Skip to content

v2.0.0 — Hardware-aware Latency NAS, N:M Sparsity, INT8

Choose a tag to compare

View all tags
@pedrovergueiro pedrovergueiro released this 01 Jul 02:14
v2.0.0
53d87f8

What's new in v2.0.0

Hardware-aware Latency NAS

  • LatencyEvolver — bi-objective NAS minimising (1−accuracy, latency_ms) instead of FLOPs. Uses real ONNX Runtime measurement; optional GBM surrogate for 80% of evaluations.
  • compress(model, data, target="latency", hw_target="rpi4") — new target parameter. Fully backwards-compatible.
  • detect_hw() / latency_scale(device) / estimate_latency(ms, from, to) — cross-device estimation without hardware. Calibrated: RPi4 (9×), RPi5 (4.5×), Jetson Nano (5×), Apple M1 (0.6×).
  • LatencyPredictor — GBM (LightGBM) surrogate. Degrades gracefully to analytical fallback.

N:M Structured Sparsity

  • apply_nm_sparsity(model, n=2, m=4) — 2:4 structured sparsity baked into weights, ONNX-ready.
  • sparsity_stats(model) — global and per-layer sparsity report.
  • compress(model, data, sparsity="2:4") — apply automatically after NAS.

Quantization-aware Search

  • QuantAwareEvolver — fitness evaluated under INT8. Filters architectures that degrade under quantization during search. ~20% overhead vs standard.
  • compress(model, data, quant_aware=True) — enable via parameter.
  • result.quantize() — apply torch.quantization.quantize_dynamic in one call.

Pandas / NumPy Auto-input

  • compress() now accepts (X, y) tuple, pandas.DataFrame, or numpy.ndarray directly — no manual DataLoader needed.

Telemetry Flywheel

  • Every compress() call appends arch/FLOPs/latency/CPU info to ~/.dnaty/telemetry.jsonl. Seeds the GBM latency predictor. Never raises.

New Scripts

  • scripts/build_latency_dataset.py — build GBM latency surrogate from measured ONNX timings.

Install

pip install dnaty==2.0.0

Benchmarks

Dataset FLOPs ↓ Val acc
Dry Bean Quality −83.4% 92.43%
Predictive Maint. (AI4I) −83.1% 96.70%
Electrical Fault Detect −83.0% 99.25%
Credit Fraud FULL (284K) −64.0% 99.96%
HAR Sensors −46.8% 99.17%
MNIST (full 70K) −41.8% 98.68%

Continual learning (Split-MNIST, 5 tasks): 4.9× less forgetting than EWC (BWT −0.204 vs −0.998).

Full changelog

See CHANGELOG.md for complete details.

Assets 2
Loading