TensorFlow model used in Frigate

[B0ndo2]

According the documentation TF lite model is used by default. Is there a way to get the original TF model ?
I am planning to use my own images to fine tune the model then convert it again to TF Lite

[NickM-27]

it comes from https://github.com/google-coral/test_data not sure if the tf models are available

[halabut69]

Here you can download the model:
https://coral.ai/models/object-detection/
Choose 'All model files' for SSDlite-mobilenetv2 (the tf v1 version). This contains the tflite.pb and a pytorch .ckpt.
The tf models seem to have been removed from the TensorFlow repository.

[NickM-27]

To be clear mobilenet v2 is different than what is used for the coral, but it should still work for the coral

[halabut69]

Ah, sorry I thought the Coral defaulted to the ssdlite mobilenet V2 like the tflite cpu detector.

[NickM-27]

Neither of them use that. They both use mobiledet

frigate/docker/main/Dockerfile

Lines 116 to 117 in 7b65bcf

	RUN wget -qO edgetpu_model.tflite https://github.com/google-coral/test_data/raw/release-frogfish/ssdlite_mobiledet_coco_qat_postprocess_edgetpu.tflite
	RUN wget -qO cpu_model.tflite https://github.com/google-coral/test_data/raw/release-frogfish/ssdlite_mobiledet_coco_qat_postprocess.tflite