I have the coral USB accelerator and I am searching to train an image classification model. The best model I can find that fits my needs is the EfficientNet-EdgeTpu (S). I am trying to find a way to retrain the model in my computer with tensoflow API for exemple, but all I can find (coral page mainly and tensorflow git readme) is how to train it using google cloud tools and VMs. Is it necesary to use google cloud sollutions to train EfficientNet-EdgeTpu models ?
Also I found how to easyly train Efficientnet models B0-B7, but I dont know if this models are the same as EfficientNet-EdgeTpu (S/M/B) for running in a coral USB.
I just want to find an easy way to train in my computer the efficientnet model that fits better the coral USB without having to create an acount and pay for google storage and VM training.
One thing you can do is to train the out of the box efficientnet (from tensorflow) on your custom data and convert that into tflite and after that compile it with the edgetpu_compiler.