Spaces:

usmar
/

tinycast

Running

Upload folder using huggingface_hub

c1295a6 verified 3 months ago

6.51 kB

/* tinycast static-Space engine.
 *
 * In the browser it overrides window.fetch for the app's own /api/* routes and
 * answers them client-side: weights come from weights.bin/meta.json in the
 * Space, the raw ETTh1 CSV comes straight from GitHub (CORS: *), and the
 * linear models run as plain matrix-vector products below. Under Node it just
 * exports the math so the Python test suite can check JS/PyTorch parity.
 */
(function (root) {
  "use strict";

  // ---------- pure forecasting math (mirrors tinycast/models.py) ----------

  function movingAverage(col, kernel) {
    const L = col.length;
    const left = Math.floor(kernel / 2);
    const right = kernel - 1 - left;
    const padded = new Float64Array(L + kernel - 1);
    for (let i = 0; i < left; i++) padded[i] = col[0];
    for (let i = 0; i < L; i++) padded[left + i] = col[i];
    for (let i = 0; i < right; i++) padded[left + L + i] = col[L - 1];
    const out = new Float64Array(L);
    let sum = 0;
    for (let i = 0; i < kernel; i++) sum += padded[i];
    out[0] = sum / kernel;
    for (let i = 1; i < L; i++) {
      sum += padded[i + kernel - 1] - padded[i - 1];
      out[i] = sum / kernel;
    }
    return out;
  }

  function matVec(w, rows, cols, v, b) {
    const out = new Float64Array(rows);
    for (let r = 0; r < rows; r++) {
      let s = b ? b[r] : 0;
      const off = r * cols;
      for (let c = 0; c < cols; c++) s += w[off + c] * v[c];
      out[r] = s;
    }
    return out;
  }

  function nlinearForward(t, col, H) {
    const L = col.length;
    const last = col[L - 1];
    const centered = new Float64Array(L);
    for (let i = 0; i < L; i++) centered[i] = col[i] - last;
    const out = matVec(t.w.data, H, L, centered, t.b.data);
    for (let h = 0; h < H; h++) out[h] += last;
    return out;
  }

  function dlinearForward(t, col, H, kernel) {
    const L = col.length;
    const trend = movingAverage(col, kernel);
    const seasonal = new Float64Array(L);
    for (let i = 0; i < L; i++) seasonal[i] = col[i] - trend[i];
    const out = matVec(t.wt.data, H, L, trend, t.bt.data);
    const s = matVec(t.ws.data, H, L, seasonal, t.bs.data);
    for (let h = 0; h < H; h++) out[h] += s[h];
    return out;
  }

  function persistence(col, H) {
    return new Float64Array(H).fill(col[col.length - 1]);
  }

  function seasonalNaive(col, H, period) {
    period = period || 24;
    const out = new Float64Array(H);
    const start = col.length - period;
    for (let h = 0; h < H; h++) out[h] = col[start + (h % period)];
    return out;
  }

  const math = { movingAverage, matVec, nlinearForward, dlinearForward, persistence, seasonalNaive };

  if (typeof module !== "undefined" && module.exports) {
    module.exports = math; // Node: parity tests only
    return;
  }

  // ---------- browser: data loading + fetch override ----------

  const realFetch = root.fetch.bind(root);
  let enginePromise = null;

  async function loadEngine() {
    const meta = await (await realFetch("meta.json")).json();
    const buf = await (await realFetch("weights.bin")).arrayBuffer();
    const tensors = {};
    for (const [model, parts] of Object.entries(meta.tensors)) {
      tensors[model] = {};
      for (const [name, t] of Object.entries(parts)) {
        const size = t.shape.reduce((a, b) => a * b, 1);
        tensors[model][name] = { data: new Float32Array(buf, t.offset * 4, size), shape: t.shape };
      }
    }
    const csv = await (await realFetch(meta.data_url)).text();
    const lines = csv.trim().split("\n");
    const header = lines[0].split(",");
    const otCol = header.indexOf("OT");
    const mean = meta.scaler_mean[meta.target_index];
    const std = meta.scaler_std[meta.target_index];
    const dates = [];
    const ot = new Float64Array(meta.test_end_row - meta.test_start_row);
    for (let r = meta.test_start_row; r < meta.test_end_row; r++) {
      const cells = lines[r + 1].split(",");
      dates.push(cells[0]);
      ot[r - meta.test_start_row] = (parseFloat(cells[otCol]) - mean) / std;
    }
    return { meta, tensors, ot, dates, mean, std };
  }

  function engine() {
    if (!enginePromise) enginePromise = loadEngine();
    return enginePromise;
  }

  async function handle(url) {
    const e = await engine();
    const { meta } = e;
    const L = meta.seq_len;
    const nWindows = (h) => e.ot.length - L - h + 1;

    const u = new URL(url, location.href);
    if (u.pathname.endsWith("/api/meta") || u.pathname.endsWith("api/meta")) {
      const n = {};
      for (const h of meta.horizons) n[String(h)] = nWindows(h);
      return {
        horizons: meta.horizons,
        n_windows: n,
        models: ["persistence", "seasonal_naive", "nlinear", "dlinear"],
        results: meta.results,
      };
    }

    const index = parseInt(u.searchParams.get("index") || "0", 10);
    const H = parseInt(u.searchParams.get("horizon") || "96", 10);
    if (!meta.horizons.includes(H)) throw new Error("horizon not exported: " + H);
    if (index < 0 || index >= nWindows(H)) throw new Error("index out of range");

    const col = e.ot.subarray(index, index + L);
    const actual = e.ot.subarray(index + L, index + L + H);
    const toC = (a) => Array.from(a, (z) => Math.round((z * e.std + e.mean) * 1000) / 1000);
    const forecasts = {
      persistence: persistence(col, H),
      seasonal_naive: seasonalNaive(col, H),
      nlinear: nlinearForward(e.tensors["nlinear_" + H], col, H),
      dlinear: dlinearForward(e.tensors["dlinear_" + H], col, H, meta.moving_avg),
    };
    const mae = {};
    for (const [k, f] of Object.entries(forecasts)) {
      let s = 0;
      for (let h = 0; h < H; h++) s += Math.abs(f[h] - actual[h]);
      mae[k] = Math.round((s / H) * e.std * 1000) / 1000;
    }
    const out = {};
    for (const [k, f] of Object.entries(forecasts)) out[k] = toC(f);
    return {
      index,
      n_windows: nWindows(H),
      horizon: H,
      t0: e.dates[index + L],
      history_ot: toC(col),
      actual_ot: toC(actual),
      forecasts: out,
      window_mae_c: mae,
    };
  }

  root.fetch = function (url, opts) {
    const u = String(url);
    if (u.startsWith("api/") || u.startsWith("/api/")) {
      return handle(u).then(
        (data) => new Response(JSON.stringify(data), { headers: { "Content-Type": "application/json" } }),
        (err) => new Response(String(err && err.message), { status: 500 })
      );
    }
    return realFetch(url, opts);
  };
})(typeof window !== "undefined" ? window : globalThis);

	/* tinycast static-Space engine.
	*
	* In the browser it overrides window.fetch for the app's own /api/* routes and
	* answers them client-side: weights come from weights.bin/meta.json in the
	* Space, the raw ETTh1 CSV comes straight from GitHub (CORS: *), and the
	* linear models run as plain matrix-vector products below. Under Node it just
	* exports the math so the Python test suite can check JS/PyTorch parity.
	*/
	(function (root) {
	"use strict";

	// ---------- pure forecasting math (mirrors tinycast/models.py) ----------

	function movingAverage(col, kernel) {
	const L = col.length;
	const left = Math.floor(kernel / 2);
	const right = kernel - 1 - left;
	const padded = new Float64Array(L + kernel - 1);
	for (let i = 0; i < left; i++) padded[i] = col[0];
	for (let i = 0; i < L; i++) padded[left + i] = col[i];
	for (let i = 0; i < right; i++) padded[left + L + i] = col[L - 1];
	const out = new Float64Array(L);
	let sum = 0;
	for (let i = 0; i < kernel; i++) sum += padded[i];
	out[0] = sum / kernel;
	for (let i = 1; i < L; i++) {
	sum += padded[i + kernel - 1] - padded[i - 1];
	out[i] = sum / kernel;
	}
	return out;
	}

	function matVec(w, rows, cols, v, b) {
	const out = new Float64Array(rows);
	for (let r = 0; r < rows; r++) {
	let s = b ? b[r] : 0;
	const off = r * cols;
	for (let c = 0; c < cols; c++) s += w[off + c] * v[c];
	out[r] = s;
	}
	return out;
	}

	function nlinearForward(t, col, H) {
	const L = col.length;
	const last = col[L - 1];
	const centered = new Float64Array(L);
	for (let i = 0; i < L; i++) centered[i] = col[i] - last;
	const out = matVec(t.w.data, H, L, centered, t.b.data);
	for (let h = 0; h < H; h++) out[h] += last;
	return out;
	}

	function dlinearForward(t, col, H, kernel) {
	const L = col.length;
	const trend = movingAverage(col, kernel);
	const seasonal = new Float64Array(L);
	for (let i = 0; i < L; i++) seasonal[i] = col[i] - trend[i];
	const out = matVec(t.wt.data, H, L, trend, t.bt.data);
	const s = matVec(t.ws.data, H, L, seasonal, t.bs.data);
	for (let h = 0; h < H; h++) out[h] += s[h];
	return out;
	}

	function persistence(col, H) {
	return new Float64Array(H).fill(col[col.length - 1]);
	}

	function seasonalNaive(col, H, period) {
	period = period \|\| 24;
	const out = new Float64Array(H);
	const start = col.length - period;
	for (let h = 0; h < H; h++) out[h] = col[start + (h % period)];
	return out;
	}

	const math = { movingAverage, matVec, nlinearForward, dlinearForward, persistence, seasonalNaive };

	if (typeof module !== "undefined" && module.exports) {
	module.exports = math; // Node: parity tests only
	return;
	}

	// ---------- browser: data loading + fetch override ----------

	const realFetch = root.fetch.bind(root);
	let enginePromise = null;

	async function loadEngine() {
	const meta = await (await realFetch("meta.json")).json();
	const buf = await (await realFetch("weights.bin")).arrayBuffer();
	const tensors = {};
	for (const [model, parts] of Object.entries(meta.tensors)) {
	tensors[model] = {};
	for (const [name, t] of Object.entries(parts)) {
	const size = t.shape.reduce((a, b) => a * b, 1);
	tensors[model][name] = { data: new Float32Array(buf, t.offset * 4, size), shape: t.shape };
	}
	}
	const csv = await (await realFetch(meta.data_url)).text();
	const lines = csv.trim().split("\n");
	const header = lines[0].split(",");
	const otCol = header.indexOf("OT");
	const mean = meta.scaler_mean[meta.target_index];
	const std = meta.scaler_std[meta.target_index];
	const dates = [];
	const ot = new Float64Array(meta.test_end_row - meta.test_start_row);
	for (let r = meta.test_start_row; r < meta.test_end_row; r++) {
	const cells = lines[r + 1].split(",");
	dates.push(cells[0]);
	ot[r - meta.test_start_row] = (parseFloat(cells[otCol]) - mean) / std;
	}
	return { meta, tensors, ot, dates, mean, std };
	}

	function engine() {
	if (!enginePromise) enginePromise = loadEngine();
	return enginePromise;
	}

	async function handle(url) {
	const e = await engine();
	const { meta } = e;
	const L = meta.seq_len;
	const nWindows = (h) => e.ot.length - L - h + 1;

	const u = new URL(url, location.href);
	if (u.pathname.endsWith("/api/meta") \|\| u.pathname.endsWith("api/meta")) {
	const n = {};
	for (const h of meta.horizons) n[String(h)] = nWindows(h);
	return {
	horizons: meta.horizons,
	n_windows: n,
	models: ["persistence", "seasonal_naive", "nlinear", "dlinear"],
	results: meta.results,
	};
	}

	const index = parseInt(u.searchParams.get("index") \|\| "0", 10);
	const H = parseInt(u.searchParams.get("horizon") \|\| "96", 10);
	if (!meta.horizons.includes(H)) throw new Error("horizon not exported: " + H);
	if (index < 0 \|\| index >= nWindows(H)) throw new Error("index out of range");

	const col = e.ot.subarray(index, index + L);
	const actual = e.ot.subarray(index + L, index + L + H);
	const toC = (a) => Array.from(a, (z) => Math.round((z * e.std + e.mean) * 1000) / 1000);
	const forecasts = {
	persistence: persistence(col, H),
	seasonal_naive: seasonalNaive(col, H),
	nlinear: nlinearForward(e.tensors["nlinear_" + H], col, H),
	dlinear: dlinearForward(e.tensors["dlinear_" + H], col, H, meta.moving_avg),
	};
	const mae = {};
	for (const [k, f] of Object.entries(forecasts)) {
	let s = 0;
	for (let h = 0; h < H; h++) s += Math.abs(f[h] - actual[h]);
	mae[k] = Math.round((s / H) * e.std * 1000) / 1000;
	}
	const out = {};
	for (const [k, f] of Object.entries(forecasts)) out[k] = toC(f);
	return {
	index,
	n_windows: nWindows(H),
	horizon: H,
	t0: e.dates[index + L],
	history_ot: toC(col),
	actual_ot: toC(actual),
	forecasts: out,
	window_mae_c: mae,
	};
	}

	root.fetch = function (url, opts) {
	const u = String(url);
	if (u.startsWith("api/") \|\| u.startsWith("/api/")) {
	return handle(u).then(
	(data) => new Response(JSON.stringify(data), { headers: { "Content-Type": "application/json" } }),
	(err) => new Response(String(err && err.message), { status: 500 })
	);
	}
	return realFetch(url, opts);
	};
	})(typeof window !== "undefined" ? window : globalThis);