| """Read-only training dashboard and paired native Asterix simulator."""
|
| import json
|
| import os
|
| from pathlib import Path
|
| from functools import lru_cache
|
| from fastapi import FastAPI,HTTPException
|
| from fastapi.responses import HTMLResponse
|
| from huggingface_hub import hf_hub_download
|
| import uvicorn
|
| from crossgame.asterix import Policy,play
|
|
|
| CONFIG=json.loads(Path(__file__).with_name('config.json').read_text())
|
| app=FastAPI()
|
|
|
|
|
| @app.get('/health')
|
| def health():return {'ok':True,'source_commit':os.environ.get('SOURCE_COMMIT'),'game':'asterix'}
|
|
|
|
|
| @app.get('/api/status')
|
| def status():
|
| try:return json.loads(Path(hf_hub_download(CONFIG['dataset'],'latest.json',repo_type='dataset')).read_text())
|
| except Exception:return {'status':'waiting','curve':[],'releases':{}}
|
|
|
|
|
| @lru_cache(maxsize=4)
|
| def model(repo,revision):return Policy.load(hf_hub_download(repo,'policy_weights.npz',revision=revision))
|
|
|
|
|
| @app.get('/api/episode')
|
| def episode(seed:int=5,variant:str='base'):
|
| if not 0<=seed<2**31 or variant not in ('base','improved'):raise HTTPException(400,'Invalid seed or policy')
|
| if variant=='base':
|
| policy=Policy.load(Path(__file__).with_name('assets')/CONFIG.get('initial_checkpoint','asterix-base.npz'));revision=CONFIG.get('continuation',{}).get('revision','upstream-pinned')
|
| else:
|
| release=status().get('releases',{}).get('improved')
|
| if not release:raise HTTPException(409,'No published candidate is available yet')
|
| policy=model(release['repo'],release['revision']);revision=release['revision']
|
| if policy.fingerprint()!=release['policy_hash']:raise HTTPException(503,'Checkpoint identity mismatch')
|
| ep=play(policy,seed,CONFIG['horizon'],trace=True)
|
|
|
| frames=[{k:s[k] for k in ['step','total','dead','player','entities']} for s in ep['worlds']+[ep['final']]]
|
| return dict(seed=seed,total=ep['total'],steps=ep['steps'],reason=ep['reason'],actions=ep['actions'],
|
| frames=frames,policy_hash=policy.fingerprint(),revision=revision)
|
|
|
|
|
| PAGE='''<!doctype html><html><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>Asterix recursive learning</title>
|
| <style>body{font:16px system-ui;background:#111925;color:#e5edf6;max-width:1100px;margin:auto;padding:32px}p{line-height:1.6;color:#abbcd0}a{color:#88d7ba}.grid{display:grid;grid-template-columns:1fr 1fr;gap:24px}section{background:#1b293b;border-radius:16px;padding:20px;margin:20px 0}canvas{width:100%;max-width:400px;background:#0a111b}button,input{padding:12px;border-radius:8px;border:0}button{background:#88d7ba;cursor:pointer}svg{width:100%;height:190px}pre{white-space:pre-wrap}small{word-break:break-all}@media(max-width:650px){.grid{grid-template-columns:1fr}}</style></head><body>
|
| <p>FAILURE IS ALL YOU NEED · CROSS-GAME EXPERIMENT</p><h1>Can the loop learn a different game?</h1>
|
| <p>Asterix: collect gold, avoid enemies. Original MinAtar rules and native sticky actions remain intact. Run 03 improves measured average return over run 02. It was released under a revised average-return rule despite individual regressions; the original strict audit result remains visible below.</p>
|
| <section><h2 id="status">Connecting…</h2><div id="curve"></div><p id="numbers"></p><p id="links"></p><pre id="summary"></pre></section>
|
| <section><h2>Watch the policies</h2><p>Green is the player, gold is treasure, red is danger. Each game uses the same starting seed. Published inference uses the declared average-return criterion; some games may score worse or end earlier.</p>
|
| <input id="seed" value="5" type="number" min="0"><button onclick="playGames()">Run paired games</button><p id="message"></p>
|
| <div class="grid"><div><h3>Starting checkpoint (run 02)</h3><canvas id="base" width="400" height="400"></canvas><p id="baseScore"></p></div><div><h3>Published candidate weights</h3><canvas id="improved" width="400" height="400"></canvas><p id="improvedScore"></p></div></div></section>
|
| <p>Development and selection gate seeds are reused. The final audit is separate and runs after selection ends. This is a 1,000-step greedy-policy pilot, not a reproduction of the upstream 10,000-step sampled-action score. Base weights: Rikiya Takehi and collaborators, CC BY 4.0. Game: Kenny Young and Tian Tian, MinAtar 1.0.15, GPL-3.0.</p>
|
| <script>
|
| const $=x=>document.getElementById(x);let generation=0;
|
| async function update(){const p=await(await fetch('/api/status')).json();$('status').textContent=p.status+(p.feedback_states!==undefined?' · '+p.feedback_states+' regression examples':'');const c=p.curve||[];if(c.length){const vals=c.flatMap(x=>[x.development_mean,x.gate_mean]),lo=Math.min(...vals)-1,hi=Math.max(...vals)+1;let lines='';for(const [key,color]of [['development_mean','#88d7ba'],['gate_mean','#e5bb72']]){const pts=c.map((x,i)=>`${30+430*i/Math.max(1,c.length-1)},${160-135*(x[key]-lo)/(hi-lo)}`).join(' ');lines+=`<polyline points="${pts}" fill="none" stroke="${color}" stroke-width="3"/>`}$('curve').innerHTML=`<svg viewBox="0 0 500 190">${lines}<text x="20" y="185" fill="white">Accepted checkpoint cycle · green: development · gold: gate</text></svg>`;$('numbers').textContent=`Development ${c.at(-1).development_mean.toFixed(3)} · Gate ${c.at(-1).gate_mean.toFixed(3)} mean return`;}$('links').innerHTML=`<a href="https://huggingface.co/datasets/__DATASET__">Trajectories and evidence</a>`+(p.wandb_url?` · <a href="${p.wandb_url}">W&B curves</a>`:'');if(p.summary)$('summary').textContent=JSON.stringify(p.summary,null,2);}
|
| function draw(id,s){const c=$(id).getContext('2d');c.fillStyle='#0a111b';c.fillRect(0,0,400,400);for(const e of s.entities){if(!e)continue;c.fillStyle=e[3]?'#e5bb72':'#e47c7c';c.fillRect(e[0]*40+5,e[1]*40+5,30,30);}c.fillStyle='#88d7ba';c.fillRect(s.player[0]*40+6,s.player[1]*40+6,28,28);$(id+'Score').textContent=`Step ${s.step} · Score ${s.total}${s.dead?' · Collision':''}`;}
|
| async function playGames(){const n=++generation;$('message').textContent='Running simulator…';const seed=Number($('seed').value);const rs=await Promise.all(['base','improved'].map(v=>fetch(`/api/episode?seed=${seed}&variant=${v}`)));const es=await Promise.all(rs.map(r=>r.json()));$('message').textContent=rs[1].ok?'Paired live inference':es[1].detail;const valid=es.filter((e,i)=>rs[i].ok);for(let t=0;t<Math.max(...valid.map(e=>e.frames.length));t++){if(n!==generation)return;for(let i=0;i<valid.length;i++)draw(i?'improved':'base',valid[i].frames[Math.min(t,valid[i].frames.length-1)]);await new Promise(r=>setTimeout(r,35));}}
|
| update();setInterval(update,20000);
|
| </script></body></html>'''.replace('__DATASET__',CONFIG['dataset'])
|
|
|
|
|
| @app.get('/',response_class=HTMLResponse)
|
| def home():return PAGE
|
|
|
|
|
| if __name__=='__main__':uvicorn.run(app,host='0.0.0.0',port=7860)
|
|
|
