Add ONNX playground for graph inspection, optimization, and benchmarking

[dchou1618]

Summary

Adds a new onnx/ playground directory for exploring ONNX internals, graph transformations, and inference benchmarking.

This PR introduces a standalone experimental environment for:

• exporting ONNX graphs
• inspecting graph structure
• benchmarking ONNX Runtime inference
• experimenting with graph optimizations
• understanding ONNX as an intermediate representation (IR)

The goal is to build familiarity with low-level ONNX mechanics rather than treating ONNX as a black-box export format.

---

Intended Structure

onnx/
├── README.md
├── requirements.txt
│
├── models/
├── scripts/

[qodo-code-review]

Code Review by Qodo

🐞 Bugs (0) 📘 Rule violations (0) 📎 Requirement gaps (0)

Great, no issues found!

Qodo reviewed your code and found no material issues that require review

[qodo-code-review]

PR Summary by Qodo

Add ONNX playground scaffolding with baseline MLP model artifact
✨ Enhancement 📝 Documentation 🕐 10-20 Minutes

Walkthroughs

Description

• Add a dedicated ONNX playground directory to support graph/IR exploration work.
• Document intended workflows (export, inspect, optimize, benchmark) and folder layout.
• Include a baseline ONNX MLP model artifact for initial experiments.

Diagram

graph TD
  A["Modeling/onnx"] --> B["README.md"] --> C[("mlp_baseline.onnx")]

Loading

High-Level Assessment

The following are alternative approaches to this PR:

1. Generate ONNX artifacts via script (don’t commit binaries)

• ➕ Avoids repository bloat and binary diff noise
• ➕ Ensures models are reproducible from source
• ➕ Easier to update/export across opsets and configs
• ➖ Adds setup friction (tooling + runtime deps) for first-time users
• ➖ Requires maintaining export scripts and pinned versions

2. Store ONNX binaries using Git LFS

• ➕ Keeps repo history slim while still versioning artifacts
• ➕ Maintains convenient local checkout workflow
• ➖ Requires Git LFS support in contributor environments and CI
• ➖ May complicate forks/mirrors depending on hosting

3. Publish artifacts as release assets (or separate data bucket)

• ➕ Keeps source repo clean and fast to clone
• ➕ Can host multiple benchmark artifacts with metadata
• ➖ More moving parts (download step, version coordination)
• ➖ Offline workflows are less convenient

Recommendation: The directory/README scaffolding is a good fit for an experimental playground, but committing ONNX binaries can become painful as the set grows. Prefer either (a) generating models from a small export script with pinned deps, or (b) using Git LFS if you want artifacts versioned in-repo. If this PR is intentionally minimal, consider adding a follow-up plan (LFS vs generation) before adding more model files.

File Changes

Enhancement (1)

mlp_baseline.onnx Add baseline ONNX MLP model artifact +0/-0

Add baseline ONNX MLP model artifact

• Introduces an ONNX model file intended as an initial baseline artifact for inspection and experimentation within the playground.

Modeling/onnx/models/mlp_baseline.onnx

---

Documentation (1)

README.md Document ONNX playground goals and intended directory layout +58/-0

Document ONNX playground goals and intended directory layout

• Adds a README describing the purpose of the ONNX playground, including learning objectives around graph structure, runtime behavior, and optimization effects. Documents an intended folder structure for models, benchmarks, scripts, and outputs.

Modeling/onnx/README.md

---