Suggestion for choosing a model

I am playing with a coral board (TF 1.X).
I would like to do a transfer learning able to distinguish small metal parts by size and orientation with great reliability.
Which network do you recommend starting from?

Take a look at:

They still don’t have a semantic segmentation retraining official tutorial/doc like for the object detection and classification but there is a feature request that you can upvote and subscribe:

@lgusm Do we have a Coral tag and Coral team subscribers?

We didn’t but I’ve just created.

I’ll ping the Coral team to get some answer here

1 Like

Thanks for the suggestion

@Cristian_Lorenzin, yes we haven’t made a official tutorial to retain a semantic segmentation model.Just because the procedure has been changing with new models. But we will consider your feedback to make one.

I mean time you can try steps below to retrain a model for edgetpu

follow steps here to train and export a model.
refer here for checkpoints ( I have tried mobilenetv2, mobilenet-edgetpu) feel free to try other backbones as shown on

The trick is to get quantization right.
Though you have trained and exported model in TF1.x you should use PTQ in TF2.x > 2.3 to quantise the model. Finally use edgetpu compiler to compile the model.

Let us know you have any questions.



@Cristian_Lorenzin, in addition to Naveen’s suggestion, if you’d like to learn more about using the Transfer Learning method, please take a look at Coral documentation on this subject using the new weight imprinting API and code samples. If the metal parts are large and changing orientation is obvious, this approach could also work using the Coral Dev Board HW and learning to recognize size and orientation changes on the fly using on-device AI/ML.