raphaelDkhn/orion-benchmark

Model benchmarking using Orion

Orion Model Benchmarking

In this repository you can find a set of benchmarkings with different model architectures from the ONNX Hub Models repository.

Results

• ONNX Operators Frequency Usage
• ONNX Quantized Operators Frequency Usage
• Orion Compatibility Percentage and Missing Operators with Quantized ONNX Hub models
• Orion Compatibility Percentage with ONNX Hub models
• Transpiler + Oion Compatibility with ONNX Hub models
• Compilation time of Transpiled models from ONNX Hub

Development

1. Clone ONNX Hub Models repository:

git clone https://github.com/onnx/models.git

2. Install Git LFS

brew install git-lfs

3. Setup Git LFS

git lfs install

4. Fetch models:

git lfs fetch

5. Checkout models:

git lfs checkout

Setup

1. Install dependencies:

poetry install
poetry run pip install keyring
poetry run pip install keyrings.google-artifactregistry-auth

2. Get smartonnx repository configuration:

gcloud artifacts print-settings python --project=giza-platform \
    --repository=smartonnx \
    --location=europe-west1

3. Install smartonnx:

pip install --index-url https://europe-west1-python.pkg.dev/giza-platform/smartonnx/simple/ smartonnx

Usage

1. Get ONNX operator usage by frequency:

python orion_benchmark/operator_analytics.py

2. Get ONNX operator usage by frequency for quantized models:

python orion_benchmark/quant_operator_analytics.py

3. Get model percentage of compatibility with Orion Operators:

python orion_benchmark/operator_coverage.py >> orion_benchmark/operator_coverage.md

4. Get model percentage of compatibility with Quant Orion Operators:

python orion_benchmark/quant_operator_coverage.py >> orion_benchmark/quant_operator_coverage.md

5. Get transpilation time of supported ONNX Hub models with Orion:

python orion_benchmark/transpile_models.py