How does OV handle state in RNN-like layers? Is every recurrent model stateful by default?

[GustawOhler]

I converted my tensorflow model to OV like this:

from openvino.runtime import serialize

ir_path = Path(model_path)/"openVINO/serialized_model.xml"
ov_model = convert_model(cloned_model, input_shape=[[1,1,224,224,3]])
serialize(ov_model, ir_path)

And then inference like this:

import openvino.runtime as ov

compiled_model = ov.compile_model(ir_path)
infer_request = compiled_model.create_infer_request()
for input_path in frame_paths:
  for state in infer_request.query_state():
      state.reset()
  # Create tensor from external memory
  input_tensor = ov.Tensor(array=get_model_input([input_path], max_sequence_len=1), shared_memory=False)
  # Set input tensor for model with one input
  infer_request.set_input_tensor(input_tensor)
  # infer_request.query_state().reset()

infer_request.start_async()
infer_request.wait()

# Get output tensor for model with one output
output = infer_request.get_output_tensor()
output_buffer = output.data

It got me different results than original TF model. But when I put infer_request = compiled_model.create_infer_request() in main loop then everything seems fine. But I can't tackle where is the difference, especially that I'm resetting query state. And when I execute infer_request.query_state() it returns with empty array.

When I tried to get input data to search for some state tensors I got only errors saying that there's only 1 tensor (the one that I supplied).

My model is LSTM layers on top of EfficientNetV2B0 so on top of bunch of convolutional layers.

[ilya-lavrenov]

CC @itikhono @mateusztabaka