Merge remote-tracking branch 'upstream/master' into debian-packages

cmake/templates/OpenVINOConfig.cmake.in

@@ -146,7 +146,6 @@ endfunction()
 set(_ov_package_prefix_dir "${PACKAGE_PREFIX_DIR}")
 
 set(THREADING "@THREADING@")
-# TODO: find system TBB instead of custom one
 if((THREADING STREQUAL "TBB" OR THREADING STREQUAL "TBB_AUTO") AND NOT TBB_FOUND)
     set(ov_system_tbb "@ENABLE_SYSTEM_TBB@")
     if(ov_system_tbb)

docs/install_guides/installing-openvino-linux.md

@@ -9,7 +9,6 @@
 
   * Ubuntu 18.04.x long-term support (LTS), 64-bit
   * Ubuntu 20.04.x long-term support (LTS), 64-bit
-  * Red Hat Enterprise Linux 8.x, 64-bit
 
   .. note::
      Since the OpenVINO™ 2022.1 release, CentOS 7.6, 64-bit is not longer supported.

docs/install_guides/installing-openvino-macos.md

@@ -66,10 +66,9 @@ This guide provides step-by-step instructions on how to install the Intel® Dist
 
    By default, the Intel® Distribution of OpenVINO™ is installed in the following directory, referred to as `<INSTALL_DIR>` elsewhere in the documentation:
 
-   * For root or administrator: `/opt/intel/openvino_<version>/`
-   * For regular users: `~/intel/openvino_<version>/`
+   `/opt/intel/openvino_<version>/`
 
-   For simplicity, a symbolic link to the latest installation is also created: `/opt/intel/openvino_2022/` or `~/intel/openvino_2022/`.
+   For simplicity, a symbolic link to the latest installation is also created: `/opt/intel/openvino_2022/`.
 
 To check **Release Notes** please visit: [Release Notes](https://software.intel.com/en-us/articles/OpenVINO-RelNotes).
 

docs/install_guides/uninstalling-openvino.md

@@ -66,13 +66,13 @@
    
   .. code-block:: sh
   
-    /home/<user>/intel/openvino/installer/installer
+    /home/<user>/intel/openvino_installer/installer
 
   or in a case of administrative installation:
 
   .. code-block:: sh
 
-    /opt/intel/openvino/installer/installer
+    /opt/intel/openvino_installer/installer
 
   2. Follow the uninstallation wizard instructions.
   
@@ -82,7 +82,7 @@
    
   .. code-block:: sh
   
-    open /opt/intel/openvino/installer/installer.app
+    open /opt/intel/openvino_installer/installer.app
 
   2. Follow the uninstallation wizard instructions.
 

src/cmake/install_tbb.cmake

@@ -50,7 +50,7 @@ if(THREADING MATCHES "^(TBB|TBB_AUTO)$" AND TBBROOT MATCHES ${TEMP})
 
     set(IE_TBB_DIR_INSTALL "3rdparty/tbb/cmake")
     install(FILES "${TBB}/cmake/TBBConfig.cmake"
-                "${TBB}/cmake/TBBConfigVersion.cmake"
+                  "${TBB}/cmake/TBBConfigVersion.cmake"
             DESTINATION runtime/${IE_TBB_DIR_INSTALL}
             COMPONENT tbb_dev)
     install(DIRECTORY "${TBB}/include"

src/common/transformations/src/transformations/common_optimizations/reshape_sequence_fusion.cpp

@@ -37,9 +37,8 @@ bool has_valid_pattern(const ov::Output<ov::Node>& node_out) {
         // Upper bound of the value
         auto ub = ngraph::evaluate_upper_bound(node_out);
         if (!ub) return false;
-        const auto ub_const_node = std::make_shared<ngraph::opset8::Constant>(ub);
-        if (!ub_const_node) return false;
 
+        const auto ub_const_node = std::make_shared<ngraph::opset8::Constant>(ub);
         const auto & ub_values = ub_const_node->cast_vector<int64_t>();
         if (lb_values.size() != ub_values.size()) return false;
 

src/common/transformations/src/transformations/common_optimizations/transpose_reshape_elimination_for_matmul.cpp

@@ -153,10 +153,12 @@ ngraph::pass::TransposeReshapeEliminationForMatmul::TransposeReshapeEliminationF
         // check transpose order before and after matmul
         auto transpose_before = std::dynamic_pointer_cast<opset1::Transpose>(pattern_value_map.at(transpose_before_pattern).get_node_shared_ptr());
         auto transpose_after = std::dynamic_pointer_cast<opset1::Transpose>(pattern_value_map.at(transpose_after_pattern).get_node_shared_ptr());
+        if (!transpose_before || !transpose_after) return false;
+
         auto transpose_before_constant = std::dynamic_pointer_cast<ngraph::opset1::Constant>(transpose_before->get_input_node_shared_ptr(1));
         auto transpose_after_constant = std::dynamic_pointer_cast<ngraph::opset1::Constant>(transpose_after->get_input_node_shared_ptr(1));
-        if (!transpose_before || !transpose_after || !transpose_before_constant || !transpose_after_constant)
-            return false;
+        if (!transpose_before_constant || !transpose_after_constant) return false;
+
         auto transpose_before_order = transpose_before_constant->cast_vector<int64_t>();
         auto transpose_after_order = transpose_after_constant->cast_vector<int64_t>();
         // need to check that input shape is correctly contracted and output shape is correctly unpacked using transposes

src/common/transformations/src/transformations/common_optimizations/transpose_sinking.cpp

@@ -210,10 +210,11 @@ ngraph::pass::TransposeFQReduction::TransposeFQReduction() {
         auto &pattern_to_output = m.get_pattern_value_map();
 
         auto transpose = pattern_to_output.at(transpose_label).get_node_shared_ptr();
+        if (!transpose) return false;
+
         auto transpose_order = std::dynamic_pointer_cast<opset6::Constant>(transpose->get_input_node_shared_ptr(1));
         auto fq = pattern_to_output.at(fq_label).get_node_shared_ptr();
-        if (!transpose || !transpose_order || !fq)
-            return false;
+        if (!transpose_order || !fq) return false;
 
         ngraph::NodeVector new_ops;
 

src/common/transformations/src/transformations/op_conversions/convert_sequences_to_tensor_iterator.cpp

@@ -152,6 +152,8 @@ namespace {
                 gnn_sequence->get_activations_beta(),
                 gnn_sequence->get_clip(),
                 gnn_sequence->get_linear_before_reset());
+        } else {
+            return false;
         }
 
         ngraph::ParameterVector body_params;

src/common/transformations/src/transformations/smart_reshape/strided_slice_squeeze.cpp

@@ -125,11 +125,11 @@ ngraph::pass::SqueezeStridedSlice::SqueezeStridedSlice() {
 
     matcher_pass_callback callback = [](pattern::Matcher& m) -> bool {
         auto slice = std::dynamic_pointer_cast<ngraph::opset5::StridedSlice>(m.get_match_root());
+        if (!slice) return false;
         auto squeeze = slice->get_input_node_shared_ptr(0);
         const auto& const_axes =
             std::dynamic_pointer_cast<ngraph::opset5::Constant>(squeeze->get_input_node_shared_ptr(1));
-        if (!const_axes || !slice)
-            return false;
+        if (!const_axes) return false;
 
         auto begin = std::dynamic_pointer_cast<ngraph::opset5::Constant>(slice->input_value(1).get_node_shared_ptr());
         auto end = std::dynamic_pointer_cast<ngraph::opset5::Constant>(slice->input_value(2).get_node_shared_ptr());

src/frontends/tensorflow/src/op/conv_3d.cpp

@@ -23,8 +23,8 @@ OutputVector translate_conv_3d_op(const NodeContext& node) {
     auto tf_data_format = node.get_attribute<std::string>("data_format");
 
     TENSORFLOW_OP_VALIDATION(node,
-                             tf_data_format == "NHWC" || tf_data_format == "NCHW",
-                             "Conv3D data format is neither NHWC nor NCHW");
+                             tf_data_format == "NDHWC" || tf_data_format == "NCDHW",
+                             "Conv3D data format is neither NDHWC nor NCDHW");
 
     bool is_ndhwc = (tf_data_format == "NDHWC");
 

src/inference/src/cpp/ie_infer_request.cpp

@@ -79,9 +79,9 @@ Blob::Ptr InferRequest::GetBlob(const std::string& name) {
     Blob::Ptr blobPtr;
     INFER_REQ_CALL_STATEMENT(blobPtr = _impl->GetBlob(name);)
     std::string error = "Internal error: blob with name `" + name + "` is not allocated!";
-    const bool remoteBlobPassed = blobPtr->is<RemoteBlob>();
     if (blobPtr == nullptr)
         IE_THROW() << error;
+    const bool remoteBlobPassed = blobPtr->is<RemoteBlob>();
     if (!remoteBlobPassed && blobPtr->buffer() == nullptr)
         IE_THROW() << error;
     return blobPtr;

src/plugins/auto_batch/auto_batch.cpp

@@ -842,6 +842,8 @@ InferenceEngine::IExecutableNetworkInternal::Ptr AutoBatchInferencePlugin::LoadN
             optBatchSize = std::max(1u, std::min(requests, optimalBatchSize));
         if (optBatchSize > 2)  // batching is usually in-efficient for batch<4 (as batch1 kernels are heavily optimized)
             metaDevice.batchForDevice = optBatchSize;
+        else
+            metaDevice.batchForDevice = 1;
     }
 
     const auto perfConfig = fullConfig.find(PluginConfigParams::KEY_PERF_COUNT);

src/plugins/intel_cpu/src/nodes/mkldnn_adaptive_pooling.cpp

@@ -68,8 +68,8 @@ void MKLDNNAdaptivePoolingNode::getSupportedDescriptors() {
 
     if (getParentEdges().size() != 2)
         IE_THROW() << errorPrefix << "has incorrect number of input edges: " << getParentEdges().size();
-    if (getChildEdges().size() != (algorithm == AdaptivePoolingMax ? 2 : 1))
-        IE_THROW() << errorPrefix << "has incorrect number of output edges: " << getParentEdges().size();
+    if (getChildEdges().size() < (algorithm == AdaptivePoolingMax ? 2 : 1))
+        IE_THROW() << errorPrefix << "has incorrect number of output edges: " << getChildEdges().size();
 
     auto srcRank = getInputShapeAtPort(0).getRank();
     if (!one_of(spatialDimsCount, 1, 2, 3)) {

src/plugins/intel_cpu/src/nodes/mkldnn_matmul_node.cpp

@@ -75,6 +75,9 @@ bool MatMulKey::operator==(const MatMulKey &rhs) const {
     return retVal;
 }
 
+bool canBeExecutedInInt8(const Precision& firstInput, const Precision& secondInput) {
+    return one_of(firstInput, Precision::U8, Precision::I8) && secondInput == Precision::I8;
+}
 } // namespace
 
 bool MKLDNNMatMulNode::isSupportedOperation(const std::shared_ptr<const ngraph::Node>& op, std::string& errorMessage) noexcept {
@@ -139,6 +142,14 @@ bool MKLDNNMatMulNode::canFuse(const MKLDNNNodePtr& node) const {
         }
     }
 
+    // Todo:
+    //  Consider the case when Matmul doesn't support execution in int8, but is getting fused with FQ with int8 output.
+    //  Then the Matmul will change its output precision to fp32, but the FQ child will still has the int8 input precision.
+    //  This information should be propagated! Note that we may need to propagate updated precision to child fused nodes.
+    if (node->getType() == FakeQuantize &&
+        one_of(node->getOriginalOutputPrecisionAtPort(0), Precision::I8, Precision::U8) &&
+        !canBeExecutedInInt8(getOriginalInputPrecisionAtPort(0), getOriginalInputPrecisionAtPort(1)))
+        return false;
     return canFuseSimpleOperation(node);
 }
 
@@ -233,10 +244,6 @@ void MKLDNNMatMulNode::getSupportedDescriptors() {
 
     withBiases = getOriginalInputsNumber() == 3;
 
-    auto canBeExecutedInInt8 = [](const Precision firstInput, const Precision secondInput) {
-        return one_of(firstInput, Precision::U8, Precision::I8) && secondInput == Precision::I8;
-    };
-
     auto firstInPortPrec = getOriginalInputPrecisionAtPort(0);
     auto secondInPortPrec = getOriginalInputPrecisionAtPort(1);
     auto outPortPrec = getOriginalOutputPrecisionAtPort(0);

thirdparty/ocl/CMakeLists.txt

@@ -32,6 +32,9 @@ function(get_lib_path OUTPUT_DIR FINAL_OUTPUT_DIR)
     set("${FINAL_OUTPUT_DIR}" "${OUTPUT_DIR}/${LIB_DIR}" PARENT_SCOPE)
 endfunction()
 
+if(WIN32)
+    set(CMAKE_DEBUG_POSTFIX "")
+endif()
 set(OPENCL_ICD_LOADER_HEADERS_DIR "${CMAKE_CURRENT_SOURCE_DIR}/cl_headers/" CACHE PATH "Path to OCL includes" FORCE)
 
 set(OPENCL_HEADERS_DIR "${CMAKE_CURRENT_SOURCE_DIR}/cl_headers/" "${CMAKE_CURRENT_SOURCE_DIR}/clhpp_headers/include/" CACHE PATH "Path to OCL (CL and CLHPP) includes" FORCE)

tools/mo/requirements.txt

@@ -11,4 +11,3 @@ urllib3>=1.26.4
 requests>=2.25.1
 fastjsonschema~=2.15.1
 openvino-telemetry>=2022.1.0
-importlib-metadata; python_version < "3.8" and sys_platform == "win32"

tools/mo/setup.py

@@ -42,16 +42,16 @@ py_modules.append(prefix.replace('/', '.') + '__main__')
 # Minimal set of dependencies
 deps_whitelist = ['networkx', 'defusedxml', 'numpy', 'openvino-telemetry']
 
-# for py37 and less on Windows need importlib-metadata in order to use entry_point *.exe files
-if sys.platform == 'win32' and sys.version_info[1] < 8:
-    deps_whitelist.append('importlib-metadata')
-
 deps = []
 with open('requirements.txt', 'rt') as req_file:
     for line in req_file.read().split('\n'):
         if line.startswith(tuple(deps_whitelist)):
             deps.append(line)
 
+# for py37 and less on Windows need importlib-metadata in order to use entry_point *.exe files
+if sys.platform == 'win32' and sys.version_info[1] < 8:
+    deps.append('importlib-metadata')
+
 
 class InstallCmd(install):
     def run(self):

tools/pot/openvino/tools/pot/algorithms/quantization/fast_bias_correction/algorithm.py

@@ -214,7 +214,9 @@ class FastBiasCorrection(Algorithm):
             calculate_input_shape[input_node.fullname] = {'shape_node': lambda x: x.shape}
         calculate_metrics = self._engine.calculate_metrics
         self._engine.calculate_metrics = False
+        self._engine.inference_for_shape = True
         _, inputs_shape = self._engine.predict(calculate_input_shape, sampler)
+        self._engine.inference_for_shape = False
         self._engine.calculate_metrics = calculate_metrics
         for node_name, shape_node in inputs_shape.items():
             inputs_shape[node_name] = shape_node['shape_node'][0]

tools/pot/openvino/tools/pot/api/engine.py

@@ -24,6 +24,7 @@ class Engine(ABC):
         self._statistic_graph_builder = StatisticGraphBuilder()
         self._stat_requests_number = self.config.get('stat_requests_number', None)
         self._eval_requests_number = self.config.get('eval_requests_number', None)
+        self.inference_for_shape = False
         self.calculate_metrics = True
 
     def set_model(self, model):

tools/pot/openvino/tools/pot/engines/ac_engine.py

@@ -248,7 +248,7 @@ class ACEngine(Engine):
         if not stats_layout:
             return
         dataset_index = kwargs['dataset_indices'][0]
-        append_stats(self._accumulated_layer_stats, stats_layout, value, dataset_index)
+        append_stats(self._accumulated_layer_stats, stats_layout, value, dataset_index, self.inference_for_shape)
 
     @staticmethod
     def _set_requests_number(params, requests_number):

tools/pot/openvino/tools/pot/engines/ie_engine.py

@@ -200,7 +200,7 @@ class IEEngine(Engine):
         :param annotations: list of annotations [(img_id, annotation)]
         """
         dataset_index = annotations[0][0] if annotations is not None and annotations[0][0] else 0
-        append_stats(self._accumulated_layer_stats, stats_layout, outputs, dataset_index)
+        append_stats(self._accumulated_layer_stats, stats_layout, outputs, dataset_index, self.inference_for_shape)
 
     def _update_metrics(self, output, annotations, need_metrics_per_sample=False):
         """ Updates metrics.

tools/pot/openvino/tools/pot/engines/simplified_engine.py

@@ -20,4 +20,4 @@ class SimplifiedEngine(IEEngine):
                               batch_annotations, batch_meta, need_metrics_per_sample):
         # Collect statistics
         if stats_layout:
-            append_stats(self._accumulated_layer_stats, stats_layout, predictions, 0)
+            append_stats(self._accumulated_layer_stats, stats_layout, predictions, 0, self.inference_for_shape)

tools/pot/openvino/tools/pot/engines/utils.py

@@ -13,10 +13,10 @@ from ..utils.utils import convert_output_key
 logger = get_logger(__name__)
 
 
-def append_stats(accumulated_layer_stats, stats_layout, value, dataset_index):
+def append_stats(accumulated_layer_stats, stats_layout, value, dataset_index, inference_for_shape):
     inplace_stats_mapping = get_inplace_stats_mapping(stats_layout)
     if isinstance(value, list):
-        value = parse_sequential_stats(value, stats_layout)
+        value = parse_sequential_stats(value, stats_layout, inference_for_shape)
     else:
         value = process_raw_output(value)
     for layer, stats in stats_layout.items():
@@ -29,7 +29,7 @@ def append_stats(accumulated_layer_stats, stats_layout, value, dataset_index):
                     (dataset_index, compute_statistic(stat_fn, value, layer_stat_name)))
 
 
-def parse_sequential_stats(value_sequential, stats_layout):
+def parse_sequential_stats(value_sequential, stats_layout, inference_for_shape):
     stat_names_by_layer, old_names_mapping = get_per_layer_stat_mapping(stats_layout)
     activation_seq = defaultdict(lambda: [])
     for value in value_sequential:
@@ -40,6 +40,9 @@ def parse_sequential_stats(value_sequential, stats_layout):
 
     for layer, act_seq in activation_seq.items():
         seq_len = len(act_seq[0].shape)
+        if inference_for_shape:
+            activation_seq[layer] = act_seq[0]
+            continue
         if not isinstance(stat_names_by_layer[layer], Statistic) or \
                 not stat_names_by_layer[layer].kwargs.get('inplace_statistics', False):
             axis = 1 if seq_len == 2 else 2
@@ -103,7 +106,8 @@ def get_sequential_activations(activations, layer, activation_seq, stats_layout,
     elif old_names_mapping.get(layer, None) in stats_layout and hasattr(stat_names_by_layer[layer], 'kwargs') \
             and not stat_names_by_layer[layer].kwargs.get('inplace_statistics', False):
         activation_seq[layer].append(activations)
-    elif old_names_mapping.get(layer, None) in stats_layout and callable(stat_names_by_layer[layer]):
+    elif old_names_mapping.get(layer, None) in stats_layout and (callable(stat_names_by_layer[layer]) \
+            or callable(stats_layout[layer][stat_names_by_layer[layer]])):
         activation_seq[layer].append(activations)
 
 

tools/pot/tests/test_engines.py

@@ -44,7 +44,8 @@ def run_append_stats_test(engine):
     fc_layer_mock = create_ng_mock(['fc_layer'])
     value = {conv_layer_mock: sample_tensor, fc_layer_mock: sample_tensor}
     ref_value = {'conv_layer': sample_tensor, 'fc_layer': sample_tensor}
-    append_stats(engine._accumulated_layer_stats, stats_layout, value, dataset_index=0)
+    append_stats(engine._accumulated_layer_stats, stats_layout, value,
+                 dataset_index=0, inference_for_shape=False)
     for layer, accumulated_value in engine._accumulated_layer_stats.items():
         assert np.array_equal(accumulated_value[stat_name][0][1], ref_value[layer])
 
@@ -57,7 +58,8 @@ def run_append_stats_test(engine):
         {'conv_layer': sample_tensor, 'fc_layer': sample_tensor},
         {'conv_layer': sample_tensor, 'fc_layer': sample_tensor},
     ]
-    append_stats(engine._accumulated_layer_stats, stats_layout, value, dataset_index=0)
+    append_stats(engine._accumulated_layer_stats, stats_layout, value,
+                 dataset_index=0, inference_for_shape=False)
     for layer, accumulated_value in engine._accumulated_layer_stats.items():
         assert np.array_equal(
             accumulated_value[stat_name][0][1][:, 0], ref_value[0][layer]

tools/pot/tests/test_sanity.py

@@ -51,10 +51,9 @@ TEST_MODELS = [
     # {'drop_type': 'relative', 'max_iter_num': 1, 'accuracy_drop': 0.005, 'metrics': [
     #     {'name': 'accuracy@top1', 'baseline_value': 0.431}]}, 'GNA'),
 
-    #TODO: Enable after the problem with the shapes will be fixed
-    # ('mtcnn', 'caffe', 'DefaultQuantization', 'performance', 1, {'recall': 0.76, 'map': 0.6844}, {}, 'CPU'),
+    ('mtcnn', 'caffe', 'DefaultQuantization', 'performance', 1, {'recall': 0.76, 'map': 0.6844}, {}, 'CPU'),
 
-    ('mtcnn', 'caffe', 'DefaultQuantization', 'performance', 2, {'recall': 0.8, 'map': 0.7445},
+    ('mtcnn', 'caffe', 'DefaultQuantization', 'performance', 2, {'recall': 0.76, 'map': 0.6638},
      {'use_fast_bias': False}, 'CPU')
 ]
 CASCADE_MAP = Dict({