Supported Model / Compatibility with USB-Accelerator

[Petros626]

Hey,

I read that TensorFlow Lite only supports the SSD models from the Zoo 2 model, for example the SSD MobileNet V2 FPNLite 640x640. My question would be do you have a tutorial to convert it to a TFLite model and can it be run with the hardware accelerator (USB stick) from Google Coral?

[TannerGilbert]

It is certainly possible to convert a model trained with the OD API to Tensorflow Lite. You can find an example in my Tensorflow-Lite-Object-Detection-with-the-Tensorflow-Object-Detection-API Repository. There is also an official example.

If you're not limited to the OD-API, I can also recommend using the TFLITE Model Maker.

[Petros626]

@TannerGilbert did you have a repo for doing this on tensorflow 1 for non quantized models? Is it possible to quantize a TF1 model post training?

For my understading you have this options:
TF1: has quantized models, which can be trained quantized
TF2: supports post training quantization and NOT award training.

[Petros626]

@TannerGilbert do you agree?