Hello,
I am currently working on Speech Recognition on Embedded system and the iMX8M Plus seems to be a fine option with its NPU. In "i.MX 8M Plus Applications Processor Datasheet for Industrial Products" it is mentionned that the NPU could be use to accelerate the inference and do so :
"Neutral Processing Unit (NPU) 2.3 TOP/s Neural Network performance
• Keyword detect, noise reduction, beamforming
• Speech recognition (i.e. Deep Speech 2)
• Image recognition (i.e. ResNet-50)"
However, I cannot seem to make it work, nor can I find an example on the topic.
Has anyone already worked on this solution ? How can I use the NPU to accelerate the inférence ? I am currently using tflite as follows :
"import tflite_runtime.interpreter as tflite
...
interpreter = tflite.Interpreter(model_path=args.model_file, experimental_delegates=ext_delegate, num_threads=args.num_threads)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
interpreter.set_tensor(input_details[0]['index'], audio_data)
interpreter.invoke()
output_data = interpreter.get_tensor(output_details[0]['index'])
"
But it does not seem to work since evrytime i ad /usr/lib/libvx_delegate.so
I have this error/warning log : Loading model from ./deepspeech-0.9.3-models.tflite
WARNING: Fallback unsupported op 32 to TfLite
WARNING: Fallback unsupported op 32 to TfLite
WARNING: dynamic shape in not support in reshape.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: hybrid data type is not supported in fullyconnected.
ERROR: String input is not supported
ERROR: String input is not supported
Model loaded
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
