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 models 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, models 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 model. For example, Keras has stateful
- a special option for RNNs, that turns on saving a state between inferences. Kaldi contains a special Offset
specifier to define time offset in a model.
OpenVINO also contains a special API to simplify work with models 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 model. 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 conversion API or create an OpenVINO model. (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.
Click to see the XML file.
<?xml version="1.0" ?>
<net name="summator" version="11">
<layers>
<layer id="0" name="init_value" type="Const" version="opset6">
<data element_type="f32" offset="0" shape="1,1" size="4"/>
<output>
<port id="1" precision="FP32">
<dim>1</dim>
<dim>1</dim>
</port>
</output>
</layer>
<layer id="1" name="read" type="ReadValue" version="opset6">
<data variable_id="id"/>
<input>
<port id="0">
<dim>1</dim>
<dim>1</dim>
</port>
</input>
<output>
<port id="1" precision="FP32">
<dim>1</dim>
<dim>1</dim>
</port>
</output>
</layer>
<layer id="2" name="input" type="Parameter" version="opset6">
<data element_type="f32" shape="1,1"/>
<output>
<port id="0" precision="FP32">
<dim>1</dim>
<dim>1</dim>
</port>
</output>
</layer>
<layer id="3" name="add_sum" type="Add" version="opset6">
<input>
<port id="0">
<dim>1</dim>
<dim>1</dim>
</port>
<port id="1">
<dim>1</dim>
<dim>1</dim>
</port>
</input>
<output>
<port id="2" precision="FP32">
<dim>1</dim>
<dim>1</dim>
</port>
</output>
</layer>
<layer id="4" name="save" type="Assign" version="opset6">
<data variable_id="id"/>
<input>
<port id="0">
<dim>1</dim>
<dim>1</dim>
</port>
</input>
</layer>
<layer id="10" name="add" type="Add" version="opset6">
<data axis="1"/>
<input>
<port id="0">
<dim>1</dim>
<dim>1</dim>
</port>
<port id="1">
<dim>1</dim>
<dim>1</dim>
</port>
</input>
<output>
<port id="2" precision="FP32">
<dim>1</dim>
<dim>1</dim>
</port>
</output>
</layer>
<layer id="5" name="output/sink_port_0" type="Result" version="opset6">
<input>
<port id="0">
<dim>1</dim>
<dim>1</dim>
</port>
</input>
</layer>
</layers>
<edges>
<edge from-layer="0" from-port="1" to-layer="1" to-port="0"/>
<edge from-layer="2" from-port="0" to-layer="3" to-port="1"/>
<edge from-layer="1" from-port="1" to-layer="3" to-port="0"/>
<edge from-layer="3" from-port="2" to-layer="4" to-port="0"/>
<edge from-layer="3" from-port="2" to-layer="10" to-port="0"/>
<edge from-layer="1" from-port="1" to-layer="10" to-port="1"/>
<edge from-layer="10" from-port="2" to-layer="5" to-port="0"/>
</edges>
<meta_data>
<MO_version value="unknown version"/>
<cli_parameters>
</cli_parameters>
</meta_data>
</net>
Example of Creating Model OpenVINO API¶
In the following example, the SinkVector
is used to create the ov::Model
. For a model with states, except inputs and outputs, the Assign
nodes should also point to the Model
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 ov::Model
.
auto arg = std::make_shared<ov::opset11::Parameter>(ov::element::f32, ov::Shape{1, 1});
auto init_const = ov::opset11::Constant::create(ov::element::f32, ov::Shape{1, 1}, {0});
// The ReadValue/Assign operations must be used in pairs in the model.
// For each such a pair, its own variable object must be created.
const std::string variable_name("variable0");
auto variable = std::make_shared<ov::op::util::Variable>(
ov::op::util::VariableInfo{ov::PartialShape::dynamic(), ov::element::dynamic, variable_name});
// Creating ov::Model
auto read = std::make_shared<ov::opset11::ReadValue>(init_const, variable);
std::vector<std::shared_ptr<ov::Node>> args = {arg, read};
auto add = std::make_shared<ov::opset11::Add>(arg, read);
auto assign = std::make_shared<ov::opset11::Assign>(add, variable);
auto add2 = std::make_shared<ov::opset11::Add>(add, read);
auto res = std::make_shared<ov::opset11::Result>(add2);
auto model =
std::make_shared<ov::Model>(ov::ResultVector({res}), ov::SinkVector({assign}), ov::ParameterVector({arg}));
OpenVINO State API¶
OpenVINO has the InferRequest::query_state
method to get the list of states from a model and ov::IVariableState
interface to operate with states. Below is a brief description of methods and the example of how to use this interface.
std::string get_name() const
- returns the name (variable_id) of a corresponding Variable.void reset()
- resets a state to a default value.void set_state(const ov::Tensor& state)
- sets a new value for a state.const ov::Tensor& get_state() const
- returns current value of state.
Example of Stateful Model 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 ov::IVariableState::set_state
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++) {
auto minput = input_tensors[0];
std::memcpy(minput.data(), &data[next_input], sizeof(float));
infer_request.infer();
// check states
auto states = infer_request.query_state();
if (states.empty()) {
throw std::runtime_error("Queried states are empty");
}
auto mstate = states[0].get_state();
if (!mstate) {
throw std::runtime_error("Can't cast state to MemoryBlob");
}
float* state = mstate.data<float>();
std::cout << state[0] << "\n";
}
// resetting state between utterances
std::cout << "Reset state\n";
for (auto&& state : infer_request.query_state()) {
state.reset();
}
// infer the second utterance
for (size_t next_input = data.size() / 2; next_input < data.size(); next_input++) {
auto minput = input_tensors[0];
std::memcpy(minput.data(), &data[next_input], sizeof(float));
infer_request.infer();
// check states
auto states = infer_request.query_state();
auto mstate = states[0].get_state();
float* state = mstate.data<float>();
std::cout << state[0] << "\n";
}
For more elaborate examples demonstrating how to work with models 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 conversion API.
ONNX and frameworks supported via ONNX format -
LSTM
,RNN
, andGRU
original layers are converted to theTensorIterator
operation. TheTensorIterator
body containsLSTM
/RNN
/GRU Cell
. ThePeepholes
andInputForget
modifications are not supported, while thesequence_lengths
optional input is.
ONNX Loop
layer is converted to the OpenVINO Loop operation.
Apache MXNet -
LSTM
,RNN
,GRU
original layers are converted toTensorIterator
operation, which body containsLSTM
/RNN
/GRU Cell
operations.TensorFlow -
BlockLSTM
is converted toTensorIterator
operation. TheTensorIterator
body containsLSTM Cell
operation, whereasPeepholes
andInputForget
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 conversion) time, the dynamic cases are not supported.
TensorFlow2 -
While
layer is converted toLoop
operation, which body can contain any supported operations.Kaldi - Kaldi models already contain
Assign
/ReadValue
(Memory) operations after model conversion. TheTensorIterator
/Loop
operations are not generated.
Note
Note that OpenVINO support for Apache MXNet, Caffe, and Kaldi is currently being deprecated and will be removed entirely in the future.