This repo builds a base image ghcr.io/greenroom-robotics/nuclio_lighting_flash:latest which is used to conveniently run flash ObjectDetection models in Nuclio / CVAT. It allows you to configure the model head, backbone and checkpoint_path from your nuclio function.yml

• Use lightningflash-efficientdet-d0 as a reference example
• Modify/create your own as you see fit
• Deploy it:

nuctl deploy --project-name cvat \
  --path example/lightningflash-efficientdet-d0 \
  --platform local

In order to develop you'll want a nuclio instance running on your local machine...

• docker-compose up to start nuclio.
• ./scripts/build.sh to build ghcr.io/greenroom-robotics/nuclio_lighting_flash:latest
• Deploy the example to your nuclio instance:

nuctl deploy --project-name cvat \
  --path example/lightningflash-efficientdet-d0 \
  --platform local

It is important to note that the num_classes (labels) differs for many pretrained model. If you are getting no detections and you think you should, your num_classes is probably wrong. See this for how the coco labels changed over time

• ./scripts/build.sh && docker run ghcr.io/greenroom-robotics/nuclio_lighting_flash:latest to build and run pytests

• Run the Release workflow on github