MediaWiki:Gadget-LabelScan.js: Unterschied zwischen den Versionen
Erscheinungsbild
Admin (Diskussion | Beiträge) Keine Bearbeitungszusammenfassung |
Admin (Diskussion | Beiträge) Keine Bearbeitungszusammenfassung |
||
| Zeile 11: | Zeile 11: | ||
maxSide: 1024, | maxSide: 1024, | ||
debug: true, | debug: true, | ||
// | forceModelWarmup: true // nur für Debug, später gern auf false | ||
}; | }; | ||
// ---------- utils ---------- | |||
const log = (...a)=>{ if(CFG.debug) console.log('[LabelScan]',...a); }; | const log = (...a)=>{ if(CFG.debug) console.log('[LabelScan]',...a); }; | ||
const warn = (...a)=>{ if(CFG.debug) console.warn('[LabelScan]',...a); }; | const warn = (...a)=>{ if(CFG.debug) console.warn('[LabelScan]',...a); }; | ||
const err = (...a)=>{ console.error('[LabelScan]',...a); }; | const err = (...a)=>{ console.error('[LabelScan]',...a); }; | ||
const qs = id => document.getElementById(id); | const qs = id => document.getElementById(id); | ||
const setStatus = t => { const el=qs('ados-scan-status'); if(el) el.textContent=t||''; }; | const setStatus = t => { const el=qs('ados-scan-status'); if(el) el.textContent=t||''; }; | ||
| Zeile 35: | Zeile 34: | ||
} | } | ||
// ---------- Index ---------- | |||
let INDEX=[], INDEX_EMB=[]; | let INDEX=[], INDEX_EMB=[]; | ||
async function loadIndex({ ui=true } = {}){ | async function loadIndex({ ui=true } = {}){ | ||
| Zeile 68: | Zeile 68: | ||
} | } | ||
// ---------- | // ---------- Transformers (Vision-only, remote models, kein Local) ---------- | ||
let _visionLoadPromise=null; | let _visionLoadPromise=null; | ||
async function ensureClipVision(){ | async function ensureClipVision(){ | ||
| Zeile 76: | Zeile 76: | ||
_visionLoadPromise = (async()=>{ | _visionLoadPromise = (async()=>{ | ||
const mod = await import(/* webpackIgnore: true */ CFG.transformersURL); | |||
// *** KRITISCHER BLOCK ***: Nur REMOTE nutzen, LOKAL deaktivieren | |||
// (sonst versucht er /models/... auf deinem Server zu holen -> 404) | |||
mod.env.allowLocalModels = false; // <— zwingt: keine lokalen Modelle prüfen | |||
mod.env.allowRemoteModels = true; | |||
mod.env.useBrowserCache = false; // für Debug: immer frisch laden (Cache-Probleme vermeiden) | |||
// Optional: Remote-Host explizit (Standard ist ohnehin HF Hub) | |||
// mod.env.remoteHost = 'https://huggingface.co'; | |||
// WASM von Transformers-CDN (enthält alle Varianten inkl. ort-wasm-simd.wasm) | |||
mod.env.backends = mod.env.backends || {}; | |||
mod.env.backends.onnx = mod.env.backends.onnx || {}; | |||
mod.env.backends.onnx.wasm = mod.env.backends.onnx.wasm || {}; | |||
mod.env.backends.onnx.wasm.wasmPaths = | |||
'https://cdn.jsdelivr.net/npm/@xenova/transformers@2.15.0/dist/'; | |||
mod.env.backends.onnx.wasm.numThreads = 1; | |||
mod.env.backends.onnx.wasm.simd = true; | |||
// (optional) WebGPU probieren | |||
try { if ('gpu' in navigator) { mod.env.backends.webgpu = { use: true }; log('Backend-Kandidat: WebGPU ist verfügbar.'); } } catch(_) {} | |||
// Vision-Pfad **explizit**: kein pipeline-Automatismus | |||
const processor = await mod.AutoProcessor.from_pretrained(CFG.modelId, { revision: 'main' }); | |||
const model = await mod.CLIPVisionModelWithProjection.from_pretrained( | |||
CFG.modelId, | |||
{ quantized: true, revision: 'main' } | |||
); | |||
// kleines Warmup | |||
try{ | |||
const dummy = new ImageData(32,32); | |||
const inputs = await processor(dummy, { return_tensors: 'pt' }); | |||
await model.forward({ pixel_values: inputs.pixel_values }); | |||
} catch(_) {} | |||
let backend='unknown'; | |||
try { backend = model?.session?.executionProvider || backend; } catch(_){} | |||
log('CLIP ready (vision):', model?.constructor?.name || 'unknown', '| Backend:', backend); | |||
return { mod, processor, model }; | |||
})(); | })(); | ||
| Zeile 162: | Zeile 153: | ||
setStatus('Bild analysieren …'); setProgress(0.38); | setStatus('Bild analysieren …'); setProgress(0.38); | ||
// Processor | // Canvas → Processor → Tensor → Modell | ||
const inputs = await processor(canvas, { return_tensors: 'pt' }); | const inputs = await processor(canvas, { return_tensors: 'pt' }); | ||
const out = await model.forward({ pixel_values: inputs.pixel_values }); | const out = await model.forward({ pixel_values: inputs.pixel_values }); | ||
const vec = out?.image_embeds?.data || out?.image_embeds; | const vec = out?.image_embeds?.data || out?.image_embeds; | ||
if (!(vec instanceof Float32Array)) | if (!(vec instanceof Float32Array)) throw new Error('Embedding-Format unerwartet (kein Float32Array).'); | ||
return normalize(vec); | return normalize(vec); | ||
} | } | ||
function normalize(v){ | // ---------- Ähnlichkeit / Rendering ---------- | ||
function normalize(v){ let n=0; for(let i=0;i<v.length;i++) n+=v[i]*v[i]; n=Math.sqrt(n)||1; const o=new Float32Array(v.length); for(let i=0;i<v.length;i++) o[i]=v[i]/n; return o; } | |||
const cosine=(a,b)=>{ let s=0,L=Math.min(a.length,b.length); for(let i=0;i<L;i++) s+=a[i]*b[i]; return s; }; | const cosine=(a,b)=>{ let s=0,L=Math.min(a.length,b.length); for(let i=0;i<L;i++) s+=a[i]*b[i]; return s; }; | ||
function rankByCosine(q){ | function rankByCosine(q){ | ||
const s=[]; | const s=[]; | ||
| Zeile 193: | Zeile 176: | ||
return s.slice(0,CFG.topK); | return s.slice(0,CFG.topK); | ||
} | } | ||
function renderResults(r){ | function renderResults(r){ | ||
const box=qs('ados-scan-results'); | const box=qs('ados-scan-results'); | ||
| Zeile 211: | Zeile 195: | ||
} | } | ||
// | // ---------- UI ---------- | ||
let BOUND=false; | let BOUND=false; | ||
function bindUI(){ | function bindUI(){ | ||
| Zeile 268: | Zeile 252: | ||
await loadIndex({ ui:true }); | await loadIndex({ ui:true }); | ||
if (CFG.forceModelWarmup) | if (CFG.forceModelWarmup) await ensureClipVision(); | ||
if(!INDEX_EMB.some(v=>v&&v.length)){ | if(!INDEX_EMB.some(v=>v&&v.length)){ | ||
Version vom 8. November 2025, 19:12 Uhr
/* global mw */
(() => {
'use strict';
const CFG = {
indexTitle: (window.LabelScanConfig && window.LabelScanConfig.indexTitle) ||
'MediaWiki:Gadget-LabelScan-index.json',
topK: 8,
transformersURL: 'https://cdn.jsdelivr.net/npm/@xenova/transformers@2.15.0',
modelId: 'Xenova/clip-vit-base-patch32',
maxSide: 1024,
debug: true,
forceModelWarmup: true // nur für Debug, später gern auf false
};
// ---------- utils ----------
const log = (...a)=>{ if(CFG.debug) console.log('[LabelScan]',...a); };
const warn = (...a)=>{ if(CFG.debug) console.warn('[LabelScan]',...a); };
const err = (...a)=>{ console.error('[LabelScan]',...a); };
const qs = id => document.getElementById(id);
const setStatus = t => { const el=qs('ados-scan-status'); if(el) el.textContent=t||''; };
const setProgress = p => {
const bar=qs('ados-scan-progress'); if(!bar) return;
if(p==null){ bar.hidden=true; bar.value=0; }
else{ bar.hidden=false; bar.value=Math.max(0,Math.min(1,p)); }
};
function showPreview(file){
const url=URL.createObjectURL(file);
const prev=qs('ados-scan-preview');
if(prev){
prev.innerHTML='<img alt="Vorschau" style="max-width:100%;height:auto;border-radius:8px;" src="'+url+'">';
prev.setAttribute('aria-hidden','false');
}
}
// ---------- Index ----------
let INDEX=[], INDEX_EMB=[];
async function loadIndex({ ui=true } = {}){
if(INDEX.length) return INDEX;
if(ui){ setStatus('Index laden …'); setProgress(0.03); }
const rawURL = mw.util.getUrl(CFG.indexTitle,{action:'raw',ctype:'application/json'});
const res = await fetch(rawURL,{cache:'reload'});
if(!res.ok) throw new Error('Index nicht ladbar: '+res.status);
const json = await res.json();
if(!Array.isArray(json)) throw new Error('Index ist keine Array-JSON');
INDEX = json;
INDEX_EMB = INDEX.map((it,i)=>{
if(typeof it.embed==='string' && it.embed.length){
try{ return base64ToFloat32(it.embed); }
catch(e){ warn('Embed-Decode',i,it.title,e); return null; }
}
return null;
});
const withEmb = INDEX_EMB.filter(v=>v && v.length).length;
log('Index geladen:', INDEX.length, 'Einträge');
log('Embeddings vorhanden:', withEmb, '/', INDEX.length);
window._LabelScan = window._LabelScan || {};
window._LabelScan.indexInfo = { total: INDEX.length, withEmbeddings: withEmb };
if(ui) setProgress(0.06);
return INDEX;
}
function base64ToFloat32(b64){
const bin=atob(b64), len=bin.length;
const buf=new ArrayBuffer(len);
const view=new Uint8Array(buf);
for(let i=0;i<len;i++) view[i]=bin.charCodeAt(i);
return new Float32Array(buf);
}
// ---------- Transformers (Vision-only, remote models, kein Local) ----------
let _visionLoadPromise=null;
async function ensureClipVision(){
if(_visionLoadPromise) return _visionLoadPromise;
setStatus('Modell laden …'); setProgress(0.08);
_visionLoadPromise = (async()=>{
const mod = await import(/* webpackIgnore: true */ CFG.transformersURL);
// *** KRITISCHER BLOCK ***: Nur REMOTE nutzen, LOKAL deaktivieren
// (sonst versucht er /models/... auf deinem Server zu holen -> 404)
mod.env.allowLocalModels = false; // <— zwingt: keine lokalen Modelle prüfen
mod.env.allowRemoteModels = true;
mod.env.useBrowserCache = false; // für Debug: immer frisch laden (Cache-Probleme vermeiden)
// Optional: Remote-Host explizit (Standard ist ohnehin HF Hub)
// mod.env.remoteHost = 'https://huggingface.co';
// WASM von Transformers-CDN (enthält alle Varianten inkl. ort-wasm-simd.wasm)
mod.env.backends = mod.env.backends || {};
mod.env.backends.onnx = mod.env.backends.onnx || {};
mod.env.backends.onnx.wasm = mod.env.backends.onnx.wasm || {};
mod.env.backends.onnx.wasm.wasmPaths =
'https://cdn.jsdelivr.net/npm/@xenova/transformers@2.15.0/dist/';
mod.env.backends.onnx.wasm.numThreads = 1;
mod.env.backends.onnx.wasm.simd = true;
// (optional) WebGPU probieren
try { if ('gpu' in navigator) { mod.env.backends.webgpu = { use: true }; log('Backend-Kandidat: WebGPU ist verfügbar.'); } } catch(_) {}
// Vision-Pfad **explizit**: kein pipeline-Automatismus
const processor = await mod.AutoProcessor.from_pretrained(CFG.modelId, { revision: 'main' });
const model = await mod.CLIPVisionModelWithProjection.from_pretrained(
CFG.modelId,
{ quantized: true, revision: 'main' }
);
// kleines Warmup
try{
const dummy = new ImageData(32,32);
const inputs = await processor(dummy, { return_tensors: 'pt' });
await model.forward({ pixel_values: inputs.pixel_values });
} catch(_) {}
let backend='unknown';
try { backend = model?.session?.executionProvider || backend; } catch(_){}
log('CLIP ready (vision):', model?.constructor?.name || 'unknown', '| Backend:', backend);
return { mod, processor, model };
})();
return _visionLoadPromise;
}
async function embedFileImage(file){
function loadImage(f){
return new Promise((res,rej)=>{
const url=URL.createObjectURL(f);
const img=new Image();
img.crossOrigin='anonymous';
img.onload=()=>{URL.revokeObjectURL(url);res(img);};
img.onerror=e=>{URL.revokeObjectURL(url);rej(e);};
img.src=url;
});
}
function scale(img,max){
const c=document.createElement('canvas');
let{width:w,height:h}=img;
const s=Math.min(1,max/Math.max(w,h));
w=Math.round(w*s); h=Math.round(h*s);
c.width=w; c.height=h;
const g=c.getContext('2d');
g.imageSmoothingEnabled=true;
g.drawImage(img,0,0,w,h);
return c;
}
const { processor, model } = await ensureClipVision();
setStatus('Bild vorbereiten …'); setProgress(0.20);
const img = await loadImage(file);
const canvas = scale(img, CFG.maxSide);
setStatus('Bild analysieren …'); setProgress(0.38);
// Canvas → Processor → Tensor → Modell
const inputs = await processor(canvas, { return_tensors: 'pt' });
const out = await model.forward({ pixel_values: inputs.pixel_values });
const vec = out?.image_embeds?.data || out?.image_embeds;
if (!(vec instanceof Float32Array)) throw new Error('Embedding-Format unerwartet (kein Float32Array).');
return normalize(vec);
}
// ---------- Ähnlichkeit / Rendering ----------
function normalize(v){ let n=0; for(let i=0;i<v.length;i++) n+=v[i]*v[i]; n=Math.sqrt(n)||1; const o=new Float32Array(v.length); for(let i=0;i<v.length;i++) o[i]=v[i]/n; return o; }
const cosine=(a,b)=>{ let s=0,L=Math.min(a.length,b.length); for(let i=0;i<L;i++) s+=a[i]*b[i]; return s; };
function rankByCosine(q){
const s=[];
for(let i=0;i<INDEX.length;i++){
const v=INDEX_EMB[i];
if(!v) continue;
s.push({i,score:cosine(q,v)});
}
s.sort((a,b)=>b.score-a.score);
return s.slice(0,CFG.topK);
}
function renderResults(r){
const box=qs('ados-scan-results');
if(!box) return;
box.innerHTML='';
if(!r.length){ box.innerHTML='<div class="empty">Keine klaren Treffer.</div>'; return; }
r.forEach(({i,score})=>{
const it=INDEX[i];
const link = mw.util.getUrl((it.title||'').replace(/ /g,'_'));
const thumb=it.thumb||'';
box.innerHTML+=
`<div class="ados-hit" style="display:grid;grid-template-columns:60px 1fr auto;gap:10px;align-items:center;padding:.35rem 0;">
${thumb?`<img src="${thumb}" style="width:60px;border-radius:6px;">`:`<div></div>`}
<div><b><a href="${link}">${mw.html.escape(it.title||'')}</a></b></div>
<div style="color:#666;font-variant-numeric:tabular-nums">${score.toFixed(3)}</div>
</div>`;
});
}
// ---------- UI ----------
let BOUND=false;
function bindUI(){
if(BOUND) return;
const btnRun=qs('ados-scan-run');
const inCam=qs('ados-scan-file-camera');
const inGal=qs('ados-scan-file-gallery');
const btnReset=qs('ados-scan-reset');
const btnCam=qs('ados-scan-btn-camera');
const btnGal=qs('ados-scan-btn-gallery');
const drop=qs('ados-scan-drop');
if(!btnRun||!inCam||!inGal) return;
btnCam && btnCam.addEventListener('click',()=>inCam.click());
btnGal && btnGal.addEventListener('click',()=>inGal.click());
const pick=e=>{ const f=e.target.files?.[0]; if(f) showPreview(f); };
inCam.addEventListener('change',pick);
inGal.addEventListener('change',pick);
if(drop){
drop.addEventListener('dragover',ev=>{ev.preventDefault();drop.classList.add('is-over');});
drop.addEventListener('dragleave',()=>drop.classList.remove('is-over'));
drop.addEventListener('drop',ev=>{
ev.preventDefault();drop.classList.remove('is-over');
const f=ev.dataTransfer?.files?.[0];
if(f){ const dt=new DataTransfer(); dt.items.add(f); inGal.files=dt.files; showPreview(f); }
});
}
btnReset && btnReset.addEventListener('click',()=>{
setStatus('Bereit.'); setProgress(null);
const p=qs('ados-scan-preview'); if(p) p.innerHTML='<div class="note">Noch keine Vorschau.</div>';
const r=qs('ados-scan-results'); if(r) r.innerHTML='<div class="empty">Hier erscheinen Treffer.</div>';
inCam.value=''; inGal.value='';
});
btnRun.addEventListener('click',onRunClick);
BOUND=true;
log('UI gebunden.');
}
async function onRunClick(){
try{
const inCam=qs('ados-scan-file-camera');
const inGal=qs('ados-scan-file-gallery');
const btnRun=qs('ados-scan-run');
const file=inCam.files?.[0]||inGal.files?.[0];
if(!file){ alert('Bitte zuerst ein Foto auswählen.'); return; }
btnRun.disabled=true;
await loadIndex({ ui:true });
if (CFG.forceModelWarmup) await ensureClipVision();
if(!INDEX_EMB.some(v=>v&&v.length)){
setStatus('Index enthält keine Embeddings – bitte Index mit Embeddings neu erzeugen.');
setProgress(null);
return;
}
const q=await embedFileImage(file);
setProgress(0.70);
setStatus('Abgleich …');
renderResults(rankByCosine(q));
setStatus('Fertig.');
setProgress(null);
}catch(e){
err('Fehler',e);
setStatus('Fehler bei Erkennung.');
setProgress(null);
}finally{
const btnRun=qs('ados-scan-run');
if(btnRun) btnRun.disabled=false;
}
}
function init(){
if(document.readyState==='loading'){
document.addEventListener('DOMContentLoaded',bindUI,{once:true});
} else bindUI();
loadIndex({ ui:false }).catch(err).finally(()=>{
setStatus('Bereit.');
setProgress(null);
});
}
log('gadget file loaded');
init();
})();