[IE][VPU]: Enables Extract Dynamic Batch Transformation (#3715)

* [IE][nGraph]: Enables begin/end iterators for PartialShape

It's convenient to be able to use STL algorithms on
PartialShape since semantically PartialShape is a
sequence of Dimensions.

* [IE][VPU][nGraph]: Introduces tree utilities

Introduces Depth-First-Search and Breadth-First-Search
utilities for tree traversal. Templated arguments
makes them extensible for different use-case scenarios.

BFS is designed in way to make it possible to guarantee
node will be visited only after all its predecessors
have been visited:

       a
      / \
     b   c
     |   |
     d   |
     \  /
       e

There with accordingly provided functors (NumEntries) it's
guaranteed node "e" will be visited after "d" and "c".
Such a property is important for nodes depth evaluation.

* [IE][VPU][nGraph]: Fixes printTo for nGraph type

For some reason if printTo for nGraph type is
usual function it's not picked up by VPU_THROW_UNLESS
triggered inside DynamicToStaticShape transformations.

Making it template specialization does the job.

* [IE][VPU]: Introduces SliceConfiguration class

SliceConfiguration is a class that's intended
to express the result of operation slicing by
batch. The result of slicing is configuration
that specifies what to do with each data object
associated with operation. There are two options
defined: Slice and Unchanged. Typical slice
scenario is Slice, when operation has the same
batch for all inputs and outputs, so all
corresponding data object will be "sliced"
(replaced with copy where batch equal to 1).

At some cases, data object should not sliced
(ex. if operation has constant input which
is the same for all input data batches and
so, does not have batch - Add of 2 tensors
with shapes [10, 1000] and [1000]). To
represent such cases there is option
"Unchanged".

At cases when operation should not be sliced
at all (ex. does not have batch, have different
batch for inputs and outputs, has static
batch and so on) SliceConfiguration object will
return false for "hasSlice" method call. In
these cases inputs and outputs methods calls
will throw an exception.

* [IE][VPU][nGraph]: Enables MatMul operation slice

In case of static batch, operation is not going to be sliced,
since for handling such cases other transformation is used.
Such approach allows both passes to co-exist while one is
being replaced with another.

If data input has other dynamic dimension than batch error
will be thrown since Myriad-X plugin does not support
convolutions (HW accelerated operations) with dynamism in
spatial dimensions.

* [IE][VPU][nGraph]: Enables Convolution operations slice

In case of static batch, operation is not going to be sliced,
since for handling such cases other transformation is used.
Such approach allows both passes to co-exist while one is
being replaced with another.

If data input has other dynamic dimension than batch error
will be thrown since Myriad-X plugin does not support
convolutions (HW accelerated operations) with dynamism in
spatial dimensions.

* [IE][VPU][nGraph]: Enables unary eltwise slice

Since extract dynamic batch transformation will handle
dynamism only by batch (so requires body loop to be static)
operations with dynamism in dimension other than batch should
not be covered by loop.

In case of dynamism in dimension other than batch eltwise
will be considered unsupported for sub-graph extraction.

* [IE][VPU][nGraph]: Enables binary eltwise slice

Since extract dynamic batch transformation will handle
dynamism only by batch (so requires body loop to be static)
operations with dynamism in dimension other than batch should
not be covered by loop.

In case of dynamism in dimension other than batch eltwise
will be considered unsupported for sub-graph extraction.

It's template function since different binary eltwise
operations have the same broadcasting rules.

* [IE][VPU][nGraph]: Enables extract dynamic batch transformation

General approach is following:

1. Extracted sub-graphs should have exactly one input and output
   operation. Otherwise, it's possible that memory consumption of
   model will be increased since loops implementation on Myriad-X
   requires to keep all inputs and outputs of loop to be alive
   along with memory used by loop body. In layout consolidation
   scenario it reflects intention to use minimized amount of
   permutations.

2. Extracted sub-graph should not have external connections (
   the only nodes that allowed to have predecessor or successor
   outside of sub-graph are input and output). Otherwise, it's
   possible that memory consumption of model will be increased
   for the same reason as in previous point.

   To make sure this restriction is met transformation looks
   for leaves in both directions, finds corresponding LCA
   (Lowest Common Ancestor) and checks if such sub-graph has
   external connections. If so, it repeats leaves search
   procedure stopping if it approaches leaves from previous
   iteration and finds LCA again. It is repeated until
   sub-graph without external connections is found (it exists,
   at least source itself forms it).

   Leaf in current context is a node which satisfies one of
   the following conditions (depending on direction):
     Top:
       1. It has no predecessors which are neither Parameter,
          nor Constant
       2. It's unknown how to slice this operation
       3. It could not be sliced (different batch for inputs and
          outputs)
     Bottom:
       1. It has no successors which are not Result
       2. It's unknown how to slice this operation
       3. It could not be sliced (different batch for inputs and
          outputs)

Signed-off-by: Gladilov, Gleb <gleb.gladilov@intel.com>

inference-engine/src/vpu/common/include/vpu/ngraph/transformations/extract_dynamic_batch/batch_extraction_configuration.hpp

@@ -0,0 +1,31 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <vector>
+
+namespace vpu {
+
+enum class SliceMode {
+    Slice,
+    Unchanged
+};
+
+class SliceConfiguration {
+public:
+    SliceConfiguration() = default;
+    SliceConfiguration(std::vector<SliceMode> inputs, std::vector<SliceMode> outputs);
+
+    bool isSliceSupported() const;
+    const std::vector<SliceMode>& inputs() const;
+    const std::vector<SliceMode>& outputs() const;
+
+private:
+    bool m_isSliceSupported = false;
+    std::vector<SliceMode> m_inputs;
+    std::vector<SliceMode> m_outputs;
+};
+
+}  // namespace vpu

inference-engine/src/vpu/common/include/vpu/ngraph/transformations/extract_dynamic_batch/extract_dynamic_batch.hpp

@@ -0,0 +1,24 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "ngraph/pass/graph_rewrite.hpp"
+
+#include <memory>
+
+namespace vpu {
+
+class ExtractBatch: public ngraph::pass::FunctionPass {
+public:
+    NGRAPH_RTTI_DECLARATION;
+
+    explicit ExtractBatch(std::unordered_set<ngraph::Node::type_info_t> targets);
+    bool run_on_function(std::shared_ptr<ngraph::Function> function) override;
+
+private:
+    std::unordered_set<ngraph::Node::type_info_t> targets;
+};
+
+}  // namespace vpu

inference-engine/src/vpu/common/include/vpu/ngraph/transformations/extract_dynamic_batch/slice_binary_eltwise.hpp

@@ -0,0 +1,14 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "ngraph/ngraph.hpp"
+#include "batch_extraction_configuration.hpp"
+
+namespace vpu {
+
+SliceConfiguration sliceBinaryEltwise(const ngraph::Node& node);
+
+}  // namespace vpu

inference-engine/src/vpu/common/include/vpu/ngraph/transformations/extract_dynamic_batch/slice_convolution.hpp

@@ -0,0 +1,14 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "ngraph/ngraph.hpp"
+#include "batch_extraction_configuration.hpp"
+
+namespace vpu {
+
+SliceConfiguration sliceConvolution(const ngraph::Node& node);
+
+}  // namespace vpu

inference-engine/src/vpu/common/include/vpu/ngraph/transformations/extract_dynamic_batch/slice_mat_mul.hpp

@@ -0,0 +1,14 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "ngraph/ngraph.hpp"
+#include "batch_extraction_configuration.hpp"
+
+namespace vpu {
+
+SliceConfiguration sliceMatMul(const ngraph::Node& node);
+
+}  // namespace vpu

inference-engine/src/vpu/common/include/vpu/ngraph/transformations/extract_dynamic_batch/slice_unary_eltwise.hpp

@@ -0,0 +1,14 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "ngraph/ngraph.hpp"
+#include "batch_extraction_configuration.hpp"
+
+namespace vpu {
+
+SliceConfiguration sliceUnaryEltwise(const ngraph::Node& node);
+
+}  // namespace vpu

inference-engine/src/vpu/common/include/vpu/ngraph/utilities.hpp

@@ -7,6 +7,11 @@
 #include "ngraph/node.hpp"
 #include "ngraph/type/element_type.hpp"
 
+#include "vpu/utils/error.hpp"
+
+#include <stack>
+#include <deque>
+
 namespace vpu {
 
 std::vector<std::int64_t> evaluateTargetShape(const ngraph::Output<ngraph::Node>& value);
@@ -15,6 +20,60 @@ std::shared_ptr<ngraph::Node> shapeToConstant(const ngraph::element::Type& type,
 
 std::shared_ptr<ngraph::Node> gatherShapeElements(const ngraph::Output<ngraph::Node>&, int startIndex, size_t elemCount);
 
-void printTo(std::ostream& stream, const ngraph::NodeTypeInfo& object);
+template<>
+inline void printTo(std::ostream& stream, const ngraph::NodeTypeInfo& object) {
+    stream << object.name << " ver. " << object.version;
+}
+
+using Nodes = std::unordered_set<ngraph::Node*>;
+
+template<class GetNext, class Visit>
+Nodes dfs(ngraph::Node* root, GetNext&& getNext, Visit&& visit) {
+    Nodes visited;
+    std::stack<ngraph::Node*> stack{{root}};
+    while (!stack.empty()) {
+        const auto current = stack.top();
+        stack.pop();
+
+        if (!visited.emplace(current).second) {
+            continue;
+        }
+
+        if (!visit(current)) {
+            continue;
+        }
+
+        for (const auto& next : getNext(current)) {
+            stack.push(next);
+        }
+    }
+    return visited;
+}
+
+template<class NumEntries, class Visit, class MoveForward>
+void bfs(ngraph::Node* root, NumEntries&& getNumEntries, Visit&& visit, MoveForward&& moveForward) {
+    std::deque<ngraph::Node*> deque{root};
+    std::unordered_map<ngraph::Node*, std::size_t> visits;
+    while (!deque.empty()) {
+        const auto current = deque.front();
+        deque.pop_front();
+
+        const auto numEntries = current == root ? 1 : getNumEntries(current);
+
+        const auto visitsCount = ++visits[current];
+        VPU_THROW_UNLESS(visitsCount <= numEntries, "Encountered loop at {}", current);
+
+        if (visitsCount < numEntries) {
+            VPU_THROW_UNLESS(!deque.empty(), "Node {} should be visited only after all predecessors, but it is not available through all of them", current);
+            continue;
+        }
+
+        if (!visit(current)) {
+            continue;
+        }
+
+        moveForward(deque, current);
+    }
+}
 
 }  // namespace vpu

inference-engine/src/vpu/common/src/ngraph/transformations/extract_dynamic_batch/batch_extraction_configuration.cpp

@@ -0,0 +1,30 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "vpu/utils/error.hpp"
+#include "vpu/ngraph/transformations/extract_dynamic_batch/batch_extraction_configuration.hpp"
+
+namespace vpu {
+
+SliceConfiguration::SliceConfiguration(std::vector<SliceMode> inputs, std::vector<SliceMode> outputs)
+   : m_isSliceSupported(true)
+   , m_inputs(std::move(inputs))
+   , m_outputs(std::move(outputs)) {}
+
+bool SliceConfiguration::isSliceSupported() const {
+    return m_isSliceSupported;
+}
+
+const std::vector<SliceMode>& SliceConfiguration::inputs() const {
+    VPU_THROW_UNLESS(m_isSliceSupported, "Encountered an attempt to access inputs slice configuration for a case when slice is unsupported");
+    return m_inputs;
+}
+
+const std::vector<SliceMode>& SliceConfiguration::outputs() const {
+    VPU_THROW_UNLESS(m_isSliceSupported, "Encountered an attempt to access outputs slice configuration for a case when slice is unsupported");
+    return m_outputs;
+}
+
+}  // namespace vpu
+

inference-engine/src/vpu/common/src/ngraph/transformations/extract_dynamic_batch/extract_dynamic_batch.cpp

@@ -0,0 +1,487 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "ngraph/opsets/opset5.hpp"
+
+#include "vpu/utils/optional.hpp"
+#include "vpu/ngraph/utilities.hpp"
+
+#include "vpu/ngraph/transformations/extract_dynamic_batch/extract_dynamic_batch.hpp"
+#include "vpu/ngraph/transformations/extract_dynamic_batch/slice_mat_mul.hpp"
+#include "vpu/ngraph/transformations/extract_dynamic_batch/slice_convolution.hpp"
+#include "vpu/ngraph/transformations/extract_dynamic_batch/slice_binary_eltwise.hpp"
+#include "vpu/ngraph/transformations/extract_dynamic_batch/slice_unary_eltwise.hpp"
+
+#include <queue>
+
+namespace vpu {
+
+NGRAPH_RTTI_DEFINITION(vpu::ExtractBatch, "ExtractBatch", 0);
+
+ExtractBatch::ExtractBatch(std::unordered_set<ngraph::Node::type_info_t> targets) : targets(std::move(targets)) {}
+
+namespace {
+
+class Slicers {
+public:
+    static bool isSupported(const ngraph::Node& node) {
+        return getSlicers().count(node.get_type_info());
+    }
+
+    static SliceConfiguration slice(const ngraph::Node& node) {
+        const auto& slicers = getSlicers();
+        const auto& type = node.get_type_info();
+        return slicers.count(type) ? slicers.at(type)(node) : SliceConfiguration{};
+    }
+
+private:
+    using Functor = std::function<SliceConfiguration(const ngraph::Node&)>;
+    static const std::unordered_map<ngraph::DiscreteTypeInfo, Functor>& getSlicers() {
+        static const std::unordered_map<ngraph::DiscreteTypeInfo, Functor>& slicers = {
+            {ngraph::opset5::MatMul::type_info,                  sliceMatMul},
+            {ngraph::opset5::Convolution::type_info,             sliceConvolution},
+            {ngraph::opset5::GroupConvolution::type_info,        sliceConvolution},
+            {ngraph::opset5::ConvolutionBackpropData::type_info, sliceConvolution},
+
+            {ngraph::opset5::Add::type_info,              sliceBinaryEltwise},
+            {ngraph::opset5::Multiply::type_info,         sliceBinaryEltwise},
+            {ngraph::opset5::Minimum::type_info,          sliceBinaryEltwise},
+            {ngraph::opset5::Maximum::type_info,          sliceBinaryEltwise},
+
+            {ngraph::opset5::Relu::type_info,             sliceUnaryEltwise},
+        };
+        return slicers;
+    }
+};
+
+struct SubGraph {
+    Nodes leaves;
+    Nodes all;
+};
+
+template<class Functor>
+Nodes getNodes(ngraph::Node* from, ngraph::Node* to, Functor&& getNext) {
+    auto visited = dfs(from, std::forward<Functor>(getNext), [to](ngraph::Node* node) { return node != to; });
+    visited.erase(from);
+    return visited;
+}
+
+template<class Functor>
+SubGraph getLeaves(ngraph::Node* source, const Nodes& blackList, Functor&& getNext) {
+    const auto isOk = [&blackList](ngraph::Node* node) { return Slicers::slice(*node).isSliceSupported() && !blackList.count(node); };
+    Nodes leaves;
+    auto visited = dfs(source, std::forward<Functor>(getNext), [isOk, getNext, &leaves](ngraph::Node* node) {
+        const auto& nextNodes = getNext(node);
+        const auto exit = nextNodes.empty() || std::any_of(nextNodes.cbegin(), nextNodes.cend(), [isOk](ngraph::Node* node) { return !isOk(node); });
+        if (exit) {
+            leaves.emplace(node);
+            return false;
+        }
+        return true;
+    });
+    visited.erase(source);
+    return {leaves, visited};
+}
+
+template<class NextForward, class NextBackward>
+void getLeavesLCA(ngraph::Node* source, ngraph::Node*& lca, Nodes& nodes, const Nodes& leaves, const Nodes& allBackward,
+                  NextForward&& getNextForward, NextBackward&& getNextBackward) {
+    std::unordered_map<ngraph::Node*, std::size_t> depths{{ source, 0}}, leavesDepths;
+    const auto less = [&depths](ngraph::Node* lhs, ngraph::Node* rhs) {
+        VPU_THROW_UNLESS(depths.count(lhs), "There is no {} in all depth", lhs);
+        VPU_THROW_UNLESS(depths.count(rhs), "There is no {} in all depth", rhs);
+        return depths.at(lhs) < depths.at(rhs);
+    };
+
+    const auto equal = [&depths](ngraph::Node* lhs, ngraph::Node* rhs) {
+        VPU_THROW_UNLESS(depths.count(lhs), "There is no {} in all depth", lhs);
+        VPU_THROW_UNLESS(depths.count(rhs), "There is no {} in all depth", rhs);
+        return depths.at(lhs) == depths.at(rhs);
+    };
+
+    Nodes visited;
+    if (leaves.size() == 1 && leaves.count(source)) {
+        lca = source;
+        nodes = visited;
+        return;
+    }
+
+    Nodes prevNodes;
+    bfs(
+        source,
+        [getNextBackward, &allBackward, &prevNodes](const ngraph::Node* current) {
+            prevNodes = getNextBackward(current);
+            for (auto it = prevNodes.begin(); it != prevNodes.end();) {
+                it = allBackward.count(*it) ? prevNodes.erase(it) : std::next(it);
+            }
+            return prevNodes.size();
+        },
+        [&](ngraph::Node* current) {
+            if (current == source) {
+                return true;
+            }
+            const auto depth = depths.at(*std::max_element(prevNodes.cbegin(), prevNodes.cend(), less)) + 1;
+            depths[current] = depth;
+
+            if (leaves.count(current)) {
+                leavesDepths[current] = depth;
+                return false;
+            }
+            return true;
+        },
+        [getNextForward](std::deque<ngraph::Node*>& deque, const ngraph::Node* current) {
+            const auto& nextNodes = getNextForward(current);
+            std::copy(nextNodes.cbegin(), nextNodes.cend(), std::back_inserter(deque));
+        });
+
+    VPU_THROW_UNLESS(leavesDepths.size() == leaves.size(), "leavesDepths and leaves have different sizes: {} vs {}", leavesDepths.size(), leaves.size());
+
+    auto lcaCandidates = leaves;
+    const auto minDepthArg = std::min_element(lcaCandidates.cbegin(), lcaCandidates.cend(), less);
+    while (!std::all_of(lcaCandidates.cbegin(), lcaCandidates.cend(), [equal, minDepthArg](ngraph::Node* end) { return equal(end, *minDepthArg); })) {
+        std::unordered_map<ngraph::Node*, ngraph::Node*> updates;
+        for (const auto& end : lcaCandidates) {
+            auto current = end;
+            while (!equal(current, *minDepthArg)) {
+                const auto& nextNodes = getNextBackward(current);
+                current = *std::max_element(nextNodes.cbegin(), nextNodes.cend(), less);
+            }
+
+            updates[end] = current;
+        }
+
+        for (const auto& update : updates) {
+            lcaCandidates.erase(update.first);
+            lcaCandidates.emplace(update.second);
+        }
+    }
+
+    while (lcaCandidates.size() != 1) {
+        std::unordered_map<ngraph::Node*, ngraph::Node*> updates;
+        for (const auto& end : lcaCandidates) {
+            const auto& nextNodes = getNextBackward(end);
+            const auto next = *std::max_element(nextNodes.cbegin(), nextNodes.cend(), less);
+
+            updates[end] = next;
+        }
+
+        for (const auto& update : updates) {
+            lcaCandidates.erase(update.first);
+            lcaCandidates.emplace(update.second);
+        }
+    }
+
+    lca = *lcaCandidates.begin();
+    nodes = getNodes(source, lca, getNextForward);
+}
+
+template<class Functor>
+std::shared_ptr<ngraph::opset5::Loop> makeLoop(ngraph::Node* root, ngraph::Node* leaf, Functor&& getNextTop) {
+    ngraph::ParameterVector parameters;
+    ngraph::ResultVector results;
+    std::unordered_map<std::shared_ptr<ngraph::opset5::Parameter>, ngraph::Output<ngraph::Node>> slicedInputs, invariantInputs;
+    std::set<ngraph::Output<ngraph::Node>> concatenatedResults;
+    std::set<ngraph::Output<ngraph::Node>> iterValueResults;
+
+    std::map<ngraph::Output<ngraph::Node>, ngraph::Output<ngraph::Node>> nodes;
+    const auto getInput = [&nodes, &parameters, &slicedInputs, &invariantInputs](const ngraph::Output<ngraph::Node>& currentInput) {
+        if (nodes.count(currentInput)) {
+            return nodes.at(currentInput);
+        } else {
+            const auto& currentInputNode = currentInput.get_node();
+            VPU_THROW_UNLESS(ngraph::op::is_constant(currentInputNode) || ngraph::op::is_parameter(currentInputNode),
+                "Encountered intermediate node {} which is not cloned yet", currentInputNode);
+
+            // assume if constant has several consumers all of them requires either Slice or Unchanged
+            const auto& targetInputs = currentInput.get_target_inputs();
+            const auto adjacentDiff = std::adjacent_find(targetInputs.cbegin(), targetInputs.cend(),
+                [](const ngraph::Input<ngraph::Node>& lhs, const ngraph::Input<ngraph::Node>& rhs) {
+                    const auto& lhsNode = lhs.get_node();
+                    const auto& rhsNode = rhs.get_node();
+
+                    const auto& lhsSplitConfig = Slicers::slice(*lhsNode);
+                    const auto& rhsSplitConfig = Slicers::slice(*rhsNode);
+                    if (!lhsSplitConfig.isSliceSupported() || !rhsSplitConfig.isSliceSupported()) {
+                        return true;
+                    }
+
+                    const auto& lhsInputSplitConfig = lhsSplitConfig.inputs();
+                    const auto& rhsInputSplitConfig = rhsSplitConfig.inputs();
+                    return lhsInputSplitConfig[lhs.get_index()] != rhsInputSplitConfig[rhs.get_index()];
+                });
+
+            VPU_THROW_UNLESS(adjacentDiff == targetInputs.cend(),
+                "Encountered constant {} that has 2 consumers ({} and {}) with different split configurations",
+                currentInput, adjacentDiff->get_node(), std::next(adjacentDiff)->get_node());
+
+            const auto& targetInput = targetInputs.begin();
+            const auto& node = targetInput->get_node();
+            const auto& index = targetInput->get_index();
+            const auto splitInputConfiguration = Slicers::slice(*node).inputs();
+
+            if (splitInputConfiguration[index] == SliceMode::Slice) {
+                auto partialShape = currentInput.get_partial_shape();
+                partialShape[0] = 1;
+                auto parameter = std::make_shared<ngraph::opset5::Parameter>(currentInput.get_element_type(), partialShape);
+                parameters.emplace_back(parameter);
+                slicedInputs[parameter] = currentInput;
+
+                nodes[currentInput] = parameter;
+                return static_cast<ngraph::Output<ngraph::Node>>(parameter);
+            } else {
+                auto argument = currentInput;
+                if (ngraph::op::is_parameter(currentInputNode)) {
+                    auto parameter = std::make_shared<ngraph::opset5::Parameter>(currentInput.get_element_type(), currentInput.get_partial_shape());
+                    parameters.emplace_back(parameter);
+                    invariantInputs[parameter] = currentInput;
+
+                    argument = parameter;
+                }
+
+                nodes[currentInput] = argument;
+                return argument;
+            }
+        }
+    };
+    const auto clone = [getInput](const ngraph::Node* source) {
+        std::vector<ngraph::Output<ngraph::Node>> newInputs;
+        newInputs.reserve(source->get_input_size());
+        const auto& currentInputs = source->input_values();
+        std::transform(currentInputs.cbegin(), currentInputs.cend(), std::back_inserter(newInputs), getInput);
+
+        auto cloned = source->copy_with_new_inputs(newInputs);
+        cloned->set_friendly_name(source->get_friendly_name());
+        VPU_THROW_UNLESS(cloned->get_output_size() == source->get_output_size(),
+                         "Encountered mismatch in output count between original node {} and copy without batch {}", source, cloned);
+        return cloned;
+    };
+
+    const auto splitInputConfiguration = Slicers::slice(*root).inputs();
+    for (std::size_t i = 0; i < root->get_input_size(); ++i) {
+        const auto& input = root->input_value(i);
+        ngraph::Output<ngraph::Node> argument;
+        if (splitInputConfiguration[i] == SliceMode::Slice) {
+            auto partialShape = input.get_partial_shape();
+            partialShape[0] = 1;
+
+            auto parameter = std::make_shared<ngraph::opset5::Parameter>(input.get_element_type(), partialShape);
+            parameters.emplace_back(parameter);
+            slicedInputs[parameter] = input;
+
+            argument = parameter;
+        } else if (!ngraph::op::is_constant(input.get_node())) {
+            auto parameter = std::make_shared<ngraph::opset5::Parameter>(input.get_element_type(), input.get_partial_shape());
+            parameters.emplace_back(parameter);
+            invariantInputs[parameter] = input;
+
+            argument = parameter;
+        } else {
+            argument = input;
+        }
+
+        nodes[input] = argument;
+    }
+
+    std::shared_ptr<ngraph::Node> bodyNode;
+    bfs(
+        root,
+        [getNextTop](const ngraph::Node* current) {
+            return getNextTop(current).size();
+        },
+        [leaf, clone, &bodyNode](const ngraph::Node* current) {
+            bodyNode = clone(current);
+            return current != leaf;
+        },
+        [&](std::deque<ngraph::Node*>& deque, ngraph::Node* current) {
+            for (std::size_t i = 0; i < current->get_output_size(); ++i) {
+                const auto& currentOutput = current->output(i);
+                const auto& bodyOutput = bodyNode->output(i);
+                const auto& currentOutputNode = currentOutput.get_node();
+                if (ngraph::op::is_output(currentOutputNode)) {
+                    const auto splitOutputConfiguration = Slicers::slice(*current).outputs();
+                    auto& outputCategory = splitOutputConfiguration[i] == SliceMode::Slice ? concatenatedResults : iterValueResults;
+                    outputCategory.emplace(bodyOutput);
+                    results.emplace_back(std::make_shared<ngraph::opset5::Result>(bodyOutput));
+                } else {
+                    nodes[currentOutput] = bodyOutput;
+                    const auto& consumers = current->get_output_target_inputs(i);
+                    std::transform(consumers.cbegin(), consumers.cend(), std::back_inserter(deque),
+                        [](const ngraph::Input<ngraph::Node>& consumer) { return consumer.get_node(); });
+                }
+            }
+        });
+
+    const auto splitOutputConfiguration = Slicers::slice(*leaf).outputs();
+    for (std::size_t i = 0; i < bodyNode->get_output_size(); ++i) {
+        const auto& output = bodyNode->output(i);
+        auto result = std::make_shared<ngraph::opset5::Result>(output);
+        auto& outputCategory = splitOutputConfiguration[i] == SliceMode::Slice ? concatenatedResults : iterValueResults;
+        outputCategory.emplace(output);
+        results.emplace_back(result);
+    }
+
+    VPU_THROW_UNLESS(!slicedInputs.empty(), "Failed to find sliced inputs for loop in extract batch");
+    const auto& slicedInput = slicedInputs.begin()->second;
+    const auto shapeOf = std::make_shared<ngraph::opset5::ShapeOf>(slicedInput);
+
+    // constant's shape has to be scalar (not empty) since if this constant has empty shape, so Gather will
+    // have empty shape as well (Gather produces scalar). When this Gather will become ScatterElementsUpdate
+    // argument ScatterElementsUpdate shape inference function will fail, since it requires indices and updates
+    // to have exactly the same shape (indices rank must be the same as rank of data input which is 1D vector,
+    // so its rank = 1 != 0)
+    const auto constant = std::make_shared<ngraph::opset5::Constant>(ngraph::element::i64, ngraph::Shape{1}, 0);
+
+    // TODO: check all other sliced inputs have the same batch?
+    const auto batchSize = std::make_shared<ngraph::opset5::Gather>(shapeOf, constant, constant);
+
+    const auto executionCondition = std::make_shared<ngraph::opset5::Constant>(ngraph::element::boolean, ngraph::Shape{}, true);
+    auto loop = std::make_shared<ngraph::opset5::Loop>(batchSize, executionCondition);
+
+    const auto iterationCondition = std::make_shared<ngraph::opset5::Constant>(ngraph::element::boolean, ngraph::Shape{}, true);
+    results.emplace_back(std::make_shared<ngraph::opset5::Result>(iterationCondition));
+    auto body = std::make_shared<ngraph::Function>(results, parameters, "body");
+    loop->set_function(body);
+    for (const auto& entry : slicedInputs) {
+        loop->set_sliced_input(entry.first, entry.second, 0, 1, 1, -1, 0);
+    }
+
+    for (const auto& entry : invariantInputs) {
+        loop->set_invariant_input(entry.first, entry.second);
+    }
+
+    for (const auto& entry : iterValueResults) {
+        loop->get_iter_value(entry, -1);
+    }
+
+    for (const auto& entry : concatenatedResults) {
+        loop->get_concatenated_slices(entry, 0, 1, 1, -1, 0);
+    }
+
+    loop->set_special_body_ports({-1, static_cast<std::int64_t>(results.size()) - 1});
+    loop->validate_and_infer_types();
+    return loop;
+}
+
+template<class Functor>
+bool updateExternals(const ngraph::Node* source, const Nodes& allForward, const Nodes& allBackward, Nodes& externals, Functor&& getNextBackward) {
+    bool updated = false;
+    for (const auto& node : allForward) {
+        const auto& nextNodes = getNextBackward(node);
+        const auto hasExternalConnection = std::any_of(nextNodes.cbegin(), nextNodes.cend(), [source, &allForward, &allBackward](ngraph::Node* next) {
+            return !allForward.count(next) && !allBackward.count(next) && next != source;
+        });
+        if (hasExternalConnection) {
+            externals.emplace(node);
+            updated = true;
+        }
+    }
+    return updated;
+}
+
+}  // namespace
+
+bool ExtractBatch::run_on_function(std::shared_ptr<ngraph::Function> functionPointer) {
+    auto& function = *functionPointer;
+    bool changed = false;
+
+    Nodes sources;
+    for (const auto& operation : function.get_ordered_ops()) {
+        if (targets.count(operation->get_type_info())) {
+            sources.emplace(operation.get());
+        }
+    }
+
+    auto getNextTop = [](const ngraph::Node* node) {
+        Nodes nextNodes;
+        for (std::size_t i = 0; i < node->get_input_size(); ++i) {
+            const auto next = node->get_input_source_output(i).get_node();
+            if (ngraph::op::is_constant(next) || ngraph::op::is_parameter(next)) {
+                continue;
+            }
+            nextNodes.emplace(next);
+        }
+        return nextNodes;
+    };
+
+    auto getNextBottom = [](const ngraph::Node* node) {
+        Nodes nextNodes;
+        for (std::size_t i = 0; i < node->get_output_size(); ++i) {
+            const auto consumers = node->get_output_target_inputs(i);
+            for (const auto consumer : consumers) {
+                const auto next = consumer.get_node();
+                if (ngraph::op::is_output(next)) {
+                    continue;
+                }
+                nextNodes.insert(next);
+            }
+        }
+        return nextNodes;
+    };
+
+    for (auto currentSource = sources.begin(); currentSource != sources.end(); currentSource = sources.erase(currentSource)) {
+        const auto& source = *currentSource;
+
+        VPU_THROW_UNLESS(Slicers::isSupported(*source),
+            "{} was requested as target operation type for batch extraction, but functor for this type is not provided.", source->get_type_info());
+
+        if (!Slicers::slice(*source).isSliceSupported()) {
+            continue;
+        }
+
+        Nodes topExternals, bottomExternals;
+
+        auto topSubGraph = getLeaves(source, topExternals, getNextTop);
+        auto bottomSubGraph = getLeaves(source, bottomExternals, getNextBottom);
+
+        auto hasNewTopExternals = updateExternals(source, topSubGraph.all, bottomSubGraph.all, topExternals, getNextBottom);
+        if (hasNewTopExternals) {
+            topSubGraph = getLeaves(source, topExternals, getNextTop);
+        }
+
+        bool hasNewBottomExternals = updateExternals(source, bottomSubGraph.all, topSubGraph.all, bottomExternals, getNextTop);
+        if (hasNewBottomExternals) {
+            bottomSubGraph = getLeaves(source, bottomExternals, getNextBottom);
+        }
+
+        ngraph::Node* top = nullptr;
+        ngraph::Node* bottom = nullptr;
+        do {
+            getLeavesLCA(source, top, topSubGraph.all, topSubGraph.leaves, bottomSubGraph.all, getNextTop, getNextBottom);
+            getLeavesLCA(source, bottom, bottomSubGraph.all, bottomSubGraph.leaves, topSubGraph.all, getNextBottom, getNextTop);
+
+            hasNewTopExternals = updateExternals(source, topSubGraph.all, bottomSubGraph.all, topExternals, getNextBottom);
+            if (hasNewTopExternals) {
+                topSubGraph = getLeaves(source, topExternals, getNextTop);
+            }
+
+            hasNewBottomExternals = updateExternals(source, bottomSubGraph.all, topSubGraph.all, bottomExternals, getNextTop);
+            if (hasNewBottomExternals) {
+                bottomSubGraph = getLeaves(source, bottomExternals, getNextBottom);
+            }
+        } while (hasNewTopExternals || hasNewBottomExternals);
+
+        for (const auto& node : topSubGraph.all) {
+            if (sources.count(node)) {
+                sources.erase(node);
+            }
+        }
+
+        for (const auto& node : bottomSubGraph.all) {
+            if (sources.count(node)) {
+                sources.erase(node);
+            }
+        }
+
+        auto loop = makeLoop(top, bottom, getNextTop);
+        auto bottomNode = bottom->shared_from_this();
+        loop->set_friendly_name(bottomNode->get_friendly_name());
+        ngraph::replace_node(bottomNode, loop);
+        function.validate_nodes_and_infer_types();
+        changed = true;
+    }
+
+    return changed;
+}
+
+}  // namespace vpu

inference-engine/src/vpu/common/src/ngraph/transformations/extract_dynamic_batch/slice_binary_eltwise.cpp

@@ -0,0 +1,79 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "vpu/utils/error.hpp"
+#include "vpu/ngraph/transformations/extract_dynamic_batch/slice_binary_eltwise.hpp"
+
+namespace vpu {
+
+SliceConfiguration sliceBinaryEltwise(const ngraph::Node& node) {
+    const auto& eltwise = dynamic_cast<const ngraph::op::util::BinaryElementwiseArithmetic&>(node);
+    VPU_THROW_UNLESS(eltwise.get_input_size() == 2,  "Expecting operation {} to have {} inputs, got {}", node, 2, eltwise.get_input_size());
+    VPU_THROW_UNLESS(eltwise.get_output_size() == 1, "Expecting operation {} to have {} outputs, got {}", node, 1, eltwise.get_output_size());
+
+    const auto& lhs = eltwise.input_value(0);
+    const auto& rhs = eltwise.input_value(1);
+    const auto& out = eltwise.output(0);
+
+    const auto& lhsPartialShape = lhs.get_partial_shape();
+    const auto& rhsPartialShape = rhs.get_partial_shape();
+    const auto& outPartialShape = out.get_partial_shape();
+
+    const auto& broadcastSpec = eltwise.get_autob();
+    auto inputPartialShape = lhsPartialShape;
+    if (broadcastSpec == ngraph::op::AutoBroadcastSpec::NONE) {
+        ngraph::PartialShape::merge_into(inputPartialShape, rhsPartialShape);
+    } else {
+        ngraph::PartialShape::broadcast_merge_into(inputPartialShape, rhsPartialShape, broadcastSpec);
+    }
+
+    const auto& inputRank = inputPartialShape.rank();
+    const auto& lhsRank   = lhsPartialShape.rank();
+    const auto& rhsRank   = rhsPartialShape.rank();
+    const auto& outRank   = outPartialShape.rank();
+
+    VPU_THROW_UNLESS(inputRank == outRank && inputRank.is_static(),
+                     "Expecting operation {} to have the same static rank for inputs and output, got merged inputs rank = {}, output rank = {}",
+                     node, inputRank, outRank);
+
+    const auto& inputRankLength = inputRank.get_length();
+    const auto& lhsRankLength   = lhsRank.get_length();
+    const auto& rhsRankLength   = rhsRank.get_length();
+    const auto& outRankLength   = outRank.get_length();
+
+    const auto& inputsBatch = inputRankLength > 0 ? inputPartialShape[0] : 0;
+    const auto& outBatch = outRankLength > 0 ? outPartialShape[0] : 0;
+    VPU_THROW_UNLESS(inputsBatch == outBatch,
+                     "Expecting operation {} to have the same batch on both inputs and output, got input batch = {}, output batch = {}",
+                     node, inputsBatch, outBatch);
+
+
+    if (inputsBatch.is_static() && inputsBatch.get_length() == 1) {
+        return {};
+    }
+
+    const auto& maxRankInputPartialShape = lhsRankLength == inputRankLength ? lhsPartialShape : rhsPartialShape;
+    const auto& minRankInputPartialShape = lhsRankLength == inputRankLength ? rhsPartialShape : lhsPartialShape;
+
+    const auto checkPartialShape = [](const ngraph::PartialShape& partialShape) {
+        const auto dynamicDimensionsCount = std::count_if(partialShape.cbegin(), partialShape.cend(),
+                                                          [](const ngraph::Dimension& dimension) { return dimension.is_dynamic(); });
+        return dynamicDimensionsCount == 0 || (dynamicDimensionsCount == 1 && partialShape[0].is_dynamic());
+    };
+
+    const auto isMaxRankInputOk = checkPartialShape(maxRankInputPartialShape);
+    const auto isMinRankInputOk = minRankInputPartialShape.rank().get_length() == maxRankInputPartialShape.rank().get_length()
+                                  ? checkPartialShape(minRankInputPartialShape)
+                                  : minRankInputPartialShape.is_static();
+    if (!isMaxRankInputOk || !isMinRankInputOk) {
+        return {};
+    }
+
+    const auto lhsSplitMode = lhsRankLength < inputRankLength || lhsPartialShape[0] != inputPartialShape[0] ? SliceMode::Unchanged : SliceMode::Slice;
+    const auto rhsSplitMode = rhsRankLength < inputRankLength || rhsPartialShape[0] != inputPartialShape[0] ? SliceMode::Unchanged : SliceMode::Slice;
+
+    return {{lhsSplitMode, rhsSplitMode}, {SliceMode::Slice}};
+}
+
+}  // namespace vpu

inference-engine/src/vpu/common/src/ngraph/transformations/extract_dynamic_batch/slice_convolution.cpp

@@ -0,0 +1,37 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "vpu/utils/error.hpp"
+#include "ngraph/opsets/opset5.hpp"
+#include "vpu/ngraph/transformations/extract_dynamic_batch/slice_convolution.hpp"
+
+namespace vpu {
+
+SliceConfiguration sliceConvolution(const ngraph::Node& node) {
+    VPU_THROW_UNLESS(node.get_input_size() == 2,  "Expecting operation {} to have {} inputs, got {}", node, 2, node.get_input_size());
+    VPU_THROW_UNLESS(node.get_output_size() == 1, "Expecting operation {} to have {} outputs, got {}", node, 1, node.get_output_size());
+    VPU_THROW_UNLESS(ngraph::op::is_constant(node.input_value(1).get_node_shared_ptr()), "Expecting operation {} to have constant kernel, got {}",
+        node, node.input_value(1));
+
+    const auto& data = node.input_value(0);
+    const auto& dataPartialShape = data.get_partial_shape();
+    const auto& dataRank = dataPartialShape.rank();
+    VPU_THROW_UNLESS(dataRank.is_static(), "Expecting operation {} to have static rank for input {}, got {}", node, data, dataPartialShape);
+    const auto& dataRankLength = dataRank.get_length();
+    VPU_THROW_UNLESS(dataRankLength >= 3 && dataRankLength <= 5, "Expecting operation {} to have rank of input {} in [{}, {}], got {}",
+        node, data, 3, 5, dataRankLength);
+
+    const auto& batch = dataPartialShape[0];
+    if (batch.is_static()) {
+        return {};
+    }
+
+    if (std::count_if(dataPartialShape.cbegin(), dataPartialShape.cend(), [](const ngraph::Dimension& dimension) { return dimension.is_dynamic(); }) > 1) {
+        return {};
+    }
+
+    return {{SliceMode::Slice, SliceMode::Unchanged}, {SliceMode::Slice}};
+}
+
+}  // namespace vpu

inference-engine/src/vpu/common/src/ngraph/transformations/extract_dynamic_batch/slice_mat_mul.cpp

@@ -0,0 +1,74 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "vpu/utils/error.hpp"
+#include "ngraph/opsets/opset5.hpp"
+#include "vpu/ngraph/transformations/extract_dynamic_batch/slice_mat_mul.hpp"
+
+namespace vpu {
+
+SliceConfiguration sliceMatMul(const ngraph::Node& node) {
+    VPU_THROW_UNLESS(node.get_input_size() == 2,  "Expecting operation {} to have {} inputs, got {}", node, 2, node.get_input_size());
+    VPU_THROW_UNLESS(node.get_output_size() == 1, "Expecting operation {} to have {} outputs, got {}", node, 1, node.get_output_size());
+
+    // target networks have MatMul only with constant second input
+    // there are tests on dynamic MatMul with non-constant second input
+    // if try to process MatMul with non-constant second input it will
+    // affect tests and they will fail, since Loop support is not ready yet
+    if (!ngraph::op::is_constant(node.input_value(1).get_node_shared_ptr())) {
+        return {};
+    }
+
+    const auto& lhs = node.input_value(0);
+    const auto& lhsPartialShape = lhs.get_partial_shape();
+    const auto& lhsRank = lhsPartialShape.rank();
+    VPU_THROW_UNLESS(lhsRank.is_static(), "Expecting operation {} to have static rank for input {}, got {}", node, lhs, lhsPartialShape);
+
+    const auto& rhs = node.input_value(0);
+    const auto& rhsPartialShape = rhs.get_partial_shape();
+    const auto& rhsRank = rhsPartialShape.rank();
+    VPU_THROW_UNLESS(rhsRank.is_static(), "Expecting operation {} to have static rank for input {}, got {}", node, rhs, rhsPartialShape);
+
+    const auto& lhsRankLength = lhsRank.get_length();
+    const auto& rhsRankLength = rhsRank.get_length();
+
+    const auto maxRankLength = std::max(lhsRankLength, rhsRankLength);
+    if (maxRankLength < 3) {
+        return {};
+    }
+
+    const auto isBatchStatic = [](const ngraph::PartialShape& shape) {
+        const auto& rank = shape.rank();
+        if (rank.is_dynamic()) {
+            return false;
+        }
+        const auto rankLength = rank.get_length();
+        if (rankLength < 3) {
+            return true;
+        }
+        return std::all_of(shape.rbegin() + 2, shape.rend(), [](const ngraph::Dimension& dimension) { return dimension.is_static(); });
+    };
+
+    if (maxRankLength > 3) {
+        VPU_THROW_UNLESS(isBatchStatic(lhsPartialShape), "Encountered multi-dimensional dynamic batch for operation {}, but it's unsupported", node);
+        VPU_THROW_UNLESS(isBatchStatic(rhsPartialShape), "Encountered multi-dimensional dynamic batch for operation {}, but it's unsupported", node);
+        return {};
+    }
+
+    if (isBatchStatic(lhsPartialShape) && isBatchStatic(rhsPartialShape)) {
+        return {};
+    }
+
+    if (std::count_if(lhsPartialShape.cbegin(), lhsPartialShape.cend(), [](const ngraph::Dimension& dimension) { return dimension.is_dynamic(); }) > 1 ||
+        std::count_if(rhsPartialShape.cbegin(), rhsPartialShape.cend(), [](const ngraph::Dimension& dimension) { return dimension.is_dynamic(); }) > 1) {
+        return {};
+    }
+
+    const auto& lhsSliceMode = lhsRankLength < 3 ? SliceMode::Unchanged : SliceMode::Slice;
+    const auto& rhsSliceMode = rhsRankLength < 3 ? SliceMode::Unchanged : SliceMode::Slice;
+
+    return {{lhsSliceMode, rhsSliceMode}, {SliceMode::Slice}};
+}
+
+}  // namespace vpu

inference-engine/src/vpu/common/src/ngraph/transformations/extract_dynamic_batch/slice_unary_eltwise.cpp

@@ -0,0 +1,50 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+
+#include "vpu/utils/error.hpp"
+#include "vpu/ngraph/transformations/extract_dynamic_batch/slice_unary_eltwise.hpp"
+
+namespace vpu {
+
+SliceConfiguration sliceUnaryEltwise(const ngraph::Node& node) {
+    VPU_THROW_UNLESS(node.get_input_size() == 1,  "Expecting unary eltwise operation {} to have {} inputs, got {}", node, 1, node.get_input_size());
+    VPU_THROW_UNLESS(node.get_output_size() == 1, "Expecting unary eltwise operation {} to have {} outputs, got {}", node, 1, node.get_output_size());
+
+    const auto& inp = node.input_value(0);
+    const auto& out = node.output(0);
+
+    const auto& inpPartialShape = inp.get_partial_shape();
+    const auto& outPartialShape = out.get_partial_shape();
+
+    const auto& inpRank = inpPartialShape.rank();
+    const auto& outRank = outPartialShape.rank();
+
+    VPU_THROW_UNLESS(inpRank == outRank,
+        "Expecting unary eltwise operation {} to have the same static rank for input and output, got input rank = {}, output rank = {}",
+        node, inpRank, outRank);
+
+    const auto& inpRankLength = inpRank.get_length();
+    const auto& outRankLength = outRank.get_length();
+
+    const auto& inpBatch = inpRankLength > 0 ? inpPartialShape[0] : 0;
+    const auto& outBatch = outRankLength > 0 ? outPartialShape[0] : 0;
+    VPU_THROW_UNLESS(inpBatch == outBatch,
+        "Expecting unary eltwise operation {} to have the same batch on input and output, got input batch = {}, output batch = {}",
+        node, inpBatch, outBatch);
+
+    if (inpBatch.is_static() && inpBatch.get_length() == 1) {
+        return {};
+    }
+
+    const auto dynamicDimensionsCount = std::count_if(inpPartialShape.cbegin(), inpPartialShape.cend(),
+        [](const ngraph::Dimension& dimension) { return dimension.is_dynamic(); });
+    if (dynamicDimensionsCount > 1 || (dynamicDimensionsCount == 1 && inpPartialShape[0].is_static())) {
+        return {};
+    }
+
+    return {{SliceMode::Slice}, {SliceMode::Slice}};
+}
+
+}  // namespace vpu

inference-engine/src/vpu/common/src/ngraph/utilities.cpp

@@ -86,8 +86,4 @@ std::shared_ptr<ngraph::Node> gatherShapeElements(const ngraph::Output<ngraph::N
         ngraph::opset5::Constant::create(ngraph::element::i64, {}, {0}));
 }
 
-void printTo(std::ostream& stream, const ngraph::NodeTypeInfo& object) {
-    stream << object.name << " ver. " << object.version;
-}
-
 }  // namespace vpu

inference-engine/src/vpu/graph_transformer/src/frontend/frontend.cpp

@@ -39,6 +39,7 @@
 #include <vpu/ngraph/operations/dynamic_shape_resolver.hpp>
 #include <legacy/ie_util_internal.hpp>
 #include <legacy/transformations/convert_opset1_to_legacy/convert_gather_to_gather_ie.hpp>
+#include <vpu/ngraph/transformations/extract_dynamic_batch/extract_dynamic_batch.hpp>
 
 namespace vpu {
 
@@ -181,6 +182,13 @@ ie::ICNNNetwork::Ptr FrontEnd::convertNetwork(ie::ICNNNetwork& network) {
     manager.register_pass<ngraph::pass::ConvertNMS3ToNMS5>();
     manager.register_pass<ngraph::pass::ConvertNMS4ToNMS5>();
     manager.register_pass<ngraph::pass::CommonOptimizations>();
+
+    manager.register_pass<vpu::ExtractBatch>(std::unordered_set<ngraph::Node::type_info_t>{
+        ngraph::opset5::MatMul::type_info,
+        ngraph::opset5::Convolution::type_info,
+        ngraph::opset5::GroupConvolution::type_info
+    });
+
     manager.register_pass<vpu::DynamicToStaticShape>();
     manager.register_pass<vpu::EliminateShapeOfAfterDSR>();
     manager.register_pass<vpu::ConvertExtractImagePatchesToReorgYolo>();

ngraph/core/include/ngraph/partial_shape.hpp

@@ -44,7 +44,14 @@ namespace ngraph
     ///     (Informal notation examples: `{1,2,3,4}`, `{6}`, `{}`)
     class NGRAPH_API PartialShape
     {
+        using Dimensions = std::vector<Dimension>;
+
     public:
+        using iterator = Dimensions::iterator;
+        using const_iterator = Dimensions::const_iterator;
+        using reverse_iterator = Dimensions::reverse_iterator;
+        using const_reverse_iterator = Dimensions::const_reverse_iterator;
+
         /// \brief Constructs a shape with static rank from an initializer list of Dimension.
         /// \param init The Dimension values for the constructed shape.
         ///
@@ -223,6 +230,54 @@ namespace ngraph
                                          const PartialShape& src,
                                          const op::AutoBroadcastSpec& autob);
 
+        /// \brief Returns a read/write iterator that points to the first
+        ///        element in the shape. Iteration is done in ordinary
+        ///        element order.
+        iterator begin() noexcept { return m_dimensions.begin(); }
+        /// \brief Returns a read-only (constant) iterator that points to the
+        ///        first element in the shape. Iteration is done in ordinary
+        ///        element order.
+        const_iterator begin() const noexcept { return cbegin(); }
+        /// \brief Returns a read/write iterator that points one past the last
+        ///        element in the shape. Iteration is done in ordinary
+        ///        element order.
+        iterator end() noexcept { return m_dimensions.end(); }
+        /// \brief Returns a read-only (constant) iterator that points one past
+        ///        the last element in the shape. Iteration is done in ordinary
+        ///        element order.
+        const_iterator end() const noexcept { return cend(); }
+        /// \brief Returns a read/write reverse iterator that points to the
+        ///        last element in the shape. Iteration is done in reverse
+        ///        element order.
+        reverse_iterator rbegin() noexcept { return m_dimensions.rbegin(); }
+        /// \brief Returns a read-only (constant) reverse iterator that points
+        ///        to the last element in the shape. Iteration is done in
+        ///        reverse element order.
+        const_reverse_iterator rbegin() const noexcept { return crbegin(); }
+        /// \brief Returns a read/write reverse iterator that points to one
+        ///        before the first element in the shape. Iteration is done
+        ///        in reverse element order.
+        reverse_iterator rend() noexcept { return m_dimensions.rend(); }
+        /// \brief Returns a read-only (constant) reverse iterator that points
+        ///        to one before the first element in the shape. Iteration
+        ///        is done in reverse element order.
+        const_reverse_iterator rend() const noexcept { return crend(); }
+        /// \brief Returns a read-only (constant) iterator that points to the
+        ///        first element in the shape. Iteration is done in ordinary
+        ///        element order.
+        const_iterator cbegin() const noexcept { return m_dimensions.cbegin(); }
+        /// \brief Returns a read-only (constant) iterator that points one past
+        ///        the last element in the shape. Iteration is done in ordinary
+        ///        element order.
+        const_iterator cend() const noexcept { return m_dimensions.cend(); }
+        /// \brief Returns a read-only (constant) reverse iterator that points
+        ///        to the last element in the shape. Iteration is done in
+        ///        reverse element order.
+        const_reverse_iterator crbegin() const noexcept { return m_dimensions.crbegin(); }
+        /// \brief Returns a read-only (constant) reverse iterator that points
+        ///        to one before the first element in the shape. Iteration
+        ///        is done in reverse element order.
+        const_reverse_iterator crend() const noexcept { return m_dimensions.crend(); }
     private:
         // Private constructor for PartialShape::dynamic().
         PartialShape(bool rank_is_static, const std::vector<Dimension>& dimensions);
@@ -250,7 +305,7 @@ namespace ngraph
         } m_shape_type{ShapeType::SHAPE_IS_UNKNOWN};
 
         // Shape dimensions. This has no meaning if m_rank_is_static is false.
-        std::vector<Dimension> m_dimensions;
+        Dimensions m_dimensions;
     };
 
     /// \brief Elementwise addition of two PartialShape objects.