ONNX-Bench

Paper Code Project Page

A large-scale benchmark of neural network architectures represented in a unified ONNX format with performance labels. Designed to train and evaluate universal, search-space-agnostic surrogate models.

✨ Highlights

• 📦 649,596 architectures across multiple NAS search spaces, standardised to ONNX.
• 🧪 Consistent evaluation on CIFAR-10 for most spaces; includes additional UnseenNAS tasks within einspace.
• 🧱 Rich architectural diversity (cell-based, hierarchical), enabling cross-space generalisation studies.
• 🔧 Pre-simplified ONNX graphs (via onnx-simplifier) for stable parsing and downstream encoding.
• 📈 Ready for performance prediction, zero-shot transfer, and universal surrogate training.

Key stats from the paper:

• Node counts span 1–3503; CIFAR-10 top-1 accuracy across [0.0, 97.03].
• Strong operational diversity within and across spaces (see paper’s JSD analysis).

🧭 What’s Inside

Search Space	Type	Evaluation	Num Architectures
NAS-Bench-101	Cell-based	CIFAR-10	423624
NAS-Bench-201	Cell-based	CIFAR-10	15625
NATS-Bench	Cell-based	CIFAR-10	32768
NAS-Bench-301	Cell-based	CIFAR-10	57189
TransNAS-Bench-101	Cell-based	Other	38895
hNAS-Bench-201	Hierarchical	CIFAR-10	8000
einspace	Hierarchical	CIFAR-10	57495
einspace	Hierarchical	UnseenNAS	16000
Total			649596