buscador-visual / source /src /worker.js
Segizu's picture
Buscador Visual: b煤squeda por texto/imagen/regi贸n con SigLIP 2 en el navegador
d18636b verified
Raw History Blame Contribute Delete
3.14 kB
// Web Worker: carga SigLIP 2 (torres de visi贸n y de texto por separado) y calcula embeddings. Nada sale del navegador.
import { AutoTokenizer, AutoProcessor, SiglipTextModel, SiglipVisionModel, RawImage, env } from '@huggingface/transformers';
import { MODEL_ID, MODEL_REVISION, VISION_DTYPE, TEXT_DTYPE, TEXT_MAX_TOKENS } from './config.js';
env.allowLocalModels = false;
env.allowRemoteModels = true;
env.useBrowserCache = true;
let tok, proc, textModel, visionModel, device;
function configureWasm(wasmBase, wantWebGPU) {
const suffix = wantWebGPU ? '.asyncify' : '';
env.backends.onnx.wasm.wasmPaths = { mjs: `${wasmBase}ort-wasm-simd-threaded${suffix}.mjs`, wasm: `${wasmBase}ort-wasm-simd-threaded${suffix}.wasm` };
const cores = self.navigator?.hardwareConcurrency || 2;
env.backends.onnx.wasm.numThreads = self.crossOriginIsolated ? Math.min(4, cores) : 1;
}
async function load({ wantWebGPU, wasmBase }) {
configureWasm(wasmBase, wantWebGPU);
const progress_callback = (p) => self.postMessage({ type: 'progress', data: { status: p.status, file: p.file, loaded: p.loaded, total: p.total } });
const opts = { revision: MODEL_REVISION, progress_callback };
tok = await AutoTokenizer.from_pretrained(MODEL_ID, opts);
proc = await AutoProcessor.from_pretrained(MODEL_ID, opts);
const dev = wantWebGPU ? 'webgpu' : 'wasm';
try {
visionModel = await SiglipVisionModel.from_pretrained(MODEL_ID, { ...opts, dtype: VISION_DTYPE, device: dev });
textModel = await SiglipTextModel.from_pretrained(MODEL_ID, { ...opts, dtype: TEXT_DTYPE, device: dev });
} catch (e) {
if (wantWebGPU) { self.postMessage({ type: 'webgpu-failed', reason: String(e?.message || e).slice(0, 200) }); return; }
throw e;
}
device = dev;
self.postMessage({ type: 'ready', device, threads: env.backends.onnx.wasm.numThreads });
}
function norm(arr) { let s = 0; for (const x of arr) s += x * x; s = Math.sqrt(s) || 1; return Float32Array.from(arr, (x) => x / s); }
async function embedImage({ id, width, height, buffer }) {
const image = new RawImage(new Uint8ClampedArray(buffer), width, height, 4).rgb();
const t0 = performance.now();
const inputs = await proc(image);
const out = await visionModel(inputs);
const vec = norm(out.pooler_output.data);
self.postMessage({ type: 'vector', id, vec, ms: performance.now() - t0, device }, [vec.buffer]);
}
async function embedText({ id, text }) {
const t0 = performance.now();
const inputs = tok([text], { padding: 'max_length', max_length: TEXT_MAX_TOKENS, truncation: true });
const out = await textModel(inputs);
const vec = norm(out.pooler_output.data);
self.postMessage({ type: 'vector', id, vec, ms: performance.now() - t0, device }, [vec.buffer]);
}
self.addEventListener('message', async (ev) => {
const msg = ev.data || {};
try {
if (msg.type === 'load') await load(msg);
else if (msg.type === 'image') await embedImage(msg);
else if (msg.type === 'text') await embedText(msg);
} catch (e) {
self.postMessage({ type: 'error', id: msg.id, stage: msg.type, message: String(e?.message || e).slice(0, 400) });
}
});