Use the Coral TPU directly from Python with no arcane libraries (TFLite, PyCoral, Edge TPU Runtime, or any other dungeon monsters)

❗ Getting rid of all dependencies means that the Coral TPU can be easily integrated with ANY device that supports USB.

Wanna run ML on a RPi Zero? on an ESP32 running MicroPython? On your custom ASIC? If you have USB, you can have ML.

Your embedded system / single board computer can use the Coral TPU if it supports:

• Python
• USB (via pyusb)

Nothing else is needed.

Install USB library:

pip install pyusb

Prepare input image: resize to 224x224 and transform to bytes:

from PIL import Image
import sys

path = sys.argv[1]

image = Image.open(path).convert('RGB').resize((224, 224))
image_bytes = image.tobytes()

with open(f'{path.split(".")[0]}.bin', 'wb') as binary_file:
    binary_file.write(image_bytes)

1. python install_firmware.py
2. python inference.py

1. python install_firmware.py
2. cd object-detection
3. python inference.py

Note: If your computer supports Pillow, you can visualze the bounding boxes after running object detection with:

python object-detection/inference.py my_image.png