I need an implementation of an SVDF layer using Tensorflow 2.* that is compatible with TFLite's SVDF operator. That is, a layer that, after conversion to TFLite, generates an SVDF operator.
TFLite has a built-in SVDF operator (as listed here), but Tensorflow doesn't have an implementation for the SVDF layer.
I tried using the implementation from Google Research, but it doesn't generate the fused SVDF operation. In fact, when listing the operators from a model with this layer (using tflite_model._get_ops_details()), I get something like: {'DEPTHWISE_CONV_2D', 'FULLY_CONNECTED', 'PAD', 'RESHAPE'}.
The only example of a converted model that I could find with the SVDF operator listed was a .tflite file from Tensorflow's Github (avaiable here), but there's no information on how this model was generated. When using tflite_model._get_ops_details() it gives out the following operators: {'FULLY_CONNECTED', 'QUANTIZE', 'SOFTMAX', 'SVDF'}.