtw package list (line ~73):
247
+
248
+
```bash
249
+
dune exec -- jtw opam astring base brr note mime_printer fpath rresult \
250
+
opam-format bos odoc.model tyxml yojson uri jsonm \
251
+
js_top_worker-widget-leaflet \
252
+
tessera-geotessera-jsoo tessera-viz-jsoo tessera-tfjs \
253
+
onnxrt -o "$SITE/_opam"
254
+
```
255
+
256
+
- [ ] **Step 2: Update the notebook setup cell to load TF.js and require tessera-tfjs**
257
+
258
+
In `site/notebooks/interactive_map.mld`, modify the setup cell (lines 15-20).
259
+
260
+
Before:
261
+
```
262
+
{@ocaml kind=setup[
263
+
#require "tessera-geotessera-jsoo";;
264
+
#require "tessera-viz-jsoo";;
265
+
#require "js_top_worker-widget-leaflet";;
266
+
Widget_leaflet.register ();;
267
+
]}
268
+
```
269
+
270
+
After:
271
+
```
272
+
{@ocaml kind=setup[
273
+
#require "tessera-geotessera-jsoo";;
274
+
#require "tessera-viz-jsoo";;
275
+
#require "tessera-tfjs";;
276
+
#require "js_top_worker-widget-leaflet";;
277
+
Widget_leaflet.register ();;
278
+
(* Load TensorFlow.js in the worker *)
279
+
let () =
280
+
let open Js_of_ocaml in
281
+
Js.Unsafe.global##importScripts
282
+
(Js.string "https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@4/dist/tf.min.js")
283
+
]}
284
+
```
285
+
286
+
- [ ] **Step 3: Replace Linalg PCA with Tfjs.pca in the fetch cell**
287
+
288
+
In `site/notebooks/interactive_map.mld`, in the fetch/visualize cell (around lines 136-138).
289
+
290
+
Before:
291
+
```ocaml
292
+
let pca = Linalg.pca_fit ~max_samples:5000 mat ~n_components:3 in
293
+
let proj = Linalg.pca_transform pca mat in
294
+
```
295
+
296
+
After:
297
+
```ocaml
298
+
let proj = Tfjs.pca mat ~n_components:3 in
299
+
```
300
+
301
+
- [ ] **Step 4: Build everything**
302
+
303
+
Run: `opam exec -- dune build`
304
+
Expected: success, no errors
305
+
306
+
- [ ] **Step 5: Commit**
307
+
308
+
```bash
309
+
git add site/notebooks/interactive_map.mld build-site.sh
310
+
git commit -m "notebook: use TF.js PCA via tessera-tfjs for faster, more accurate PCA"
311
+
```
312
+
313
+
---
314
+
315
+
### Task 4: Test in browser
316
+
317
+
**Files:** (none modified — manual testing)
318
+
319
+
- [ ] **Step 1: Rebuild the site universe**
320
+
321
+
```bash
322
+
rm -rf _site/_opam
323
+
./build-site.sh --fresh
324
+
```
325
+
326
+
Or for faster iteration during development:
327
+
328
+
```bash
329
+
dune exec -- jtw opam astring base brr note mime_printer fpath rresult \
330
+
opam-format bos odoc.model tyxml yojson uri jsonm \
331
+
js_top_worker-widget-leaflet \
332
+
tessera-geotessera-jsoo tessera-viz-jsoo tessera-tfjs \
333
+
onnxrt -o _site/_opam
334
+
```
335
+
336
+
- [ ] **Step 2: Serve and test**
337
+
338
+
```bash
339
+
cd _site && python3 -m http.server 8080
340
+
```
341
+
342
+
Open `http://localhost:8080/notebooks/interactive_map.html`:
343
+
1. Wait for setup cell to load (should load TF.js without errors)
344
+
2. Draw a bounding box on the map
345
+
3. Run the fetch cell
346
+
4. Verify: PCA overlay appears on the map
347
+
5. Check browser console for any TF.js errors
348
+
349
+
- [ ] **Step 3: Verify TF.js is actually being used**
350
+
351
+
Open the browser console. After running the fetch cell, you should see TF.js
352
+
initialization messages (e.g., "Initializing backend 'cpu'" or "'webgl'").
353
+
No messages about "TensorFlow.js not loaded" errors.
354
+
355
+
---
356
+
357
+
### Notes
358
+
359
+
**TF.js SVD approach:** We use covariance matrix eigendecomposition rather than
360
+
direct SVD on the full data matrix. For the typical GeoTessera case (40k samples
361
+
× 128 features), computing the 128×128 covariance matrix and SVD-ing it is much
362
+
faster than SVD on a 40k×128 matrix.
363
+
364
+
**TF.js loading:** The CDN script is loaded via `importScripts` in the worker's
365
+
setup cell. This is synchronous and blocks until loaded (~1-2 seconds). The TF.js
366
+
global `tf` object is then available for all subsequent cells.
367
+
368
+
**Memory management:** All intermediate TF.js tensors are explicitly disposed
369
+
after use. TF.js tracks tensors for manual memory management — failing to dispose
370
+
causes GPU/CPU memory leaks.
371
+
372
+
**Fallback:** The existing `Linalg.pca_fit`/`pca_transform` remains available as
373
+
a pure-OCaml fallback for environments without TF.js.
374
+
375
+
**TF.js SVD on symmetric matrices:** For a symmetric positive semi-definite matrix
376
+
(like a covariance matrix), SVD is equivalent to eigendecomposition. The right
377
+
singular vectors V are the eigenvectors, and the singular values S are the
378
+
eigenvalues. `tf.linalg.svd` with `fullMatrices=true` returns `{s, u, v}`.
+8
-3
site/notebooks/interactive_map.mld
+8
-3
site/notebooks/interactive_map.mld
···
15
15
{@ocaml kind=setup[
16
16
#require "tessera-geotessera-jsoo";;
17
17
#require "tessera-viz-jsoo";;
18
+
#require "tessera-tfjs";;
18
19
#require "js_top_worker-widget-leaflet";;
19
20
Widget_leaflet.register ();;
21
+
(* Load TensorFlow.js in the worker *)
22
+
let () =
23
+
let open Js_of_ocaml in
24
+
Js.Unsafe.global##importScripts
25
+
(Js.string "https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@4/dist/tf.min.js")
20
26
]}
21
27
22
28
{1 Select region of interest}
···
133
139
let mosaic_east = mosaic_bbox.Geotessera.max_lon in
134
140
Widget.update ~id:"status"
135
141
(status_view (Printf.sprintf "Working at %d×%d (%d pixels). Computing PCA..." h w (h * w)));
136
-
(* PCA to 3 components for RGB visualisation *)
137
-
let pca = Linalg.pca_fit ~max_samples:5000 mat ~n_components:3 in
138
-
let proj = Linalg.pca_transform pca mat in
142
+
(* PCA to 3 components for RGB visualisation (via TensorFlow.js SVD) *)
143
+
let proj = Tfjs.pca mat ~n_components:3 in
139
144
projected := Some proj;
140
145
let pca_img = Viz.pca_to_rgba ~width:w ~height:h proj in
141
146
let url = Viz_jsoo.to_data_url pca_img in
+13
tessera-tfjs/dune-project
+13
tessera-tfjs/dune-project
···
1
+
(lang dune 3.17)
2
+
(name tessera-tfjs)
3
+
(generate_opam_files true)
4
+
(license ISC)
5
+
(package
6
+
(name tessera-tfjs)
7
+
(synopsis "TensorFlow.js PCA bindings for OCaml")
8
+
(description "Minimal js_of_ocaml bindings to TensorFlow.js for SVD-based PCA. Operates on tessera-linalg mat types.")
9
+
(depends
10
+
(ocaml (>= 5.2))
11
+
(tessera-linalg (>= 0.1))
12
+
(js_of_ocaml (>= 5.0))
13
+
(js_of_ocaml-ppx (>= 5.0))))
+5
tessera-tfjs/lib/dune
+5
tessera-tfjs/lib/dune
+83
tessera-tfjs/lib/tfjs.ml
+83
tessera-tfjs/lib/tfjs.ml
···
1
+
open Js_of_ocaml
2
+
3
+
(* Access the global tf object loaded via importScripts/CDN *)
4
+
let tf () : Js.Unsafe.any =
5
+
let v = Js.Unsafe.global##.tf in
6
+
if Js.Optdef.test v then v
7
+
else failwith "TensorFlow.js not loaded. Load it via importScripts before calling Tfjs.pca."
8
+
9
+
(* Helper: call a method on a JS object *)
10
+
let call obj meth args =
11
+
Js.Unsafe.meth_call obj meth (Array.of_list args)
12
+
13
+
(* Helper: get a nested property *)
14
+
let get obj prop =
15
+
Js.Unsafe.get obj (Js.string prop)
16
+
17
+
(* Convert Linalg.mat (float32 bigarray, row-major) to tf.tensor2d *)
18
+
let mat_to_tensor (mat : Linalg.mat) : Js.Unsafe.any =
19
+
let tf = tf () in
20
+
let len = mat.rows * mat.cols in
21
+
let js_arr = new%js Typed_array.float32Array len in
22
+
for i = 0 to len - 1 do
23
+
Typed_array.set js_arr i (Js.number_of_float (Bigarray.Array1.get mat.data i))
24
+
done;
25
+
Js.Unsafe.meth_call tf "tensor2d"
26
+
[| Js.Unsafe.inject js_arr;
27
+
Js.Unsafe.inject (Js.array [| mat.rows; mat.cols |]) |]
28
+
29
+
(* Convert tf.tensor2d back to Linalg.mat *)
30
+
let tensor_to_mat (t : Js.Unsafe.any) : Linalg.mat =
31
+
let shape = Js.to_array (get t "shape") in
32
+
let rows = shape.(0) in
33
+
let cols = shape.(1) in
34
+
let result = Linalg.create_mat ~rows ~cols in
35
+
let data : Typed_array.float32Array Js.t = call t "dataSync" [] in
36
+
let len = rows * cols in
37
+
for i = 0 to len - 1 do
38
+
let v = Js.Optdef.get (Typed_array.get data i) (fun () -> Js.number_of_float 0.0) in
39
+
Bigarray.Array1.set result.data i (Js.float_of_number v)
40
+
done;
41
+
result
42
+
43
+
(* Dispose a tensor to free GPU/CPU memory *)
44
+
let dispose t = ignore (call t "dispose" [])
45
+
46
+
let pca (mat : Linalg.mat) ~n_components : Linalg.mat =
47
+
let tf = tf () in
48
+
(* 1. Convert input to tensor *)
49
+
let x = mat_to_tensor mat in
50
+
(* 2. Compute mean along axis 0 *)
51
+
let mean = call x "mean" [Js.Unsafe.inject 0] in
52
+
(* 3. Center the data: x_centered = x - mean *)
53
+
let x_centered = call x "sub" [Js.Unsafe.inject mean] in
54
+
(* 4. Covariance matrix: X^T X / (n-1), shape (n_features, n_features).
55
+
For n_samples >> n_features (typical: 40k x 128), this is much faster
56
+
than SVD on the full data matrix. *)
57
+
let n_samples = Float.of_int mat.rows in
58
+
let xt = call x_centered "transpose" [] in
59
+
let xtx = Js.Unsafe.meth_call tf "matMul"
60
+
[| Js.Unsafe.inject xt; Js.Unsafe.inject x_centered |] in
61
+
let scale = Js.Unsafe.meth_call tf "scalar"
62
+
[| Js.Unsafe.inject (n_samples -. 1.0) |] in
63
+
let cov = call xtx "div" [Js.Unsafe.inject scale] in
64
+
(* 5. SVD of symmetric covariance matrix.
65
+
For symmetric PSD matrices, SVD gives eigendecomposition:
66
+
V columns are eigenvectors, S values are eigenvalues. *)
67
+
let svd_result = Js.Unsafe.meth_call (get tf "linalg") "svd"
68
+
[| Js.Unsafe.inject cov; Js.Unsafe.inject Js._true |] in
69
+
let v_full = get svd_result "v" in
70
+
(* 6. Take top n_components columns of V *)
71
+
let v_top = Js.Unsafe.meth_call v_full "slice"
72
+
[| Js.Unsafe.inject (Js.array [| 0; 0 |]);
73
+
Js.Unsafe.inject (Js.array [| mat.cols; n_components |]) |] in
74
+
(* 7. Project: result = x_centered @ v_top, shape (n_samples, n_components) *)
75
+
let projected = Js.Unsafe.meth_call tf "matMul"
76
+
[| Js.Unsafe.inject x_centered; Js.Unsafe.inject v_top |] in
77
+
(* 8. Convert back to Linalg.mat *)
78
+
let result = tensor_to_mat projected in
79
+
(* 9. Clean up all tensors *)
80
+
List.iter dispose [x; mean; x_centered; xt; xtx; scale; cov;
81
+
get svd_result "s"; get svd_result "u"; v_full;
82
+
v_top; projected];
83
+
result
+15
tessera-tfjs/lib/tfjs.mli
+15
tessera-tfjs/lib/tfjs.mli
···
1
+
(** TensorFlow.js PCA via SVD.
2
+
3
+
Requires TensorFlow.js to be loaded in the JavaScript environment
4
+
(e.g., via [importScripts] in a web worker). *)
5
+
6
+
val pca : Linalg.mat -> n_components:int -> Linalg.mat
7
+
(** [pca mat ~n_components] computes PCA on a matrix of shape
8
+
[(n_samples, n_features)] and returns the projected data of shape
9
+
[(n_samples, n_components)].
10
+
11
+
Uses TF.js SVD on the covariance matrix for accurate principal
12
+
component extraction. All intermediate tensors are disposed to
13
+
avoid memory leaks.
14
+
15
+
@raise Failure if TensorFlow.js is not loaded. *)
+28
tessera-tfjs/tessera-tfjs.opam
+28
tessera-tfjs/tessera-tfjs.opam
···
1
+
# This file is generated by dune, edit dune-project instead
2
+
opam-version: "2.0"
3
+
synopsis: "TensorFlow.js PCA bindings for OCaml"
4
+
description:
5
+
"Minimal js_of_ocaml bindings to TensorFlow.js for SVD-based PCA. Operates on tessera-linalg mat types."
6
+
license: "ISC"
7
+
depends: [
8
+
"dune" {>= "3.17"}
9
+
"ocaml" {>= "5.2"}
10
+
"tessera-linalg" {>= "0.1"}
11
+
"js_of_ocaml" {>= "5.0"}
12
+
"js_of_ocaml-ppx" {>= "5.0"}
13
+
"odoc" {with-doc}
14
+
]
15
+
build: [
16
+
["dune" "subst"] {dev}
17
+
[
18
+
"dune"
19
+
"build"
20
+
"-p"
21
+
name
22
+
"-j"
23
+
jobs
24
+
"@install"
25
+
"@runtest" {with-test}
26
+
"@doc" {with-doc}
27
+
]
28
+
]