Coop Cam Bird Classifier

Int8 quantized MobileNetV2 for on-device bird species classification on ESP32-S3.

Model Details

  • Architecture: MobileNetV2 (width_mult=0.25)
  • Input: 96x96 grayscale (int8), shape [1, 96, 96, 1]
  • Output: 168 North American bird species (int8)
  • Size: 612 KB (TFLite int8)
  • Validation accuracy: 58.8% (post-quantization)

Training

Knowledge distillation from chriamue/bird-species-classifier (EfficientNet-B2, 525 classes) using the chriamue/bird-species-dataset.

  • Distillation temperature: 4.0
  • Alpha (KL weight): 0.7
  • Optimizer: AdamW (lr=1e-3, weight_decay=1e-4)
  • Scheduler: Cosine annealing over 50 epochs
  • Early stopping: Patience 10

Files

File Description
Int8 quantized TFLite model (deploy to ESP32)
C array header for embedding in firmware
C header with species name array
PyTorch checkpoint (for retraining)
ONNX intermediate (for re-export)

Usage

For ESP32-S3 firmware, include and in your build.

To regenerate from the TFLite file:

License

MIT (derived from MIT-licensed teacher model and CC0 dataset)

Downloads last month
1
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train mvdmakesthings/coopcam-bird-classifier