[Edit for relevance]
After compiling this model with this config:
converter = tf.lite.TFLiteConverter.from_frozen_graph(
path/to/graph.pb,
input_arrays=["Cast"],
output_arrays=["features"],
input_shapes={"Cast":[1, 128, 64, 3]})
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS,
tf.lite.OpsSet.SELECT_TF_OPS]
converter.allow_custom_ops = True
converter.experimental_new_converter = True
converter.optimizations = [tf.lite.Optimize.DEFAULT]
tflite_model = converter.convert()
a tflite model is produced, but calling allocate_tensors on the interpreter causes this error:
RuntimeError: Regular TensorFlow ops are not supported by this interpreter.
Make sure you apply/link the Flex delegate before
inference.Node number 0 (FlexTensorArrayV3) failed to prepare.
So it would seem I have build tflite_convert with support for tensorflow ops...