@thi.ng/tsne 0.1.14 → 0.1.16

package/CHANGELOG.md

@@ -1,6 +1,6 @@
 # Change Log
 
-- **Last updated**: 2025-05-28T12:02:40Z
+- **Last updated**: 2025-06-05T12:59:44Z
 - **Generator**: [thi.ng/monopub](https://thi.ng/monopub)
 
 All notable changes to this project will be documented in this file.
@@ -11,6 +11,15 @@ See [Conventional Commits](https://conventionalcommits.org/) for commit guidelin
 **Note:** Unlisted _patch_ versions only involve non-code or otherwise excluded changes
 and/or version bumps of transitive dependencies.
 
+### [0.1.15](https://github.com/thi-ng/umbrella/tree/@thi.ng/tsne@0.1.15) (2025-06-03)
+
+#### ⏱ Performance improvements
+
+- avoid allocations, optimize computeGradient() ([8888892](https://github.com/thi-ng/umbrella/commit/8888892))
+  - pre-allocate arrays for gradient & yMean
+  - skip gradient update if same rows
+  - use destructured Math fns
+
 ### [0.1.9](https://github.com/thi-ng/umbrella/tree/@thi.ng/tsne@0.1.9) (2025-04-16)
 
 #### ♻️ Refactoring

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@thi.ng/tsne",
-  "version": "0.1.14",
+  "version": "0.1.16",
   "description": "Highly configurable t-SNE implementation for arbitrary dimensions",
   "type": "module",
   "module": "./index.js",
@@ -51,7 +51,16 @@
     "typescript": "^5.8.3"
   },
   "keywords": [
-    "typescript"
+    "2d",
+    "3d",
+    "cluster",
+    "compression",
+    "dataviz",
+    "nd",
+    "tsne",
+    "typescript",
+    "vector",
+    "visualization"
   ],
   "publishConfig": {
     "access": "public"
@@ -82,5 +91,5 @@
     "status": "alpha",
     "year": 2021
   },
-  "gitHead": "61c3833b7ef7d044621454b5ea4af885d39f065e\n"
+  "gitHead": "5a1224bbd501c0dad203fa231749f21740350db7\n"
 }

package/tsne.d.ts

@@ -13,6 +13,8 @@ export declare class TSNE {
     points: number[][];
     steps: number[][];
     gains: number[][];
+    gradient: FloatArray[];
+    ymean: FloatArray;
     opDist: DistanceFn;
     opDivN: VecOpVN;
     opSub: VecOpVV;
@@ -21,7 +23,7 @@ export declare class TSNE {
     update(): number;
     computeGradient(): {
         cost: number;
-        gradient: ReadonlyVec[];
+        gradient: FloatArray[];
     };
 }
 //# sourceMappingURL=tsne.d.ts.map

package/tsne.js

@@ -7,10 +7,10 @@ import { repeatedly } from "@thi.ng/transducers";
 import {
   distSq,
   divN,
-  sub,
-  zeroes
+  sub
 } from "@thi.ng/vectors";
 const EPS = Number.EPSILON;
+const { exp, log, max, sign } = Math;
 const DEFAULT_OPTS = {
   rnd: SYSTEM,
   dist: distSq,
@@ -44,6 +44,8 @@ class TSNE {
   points;
   steps;
   gains;
+  gradient;
+  ymean;
   opDist;
   opDivN;
   opSub;
@@ -70,11 +72,13 @@ class TSNE {
     this.gains = initMatrix(n, dim, () => 1);
     this.q = new Float64Array(n * n);
     this.qu = new Float64Array(n * n);
+    this.ymean = new Float64Array(dim);
+    this.gradient = [...repeatedly(() => new Float64Array(dim), n)];
     this.iter = 0;
   }
   update() {
     if (++this.iter >= this.opts.maxIter) return 0;
-    const { n, dim, points, steps, gains, opDivN, opSub } = this;
+    const { n, dim, points, steps, gains, ymean, opDivN, opSub } = this;
     const {
       rate,
       minGain,
@@ -84,7 +88,7 @@ class TSNE {
     } = this.opts;
     const { cost, gradient } = this.computeGradient();
     const momentum = tweenParam($momentum, this.iter);
-    const ymean = zeroes(dim);
+    ymean.fill(0);
     let i, d;
     for (i = 0; i < n; i++) {
       const row = points[i];
@@ -93,9 +97,9 @@ class TSNE {
       const rowGains = gains[i];
       for (d = 0; d < dim; d++) {
         let step = rowStep[d];
-        const newGain = Math.max(
+        const newGain = max(
           minGain,
-          Math.sign(rowGrad[d]) === Math.sign(step) ? (
+          sign(rowGrad[d]) === sign(step) ? (
             // rowGrad[d] * step < 0
             rowGains[d] * gainDecay
           ) : rowGains[d] + gainBias
@@ -112,7 +116,7 @@ class TSNE {
     return cost;
   }
   computeGradient() {
-    const { n, dim, points: y, p, q, qu, opDist } = this;
+    const { n, dim, gradient, points: y, p, q, qu, opDist } = this;
     let i, j, rowIdx, d;
     let rowI, rowJ;
     let qsum = 0;
@@ -128,22 +132,25 @@ class TSNE {
     }
     qsum = 1 / qsum;
     for (i = n * n; i-- > 0; ) {
-      q[i] = Math.max(qu[i] * qsum, EPS);
+      q[i] = max(qu[i] * qsum, EPS);
     }
     let cost = 0;
-    const gradient = new Array(n);
     const gscale = tweenParam(this.opts.gradientScale, this.iter);
     for (i = 0; i < n; i++) {
-      rowIdx = i * n;
+      const g = gradient[i].fill(0);
       rowI = y[i];
-      const g = gradient[i] = zeroes(dim);
+      rowIdx = i * n;
       for (j = 0; j < n; j++) {
-        rowJ = y[j];
         const ij = rowIdx + j;
-        cost += -p[ij] * Math.log(q[ij]);
-        const s = 4 * (gscale * p[ij] - q[ij]) * qu[ij];
-        for (d = 0; d < dim; d++) {
-          g[d] += s * (rowI[d] - rowJ[d]);
+        const pij = p[ij];
+        const qij = q[ij];
+        cost += -pij * log(qij);
+        const s = 4 * (gscale * pij - qij) * qu[ij];
+        if (i !== j) {
+          rowJ = y[j];
+          for (d = 0; d < dim; d++) {
+            g[d] += s * (rowI[d] - rowJ[d]);
+          }
         }
       }
     }
@@ -166,7 +173,7 @@ const pairwiseDistances = (points, distFn) => {
   return dist;
 };
 const initProbabilities = (distances, n, perplexity, eps, iter) => {
-  const htarget = Math.log(perplexity);
+  const htarget = log(perplexity);
   const p = new Float64Array(n * n);
   for (let i = 0; i < n; i++) {
     distProbRow(
@@ -184,7 +191,7 @@ const initProbabilities = (distances, n, perplexity, eps, iter) => {
   for (let i = 0; i < n; i++) {
     const ii = i * n;
     for (let j = 0; j < n; j++) {
-      res[ii + j] = Math.max((p[ii + j] + p[j * n + i]) * invN2, EPS);
+      res[ii + j] = max((p[ii + j] + p[j * n + i]) * invN2, EPS);
     }
   }
   return res;
@@ -210,7 +217,7 @@ const rowEntropy = (distances, row, n, i, beta) => {
   let psum = 0;
   for (let j = 0; j < n; j++) {
     if (i !== j) {
-      psum += row[j] = Math.exp(-distances[ii + j] * beta);
+      psum += row[j] = exp(-distances[ii + j] * beta);
     } else {
       row[j] = 0;
     }
@@ -223,7 +230,7 @@ const normalizeRow = (row, n, psum) => {
     psum = 1 / psum;
     for (let i = 0; i < n; i++) {
       const p = row[i] *= psum;
-      if (p > 1e-7) h -= p * Math.log(p);
+      if (p > 1e-7) h -= p * log(p);
     }
   } else {
     row.fill(0);