[IE] Add batched blob support (#2203)

* [IE] Add batched blob support

New `class BatchedBlob : public CompoundBlob` defined to allow to pass multiple blobs as 1 InferRequest input.

Motivation: There is the special user case when a number of plain images (e.g. `NV12Blob`) should be passed as one input for network which batch size > 1.

`class CompoundBlob` is not applicable for such cases due to:
1. `NV12Blob` is `CompoundBlob` which prevents to combine multiple NV12 images to a CompoundBlob
2. The default behavior in most of plugins - do not accept generic CompoundBlob as `SetBlob()` argument

Adding `SetBlob(name, vector<Blob::Ptr>...)` to `class IInferRequest`, `class InferRequest`, `class IInferRequestInternal`, ...  - is not effective solution due to limited and specific use cases for `batched inputs`.

+ Apply rule-of-zero to CompoundBlob and inherited classes.

* Add "BATCHED_BLOB" optimization capability metric

* Add BatchedBlob usage to hello_nv12_input_classification

* Apply offline code review outcome:

1. Revert CompoundBlob public .ctors signatures
2. Remove 'workaround' .ctor for `BatchedBlob`
3. Revert tensor descriptors of `I420Blob` `NV12Blob` back to the 'fake' value.

* Code review fix

* Add functional tests for CPU, GPU, MULTI, HETERO

* update doc comment

* Apply code review change requests.

inference-engine/include/ie_compound_blob.h

@@ -34,31 +34,6 @@ public:
      */
     using CPtr = std::shared_ptr<const CompoundBlob>;
 
-    /**
-     * @brief A virtual destructor
-     */
-    virtual ~CompoundBlob() = default;
-
-    /**
-     * @brief A copy constructor
-     */
-    CompoundBlob(const CompoundBlob& blob);
-
-    /**
-     * @brief A copy assignment operator
-     */
-    CompoundBlob& operator=(const CompoundBlob& blob) = default;
-
-    /**
-     * @brief A move constructor
-     */
-    CompoundBlob(CompoundBlob&& blob);
-
-    /**
-     * @brief A move assignment operator
-     */
-    CompoundBlob& operator=(CompoundBlob&& blob) = default;
-
     /**
      * @brief Constructs a compound blob from a vector of blobs
      *
@@ -121,9 +96,11 @@ public:
 
 protected:
     /**
-     * @brief A default constructor
+     * @brief Constructs a compound blob with specified descriptor
+     *
+     * @param tensorDesc A tensor descriptor for the compound blob
      */
-    CompoundBlob();
+    explicit CompoundBlob(const TensorDesc& tensorDesc);
 
     /**
      * @brief Compound blob container for underlying blobs
@@ -156,11 +133,6 @@ public:
      */
     using CPtr = std::shared_ptr<const NV12Blob>;
 
-    /**
-     * @brief A deleted default constructor
-     */
-    NV12Blob() = delete;
-
     /**
      * @brief Constructs NV12 blob from two planes Y and UV
      *
@@ -177,31 +149,6 @@ public:
      */
     NV12Blob(Blob::Ptr&& y, Blob::Ptr&& uv);
 
-    /**
-     * @brief A virtual destructor
-     */
-    virtual ~NV12Blob() = default;
-
-    /**
-     * @brief A copy constructor
-     */
-    NV12Blob(const NV12Blob& blob) = default;
-
-    /**
-     * @brief A copy assignment operator
-     */
-    NV12Blob& operator=(const NV12Blob& blob) = default;
-
-    /**
-     * @brief A move constructor
-     */
-    NV12Blob(NV12Blob&& blob) = default;
-
-    /**
-     * @brief A move assignment operator
-     */
-    NV12Blob& operator=(NV12Blob&& blob) = default;
-
     /**
      * @brief Returns a shared pointer to Y plane
      */
@@ -240,11 +187,6 @@ public:
      */
     using CPtr = std::shared_ptr<const I420Blob>;
 
-    /**
-     * @brief A deleted default constructor
-     */
-    I420Blob() = delete;
-
     /**
      * @brief Constructs I420 blob from three planes Y, U and V
      * @param y Blob object that represents Y plane in I420 color format
@@ -261,32 +203,6 @@ public:
      */
     I420Blob(Blob::Ptr&& y, Blob::Ptr&& u, Blob::Ptr&& v);
 
-    /**
-     * @brief A virtual destructor. It is made out of line for RTTI to
-     * work correctly on some platforms.
-     */
-    virtual ~I420Blob();
-
-    /**
-     * @brief A copy constructor
-     */
-    I420Blob(const I420Blob& blob) = default;
-
-    /**
-     * @brief A copy assignment operator
-     */
-    I420Blob& operator=(const I420Blob& blob) = default;
-
-    /**
-     * @brief A move constructor
-     */
-    I420Blob(I420Blob&& blob) = default;
-
-    /**
-     * @brief A move assignment operator
-     */
-    I420Blob& operator=(I420Blob&& blob) = default;
-
     /**
      * @brief Returns a reference to shared pointer to Y plane
      *
@@ -349,4 +265,48 @@ public:
 
     Blob::Ptr createROI(const ROI& roi) const override;
 };
+
+/**
+ * @brief This class represents a blob that contains other blobs - one per batch
+ * @details Plugin which supports BatchedBlob input should report BATCHED_BLOB
+ * in the OPTIMIZATION_CAPABILITIES metric.
+ */
+class INFERENCE_ENGINE_API_CLASS(BatchedBlob) : public CompoundBlob {
+ public:
+    /**
+     * @brief A smart pointer to the BatchedBlob object
+     */
+    using Ptr = std::shared_ptr<BatchedBlob>;
+
+    /**
+     * @brief A smart pointer to the const BatchedBlob object
+     */
+    using CPtr = std::shared_ptr<const BatchedBlob>;
+
+    /**
+     * @brief Constructs a batched blob from a vector of blobs
+     * @details All passed blobs should meet following requirements:
+     * - all blobs have equal tensor descriptors,
+     * - blobs layouts should be one of: NCHW, NHWC, NCDHW, NDHWC, NC, CN, C, CHW
+     * - batch dimensions should be equal to 1 or not defined (C, CHW).
+     * Resulting blob's tensor descriptor is constructed using tensor descriptors
+     * of passed blobs by setting batch dimension to blobs.size()
+     *
+     * @param blobs A vector of blobs that is copied to this object
+     */
+    explicit BatchedBlob(const std::vector<Blob::Ptr>& blobs);
+
+    /**
+     * @brief Constructs a batched blob from a vector of blobs
+     * @details All passed blobs should meet following requirements:
+     * - all blobs have equal tensor descriptors,
+     * - blobs layouts should be one of: NCHW, NHWC, NCDHW, NDHWC, NC, CN, C, CHW
+     * - batch dimensions should be equal to 1 or not defined (C, CHW).
+     * Resulting blob's tensor descriptor is constructed using tensor descriptors
+     * of passed blobs by setting batch dimension to blobs.size()
+     *
+     * @param blobs A vector of blobs that is moved to this object
+     */
+    explicit BatchedBlob(std::vector<Blob::Ptr>&& blobs);
+};
 }  // namespace InferenceEngine

inference-engine/include/ie_plugin_config.hpp

@@ -95,6 +95,7 @@ DECLARE_METRIC_KEY(FULL_DEVICE_NAME, std::string);
  *  - "INT8" - device can support models with INT8 layers
  *  - "BIN" - device can support models with BIN layers
  *  - "WINOGRAD" - device can support models where convolution implemented via Winograd transformations
+ *  - "BATCHED_BLOB" - device can support BatchedBlob
  */
 DECLARE_METRIC_KEY(OPTIMIZATION_CAPABILITIES, std::vector<std::string>);
 
@@ -104,6 +105,7 @@ DECLARE_METRIC_VALUE(FP16);
 DECLARE_METRIC_VALUE(INT8);
 DECLARE_METRIC_VALUE(BIN);
 DECLARE_METRIC_VALUE(WINOGRAD);
+DECLARE_METRIC_VALUE(BATCHED_BLOB);
 
 /**
  * @brief Metric to provide information about a range for streams on platforms where streams are supported.

inference-engine/samples/hello_nv12_input_classification/main.cpp

@@ -14,6 +14,15 @@
 
 #include <samples/common.hpp>
 #include <samples/classification_results.h>
+#include <samples/slog.hpp>
+
+#include <sys/stat.h>
+#ifdef _WIN32
+#include <os/windows/w_dirent.h>
+#else
+#include <dirent.h>
+#endif
+
 
 using namespace InferenceEngine;
 
@@ -49,29 +58,80 @@ std::pair<size_t, size_t> parseImageSize(const std::string& size_string) {
     return {width, height};
 }
 
+// Comparing to samples/args_helper.hpp, this version filters files by ".yuv" extension
+/**
+* @brief This function checks input args and existence of specified files in a given folder
+* @param path path to a file to be checked for existence
+* @return files updated vector of verified input files
+*/
+std::vector<std::string> readInputFileNames(const std::string& path) {
+    struct stat sb;
+    if (stat(path.c_str(), &sb) != 0) {
+        slog::warn << "File " << path << " cannot be opened!" << slog::endl;
+        return {};
+    }
+
+    std::vector<std::string> files;
+
+    if (S_ISDIR(sb.st_mode)) {
+        DIR *dp = opendir(path.c_str());
+        if (dp == nullptr) {
+            slog::warn << "Directory " << path << " cannot be opened!" << slog::endl;
+            return {};
+        }
+
+        for (struct dirent* ep = readdir(dp); ep != nullptr; ep = readdir(dp)) {
+            std::string fileName = ep->d_name;
+            if (fileName == "." || fileName == ".." || fileName.substr(fileName.size() - 4) != ".yuv") continue;
+            files.push_back(path + "/" + ep->d_name);
+        }
+        closedir(dp);
+    } else {
+        files.push_back(path);
+    }
+
+    if (files.size() < 20) {
+        slog::info << "Files were added: " << files.size() << slog::endl;
+        for (std::string filePath : files) {
+            slog::info << "    " << filePath << slog::endl;
+        }
+    } else {
+        slog::info << "Files were added: " << files.size() << ". Too many to display each of them." << slog::endl;
+    }
+
+    return files;
+}
+
+using UString = std::basic_string<uint8_t>;
+
 /**
  * \brief Read image data from file
- * @return buffer containing the image data
+ * @return buffers containing the images data
  */
-std::unique_ptr<unsigned char[]> readImageDataFromFile(const std::string& image_path, size_t size) {
+std::vector<UString> readImagesDataFromFiles(const std::vector<std::string>& files, size_t size) {
-    std::ifstream file(image_path, std::ios_base::ate | std::ios_base::binary);
+    std::vector<UString> result;
-    if (!file.good() || !file.is_open()) {
-        std::stringstream err;
-        err << "Cannot access input image file. File path: " << image_path;
-        throw std::runtime_error(err.str());
-    }
 
-    const size_t file_size = file.tellg();
+    for (const auto& image_path : files) {
-    if (file_size < size) {
+        std::ifstream file(image_path, std::ios_base::ate | std::ios_base::binary);
-        std::stringstream err;
+        if (!file.good() || !file.is_open()) {
-        err << "Invalid read size provided. File size: " << file_size << ", to read: " << size;
+            std::stringstream err;
-        throw std::runtime_error(err.str());
+            err << "Cannot access input image file. File path: " << image_path;
-    }
+            throw std::runtime_error(err.str());
-    file.seekg(0);
+        }
 
-    std::unique_ptr<unsigned char[]> data(new unsigned char[size]);
+        const size_t file_size = file.tellg();
-    file.read(reinterpret_cast<char*>(data.get()), size);
+        if (file_size < size) {
-    return data;
+            std::stringstream err;
+            err << "Invalid read size provided. File size: " << file_size << ", to read: " << size;
+            throw std::runtime_error(err.str());
+        }
+        file.seekg(0);
+
+        UString data(size, 0);
+        file.read(reinterpret_cast<char*>(&data[0]), size);
+        result.push_back(std::move(data));
+    }
+    return result;
 }
 
 /**
@@ -89,6 +149,49 @@ void setBatchSize(CNNNetwork& network, size_t batch) {
     network.reshape(inputShapes);
 }
 
+std::vector<Blob::Ptr> readInputBlobs(std::vector<UString>& data, size_t width, size_t height) {
+    // read image with size converted to NV12 data size: height(NV12) = 3 / 2 * logical height
+
+    // Create tensor descriptors for Y and UV blobs
+    const InferenceEngine::TensorDesc y_plane_desc(InferenceEngine::Precision::U8, {1, 1, height, width},
+                                                   InferenceEngine::Layout::NHWC);
+    const InferenceEngine::TensorDesc uv_plane_desc(InferenceEngine::Precision::U8, {1, 2, height / 2, width / 2},
+                                                    InferenceEngine::Layout::NHWC);
+    const size_t offset = width * height;
+
+    std::vector<Blob::Ptr> blobs;
+    for (auto& buf : data) {
+        // --------------------------- Create a blob to hold the NV12 input data -------------------------------
+        auto ptr = &buf[0];
+
+        // Create blob for Y plane from raw data
+        Blob::Ptr y_blob = make_shared_blob<uint8_t>(y_plane_desc, ptr);
+        // Create blob for UV plane from raw data
+        Blob::Ptr uv_blob = make_shared_blob<uint8_t>(uv_plane_desc, ptr + offset);
+        // Create NV12Blob from Y and UV blobs
+        blobs.emplace_back(make_shared_blob<NV12Blob>(y_blob, uv_blob));
+    }
+
+    return blobs;
+}
+
+bool isBatchedBlobSupported(const Core& ie, const std::string& device_name) {
+    const std::vector<std::string> supported_metrics =
+        ie.GetMetric(device_name, METRIC_KEY(SUPPORTED_METRICS));
+
+    if (std::find(supported_metrics.begin(), supported_metrics.end(),
+                METRIC_KEY(OPTIMIZATION_CAPABILITIES)) ==
+        supported_metrics.end()) {
+    return false;
+    }
+
+    const std::vector<std::string> optimization_caps =
+        ie.GetMetric(device_name, METRIC_KEY(OPTIMIZATION_CAPABILITIES));
+
+    return std::find(optimization_caps.begin(), optimization_caps.end(),
+                   METRIC_VALUE(BATCHED_BLOB)) != optimization_caps.end();
+}
+
 /**
 * @brief The entry point of the Inference Engine sample application
 */
@@ -96,7 +199,7 @@ int main(int argc, char *argv[]) {
     try {
         // ------------------------------ Parsing and validatiing input arguments------------------------------
         if (argc != 5) {
-            std::cout << "Usage : ./hello_nv12_input_classification <path_to_model> <path_to_image> <image_size> <device_name>"
+            std::cout << "Usage : " << argv[0] << " <path_to_model> <path_to_image(s)> <image_size> <device_name>"
                       << std::endl;
             return EXIT_FAILURE;
         }
@@ -108,13 +211,26 @@ int main(int argc, char *argv[]) {
         const std::string device_name{argv[4]};
         // -----------------------------------------------------------------------------------------------------
 
+        // --------------------------- 0. Read image names -----------------------------------------------------
+        auto image_names = readInputFileNames(input_image_path);
+
+        if (image_names.empty()) {
+            throw std::invalid_argument("images not found");
+        }
+        // -----------------------------------------------------------------------------------------------------
+
         // --------------------------- 1. Load inference engine ------------------------------------------------
         Core ie;
         // -----------------------------------------------------------------------------------------------------
 
         // 2. Read a model in OpenVINO Intermediate Representation (.xml and .bin files) or ONNX (.onnx file) format
         CNNNetwork network = ie.ReadNetwork(input_model);
-        setBatchSize(network, 1);
+        // -----------------------------------------------------------------------------------------------------
+
+        // --------------------------- 2. Set model batch size -------------------------------------------------
+        size_t batch_size = isBatchedBlobSupported(ie, device_name) ? image_names.size() : 1;
+        std::cout << "Setting network batch size to " << batch_size << std::endl;
+        setBatchSize(network, batch_size);
         // -----------------------------------------------------------------------------------------------------
 
         // --------------------------- 3. Configure input and output -------------------------------------------
@@ -154,40 +270,54 @@ int main(int argc, char *argv[]) {
         // -----------------------------------------------------------------------------------------------------
 
         // --------------------------- 6. Prepare input --------------------------------------------------------
-        // read image with size converted to NV12 data size: height(NV12) = 3 / 2 * logical height
+        auto image_bufs = readImagesDataFromFiles(image_names, input_width * (input_height * 3 / 2));
-        auto image_buf = readImageDataFromFile(input_image_path, input_width * (input_height * 3 / 2));
 
-        // --------------------------- Create a blob to hold the NV12 input data -------------------------------
+        auto inputs = readInputBlobs(image_bufs, input_width, input_height);
-        // Create tensor descriptors for Y and UV blobs
-        InferenceEngine::TensorDesc y_plane_desc(InferenceEngine::Precision::U8,
-            {1, 1, input_height, input_width}, InferenceEngine::Layout::NHWC);
-        InferenceEngine::TensorDesc uv_plane_desc(InferenceEngine::Precision::U8,
-            {1, 2, input_height / 2, input_width / 2}, InferenceEngine::Layout::NHWC);
-        const size_t offset = input_width * input_height;
 
-        // Create blob for Y plane from raw data
+        // If batch_size > 1 => batched blob supported => replace all inputs by a BatchedBlob
-        Blob::Ptr y_blob = make_shared_blob<uint8_t>(y_plane_desc, image_buf.get());
+        if (batch_size > 1) {
-        // Create blob for UV plane from raw data
+            assert(batch_size == inputs.size());
-        Blob::Ptr uv_blob = make_shared_blob<uint8_t>(uv_plane_desc, image_buf.get() + offset);
+            std::cout << "Infer using BatchedBlob of NV12 images." << std::endl;
-        // Create NV12Blob from Y and UV blobs
+            Blob::Ptr batched_input = make_shared_blob<BatchedBlob>(inputs);
-        Blob::Ptr input = make_shared_blob<NV12Blob>(y_blob, uv_blob);
+            inputs = {batched_input};
+        }
 
-        // --------------------------- Set the input blob to the InferRequest ----------------------------------
+        /** Read labels from file (e.x. AlexNet.labels) **/
-        infer_request.SetBlob(input_name, input);
+        std::string labelFileName = fileNameNoExt(input_model) + ".labels";
-        // -----------------------------------------------------------------------------------------------------
+        std::vector<std::string> labels;
 
-        // --------------------------- 7. Do inference ---------------------------------------------------------
+        std::ifstream inputFile;
-        /* Running the request synchronously */
+        inputFile.open(labelFileName, std::ios::in);
-        infer_request.Infer();
+        if (inputFile.is_open()) {
-        // -----------------------------------------------------------------------------------------------------
+            std::string strLine;
+            while (std::getline(inputFile, strLine)) {
+                trim(strLine);
+                labels.push_back(strLine);
+            }
+        }
 
-        // --------------------------- 8. Process output -------------------------------------------------------
+        for (size_t i = 0; i < inputs.size(); i++) {
-        Blob::Ptr output = infer_request.GetBlob(output_name);
+            const auto& input = inputs[i];
+            // --------------------------- Set the input blob to the InferRequest ------------------------------
+            infer_request.SetBlob(input_name, input);
+            // -------------------------------------------------------------------------------------------------
 
-        // Print classification results
+            // --------------------------- 7. Do inference -----------------------------------------------------
-        ClassificationResult classificationResult(output, {input_image_path});
+            /* Running the request synchronously */
-        classificationResult.print();
+            infer_request.Infer();
-        // -----------------------------------------------------------------------------------------------------
+            // -------------------------------------------------------------------------------------------------
+
+            // --------------------------- 8. Process output ---------------------------------------------------
+            Blob::Ptr output = infer_request.GetBlob(output_name);
+
+            // Print classification results
+            const auto names_offset = image_names.begin() + batch_size * i;
+            std::vector<std::string> names(names_offset, names_offset + batch_size);
+
+            ClassificationResult classificationResult(output, names, batch_size, 10, labels);
+            classificationResult.print();
+            // -------------------------------------------------------------------------------------------------
+        }
     } catch (const std::exception & ex) {
         std::cerr << ex.what() << std::endl;
         return EXIT_FAILURE;

inference-engine/src/inference_engine/ie_compound_blob.cpp

@@ -18,7 +18,7 @@ namespace InferenceEngine {
 
 namespace {
 
-void verifyNV12BlobInput(const Blob::Ptr& y, const Blob::Ptr& uv) {
+TensorDesc verifyNV12BlobInput(const Blob::Ptr& y, const Blob::Ptr& uv) {
     // Y and UV must be valid pointers
     if (y == nullptr || uv == nullptr) {
         THROW_IE_EXCEPTION << "Y and UV planes must be valid Blob objects";
@@ -91,9 +91,11 @@ void verifyNV12BlobInput(const Blob::Ptr& y, const Blob::Ptr& uv) {
         THROW_IE_EXCEPTION << "The width of the Y plane must be equal to (2 * the width of the UV plane), actual: "
                            << yDims[3] << "(Y plane) and " << uvDims[3] << "(UV plane)";
     }
+
+    return {Precision::U8, {}, Layout::NCHW};
 }
 
-void verifyI420BlobInput(const Blob::Ptr& y, const Blob::Ptr& u, const Blob::Ptr& v) {
+TensorDesc verifyI420BlobInput(const Blob::Ptr& y, const Blob::Ptr& u, const Blob::Ptr& v) {
     // Y and UV must be valid pointers
     if (y == nullptr || u == nullptr || v == nullptr) {
         THROW_IE_EXCEPTION << "Y, U and V planes must be valid Blob objects";
@@ -190,21 +192,85 @@ void verifyI420BlobInput(const Blob::Ptr& y, const Blob::Ptr& u, const Blob::Ptr
         THROW_IE_EXCEPTION << "The width of the Y plane must be equal to (2 * the width of the UV plane), actual: "
                            << yDims[3] << "(Y plane) and " << vDims[3] << "(V plane)";
     }
+
+    return {Precision::U8, {}, Layout::NCHW};
+}
+
+TensorDesc getBlobTensorDesc(const Blob::Ptr& blob) {
+    if (auto nv12 = dynamic_cast<NV12Blob*>(blob.get())) {
+        auto yDesc = nv12->y()->getTensorDesc();
+        yDesc.getDims()[1] += 2;
+        return yDesc;
+    }
+
+    if (auto i420 = dynamic_cast<I420Blob*>(blob.get())) {
+        auto yDesc = i420->y()->getTensorDesc();
+        yDesc.getDims()[1] += 2;
+        return yDesc;
+    }
+
+    return blob->getTensorDesc();
+}
+
+TensorDesc verifyBatchedBlobInput(const std::vector<Blob::Ptr>& blobs) {
+    // verify invariants
+    if (blobs.empty()) {
+        THROW_IE_EXCEPTION << "BatchedBlob cannot be created from empty vector of Blob, Please, make sure vector contains at least one Blob";
+    }
+
+    // Cannot create a compound blob from nullptr Blob objects
+    if (std::any_of(blobs.begin(), blobs.end(), [](const Blob::Ptr& blob) {
+            return blob == nullptr;
+        })) {
+        THROW_IE_EXCEPTION << "Cannot create a compound blob from nullptr Blob objects";
+    }
+
+    const auto subBlobDesc = getBlobTensorDesc(blobs[0]);
+
+    if (std::any_of(blobs.begin(), blobs.end(),
+                    [&subBlobDesc](const Blob::Ptr& blob) {
+                        return getBlobTensorDesc(blob) != subBlobDesc;
+                    })) {
+        THROW_IE_EXCEPTION << "All blobs tensors should be equal";
+    }
+
+    auto subBlobLayout = subBlobDesc.getLayout();
+
+    auto blobLayout = Layout::ANY;
+    SizeVector blobDims = subBlobDesc.getDims();
+    switch (subBlobLayout) {
+    case NCHW:
+    case NHWC:
+    case NCDHW:
+    case NDHWC:
+    case NC:
+    case CN:
+        blobLayout = subBlobLayout;
+        if (blobDims[0] != 1) {
+            THROW_IE_EXCEPTION << "All blobs should be batch 1";
+        }
+        blobDims[0] = blobs.size();
+        break;
+    case C:
+        blobLayout = NC;
+        blobDims.insert(blobDims.begin(), blobs.size());
+        break;
+    case CHW:
+        blobLayout = NCHW;
+        blobDims.insert(blobDims.begin(), blobs.size());
+        break;
+    default:
+        THROW_IE_EXCEPTION << "Unsupported sub-blobs layout - to be one of: [NCHW, NHWC, NCDHW, NDHWC, NC, CN, C, CHW]";
+    }
+
+    return TensorDesc{subBlobDesc.getPrecision(), blobDims, blobLayout};
 }
 
 }  // anonymous namespace
 
-CompoundBlob::CompoundBlob(): Blob(TensorDesc(Precision::UNSPECIFIED, {}, Layout::ANY)) {}
+CompoundBlob::CompoundBlob(const TensorDesc& tensorDesc): Blob(tensorDesc) {}
 
-CompoundBlob::CompoundBlob(const CompoundBlob& blob): CompoundBlob() {
+CompoundBlob::CompoundBlob(const std::vector<Blob::Ptr>& blobs): CompoundBlob(TensorDesc{}) {
-    this->_blobs = blob._blobs;
-}
-
-CompoundBlob::CompoundBlob(CompoundBlob&& blob): CompoundBlob() {
-    this->_blobs = std::move(blob._blobs);
-}
-
-CompoundBlob::CompoundBlob(const std::vector<Blob::Ptr>& blobs): CompoundBlob() {
     // Cannot create a compound blob from nullptr Blob objects
     if (std::any_of(blobs.begin(), blobs.end(), [](const Blob::Ptr& blob) {
             return blob == nullptr;
@@ -223,7 +289,7 @@ CompoundBlob::CompoundBlob(const std::vector<Blob::Ptr>& blobs): CompoundBlob()
     this->_blobs = blobs;
 }
 
-CompoundBlob::CompoundBlob(std::vector<Blob::Ptr>&& blobs): CompoundBlob() {
+CompoundBlob::CompoundBlob(std::vector<Blob::Ptr>&& blobs): CompoundBlob(TensorDesc{}) {
     // Cannot create a compound blob from nullptr Blob objects
     if (std::any_of(blobs.begin(), blobs.end(), [](const Blob::Ptr& blob) {
             return blob == nullptr;
@@ -295,22 +361,14 @@ void* CompoundBlob::getHandle() const noexcept {
     return nullptr;
 }
 
-NV12Blob::NV12Blob(const Blob::Ptr& y, const Blob::Ptr& uv) {
+NV12Blob::NV12Blob(const Blob::Ptr& y, const Blob::Ptr& uv)
-    // verify data is correct
+    : CompoundBlob(verifyNV12BlobInput(y, uv)) {
-    verifyNV12BlobInput(y, uv);
+    this->_blobs = {y, uv};
-    // set blobs
-    _blobs.emplace_back(y);
-    _blobs.emplace_back(uv);
-    tensorDesc = TensorDesc(Precision::U8, {}, Layout::NCHW);
 }
 
-NV12Blob::NV12Blob(Blob::Ptr&& y, Blob::Ptr&& uv) {
+NV12Blob::NV12Blob(Blob::Ptr&& y, Blob::Ptr&& uv)
-    // verify data is correct
+    : CompoundBlob(verifyNV12BlobInput(y, uv)) {
-    verifyNV12BlobInput(y, uv);
+    this->_blobs = {std::move(y), std::move(uv)};
-    // set blobs
-    _blobs.emplace_back(std::move(y));
-    _blobs.emplace_back(std::move(uv));
-    tensorDesc = TensorDesc(Precision::U8, {}, Layout::NCHW);
 }
 
 Blob::Ptr& NV12Blob::y() noexcept {
@@ -346,28 +404,16 @@ Blob::Ptr NV12Blob::createROI(const ROI& roi) const {
     return std::make_shared<NV12Blob>(yRoiBlob, uvRoiBlob);
 }
 
-I420Blob::I420Blob(const Blob::Ptr& y, const Blob::Ptr& u, const Blob::Ptr& v) {
+I420Blob::I420Blob(const Blob::Ptr& y, const Blob::Ptr& u, const Blob::Ptr& v)
-    // verify data is correct
+    : CompoundBlob(verifyI420BlobInput(y, u, v)) {
-    verifyI420BlobInput(y, u, v);
+    this->_blobs = {y, u, v};
-    // set blobs
-    _blobs.emplace_back(y);
-    _blobs.emplace_back(u);
-    _blobs.emplace_back(v);
-    tensorDesc = TensorDesc(Precision::U8, {}, Layout::NCHW);
 }
 
-I420Blob::I420Blob(Blob::Ptr&& y, Blob::Ptr&& u, Blob::Ptr&& v) {
+I420Blob::I420Blob(Blob::Ptr&& y, Blob::Ptr&& u, Blob::Ptr&& v)
-    // verify data is correct
+    : CompoundBlob(verifyI420BlobInput(y, u, v)) {
-    verifyI420BlobInput(y, u, v);
+    this->_blobs = {std::move(y), std::move(u), std::move(v)};
-    // set blobs
-    _blobs.emplace_back(std::move(y));
-    _blobs.emplace_back(std::move(u));
-    _blobs.emplace_back(std::move(v));
-    tensorDesc = TensorDesc(Precision::U8, {}, Layout::NCHW);
 }
 
-I420Blob::~I420Blob() {}
-
 Blob::Ptr& I420Blob::y() noexcept {
     // NOTE: Y plane is a memory blob, which is checked in the constructor
     return _blobs[0];
@@ -412,4 +458,14 @@ Blob::Ptr I420Blob::createROI(const ROI& roi) const {
     return std::make_shared<I420Blob>(yRoiBlob, uRoiBlob, vRoiBlob);
 }
 
+BatchedBlob::BatchedBlob(const std::vector<Blob::Ptr>& blobs)
+    : CompoundBlob(verifyBatchedBlobInput(blobs)) {
+    this->_blobs = blobs;
+}
+
+BatchedBlob::BatchedBlob(std::vector<Blob::Ptr>&& blobs)
+    : CompoundBlob(verifyBatchedBlobInput(blobs)) {
+    this->_blobs = std::move(blobs);
+}
+
 }  // namespace InferenceEngine

inference-engine/tests/functional/plugin/cpu/shared_tests_instances/behavior/set_blob_of_kind.cpp

@@ -0,0 +1,39 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "behavior/set_blob_of_kind.hpp"
+#include "common_test_utils/test_constants.hpp"
+#include "multi-device/multi_device_config.hpp"
+
+using namespace BehaviorTestsDefinitions;
+using namespace InferenceEngine;
+
+const std::vector<FuncTestUtils::BlobKind> blobKinds = {
+    FuncTestUtils::BlobKind::Simple,
+    FuncTestUtils::BlobKind::Compound,
+    FuncTestUtils::BlobKind::BatchOfSimple
+};
+
+const SetBlobOfKindConfig cpuConfig{}; //nothing special
+const SetBlobOfKindConfig multiConfig{{ MULTI_CONFIG_KEY(DEVICE_PRIORITIES) , CommonTestUtils::DEVICE_CPU}};
+const SetBlobOfKindConfig heteroConfig{{ "TARGET_FALLBACK", CommonTestUtils::DEVICE_CPU }};
+
+INSTANTIATE_TEST_CASE_P(smoke_SetBlobOfKindCPU, SetBlobOfKindTest,
+    ::testing::Combine(::testing::ValuesIn(blobKinds),
+                       ::testing::Values(CommonTestUtils::DEVICE_CPU),
+                       ::testing::Values(cpuConfig)),
+    SetBlobOfKindTest::getTestCaseName);
+
+
+INSTANTIATE_TEST_CASE_P(smoke_SetBlobOfKindMULTI, SetBlobOfKindTest,
+    ::testing::Combine(::testing::ValuesIn(blobKinds),
+                       ::testing::Values(CommonTestUtils::DEVICE_MULTI),
+                       ::testing::Values(multiConfig)),
+    SetBlobOfKindTest::getTestCaseName);
+
+INSTANTIATE_TEST_CASE_P(smoke_SetBlobOfKindHETERO, SetBlobOfKindTest,
+    ::testing::Combine(::testing::ValuesIn(blobKinds),
+                       ::testing::Values(CommonTestUtils::DEVICE_HETERO),
+                       ::testing::Values(heteroConfig)),
+    SetBlobOfKindTest::getTestCaseName);

inference-engine/tests/functional/plugin/gpu/shared_tests_instances/behavior/set_blob_of_kind.cpp

@@ -0,0 +1,23 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "behavior/set_blob_of_kind.hpp"
+#include "common_test_utils/test_constants.hpp"
+
+using namespace BehaviorTestsDefinitions;
+using namespace InferenceEngine;
+
+const std::vector<FuncTestUtils::BlobKind> blobKinds = {
+    FuncTestUtils::BlobKind::Simple,
+    FuncTestUtils::BlobKind::Compound,
+    FuncTestUtils::BlobKind::BatchOfSimple
+};
+
+const SetBlobOfKindConfig gpuConfig{}; //nothing special
+
+INSTANTIATE_TEST_CASE_P(smoke_SetBlobOfKindGPU, SetBlobOfKindTest,
+    ::testing::Combine(::testing::ValuesIn(blobKinds),
+                       ::testing::Values(CommonTestUtils::DEVICE_GPU),
+                       ::testing::Values(gpuConfig)),
+    SetBlobOfKindTest::getTestCaseName);

inference-engine/tests/functional/plugin/shared/include/behavior/set_blob_of_kind.hpp

@@ -0,0 +1,32 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "functional_test_utils/layer_test_utils.hpp"
+#include "common_test_utils/common_utils.hpp"
+
+namespace BehaviorTestsDefinitions {
+
+using SetBlobOfKindConfig = std::remove_reference<decltype(((LayerTestsUtils::LayerTestsCommon*)0)->GetConfiguration())>::type;
+
+using SetBlobOfKindParams = std::tuple<FuncTestUtils::BlobKind, // The kind of blob
+                                       std::string,             // Device name
+                                       SetBlobOfKindConfig>;    // configuration
+
+class SetBlobOfKindTest : public testing::WithParamInterface<SetBlobOfKindParams>, virtual public LayerTestsUtils::LayerTestsCommon {
+public:
+    InferenceEngine::Blob::Ptr GenerateInput(const InferenceEngine::InputInfo &info) const override;
+    void Run() override;
+    static std::string getTestCaseName(testing::TestParamInfo<SetBlobOfKindParams> obj);
+    void ExpectSetBlobThrow();
+
+protected:
+    void SetUp() override;
+
+private:
+    FuncTestUtils::BlobKind blobKind;
+};
+
+} // namespace BehaviorTestsDefinitions

inference-engine/tests/functional/plugin/shared/src/behavior/set_blob_of_kind.cpp

@@ -0,0 +1,106 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "behavior/set_blob_of_kind.hpp"
+
+#include <single_layer_tests/cum_sum.hpp>
+#include <functional_test_utils/plugin_config.hpp>
+
+#include <ie_compound_blob.h>
+
+using namespace InferenceEngine;
+
+namespace BehaviorTestsDefinitions {
+
+std::string SetBlobOfKindTest::getTestCaseName(testing::TestParamInfo<SetBlobOfKindParams> obj) {
+    FuncTestUtils::BlobKind blobKind;
+    std::string targetDevice;
+    std::map<std::string, std::string> configuration;
+    std::tie(blobKind, targetDevice, configuration) = obj.param;
+
+    std::ostringstream result;
+    result << "Kind=" << blobKind;
+    result << " Device="<< targetDevice;
+    return result.str();
+}
+
+namespace {
+
+bool isBatchedBlobSupported(const std::shared_ptr<Core>& core, const LayerTestsUtils::TargetDevice& targetDevice) {
+    const std::vector<std::string> supported_metrics = core->GetMetric(targetDevice, METRIC_KEY(SUPPORTED_METRICS));
+
+    if (std::find(supported_metrics.begin(), supported_metrics.end(),
+                 METRIC_KEY(OPTIMIZATION_CAPABILITIES)) == supported_metrics.end()) {
+        return false;
+    }
+
+    const std::vector<std::string> optimization_caps =
+        core->GetMetric(targetDevice, METRIC_KEY(OPTIMIZATION_CAPABILITIES));
+
+    return std::find(optimization_caps.begin(), optimization_caps.end(),
+                    METRIC_VALUE(BATCHED_BLOB)) != optimization_caps.end();
+}
+
+bool isBlobKindSupported(const std::shared_ptr<Core>& core,
+                         const LayerTestsUtils::TargetDevice& targetDevice,
+                         FuncTestUtils::BlobKind blobKind) {
+    switch (blobKind) {
+    case FuncTestUtils::BlobKind::Simple:
+        return true;
+    case FuncTestUtils::BlobKind::Compound:
+        return false;
+    case FuncTestUtils::BlobKind::BatchOfSimple:
+        return isBatchedBlobSupported(core, targetDevice);
+    default:
+        THROW_IE_EXCEPTION << "Test does not support the blob kind";
+    }
+}
+
+} // namespace
+
+Blob::Ptr SetBlobOfKindTest::GenerateInput(const InferenceEngine::InputInfo& info) const {
+    return makeBlobOfKind(info.getTensorDesc(), blobKind);
+}
+
+void SetBlobOfKindTest::Run() {
+    SKIP_IF_CURRENT_TEST_IS_DISABLED()
+
+    LoadNetwork();
+
+    if (isBlobKindSupported(core, targetDevice, blobKind)) {
+        Infer();
+    } else {
+        ExpectSetBlobThrow();
+    }
+}
+
+void SetBlobOfKindTest::ExpectSetBlobThrow() {
+    inferRequest = executableNetwork.CreateInferRequest();
+
+    for (const auto &input : executableNetwork.GetInputsInfo()) {
+        const auto &info = input.second;
+        auto blob = GenerateInput(*info);
+        EXPECT_THROW(inferRequest.SetBlob(info->name(), blob),
+                     InferenceEngine::details::InferenceEngineException);
+    }
+}
+
+void SetBlobOfKindTest::SetUp() {
+    SizeVector IS{4, 3, 6, 8};
+    std::tie(blobKind, targetDevice, configuration) = this->GetParam();
+
+    auto ngPrc = FuncTestUtils::PrecisionUtils::convertIE2nGraphPrc(Precision::FP32);
+    auto params = ngraph::builder::makeParams(ngPrc, {IS});
+    auto paramOuts = ngraph::helpers::convert2OutputVector(ngraph::helpers::castOps2Nodes<ngraph::op::Parameter>(params));
+    auto axisNode = std::make_shared<ngraph::op::Constant>(ngraph::element::Type_t::i64, ngraph::Shape{}, std::vector<int64_t>{-1})->output(0);
+    auto cumSum = std::dynamic_pointer_cast<ngraph::opset4::CumSum>(ngraph::builder::makeCumSum(paramOuts[0], axisNode, false, false));
+    ngraph::ResultVector results{std::make_shared<ngraph::opset4::Result>(cumSum)};
+    function = std::make_shared<ngraph::Function>(results, params, "InferSetBlob");
+}
+
+TEST_P(SetBlobOfKindTest, CompareWithRefs) {
+    Run();
+}
+
+} // namespace BehaviorTestsDefinitions

inference-engine/tests/ie_test_utils/functional_test_utils/blob_utils.hpp

@@ -13,6 +13,7 @@
 #include <gtest/gtest.h>
 #include "blob_factory.hpp"
 #include "blob_transform.hpp"
+#include "ie_compound_blob.h"
 #include "precision_utils.h"
 #include "common_test_utils/data_utils.hpp"
 #include "common_test_utils/test_constants.hpp"
@@ -599,4 +600,46 @@ static short reducePrecisionBitwiseS(const float in) {
     return s;
 }
 }  // namespace Bf16TestUtils
+
+enum class BlobKind {
+    Simple,
+    Compound,
+    BatchOfSimple
+};
+
+inline std::ostream& operator<<(std::ostream& os, BlobKind kind) {
+    switch (kind) {
+    case BlobKind::Simple:
+        return os << "Simple";
+    case BlobKind::Compound:
+        return os << "Compound";
+    case BlobKind::BatchOfSimple:
+        return os << "BatchOfSimple";
+    default:
+        THROW_IE_EXCEPTION << "Test does not support the blob kind";
+  }
+}
+
+inline InferenceEngine::Blob::Ptr makeBlobOfKind(const InferenceEngine::TensorDesc& td, BlobKind blobKind) {
+    using namespace ::InferenceEngine;
+    switch (blobKind) {
+    case BlobKind::Simple:
+        return createAndFillBlob(td);
+    case BlobKind::Compound:
+        return make_shared_blob<CompoundBlob>(std::vector<Blob::Ptr>{});
+    case BlobKind::BatchOfSimple: {
+        const auto subBlobsNum = td.getDims()[0];
+        auto subBlobDesc = td;
+        subBlobDesc.getDims()[0] = 1;
+        std::vector<Blob::Ptr> subBlobs;
+        for (size_t i = 0; i < subBlobsNum; i++) {
+            subBlobs.push_back(makeBlobOfKind(subBlobDesc, BlobKind::Simple));
+        }
+        return make_shared_blob<BatchedBlob>(subBlobs);
+    }
+    default:
+        THROW_IE_EXCEPTION << "Test does not support the blob kind";
+    }
+}
+
 }  // namespace FuncTestUtils