# Stateful models¶

Several use cases require processing of data sequences. When length of a sequence is known and small enough, it can be processed with RNN like networks that contain a cycle inside. However, in some cases (e.g., online speech recognition of time series forecasting) length of data sequence is unknown. Then, data can be divided in small portions and processed step-by-step. The dependency between data portions should be addressed. For that, networks save some data between inferences - a state. When one dependent sequence is over, a state should be reset to initial value and a new sequence can be started.

Several frameworks have special APIs for states in networks. For example, Keras has stateful - a special option for RNNs, that turns on saving a state between inferences. Kaldi contains special Offset specifier to define time offset in a network.

OpenVINO also contains a special API to simplify work with networks with states. A state is automatically saved between inferences, and there is a way to reset a state when needed. A state can also be read or set to some new value between inferences.

## OpenVINO State Representation¶

OpenVINO contains the Variable, a special abstraction to represent a state in a network. There are two operations: Assign - to save a value in a state and ReadValue - to read a value saved on previous iteration.

To get a model with states ready for inference, convert a model from another framework to OpenVINO IR with Model Optimizer or create an nGraph function. (For more information, refer to the Build OpenVINO Model section).

Below is the graph in both forms:

### Example of IR with State¶

The bin file for this graph should contain float 0 in binary form. The content of the xml file is as follows.

### Example of Creating Model nGraph API¶

In the following example, the SinkVector is used to create the ngraph::Function. For a network with states, except inputs and outputs, the Assign nodes should also point to the Function to avoid deleting it during graph transformations. Use the constructor to do it, as shown in the example, or with the special add_sinks(const SinkVector& sinks) method. After deleting the node from the graph with the delete_sink() method, a sink can be deleted from ngraph::Function.

#include <ngraph/opsets/opset6.hpp>
#include <ngraph/op/util/variable.hpp>
// ...

auto arg = make_shared<ngraph::opset6::Parameter>(element::f32, Shape{1, 1});
auto init_const = ngraph::opset6::Constant::create(element::f32, Shape{1, 1}, {0});

// The ReadValue/Assign operations must be used in pairs in the network.
// For each such a pair, its own variable object must be created.
const std::string variable_name("variable0");
auto variable = std::make_shared<ngraph::Variable>(VariableInfo{PartialShape::dynamic(), element::dynamic, variable_name});

// Creating ngraph::function
auto read = make_shared<ngraph::opset6::ReadValue>(init_const, variable);
std::vector<shared_ptr<ngraph::Node>> args = {arg, read};
auto add = make_shared<ngraph::opset6::Add>(arg, read);
auto assign = make_shared<ngraph::opset6::Assign>(add, variable);
auto add2 = make_shared<ngraph::opset6::Add>(add, read);
auto res = make_shared<ngraph::opset6::Result>(add2);

auto f = make_shared<Function>(ResultVector({res}), ParameterVector({arg}), SinkVector({assign}));

## OpenVINO State API¶

Inference Engine has the InferRequest::QueryState method to get the list of states from a network and IVariableState interface to operate with states. Below is a brief description of methods and the example of how to use this interface.

• std::string GetName() const - returns the name (variable_id) of a corresponding Variable.

• void Reset() - resets a state to a default value.

• void SetState(Blob::Ptr newState) - sets a new value for a state.

• Blob::CPtr GetState() const - returns current value of state.

## Example of Stateful Network Inference¶

Based on the IR from the previous section, the example below demonstrates inference of two independent sequences of data. A state should be reset between these sequences.

One infer request and one thread will be used in this example. Using several threads is possible if there are several independent sequences. Then, each sequence can be processed in its own infer request. Inference of one sequence in several infer requests is not recommended. In one infer request, a state will be saved automatically between inferences, but if the first step is done in one infer request and the second in another, a state should be set in a new infer request manually (using the IVariableState::SetState method).

  // input data
std::vector<float> data = { 1,2,3,4,5,6};
// infer the first utterance
for (size_t next_input = 0; next_input < data.size()/2; next_input++) {
MemoryBlob::Ptr minput = as<MemoryBlob>(ptrInputBlobs[0]);
auto minputHolder = minput->wmap();

std::memcpy(minputHolder.as<void \*>(),
&data[next_input],
sizeof(float));

inferRequest.Infer();
// check states
auto states = inferRequest.QueryState();
if (states.empty()) {
throw std::runtime_error("Queried states are empty");
}
auto mstate = as<MemoryBlob>(states[0].GetState());
if (mstate == nullptr) {
throw std::runtime_error("Can't cast state to MemoryBlob");
}
auto state_buf = mstate->rmap();
float \* state =state_buf.as<float\*>();
std::cout << state[0] << "\n";
}

// resetting state between utterances
std::cout<<"Reset state\n";
for (auto &&state : inferRequest.QueryState()) {
state.Reset();
}

// infer the second utterance
for (size_t next_input = data.size()/2; next_input < data.size(); next_input++) {
MemoryBlob::Ptr minput = as<MemoryBlob>(ptrInputBlobs[0]);
auto minputHolder = minput->wmap();

std::memcpy(minputHolder.as<void \*>(),
&data[next_input],
sizeof(float));

inferRequest.Infer();
// check states
auto states = inferRequest.QueryState();
auto mstate = as<MemoryBlob>(states[0].GetState());
auto state_buf = mstate->rmap();
float \* state =state_buf.as<float\*>();
std::cout << state[0] << "\n";
}

For more elaborate examples demonstrating how to work with networks with states, refer to the speech sample and a demo in the Samples Overview.

## LowLatency Transformations¶

If the original framework does not have a special API for working with states, OpenVINO representation will not contain Assign / ReadValue layers after importing the model. For example, if the original ONNX model contains RNN operations, OpenVINO IR will contain TensorIterator operations and the values will be obtained only after execution of the whole TensorIterator primitive. Intermediate values from each iteration will not be available. Working with these intermediate values of each iteration is enabled by special LowLatency and LowLatency2 transformations, which also help receive these values with a low latency after each infer request.

Note

It is recommended to use LowLatency2, as LowLatency transformation has already been deprecated.

### TensorIterator/Loop operations¶

You can get the TensorIterator/Loop operations from different frameworks via Model Optimizer.

• ONNX and frameworks supported via ONNX format - LSTM, RNN, and GRU original layers are converted to the TensorIterator operation. The TensorIterator body contains LSTM / RNN / GRU Cell. The Peepholes and InputForget modifications are not supported, while the sequence_lengths optional input is. ONNX Loop layer is converted to the OpenVINO Loop operation.

• Apache MXNet - LSTM, RNN, GRU original layers are converted to TensorIterator operation, which body contains LSTM / RNN / GRU Cell operations.

• TensorFlow - BlockLSTM is converted to TensorIterator operation. The TensorIterator body contains LSTM Cell operation, whereas Peepholes and InputForget modifications are not supported. The While layer is converted to TensorIterator, which body can contain any supported operations. However, when count of iterations cannot be calculated in shape inference (Model Optimizer conversion) time, the dynamic cases are not supported.

• TensorFlow2 - While layer is converted to Loop operation, which body can contain any supported operations.

• Kaldi - Kaldi models already contain Assign / ReadValue (Memory) operations after model conversion. The TensorIterator / Loop operations are not generated.