scripts/cloud/skypilot_annotate_test.yaml

# Test annotation with 1000 real Orpheus samples
name: ensemble-annotate-test

resources:
  use_spot: true
  accelerators: {A100:1, V100:1, T4:1}
  memory: 32+
  disk_size: 100

setup: |
  echo "=================================================="
  echo "🧪 ENSEMBLE ANNOTATION TEST"
  echo "=================================================="

  pip install -q torch transformers librosa soundfile datasets huggingface_hub tqdm

  if [ ! -d "ensemble-tts-annotation" ]; then
    git clone -q https://huggingface.co/marcosremar2/ensemble-tts-annotation
  else
    cd ensemble-tts-annotation && git pull -q && cd ..
  fi

  echo "✅ Setup complete!"

run: |
  cd ensemble-tts-annotation

  echo "📥 Downloading 1000 Orpheus samples..."
  python3 -c 'from datasets import load_dataset; from pathlib import Path; import soundfile as sf; ds = load_dataset("marcosremar2/orpheus-tts-portuguese-dataset", split="train", streaming=True); out = Path("data/raw/orpheus_test"); out.mkdir(parents=True, exist_ok=True); [(sf.write(out / f"orpheus_{i:05d}.wav", s["audio"]["array"], s["audio"]["sampling_rate"]), print(f"  {i+1}/1000") if (i+1) % 100 == 0 else None) for i, s in enumerate(ds) if i < 1000]; print("✅ Downloaded 1000 samples")'

  echo ""
  echo "🤖 Annotating with ensemble (quick mode)..."
  python3 -c 'import sys; sys.path.insert(0, "."); from ensemble_tts.annotator import EnsembleAnnotator; from pathlib import Path; import json; from tqdm import tqdm; import time; print("Loading..."); ann = EnsembleAnnotator(mode="quick", device="cuda", enable_events=False); print("✅ Loaded\n"); files = sorted(Path("data/raw/orpheus_test").glob("*.wav")); print(f"Annotating {len(files)} files...\n"); start = time.time(); results = [{"file": f.name, "emotion": (r := ann.annotate(str(f))).get("emotion", {}).get("label", "error"), "confidence": r.get("emotion", {}).get("confidence", 0.0)} if not (i % 100) else {"file": f.name, "emotion": (r := ann.annotate(str(f))).get("emotion", {}).get("label", "error"), "confidence": r.get("emotion", {}).get("confidence", 0.0)} for i, f in enumerate(tqdm(files))]; out = Path("data/annotations/orpheus_test_annotations.json"); out.parent.mkdir(parents=True, exist_ok=True); json.dump(results, open(out, "w"), indent=2); elapsed = time.time() - start; ok = sum(1 for r in results if r["emotion"] != "error"); print(f"\n✅ COMPLETE\nTotal: {len(results)}\nSuccess: {ok}\nTime: {elapsed/60:.1f} min\nRate: {len(results)/elapsed:.2f} files/s")'

  echo ""
  echo "📊 Sample results:"
  head -30 data/annotations/orpheus_test_annotations.json