File size: 8,374 Bytes
4cefcfa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
/* nlrt.js: browser and Node twin of nlrt.py. Batch size 1, float64 math. */
(function (root, factory) {
  if (typeof module === "object" && module.exports) module.exports = factory();
  else root.NLRT = factory();
})(typeof self !== "undefined" ? self : this, function () {
  "use strict";

  function b64d(s) {
    let bin;
    if (typeof atob === "function") bin = atob(s);
    else bin = Buffer.from(s, "base64").toString("binary");
    const bytes = new Uint8Array(bin.length);
    for (let i = 0; i < bin.length; i++) bytes[i] = bin.charCodeAt(i);
    return new Float32Array(bytes.buffer);
  }

  function decode(spec) {
    for (const name in spec.nets || {}) {
      for (const L of spec.nets[name]) {
        if (typeof L.w === "string") L.w = b64d(L.w);
        if (typeof L.b === "string") L.b = b64d(L.b);
      }
    }
    for (const name in spec.tensors || {}) {
      const t = spec.tensors[name];
      if (typeof t.data === "string") t.data = b64d(t.data);
    }
    return spec;
  }

  async function loadModel(url) {
    const r = await fetch(url);
    if (!r.ok) throw new Error("cannot load " + url);
    return decode(await r.json());
  }

  function pair(v) { return Array.isArray(v) ? v : [v, v]; }

  // Tensor: {shape: [D]} or {shape: [C, H, W]} or {shape: [C, L]}, data: Float64Array
  function T(shape, data) {
    const n = shape.reduce((a, b) => a * b, 1);
    return { shape: shape.slice(), data: data || new Float64Array(n) };
  }

  function as3(x) {
    if (x.shape.length === 2) return { t: T([x.shape[0], 1, x.shape[1]], x.data), sq: true };
    return { t: x, sq: false };
  }

  function conv2d(x0, L) {
    const { t: x, sq } = as3(x0);
    const [C, H, W] = x.shape;
    const [O, I, KH, KW] = L.w_shape;
    const [sh, sw] = pair(L.stride), [ph, pw] = pair(L.pad);
    const OH = Math.floor((H + 2 * ph - KH) / sh) + 1;
    const OW = Math.floor((W + 2 * pw - KW) / sw) + 1;
    const out = new Float64Array(O * OH * OW);
    const w = L.w, b = L.b, xd = x.data;
    for (let o = 0; o < O; o++) {
      for (let oy = 0; oy < OH; oy++) {
        for (let ox = 0; ox < OW; ox++) {
          let s = b[o];
          for (let c = 0; c < I; c++) {
            for (let ky = 0; ky < KH; ky++) {
              const iy = oy * sh + ky - ph;
              if (iy < 0 || iy >= H) continue;
              const wb = ((o * I + c) * KH + ky) * KW, xb = (c * H + iy) * W;
              for (let kx = 0; kx < KW; kx++) {
                const ix = ox * sw + kx - pw;
                if (ix < 0 || ix >= W) continue;
                s += w[wb + kx] * xd[xb + ix];
              }
            }
          }
          out[(o * OH + oy) * OW + ox] = s;
        }
      }
    }
    return sq ? T([O, OW], out) : T([O, OH, OW], out);
  }

  function maxpool(x0, L) {
    const { t: x, sq } = as3(x0);
    const [C, H, W] = x.shape, [kh, kw] = pair(L.k);
    const OH = Math.floor(H / kh), OW = Math.floor(W / kw);
    const out = new Float64Array(C * OH * OW);
    for (let c = 0; c < C; c++)
      for (let oy = 0; oy < OH; oy++)
        for (let ox = 0; ox < OW; ox++) {
          let m = -Infinity;
          for (let i = 0; i < kh; i++)
            for (let j = 0; j < kw; j++) {
              const v = x.data[(c * H + oy * kh + i) * W + ox * kw + j];
              if (v > m) m = v;
            }
          out[(c * OH + oy) * OW + ox] = m;
        }
    return sq ? T([C, OW], out) : T([C, OH, OW], out);
  }

  function linear(x, L) {
    const [O, I] = L.w_shape;
    const out = new Float64Array(O);
    for (let o = 0; o < O; o++) {
      let s = L.b[o];
      const base = o * I;
      for (let i = 0; i < I; i++) s += L.w[base + i] * x.data[i];
      out[o] = s;
    }
    return T([O], out);
  }

  function map(x, f) {
    const d = new Float64Array(x.data.length);
    for (let i = 0; i < d.length; i++) d[i] = f(x.data[i]);
    return T(x.shape, d);
  }

  function forward(layers, x) {
    if (!(x.data instanceof Float64Array)) x = T(x.shape, Float64Array.from(x.data));
    for (const L of layers) {
      switch (L.type) {
        case "linear": x = linear(x, L); break;
        case "conv2d": x = conv2d(x, L); break;
        case "relu": x = map(x, v => (v > 0 ? v : 0)); break;
        case "tanh": x = map(x, Math.tanh); break;
        case "sigmoid": x = map(x, v => 1 / (1 + Math.exp(-v))); break;
        case "maxpool": x = maxpool(x, L); break;
        case "gap": {
          const C = x.shape[0], n = x.data.length / C, d = new Float64Array(C);
          for (let c = 0; c < C; c++) { let s = 0; for (let i = 0; i < n; i++) s += x.data[c * n + i]; d[c] = s / n; }
          x = T([C], d); break;
        }
        case "flatten": x = T([x.data.length], x.data); break;
        case "reshape": x = T(L.shape, x.data); break;
        case "upsample": {
          const s = L.s;
          if (x.shape.length === 3) {
            const [C, H, W] = x.shape, d = new Float64Array(C * H * s * W * s);
            for (let c = 0; c < C; c++)
              for (let y = 0; y < H * s; y++)
                for (let xx = 0; xx < W * s; xx++)
                  d[(c * H * s + y) * W * s + xx] = x.data[(c * H + Math.floor(y / s)) * W + Math.floor(xx / s)];
            x = T([C, H * s, W * s], d);
          } else {
            const [C, Ln] = x.shape, d = new Float64Array(C * Ln * s);
            for (let c = 0; c < C; c++) for (let i = 0; i < Ln * s; i++) d[c * Ln * s + i] = x.data[c * Ln + Math.floor(i / s)];
            x = T([C, Ln * s], d);
          }
          break;
        }
        case "l2norm": {
          let s = 0; for (const v of x.data) s += v * v;
          const n = Math.max(Math.sqrt(s), 1e-12);
          x = map(x, v => v / n); break;
        }
        default: throw new Error("unknown layer " + L.type);
      }
    }
    return x;
  }

  function mulberry32(seed) {
    let a = seed >>> 0;
    const random = function () {
      a = (a + 0x6D2B79F5) >>> 0;
      let t = a;
      t = Math.imul(t ^ (t >>> 15), t | 1) >>> 0;
      t = (t ^ ((t + (Math.imul(t ^ (t >>> 7), t | 61) >>> 0)) >>> 0)) >>> 0;
      return ((t ^ (t >>> 14)) >>> 0) / 4294967296;
    };
    return {
      random,
      uniform: (lo, hi) => lo + (hi - lo) * random(),
      randint: n => Math.floor(random() * n),
      normal: () => { const u = Math.max(random(), 1e-12), v = random(); return Math.sqrt(-2 * Math.log(u)) * Math.cos(2 * Math.PI * v); },
    };
  }

  function softmax(arr) {
    let m = -Infinity; for (const v of arr) if (v > m) m = v;
    const e = Array.from(arr, v => Math.exp(v - m)); const s = e.reduce((a, b) => a + b, 0);
    return e.map(v => v / s);
  }

  function argmax(arr) { let k = 0; for (let i = 1; i < arr.length; i++) if (arr[i] > arr[k]) k = i; return k; }

  // Deep comparison used by parity tests: numbers within tol, everything else exact.
  function compare(got, want, tol, path, acc) {
    acc = acc || { maxDiff: 0, mismatches: [] };
    path = path || "$";
    if (typeof want === "number") {
      const d = Math.abs(Number(got) - want);
      if (!(d <= tol * Math.max(1, Math.abs(want)))) acc.mismatches.push(path + ": got " + got + " want " + want);
      if (isFinite(d)) acc.maxDiff = Math.max(acc.maxDiff, d);
    } else if (Array.isArray(want)) {
      if (!got || got.length !== want.length) acc.mismatches.push(path + ": length " + (got && got.length) + " vs " + want.length);
      else for (let i = 0; i < want.length; i++) compare(got[i], want[i], tol, path + "[" + i + "]", acc);
    } else if (want && typeof want === "object") {
      for (const k in want) compare(got ? got[k] : undefined, want[k], tol, path + "." + k, acc);
    } else if (got !== want) acc.mismatches.push(path + ": got " + got + " want " + want);
    return acc;
  }

  function runParity(core, ctx, parity) {
    const results = [];
    let maxDiff = 0, failed = 0;
    for (const c of parity.cases) {
      const out = core.runCase(ctx, c.input);
      const r = compare(out, c.expected, c.tol != null ? c.tol : parity.tol);
      maxDiff = Math.max(maxDiff, r.maxDiff);
      if (r.mismatches.length) failed++;
      results.push({ name: c.name, maxDiff: r.maxDiff, mismatches: r.mismatches.slice(0, 5) });
    }
    return { cases: parity.cases.length, failed, maxDiff, results };
  }

  return { b64d, decode, loadModel, T, forward, mulberry32, softmax, argmax, compare, runParity };
});