🛠 Inférence de réseau de neurones sur le C7x+MMA avec edgeai-tidl-tools

Objectives

Move beyond CPU emulation: install the TI edgeai-tidl-tools stack (the J721E/TDA4VM is an officially supported device), retrieve pre-quantized models from the Edge AI Model Zoo, and run classification/detection on the C7x+MMA accelerator (~8 TOPS INT8) using the TIDL runtime. This tutorial justifies the purchase of this board.

Hardware

• 1 BeagleBone AI-64 with a recent BeagleBoard Debian image or TI Edge AI image
• Generous storage: 64 GB+ microSD or NVMe via M.2 — the stack and models weigh several GB; the 16 GB eMMC is not enough
• Network access to download models and dependencies

Complete Code

sudo apt update && sudo apt install -y python3-pip git python3-venv
git clone --depth 1 https://github.com/TexasInstruments/edgeai-tidl-tools
cd edgeai-tidl-tools
# Suivre le README pour le setup aarch64 ON-DEVICE (script osd/setup.sh) ;
# les versions exactes d'API bougent, rester sur un tag stable (ex. 10_01_00_02).
export OS_ARCH=aarch64
cd examples/osd && sudo ./setup.sh && cd ../..
sudo ./get_models.sh            # modèles pré-quantisés du Model Zoo

# Exemple de classification sur le C7x-MMA (cible SoC, pas CPU)
sudo ./tidl_classification \
  -t 1 -m mobilenet-v2-12 \
  -l ../../dl_test_lists/labels.txt \
  -i ../../test_images/ -b ../../model_zoo/

# Vérifier que l'accélérateur travaille (pas les A72)
htop   # les 2 Cortex-A72 doivent rester majoritairement inactifs

Expected Verification

The benchmark displays a throughput (FPS or inferences/s) significantly higher than what the CPU would achieve, and the runtime log indicates execution on the target SoC with C7x/MMA offload. The classes and scores for the test image are output correctly.

Common Errors

• Believing you are accelerating when everything is running on the A72s: check the execution mode in the log and CPU usage (htop) — an unsupported model silently falls back to the CPU.
• Running out of space on the eMMC: installation fails at 99% without a clear message; migrate the root filesystem or work from microSD (see the storage tutorial).
• Mixing versions: the TIDL API changes between tags; pin a specific tag (e.g., 10_01_00_02) and its associated Model Zoo.
• Pure mainline image: the C7x-MMA libraries come from the TI downstream stack; without it, nothing is accelerated.
• git.beagleboard.org repository behind Anubis anti-bot: some Yocto fetch scripts fail (issue #45) — use GitHub mirrors.