PNW-Cnet-5 ONNX Conversion Tutorial

This repository documents converting the PNW-Cnet-5 bioacoustic classifier from Keras/TensorFlow to ONNX format, achieving 8-12x faster inference on Apple Silicon (CoreML) and NVIDIA GPUs (CUDA).

This is meant to be informative and potentially help you if you are working with a similar model.

The discussion forum on this repo is open and let me know if you have any questions or suggestions!

Results Summary

NVIDIA A100 (Linux)

Metric	Keras (GPU)	ONNX (CUDA)
Time per image	4.52ms	0.53ms
3300 images	14.9s	1.76s
Speedup	1x	8.5x

• Hardware: Lambda Labs 1x A100 (NVIDIA A100-SXM4-40GB, 30 CPU cores, 200GB RAM)
• OS: Ubuntu Linux
• ONNX Runtime: 1.20+ with CUDAExecutionProvider, cuDNN 9.x

Apple Silicon (M2 Pro)

Metric	Keras (CPU)	ONNX (CoreML)
Time per image	16.6ms	1.4ms
3300 images	55s	4.7s
Speedup	1x	11.6x

• Hardware: Apple M2 Pro (10-core CPU, 16-core GPU, 16-core Neural Engine)
• OS: macOS 15 (Sequoia)
• ONNX Runtime: 1.20+ with CoreMLExecutionProvider

Note: On Apple Silicon, TensorFlow's Metal GPU plugin produces incorrect results for this model architecture. The comparison script automatically runs Keras in CPU-only mode. ONNX uses CoreML which correctly leverages the Neural Engine and GPU.

Classification Accuracy

Both platforms show 100% classification match between Keras and ONNX models. Minor floating-point differences (max ~0.001) do not affect predictions.

Quick Start

Prerequisites

• Python 3.10-3.12
• sox for audio processing
• macOS with Apple Silicon, or Linux with NVIDIA GPU (CUDA 12.x, cuDNN 9.x)

# Install sox
# macOS
brew install sox
# Ubuntu/Debian
sudo apt install sox

# Install Python dependencies
uv sync

Download Audio Data

Download .wav files into a recordings/ directory. You can use recordings from the PNW-Cnet dataset or your own field recordings.

# Download dataset part 1 (extracts .wav files to recordings/)
uv run download_dataset.py

# List available dataset parts
uv run download_dataset.py --list

# Download a specific part
uv run download_dataset.py --part part2

Convert Audio to Spectrograms

# Single file
uv run wav_to_spectrogram.py recording.wav output_spectrograms/

# Entire directory
uv run wav_to_spectrogram.py recordings/ output_spectrograms/

Audio files are split into 12-second clips, each producing a 257x1000 grayscale spectrogram.

Convert Model to ONNX

uv run convert_to_onnx.py

This converts model/Final_Model.h5 to model/Final_Model.onnx.

ONNX-slim (Optional)

You can optionally optimize the ONNX model with onnxslim for a smaller file size:

uv run onnxslim model/Final_Model.onnx model/Final_Model_slim.onnx

In our benchmarks, ONNX-slim produces identical results but was slightly slower than the original ONNX model with CoreML. The file size reduction may still be useful for deployment.

Run Inference

With ONNX (fast):

uv run run_inference_onnx.py output_spectrograms/

# Export to CSV
uv run run_inference_onnx.py output_spectrograms/ --output predictions.csv

With Keras (for comparison):

uv run run_inference.py output_spectrograms/

Compare Models

Verify the ONNX conversion produces identical results:

# Basic comparison
uv run compare_models.py output_spectrograms/

# Compare all three models
uv run compare_models.py output_spectrograms/ --onnx-slim-model model/Final_Model_slim.onnx

Output shows numerical differences, classification agreement, and performance comparison.

Command Reference

download_dataset.py

uv run download_dataset.py [options]
  --list                List available dataset files
  --part PART           Which part to download: part1, part2, part3, part4 (default: part1)
  --output-dir PATH     Output directory (default: recordings)
  --keep-zip            Keep the zip file after extraction

wav_to_spectrogram.py

uv run wav_to_spectrogram.py <input> <output_dir> [--clip-duration 12]

convert_to_onnx.py

uv run convert_to_onnx.py [--input model.h5] [--output model.onnx] [--opset 13]

run_inference_onnx.py

uv run run_inference_onnx.py <input_dir> [options]
  --model PATH        ONNX model path (default: model/Final_Model.onnx)
  --threshold FLOAT   Confidence threshold (default: 0.5)
  --top-k INT         Top predictions to show (default: 5)
  --batch-size INT    Batch size (default: 32)
  --output PATH       Export results to CSV

compare_models.py

uv run compare_models.py <input_dir> [options]
  --h5-model PATH          Keras model path
  --onnx-model PATH        ONNX model path
  --onnx-slim-model PATH   ONNX-slim model path (optional)
  --threshold FLOAT        Classification threshold

References

• Dataset: https://zenodo.org/records/10895837
• Original Model: https://github.com/zjruff/PNW-Cnet-5
• BirdNET ONNX Conversion: birdnet-team/BirdNET-Analyzer#177

Reference Scripts

The reference_scripts/ directory contains the original model architecture and training scripts from PNW-Cnet-5.

---

This tutorial was created by Bread Board Foundry. We're building Pandoro, GPU infrastructure for research teams who need predictable compute access without the overhead. If you're running ML experiments and struggling with compute access, we'd love to chat: hello@breadboardfoundry.com