Error occurs during face recognition on ARM

[Felix-Franz]

Describe the bug
Face recognition is not working.

To Reproduce
Steps to reproduce the behavior:

1. Initially start photonix
2. Add large media folder
3. Wait for processing the images
4. Have a look at the log files to see that this error occurs multiple times

Expected behavior
It looks like an import of a external library fails.
That is not how it is supposed to be.

Server (please complete the following information):

• Device: Raspberry Pi 4B
• OS: Raspberry Pi OS 5.6.8-1 (Kernel 5.10.17-v8+)
• Architecture: arm64

Desktop (please complete the following information):

• OS: Windows 10
• Browser Chrome
• Version 91.0.4472.124

Additional context

Console log:

photonix_1  | ImportError: `load_weights` requires h5py when loading weights from HDF5.
photonix_1  | Traceback (most recent call last):
photonix_1  |   File "/srv/photonix/photos/utils/classification.py", line 67, in __process_task
photonix_1  |     self.runner(task.subject_id)
photonix_1  |   File "/srv/photonix/classifiers/face/model.py", line 225, in run_on_photo
photonix_1  |     model = FaceModel(library_id=photo and photo.library_id)
photonix_1  |   File "/srv/photonix/classifiers/face/model.py", line 41, in __init__
photonix_1  |     self.graph = self.load_graph(graph_file)
photonix_1  |   File "/srv/photonix/classifiers/face/model.py", line 61, in load_graph
photonix_1  |     facenet_graph = build_model('Facenet')
photonix_1  |   File "/srv/photonix/classifiers/face/deepface/DeepFace.py", line 37, in build_model
photonix_1  |     model = model()
photonix_1  |   File "/srv/photonix/classifiers/face/deepface/basemodels/Facenet.py", line 540, in loadModel
photonix_1  |     model.load_weights(weights)
photonix_1  |   File "/usr/local/lib/python3.8/site-packages/tensorflow/python/keras/engine/training.py", line 2219, in load_weights

docker-compose.yaml:

version: '3'

services:
  postgres:
    image: postgres:11.1-alpine
    environment:
      POSTGRES_DB: photonix
      POSTGRES_PASSWORD: password
    volumes:
      - ./data/db:/var/lib/postgresql/data
    networks:
      - photonix
    restart: always

  redis:
    image: arm32v7/redis:3.2.12
    networks:
      - photonix
    restart: always

  photonix:
    image: photonixapp/photonix:latest
    ports:
      - '85:80'
    environment:
      ENV: prd
      POSTGRES_HOST: postgres
      POSTGRES_DB: photonix
      POSTGRES_USER: postgres
      POSTGRES_PASSWORD: password
      REDIS_HOST: redis
      ALLOWED_HOSTS: '*'
    volumes:
      - /path/to/images/:/data/media:ro
      - ./data/raw-photos-processed:/data/raw-photos-processed
      - ./data/cache:/data/cache
      - ./data/models:/data/models
    networks:
      - photonix
    restart: always

networks:
  photonix:

[damianmoore]

Hi @Felix-Franz, thanks for reporting this. I tried your docker-compose.yml (just changed a volume, port and Redis image) with latest version (v0.14.0) and it detected photos fine. I assume this means the problem is with ARM 64-bit version. I'll try on my Pi 4 next.

[Felix-Franz]

I have update to postgres:13-alpine, arm64v8/redis:6-alpine and photonixapp/photonix:0.14.0, but the error still persists:

Console log:

photonix_1  | 2021-07-13 15:54:14,286 ERROR    Error processing task: classify.face - ee15d4b5-6d2d-4cf2-d030-8e2141943913
photonix_1  | Traceback (most recent call last):
photonix_1  |   File "/srv/photonix/photos/utils/classification.py", line 70, in __process_task
photonix_1  |     self.runner(task.subject_id)
photonix_1  |   File "/srv/photonix/classifiers/face/model.py", line 225, in run_on_photo
photonix_1  |     model = FaceModel(library_id=photo and photo.library_id)
photonix_1  |   File "/srv/photonix/classifiers/face/model.py", line 41, in __init__
photonix_1  |     self.graph = self.load_graph(graph_file)
photonix_1  |   File "/srv/photonix/classifiers/face/model.py", line 61, in load_graph
photonix_1  |     facenet_graph = build_model('Facenet')
photonix_1  |   File "/srv/photonix/classifiers/face/deepface/DeepFace.py", line 37, in build_model
photonix_1  |     model = model()
photonix_1  |   File "/srv/photonix/classifiers/face/deepface/basemodels/Facenet.py", line 540, in loadModel
photonix_1  |     model.load_weights(weights)
photonix_1  |   File "/usr/local/lib/python3.8/site-packages/tensorflow/python/keras/engine/training.py", line 2219, in load_weights
photonix_1  |     raise ImportError(
photonix_1  | ImportError: `load_weights` requires h5py when loading weights from HDF5.

[damianmoore]

@Felix-Franz This is now fixed as of 618eeb7. Just building a new release with the fix which is v0.16.0. The problem was a missing shared library for h5py (libhdf5_serial.so provided by libhdf5-dev) which was in the wrong stage of the Docker multi-stage build. I'm assuming the amd64 pre-compiled wheel includes this file as there was no problem on this platform.

[Felix-Franz]

Thanks, face recognition is working now! 🎉