Can't use tflite backend

[hillz2]

I tried to test the tflite backend:

root@at-work:/mnt/mmcblk2p4/stacks/other-stack # curl -X POST 'http://192.168.2.1:8000/api/v1/detect?backend=tflite&confidence_threshold=0.50&return_image=false' -F 'file=@/tmp/img.jpeg'
{"detail":"Backend initialization error: Backend 'tflite' not found. Available backends: ['onnx', 'opencv']"}#                                                                                                     root@at-work:/mnt/mmcblk2p4/stacks/other-stack # docker exec -it light-object-detect bash
root@86c0ed470f79:/app# python scripts/download_model.py
Ensuring default TFLite model for object detection...
Model and labels already exist.
Done.
root@86c0ed470f79:/app# ls -lh backends/tflite/models/
total 4.0M
-rw-r--r-- 1 root root  621 May 22 13:31 labelmap.txt
-rw-r--r-- 1 root root 4.0M May 22 13:31 ssd_mobilenet_v1.tflite
root@86c0ed470f79:/app# 

I'm running this on an android tv box that's installed with OpenWrt and docker, I can only use onnx at the moment and it's very CPU heavy for my tv box. This is my compose.yaml file:

  light-object-detect:
    build: https://github.com/opensensor/light-object-detect.git
    container_name: light-object-detect
    dns:
      - 8.8.8.8
      - 1.1.1.1
    ports:
      - "8000:8000"
    environment:
      - BACKEND=tflite
      - TFLITE_MODEL_PATH=backends/tflite/models/ssd_mobilenet_v1.tflite
      - TFLITE_LABELS_PATH=backends/tflite/models/labelmap.txt
      - TFLITE_CONFIDENCE_THRESHOLD=0.5
    volumes:
      - ./lightnvr/data/models:/app/backends/opencv/models/
    restart: unless-stopped
    deploy:
      resources:
        limits:
          memory: 750M

How can I use tflite ?

[tomo90]

Do you have tflite (tensorflow) support in your image? This was my case. The following helped me to resolve it.

go to your container: docker exec -it <your-container> /bin/bash
go to /app directory: cd /app
run: python -c "import backends.tflite.backend"

If this import fails fails your image has been built without tflite support. Look into your Dockerfile.

Try building your image with INSTALL_TENSORFLOW=1 (and other parameters you may be already using):

docker build \
  --build-arg INSTALL_TENSORFLOW=1 \
  -t light-object-detect:local .

In you compose file in the environment block you should also send DEFAULT_BACKEND instead BACKEND - when you are in your container, take a look at /app/config.py.