I am trying to use Tensorflow object detection API models in another custom model I built (in the same codebase). Specifically, I am trying to figure out how below can be handled (mutually exclusive points):
- jointly train Tensorflow object detection model Y with another Custom auto-encoder model X - Possibility 1
- or just train custom auto-encoder model X while obtaining object detection Y predictions and incorporating it into X - Possibility 2
- or object detection Y input is (a tensor) from another custom auto-encoder X's intermediate layer (not a tfrecord or RGB image) - Possibility 3
- please note the above image shows only one architectural variation (Possibility 3). Essentially, if I can figure out how to do joint training with Tensorflow object detection API, I can do them all.
I have gone through the official TF object detection API docs and scoured the net trying to find good examples where TF object detection API was customized for cases beyond just object detection. I haven't found any. Any help or links would be appreciated.
p.s.: some relevant points
- FYI, I can run/train Tensorflow OD API independently
- Stackoverflow thrives on "show-me-what-you-did" culture, but as this question is preliminary and something that I haven't found an answer to in their documentation or on the web, hence shaking the community to find if someone has some thoughts on this.
- I had posted a similar question on datascience a few days back but no response.
- TF object detection API github new issues encourages posting to stackoverflow for help and support.