n2b8/backyard-birds

🐦 EfficientNet-B7 — Backyard Feeder Bird Classifier

Custom ONNX-based bird classification model trained on a filtered subset of the NABirds dataset, optimized for backyard bird feeders.

Designed to run on a Raspberry Pi + Hailo-8 setup in real-time as part of the Birdwatcher Project

🧠 Model Details

• Architecture: EfficientNet-B7
• Resolution: 600×600
• Precision: Mixed (AMP)
• Format: ONNX
• Classes: 95 backyard species + not_a_bird
• Validation Accuracy: 93.14% @ Epoch 23
• Optimized for inference on edge devices (e.g., Raspberry Pi 5 + Hailo-8)

📁 Files

• efficientnet_b7_backyard_feeder_birds.onnx — Trained ONNX model
• class_labels_v3.txt — One class label per line

🛠️ Intended Use

• Use Case: Real-time fine-grained classification of birds visiting backyard feeders
• Hardware: Optimized for Raspberry Pi + Hailo-8 AI accelerator
• Pipeline: Captures images via YOLOv8 detection, then classifies via this model

🐍 Example Inference Code

import onnxruntime as ort
import numpy as np
from PIL import Image

# Load model
session = ort.InferenceSession("efficientnet_b7_backyard_feeder_birds.onnx")

# Preprocess
img = Image.open("bird.jpg").resize((600, 600))
x = np.array(img).astype(np.float32) / 255.0
x = np.transpose(x, (2, 0, 1))[np.newaxis, :]  # CHW + batch

# Predict
outputs = session.run(None, {"input": x})
pred_idx = np.argmax(outputs[0])

📊 Training

• Dataset: Subset of NABirds (filtered for common backyard species)
• Augmentations: Flip, rotation, brightness
• Regularization: Dropout, label smoothing
• Loss: Cross-entropy
• Optimizer: AdamW
• Early stopping enabled

🔗 Related Repos

• Birdwatcher Project (GitHub)
• NABirds Dataset (Kaggle)

📜 License

GPLv3