MynaNet

Lightweight CNN for bird call classification targeting deployment on Arduino Portenta H7 (Cortex-M7, 512 KB flash).

Adopted Model: MynaNet = 1j (MBV3-SE)

1j_mbv3_se.py is the production MynaNet model.

• Architecture: MobileNetV3-style inverted residual blocks with 5×5 depthwise convolutions and hard-sigmoid Squeeze-Excitation
• 94.91% INT8 mean accuracy (3 seeds, Linux/CUDA authoritative)
• 267 KB INT8 — well within the 512 KB H7 flash limit
• MCU-deployable: all ops supported by TFLite Micro on Portenta H7

Dataset

mygardenbird16khz — 12 garden bird species, 16 kHz
Fixed 80:10:10 train/val/test split (CSV-based, no leakage)

Repository layout

deploy/          ← production: train MynaNet, convert to firmware C array
  train.py           MynaNet (1j) training + INT8 quantization
  convert_xxd.sh     TFLite → alignas(8) C array for Portenta H7 firmware
  README.md          End-to-end guide: download → train → quantize → deploy

develop/         ← ablation: all Series 1 model scripts + sweep runner
  1a_baseline_2dcnn.py … 1n_efficientnetb0.py
  run_seabird12_ablation.sh
  README.md          Full ablation results and model comparison

Quick start (deploy)

See deploy/README.md for the full end-to-end guide.

# Train MynaNet on mygardenbird16khz
python deploy/train.py \
  --flat_dir /path/to/mygardenbird16khz \
  --splits_csv /path/to/metadata16khz/splits_mip_80_10_10.csv

# Convert trained INT8 TFLite → firmware C array
bash deploy/convert_xxd.sh model_int8.tflite src/mynanet_model_data g_mynanet_model_data

Ablation

See develop/README.md for the full model comparison, results table, and key findings.

License

MIT