OpenVINO Blog

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
Sort By:

Joint Pruning, Quantization and Distillation for Efficient Inference of Transformers

April 23, 2023


Pre-trained transformer models are widely deployed for various NLP tasks such as text classification, question answering, and generation task. The recent trend is that models continue to scale while yielding improved performance. However, growth of transformers also leads to great amount of compute resources and energy need for deployment. The goal of model compression is to achieve model simplification from the original without significantly diminished accuracy. Pruning, quantization, and knowledge distillation are three most popular model compression techniques for deep learning models. Pruning is a technique for reducing the size of a model to improve efficiency or performance. By reducing the number of bits needed to represent data, quantization can significantly reduce storage and computational requirements. Knowledge distillation involves training a small model to imitate the behavior of a larger model.

Figure 1: End-to-End Joint Optimization in One Pipeline

OpenVINOTM Neural Network Compression Framework (NNCF) develops Joint Pruning, Quantization and Distillation (JPQD) as a single joint-optimization pipeline to improve transformer inference performance by pruning, quantization, and distillation in parallel during transfer learning of a pretrained transformer. JPQD alleviates the developer complexity of sequential optimization of different compression techniques, resulting in optimized model with significant efficiency improvement while preserving good task accuracy. The output of JPQD is a structurally pruned, quantized model in OpenVINOTM IR, which is ready to deploy with OpenVINOTM runtimes optimized on Intel platforms. Optimum intel provides simple API to integrate JPQD into training pipeline for Hugging Face Transformers.

JPQD of BERT-base Model with Optimum Intel

In this blog, we introduce how to apply JPQD to BERT-base model on GLUE benchmark for SST-2 text classification task.

Here is a compression config example with the format that follows NNCF specifications. We specify pruning and quantization in a list of compression algorithms with hyperparameters. The pruning method closely resembles the work of Movement Pruning (Sanh et al., 2020) and Block Pruning For Faster Transformers (Lagunas et al., 2021) for unstructured and structured movement sparsity. Quantization refers to Quantization-aware Training (QAT), see details for QAT in previous blog. At the beginning of training, the model under optimization will be initialized with pruning and quantization operators with this configuration.

compression_config = [
        "algorithm":  "movement_sparsity",
        "params": {
            "warmup_start_epoch":  1,
            "warmup_end_epoch":    4,
            "importance_regularization_factor":  0.01,
            "enable_structured_masking":  True
        "sparse_structure_by_scopes": [
            {"mode":  "block",   "sparse_factors": [32, 32], "target_scopes": "{re}.*BertAttention.*"},
            {"mode":  "per_dim", "axis":  0,                 "target_scopes": "{re}.*BertIntermediate.*"},
            {"mode":  "per_dim", "axis":  1,                 "target_scopes": "{re}.*BertOutput.*"},
        "ignored_scopes": ["{re}.*NNCFEmbedding", "{re}.*pooler.*", "{re}.*LayerNorm.*"]
        "algorithm": "quantization",
        "weights": {"mode": "symmetric"}
        "activations": { "mode": "symmetric"},
Figure 2: Sparsity level of BERT-base model in two major stages

Figure 2 shows the sparsity level of BERT-base model over the optimization lifecycle, including two major stages:

  • Unstructured sparsification: In the first stage, model weights are gradually sparsified in the grain size specified by "sparse_structure_by_scopes". The BertAttention layers (Multi-Head Attention: MHA) will be sparsified in 32x32 block size, while BertIntermediate, and BertOutput layers (Feed-Forward Network: FFN) will be sparsified in its row or column respectively. The first stage serves as a warmup stage defined by parameter “warmup_start_epoch” and “warmup_end_epoch”. The “importance_regularization_factor” defines regularization factor onweight importance scores. The factor stays zero before warmup stage, and gradually increases during warmup, finally stays at the fixed value after warmup, users might need some heuristics to find a satisfactory trade-off between sparsity and task accuracy.
  • Structured masking and fine-tuning: The first warm-up stage will produce the unstructured sparsified model. Currently, unstructured sparsity optimized inference is only supported on 4th Gen Intel® Xeon® Scalable Processors with OpenVINO 2022.3 or a later version, for details, please refer to Accelerate Inference of Sparse Transformer Models with OpenVINO™ and 4th Gen Intel® Xeon®Scalable Processors. But it is possible to discard some sparse structure entirely from the model to save compute and memory footprint. NNCF provides a mechanism to achieve structured masking by “enable_structured_masking”: true, where it automatically resolves the structured masking between dependent layers and rewinds the sparsified parameters that do not participate in acceleration for task modeling. As Figure 2 shows, the sparsity level has dropped after “warmup_end_epoch” due to structured masking and the model will continue to be fine-tuned.

Known limitation: currently structured pruning with movement sparsity only supports BERT, Wav2vec2, and Swin family of models. See here for more information.

For distillation, the teacher model can be loaded with transformer API, e.g., a BERT-large pre-trained model from Hugging Face Hub. OVTrainingArguments extends transformers’ TrainingArguments with distillation hyperparameters, i.e., distillation weight and temperature for ease of use. The snippet below shows how we load a teacher model and create training arguments with OVTrainingArguments. Subsequently, the teacher model, with the instantiated OVConfig and OVTrainingArguments is fed to OVTrainer. The rest of the pipeline is identical to the native transformers' training, while internally the training is applied with pruning, quantization, and distillation.

from import OVConfig, OVTrainer, OVTrainingArguments

# Load teacher model
teacher_model = AutoModelForSequenceClassification.from_pretrained(

ov_config = OVConfig(compression=compression_config)

trainer = OVTrainer(
    args=OVTrainingArguments(save_dir, num_train_epochs=1.0, do_train=True,
                             do_eval=True, distillation_temperature=3, distillation_weight=0.9),

# Train the model like usual, internally the training is applied with pruning, quantization and distillation
train_result = trainer.train()
metrics = trainer.evaluate()
# Export the quantized model to OpenVINO IR format and save it

Besides, NNCF provides JPQD examples of othertasks, e.g., question answering. Please refer to the examples provided here.

End-to-End JPQD of BERT-base Demo

Set up Python environment with necessary dependencies.

conda create -n optimum_intel python=3.8
conda activate optimum_intel
python -m pip install --upgrade pip
python -m pip install "optimum-intel[openvino,nncf]"@git+
git clone
cd optimum-intel/examples/openvino/text-classification
python -m pip install -r requirements.txt

Run text classification example with JPQD of BERT on GLUE

python \
    --model_name_or_path bert-base-uncased \
    --task_name $TASK_NAME \
    --teacher_model_name_or_path yoshitomo-matsubara/bert-large-uncased-sst2 \
    --nncf_compression_config ./configs/bert-base-jpqd.json \
    --distillation_weight 0.9 \
    --output_dir ./jpqd-bert-base-ft-$TASK_NAME \
    --overwrite_output_dir \
    --do_train \
    --do_eval \
    --max_seq_length 128 \
    --per_device_train_batch_size 32 \
    --learning_rate 2e-5 \
    --optim adamw_torch \
    --num_train_epochs 5 \
    --logging_steps 10 \
    --evaluation_strategy steps \
    --eval_steps 250 \
    --save_strategy epoch \
    --save_total_limit 3 \
    --fp16 \
    --seed 42

All JPQD configurations and results are saved in ./jpqd-bert-base-ft-$TASK_NAME directory. Optimized OpenVINOTM IR is generated for efficient inference on intel platforms.

BERT-base Performance Evaluation and Accuracy Verification on Xeon

Table 1: BERT-base model for text classification task performance and accuracy verification results

Table 1 shows BERT-base model for text classification task performance evaluation and accuracy verification results on 4th Gen Intel® Xeon® Scalable Processors. BERT-base FP32 model serves as the baseline. BERT-base INT8 (QAT) refers to the model optimized with the 8-bit quantization method only. BERT-base INT8 (JPQD) refers to the model optimized by pruning, quantization, and distillation method.

Here we use benchmark app with performance hint “throughput” to evaluate model performance with input sequence length=128.

As results shows, BERT-base INT8 (QAT) can already reach a 2.39x compression rate and 3.17x performance gain without significant accuracy drop (1.3%) on SST-2 compared with baseline. BERT-base INT8 (JPQD) can further increase compression rate to 5.24x to reach 4.19x performance improvement while keeping minimal accuracy drop (<1%) on SST-2 compared with baseline.

Table 2: BERT-base model for question answering task performance and accuracy verification results

With proper fine-tuning, JPQD can even improve model accuracy while increasing performance in the meantime. Table 2 shows BERT-base model for question answering task performance evaluation and accuracy verification results on 4th Gen Intel® Xeon® Scalable Processors. BERT-base INT8 (JPQD) can increase compression rate to 5.15x to reach 4.25x performance improvement while improving Exact Match (1.35%) and F1 score (1.15%) metric on SQuAD compared with FP32 baseline.

Figure 3: Visualization of BERT-base model parameter counts per layer after JPQD

Figure 3 shows the visualization of parameter counts per layer in the BERT-base model optimized by JPQD for the text classification task. You can find that fully connected layers are actually “dense”, while most (Multi-Head Attention) MHA layers will be much sparser compared to the original model.

Figure 4: BERT-base Multi-Head Attention (MHA) per layer after JPQD

Figure 4 shows MHA head counts per layer in the BERT-base model optimized by JPQD for the text classification task, where active (blue) refer to remaining MHA head counts, while pruned (grey) refers to removed MHA head counts. Instead of pruning uniformly across all MHA heads in transformer layers, we observed that JPQD tends to preserve the weight to the lower layers while heavily pruning the highest layers, similar to experimental results from Movement Pruning (Sanh et al., 2020).


In this blog, we introduce a Joint Pruning, Quantization, and Distillation (JPQD) method to accelerate transformers inference on intel platforms. Here are three key takeaways:

  • Optimum Intel provides simple API to integrate JPQD into training pipeline to enable pruning, quantization, and distillation in parallel during transfer learning of a pre-trained transformer. Optimized OpenVINOTM IR will be generated for efficient inference on intel architecture.
  • BERT-base INT8 (JPQD) model for text classification task can reach 5.24x compression rate, leading to 4.19x performance improvement on 4th Gen Intel® Xeon® Scalable Processors while keeping minimal accuracy drop (<1%) on SST-2 compared with BERT-base FP32 models.
  • BERT-base INT8 (JPQD) model for question answering task can reach 5.15x compression rate to achieve 4.25x performance improvement on 4th Gen Intel® Xeon® Scalable Processors while improving Exact Match (1.35%) and F1 score (1.15%) metric on SQuAD compared with BERT-base FP32 model.



OpenVINO™ Frontend Extension Samples with ConversionExtension  

March 7, 2023

Authors: Wenyi Zou, Su Yang

The OpenVINO™ Frontend extension API enables the mapping of custom operations from framework model representation to OpenVINO representation. In this blog, two samples focus on the mapping to multiple operations with the ConversionExtension in practice.

Sample One: grid_sampler

This sample explains how to use Frontend ConversionExtension classes to facilitate the mapping of custom operations from ONNX model representation to OpenVINO™ representation. It enables writing arbitrary code to replace a single framework operation with multiple connected OpenVINO™ operations constructing dependency graph of any complexity.

When convert the ONNX model BEVFormer tiny to OpenVINO IR, the following error will occur.

Figure 1.0 error

Network BEVFormer tiny viewing with Netron, we can see the node of grid_sampler.  As shown in Figure 1.1.

Figure 1.1 grid_sampler node of BEVFormer tiny

ONNX Nodes

Computation nodes are comprised of a name, the name of an operator that it invokes, a list of named inputs, a list of named outputs, and a list of attributes.

Input and outputs are positionally associated with operator inputs and outputs. Attributes are associated with operator attributes by name.

They have the following properties:

Figure 1.2 node properties

According to the node properties of ONNX, the node grid_sampler_631 op_type is grid_sampler, the domain is mmdeploy. We can use ov::frontend::onnx::ConversionExtension to set the domain paramerter.

#include <map>
#include <iterator>
#include <memory>
#include <sstream>
#include <string>
#include <vector>

#include "openvino/openvino.hpp"
#include <openvino/core/extension.hpp>
#include <openvino/core/op_extension.hpp>
#include <openvino/frontend/extension.hpp>
#include <openvino/opsets/opset9.hpp>
#include <openvino/frontend/node_context.hpp>
#include <openvino/frontend/onnx/extension/conversion.hpp>

int tmain(int argc, tchar* argv[]) {
    // -------- Step 1. Initialize OpenVINO Runtime Core --------
    ov::Core core;
    // -------- Step 2. Add Extension --------
    ov::frontend::onnx::ConversionExtension("grid_sampler", "mmdeploy", [](const ov::frontend::NodeContext& node) {
        ov::opset9::GridSample::Attributes attributes{};
        std::map<int, std::string> mapping, padmapping;
        mapping.insert(std::make_pair(0, "bilinear"));
        mapping.insert(std::make_pair(1, "bicubic"));
        mapping.insert(std::make_pair(2, "nearest"));
        padmapping.insert(std::make_pair(0, "zeros"));
        padmapping.insert(std::make_pair(1, "border"));
        padmapping.insert(std::make_pair(2, "reflection"));
        attributes.align_corners = node.get_attribute<int64_t>("align_corners");
        std::string interp_str = mapping.find(node.get_attribute<int64_t>("interpolation_mode"))->second;
        std::string pad_str = padmapping.find(node.get_attribute<int64_t>("padding_mode"))->second;
        attributes.mode = ov::EnumNames<ov::opset9::GridSample::InterpolationMode>::as_enum(interp_str);
        attributes.padding_mode = ov::EnumNames<ov::opset9::GridSample::PaddingMode>::as_enum(pad_str);
        return ov::OutputVector{
            std::make_shared<ov::opset9::GridSample>(node.get_input(0), node.get_input(1), attributes)};

    // -------- Step 3. Read an ONNX model --------
    std::string model_path;
    std::shared_ptr<ov::Model> model = core.read_model(model_path=”./ bevformer_tiny_epoch_24.onnx”);
    //-------- Step 4. Serialize network to OpenVINO IR and weights files--------
    serialize(model, xml_path="./bevformer_tiny_epoch_24.xml");
    return EXIT_SUCCESS;

Sample Two: aten::uniform

In the OpenVINO™ documentation, the example illustrates basic knowledge of ConversionExtension, like node object of type NodeContext. Real mapping issues like different node modules(or domains), different input types, and missing attributes are under discussion and solved with the workaround.

To support the VectorNet model, try to export the ONNX model from PyTorch. Unfortunately, aten::uniform (ATen is PyTorch’s built-in tensor library) isn’t yet supported by onnx. But OpenVINO™ has RandomUniform operation. Comparing the PyTorch Uniform operation with the RandomUniform operation (generates random numbers from a uniform distribution in the range [minval, maxval)), it shows the same math task with the different input types. Therefore, It’s possible to use Frontend Extensions to map this uniform distribution operation with the onnx model if solving the potential mapping issues. As one-to-one mapping is impossible, decomposition to multiple operations (at least Op Convert additionally) should be considered.

Export Model with Fallback

Because support has not been added to convert a particular torch op to ONNX, we cannot export each ATen op (in the TorchScript namespace “aten”) as a regular ONNX op. So, we fall back to exporting an ATen op with OperatorExportTypes.ONNX_ATEN_FALLBACK.

To optimize the onnx model with OpenVINO™ , create a new sample based on the C++ hello_classification in Linux.

 ~/workspace/openvino22.3/openvino/install/samples/cpp$ ./ -b .
$ ./intel64/Release/hello_extension ./hello_extension/vectornet1.onnx 

Error: Check 'unknown_operators.empty()' failed at src/frontends/onnx/frontend/src/core/graph.cpp:213: OpenVINO™ does not support the following ONNX operations: org.pytorch.aten.Aten.

Visualize Graph for Mapping

In Netron, we could find 6 ATen nodes with the same input values. The obvious mapping problem is that the attribute uniform of node aten should be the node type, while the additional node’s domain is org.pytorch.aten. So, we use ov::frontend::onnx::conversion to set domain parameter, which is similar to the sample one.

Figure 2.1 node properties

As below, real attributes of PyTorch uniform operation aren’t available in the ONNX. The needed attributes of OpenVINO™ RandomUniform operation are output_type, global_seed, and op_seed.

Note: Types are int32 or int64, while uniform op is float64 in the figure.

Figure 2.2 node attributes

As a workaround, we set the seed of attributes as a constant because of the missing aten::uniform attributes.

To solve the difference between aten::uniform and RandomUniform, the mapping issue could be solved as below:

  • Use Op ShapeOf to get the 1D tensor of the input shape.
  • Use Op Convert to convert the input types from aten::uniform’s f64 to RandomUniform’s i64.
  • Use Op Add the input with the Op Constant “117” and Op Multiply with the Op Constant “0.001”, because the output value of the upstream Op ConstantOfShape_output_0 is “0” and the real inputs of all six aten::uniform’s “minval” and “maxval” are “-0.11785113…” and “0.11785113…”.

Add Extension in Practice

Debug steps of the Frontend extension on Windows Visual Studio:

  1. Add add_extension code into C++ sample and build project
  2. Debug with onnx file path

Thanks to the NODE_VALIDATION_CHECK from random_uniform Op, the debug is friendly to the new user.

Code sample of the core.add_extension function

    ov::frontend::onnx::ConversionExtension("ATen", "org.pytorch.aten", [](const ov::frontend::NodeContext& node) {
        ov::element::Type type;
        type = ov::element::Type_t::i64;
        auto input_0 = std::make_shared<ov::opset9::ShapeOf>(node.get_input(0), ov::element::i64);
        auto input_1 = std::make_shared<ov::opset9::Convert>(node.get_input(1), ov::element::i64);
        auto add_constant_1 = ov::opset9::Constant::create(ov::element::i64, ov::Shape{1}, {-117});
        auto input_1_a = std::make_shared<ov::opset9::Add>(input_1, add_constant_1);
        auto input_2 = std::make_shared<ov::opset9::Convert>(node.get_input(2), ov::element::i64);
        auto add_constant_2 = ov::opset9::Constant::create(ov::element::i64, ov::Shape{1}, {117});
        auto input_2_a = std::make_shared<ov::opset9::Add>(input_2, add_constant_2);
        auto output_i64 = std::make_shared<ov::opset9::RandomUniform>(input_0, input_1_a, input_2_a, type, 1, 1);
        auto output_f64 = std::make_shared<ov::opset9::Convert>(output_i64, ov::element::f64);
        auto mul_constant = ov::opset9::Constant::create(ov::element::f64, ov::Shape{1}, {0.001});
        return ov::OutputVector{std::make_shared<ov::opset9::Multiply>(output_f64, mul_constant)};

See Also


Reduce OpenVINO Model Server Latency with In-Process C-API

January 31, 2023

Starting with the 2022.3 release, OpenVINO Model Server (OVMS) provides a C-API that allows OVMS to be linked directly into a C/C++ application as a dynamic library. Existing AI applications can leverage serving functionalities while running inference locally without networking latency overhead.  

The ability to bypass gRPC/REST endpoints and send input data directly from in-process memory creates new opportunities to use OpenVINO locally while maintaining the benefits of model serving. For example, we can combine the benefits of using OpenVINO Runtime with model configuration, version management and support for both local and cloud model storage.

Figure 1. High Level Diagram of C-API Usage

OpenVINO Model Server is typically started as a separate process or run in a container where the client application communicates over a network connection. Now, as you can see above, it is possible to link the model server as a shared library inside the client application and use the internal C API to execute internal inference methods.

We demonstrate the concept in a simple example below and show the impact on latency.

Example C-API Usage

NOTE: complete end to end inference demonstration via C-API with example app can be found here:  

To start using the Model Server C-API, we need to prepare a model and configuration file. Download an example dummy model from our GitHub repo and prepare a config.json file to serve this model. “Dummy” model adds value 1 to all numbers inside an input.

Download Model

wget{xml,bin} -P models/dummy/1

Create Config File

    "model_config_list": [ 
        {"config": { 
                "name": "dummy", 
                "base_path": "./models/dummy"}} 


Next, download and unpack the OVMS library. The library can be obtained from GitHub release page. There are 2 packages – one for Ubuntu 20 and one for RedHat 8.7. There is also documentation showing how to build the library from source. For purpose of this demo, we will use the Ubuntu version:

wget && tar -xvf ovms_ubuntu.tar.gz

Start Server

To start the server, use ServerStartFromConfigurationFile. There are many options, all of which are documented in the header file. Let’s launch the server with configuration file and optional log level error:

OVMS_ServerSettings* serverSettings; 
OVMS_ModelsSettings* modelsSettings; 
OVMS_Server* srv; 
OVMS_ServerSettingsSetLogLevel(serverSettings, OVMS_LOG_ERROR);  // Make the serving silent 
OVMS_ModelsSettingsSetConfigPath(modelsSettings, "./config.json");  // Previously created file 
OVMS_ServerStartFromConfigurationFile(srv, serverSettings, modelsSettings);  // Start the server 

Input Data Preparation

Use OVMS_InferenceRequestInputSetData call, to provide input data with no additional copy operation. In InferenceRequestNew call, we can specify model name (the same as defined in config.json) and specific version (or 0 to use default). We also need to pass input names, data precision and shape information. In the example we provide 10 subsequent floating-point numbers, starting from 0.

const char* MODEL_NAME = "dummy"; 
const uint64_t MODEL_VERSION = 1; 
const char* INPUT_NAME = "b"; 
constexpr size_t NUM_OF_ELEMENTS = 10; 
constexpr std::array SHAPE = {1, NUM_OF_ELEMENTS}; 
OVMS_InferenceRequest* request; 
OVMS_InferenceRequestNew(&request, srv, MODEL_NAME, MODEL_VERSION); 
OVMS_InferenceRequestAddInput(request, INPUT_NAME, OVMS_DATATYPE_FP32,, SHAPE.size()); 
std::array data{0, 1, 2, 3, 4, 5, 6, 7, 8, 9}; 
OVMS_InferenceRequestInputSetData(request, INPUT_NAME,, sizeof(data), OVMS_BUFFERTYPE_CPU, 0); 

Invoke Synchronous Inference

Simply call OVMS_Inference. This is required to pass response pointer and receive results in the next steps.

OVMS_InferenceResponse* response; 
OVMS_Inference(srv, request, &response); 

Read Results

Use call OVMS_InferenceResponseGetOutput API call to read the results. There are bunch of metadata we can read optionally, such as: precision, shape, buffer type and device ID. The expected output after addition should be:

1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
const char* outputName; 
OVMS_DataType dataType; 
const uint64_t* shape; 
uint32_t dimCount; 
const void* outputData; 
size_t byteSize; 
OVMS_BufferType bufferType; 
uint32_t deviceId; 
OVMS_InferenceResponseGetOutput(response, 0, 
        &outputName, &dataType, &shape, &dimCount, &outputData, &byteSize, &bufferType, &deviceId); 
for (int i = 0; i < NUM_OF_ELEMENTS; i++) 
std::cout << ((float*)outputData)[i] << ", "; 
std::cout << std::endl;

Check the header file to learn more about the supported methods and their parameters.

Compile and Run Application

In this example we omitted error handling and resource cleanup upon failure. Please refer to the full demo instructions for a more complete example.

Performance Analysis

Using benchmarking tools from OpenVINO Runtime and both the C-API and gRPC API in OpenVINO Model Server, we can compare inference results via C-API to typical scenario of gRPC or direct integration of OpenVINO Runtime. The Resnet-50-tf model from Open Model Zoo was used for the testing below.

Figure 2. Inference Latency Measurement for ResNet-50 with each deployment option (lower is better)
Figure 2. Inference Latency Measurement for ResNet-50 with each deployment option (lower is better)

Hardware configuration used:

- 1-node, Intel Xeon Gold 6252 @ 2.10GHz processor with 256GB (8 slots/16GB/2666) total DDR memory, HT on, Turbo on, Ubuntu 20.04.2 LTS,5.4.0-109-generic kernel

- Intel S2600WFT motherboard

Tested by Intel on 01/31/2023.


With the new method of embedding OVMS into C++ applications, users can decrease inference latency even further by entirely skipping the networking part of model serving. The C-API is still in preview and has some limitations, but in its current state is ready to integrate into C++ applications. If you have questions or feedback, please file an issue on GitHub.

Read more:


Make Your Own YOLOv8 OpenVINO™ Model from Any Data Format with Datumaro

May 3, 2023

Authors: Vinnam Kim, Wonju Lee, Mark Byun, Minje Park

Dataset formats and trainer


OpenVINO™ provides an easy way to deploy your model with the best inference performance on any Intel hardwares. However, to train your own model for deployment you need to prepare a training framework and dataset. Fortunately, there are many ready-to-use training frameworks and implementations. Then, what about the dataset? A specific training framework requires a specific data format, but there are many data formats in the world. For example, in object detection tasks there are data formats such as YOLO, COCO, and Pascal VOC that are widely used. These formats have different directory structures and annotation file formats as well as different extensions such as txt, json, and, xml, respectively. It's tedious task to convert dataset from one format to another whenever you adopt different training framework.

Let's assume you choose Detectron2, which only supports COCO format datasets. If your dataset is formatted as VOC, you have to convert it into COCO format. Below, we compare the directory structures and annotation file formats of both datasets, VOC and COCO. These datasets have distinct formats and you need to implement codes for format conversion at each time of handling different formats. Of course, this is not technically challenging but this may require tedious code work and debugging for several days. It won't be good to repeat this process if you intend to add more datasets with different formats.

Directory structures of VOC and COCO (Click me!)

└─ Dataset/
   ├── Annotations/
   │     ├── ann1.xml
   │     ├── ann2.xml
   │     └── ...
   ├── JPEGImages/
   │    ├── img1.jpg
   │    ├── img2.jpg
   │    └── ...
   ├── SegmentationClass/
   │    ├── img1.png
   │    ├── img2.png
   │    └── ...
   ├── SegmentationObject/
   │    ├── img1.png
   │    ├── img2.png
   │    └── ...
   └── ImageSets/
        ├── Main/
        │   ├── test.txt
        |   ├── train.txt
        |   └── ...
        ├── Layout/
        │   ├── test.txt
        |   ├── train.txt
        |   └── ...
        ├── Action/
        │   ├── test.txt
        |   ├── train.txt
        |   └── ...
        └── Segmentation/
            ├── test.txt
            ├── train.txt
            └── ...

└─ Dataset/
    ├── images/
    │   ├── train/
    │   │   ├── img1.jpg
    │   │   ├── img2.jpg
    │   │   └── ...
    │   └── val/
    │       ├── img1.jpg
    │       ├── img2.jpg
    │       └── ...
    └── annotations/
        ├── instances_train.json
        └── ...
Annotation file formats of VOC and COCO (Click me!)
VOC (xml) COCO (json)
<?xml version="1.0" encoding="UTF-8"?>
  "licenses": [
      "name": "",
      "id": 0,
      "url": ""
  "info": {
    "contributor": "",
    "date_created": "",
    "description": "",
    "url": "",
    "version": "",
    "year": ""
  "categories": [
      "id": 1,
      "name": "cat",
      "supercategory": ""
      "id": 2,
      "name": "dog",
      "supercategory": ""
  "images": [
      "id": 5,
      "width": 10,
      "height": 5,
      "file_name": "img1.jpg",
      "license": 0,
      "flickr_url": "",
      "coco_url": "",
      "date_captured": 0
  "annotations": [
      "id": 1,
      "image_id": 5,
      "category_id": 2,
      "segmentation": [],
      "area": 3.0,
      "bbox": [
      "iscrowd": 0

Dataset Management Framework (Datumaro) is a framework that provides Python API and CLI tools to convert, transform, and analyze datasets. Among the many features of Datumaro, we would like to introduce the data format conversion feature on this blog, which is one of the fundamental feature for handling many datasets with different training frameworks. Datumaro supports the import and export of over 40 computer vision data formats (please take a look at supported formats for details!). This means that you can easily change your data format through Datumaro. If your model training framework can only read specific formats, don't worry. Use Datumaro and convert it!

Train YOLOv8 model and export it to OpenVINO™ model

  • Prepare dataset
  • Convert dataset with Datumaro
  • Train with YOLOv8 and export to OpenVINO™ IR

YOLOv8 is a well-known model training framework for object detection and tracking, instance segmentation, image classification, and pose estimation tasks. It provides simple CLI commands to train, test, and export a model to OpenVINO™ Intermediate Representation (IR). However, the data format consumed by YOLOv8 is slightly different from the YOLO format itself. Datumaro named it refers to it as YOLO-Ultralytics format. As you can see here, it requires a special meta file to indicate annotation files for each subset and subset files to list subset image files. It further requires them to be placed in an appropriate directory structure. It can be very tedious to go through these details and implement dataset preprocessing when you want to train a model on your custom dataset.

On this blog, we provide an end-to-end example that covers the complete process of converting your dataset, training a model with the converted dataset, and exporting the trained model to OpenVINO™ IR. We understand that dataset conversion can be a tricky process, especially if you have annotated and built your own dataset. Therefore, we will provide an example of converting the dataset created by the popular CVAT annotation tool. By following our step-by-step guide, you will be able to convert your data format easily and accelerate the inference of your trained model with OpenVINO™.

Prepare dataset

In this section, we introduce the steps to export the project annotated by CVAT for the following workflows. You can skip this section if your dataset is formatted as a different data format and is ready to be imported by Datumaro.

CVAT project for object detection task with train, val, and test subsets

NOTE: We used the cats-and-dogs dataset for this example. You can find the reference for this dataset here.

NOTE: You should have three subsets in your project: "train", "val", and "test" (optional). If your dataset has different subset names, you have to rename them. You can do this by using Datumaro's MapSubsets transform.

We export this project to CVAT for images 1.1 data format. Datumaro can import this data format and export it to YOLO-Ultralytics format which can be consumed by YOLOv8.

Export CVAT project to CVAT for images 1.1 data format

Export CVAT project to CVAT for images 1.1 data format. After exporting the dataset, extract it to the cvat_dataset directory.

ls yolo_v8_dataset

You can see the following directory structure:

annotations.xml  images

Convert your dataset using Datumaro

You can convert the dataset located in cvat_dataset using Datumaro's CLI command as follows. For a detailed explanation of the input arguments, see here.

datum convert -i cvat_dataset -if cvat -f yolo_ultralytics -o yolo_v8_dataset -- --save-media

NOTE: If your dataset is not CVAT for images 1.1 format, you can replace -if cvat with the different input format as -if INPUT_FORMAT. Use datum detect CLI command to figure out what format your dataset is.

After the conversion, you can see that yolo_v8_dataset directory is created.

ls yolo_v8_dataset

This directory is structured as follows.

data.yaml  images  labels  test.txt  train.txt  val.txt

Train with YOLOv8 Trainer and Export to OpenVINO™ IR

In this section, we will train the YOLOv8 detector with the dataset converted in the previous section. To train a YOLOv8 detector, please execute the following command.

yolo detect train data=$(realpath yolo_v8_dataset/data.yaml) project=my-project

NOTE: We use data=$(realpath yolo_v8_dataset/data.yaml) to convert the relative path yolo_v8_dataset/data.yaml to the absolute path. This is because YOLOv8 needs the absolute path for the custom dataset.

After the training, the following command enables testing on the test dataset.

yolo detect val model=my-project/train/weights/ data=$(realpath yolo_v8_dataset/data.yaml) split=test

Lastly, we will export your YOLOv8 detector to OpenVINO™ IR for inference acceleration on Intel devices.

yolo detect export model=my-project/train/weights/ format=openvino

Using this command, the exported IR is created at this directory path, my-project/train/weights/best_openvino_model.

ls my-project/train/weights/best_openvino_model
best.bin  best.xml  metadata.yaml


This post provided an example of training a YOLOv8 detector on an arbitrary data format by utilizing the data format conversion feature of Datumaro and exporting the model to OpenVINO™ IR. You can refer to the executable Jupyter notebook example provided on this blog post here for step-by-step guide. Datumaro offers a range of useful features for managing datasets beyond data format conversion. You can find examples of other Datumaro features, such as noisy label detection during training with OpenVINO™ Training Extensions, in the Jupyter examples directory. For more information about Datumaro and its capabilities, you can visit the Datumaro documentation page. If you have any questions or requests about using Datumaro, feel free to open an issue here.


OpenVINO optimizer Latent Diffusion Models (LDM) for super-resolution

April 23, 2023

OpenVINO optimizer Latent Diffusion Models(LDM) for super-resolution


A computer vision approach called image super-resolution aims to increase the resolution of low-resolution images so that they are clearer and more detailed. Applicationsfor super-resolution include the processing of medical images, surveillancefootage, and satellite images. 

Figure1-super-resolution effect display

The LDM (LatentDiffusion Models) Super Resolution model, a deep learning-based approach to photo super-resolution, was developed by the Hugging Face Research team. The residual network (ResNet) architecture, a type of convolutional neural network(CNN) created to address the issue of vanishing gradients in deep neuralnetworks.

Diffusion models are generative models,meaning that they are used to generate data similar to the data on which they are trained. Fundamentally, Diffusion Models work by destroying training data through the successive addition of Gaussian noise, andthen learning to recover the data by reversing this noising process. After training, we can use the Diffusion Model to generatedata by simply passing randomly sampled noise through the learned denoising process.

Figure2-Diffusion Models can be used to generate images from noise

Diffusion Model is a latent variable model which maps to the latent space using a fixed Markov chain. This chain gradually adds noise to thedata in order to obtain the approximate posterior.

Figure3-The Markov chain is manifested for image data

Ultimately, the image is asymptotically transformed to pure Gaussian noise. The goal of training a diffusion model is to learn the reverse process. By traversing backward along this chain, we can generate new data.

Figure4-The process for new image data generation by diffusion


-      Optimum-intel Optimum Intel is the interface betweenthe HuggingFace Transformers and Diffusers libraries and the differenttools and libraries provided by Intel to accelerate end-to-end pipelines onIntel architectures.
Intel Neural Compressor is an open-source library enabling the usageof the most popular compression techniques such as quantization, pruning and knowledge distillation

-      OpenVINO™ is an open-sourcetoolkit for optimizing and deploying AI inference which can boost deep learningperformance in computer vision, automatic speech recognition, natural language processing and other common task.

-      optimum-intel==1.5.2(include openvino)

   - openvino

    - openvino-dev

-      diffusers

-      pytorch >= 1.9.1

-      onnx >= 1.13.0

Reference: optimum-intel-ldm-super-resolution-4x

QuickStart Demo

Original repo is from HuggingFace CompVis/ldm-super-resolution-4x-openimages,we are reference to build our pipeline to implement super-resolution related function.

Figure5-The super resolution pipeline from huggingface

To transformand acceleration optimize the pipeline by openvino, there are 3 steps need to do.

-      Step1. Install the requirement package and initial environment.

-      Step2. Convert original model to openvino IR model.

-      Step3. Build OpenVINO super resolution pipeline.

Now, Let’s start with the content of our tutorial.


Step 1. Install the requirementpackage and initial environment  

OpenVINO has the standard installation process, we can directly refer tothe official OpenVINO documentation to install.

Reference: Install OpenVINO by source code for Linux

Reference: Install OpenVINO by release package

Optimum Intel also can refer the standard guide.

Reference: Optimum-intel install guide

(Optional) Install the latest stable release by pipe :

   # pip install openvino, openvino-dev

   # pip install"optimum[openvino,nncf]"

Step 2. Convert originalmodel to OpenVINO IR model

Firstly, run pipe the HuggingFace pipeline, it will automate download the models, and we need to convert them from pytorch->onnx->IR, to enable the model by OpenVINO.

Figure6-OpenVINO enable HuggingFace model (pytorch base) workflow

The LDM (LatentDiffusion Models) Super Resolution model has two part of sub-models: unet and vqvae,we should convert each of them in to IR model.

Figure7-OpenVINO enable super resolution pipeline workflow

The reference source code for model convert,also we provide the script in the GitHub repo :

Initial parameter and the ov-pipeline

Figure8-Initial parameter and ov-pipeline

Unet sub-model convert to IR

Figure9-Unet sub-model convert to IR

Vqvae sub-model convert to IR

Figure10-Vqvae sub-model convert to IR

Step 3. Build OpenVINOsuper resolution pipeline

The LDM (Latent Diffusion Models) Super Resolution OpenVINO pipeline main function part code, the whole pipeline script is provided in GitHub repo:

Figure11-ov pipeline main function code

Inference Result  

Figure12-Super resolution effect display


Deploy End to End Super-Resolution Pipeline with OpenVINO™ Model Server

April 6, 2023


In this blog, we will show how to deploy an end-to-end super-resolution pipeline by leveraging OpenVINOTM Model Server with Demultiplexing in DAG and Custom Node features.

OpenVINOTM Model Server (OVMS) is a high-performance system for serving models that uses the same architecture and API as TensorFlow Serving and KServe while applying OpenVINOTM for inference execution. It is implemented in C++ for scalability and optimized for deployment on intel architectures.

Directed Acyclic Graph (DAG) is an OVMS feature that controls the execution of an entire graph of interconnected models defined within the OVMS configuration. The DAG scheduler makes it possible to create a pipeline of models for execution in the server with a single client request.

During the pipeline execution, it is possible to split a request with multiple batches into a set of branches with a single batch. Internally, OVMS demultiplexer will divide the data, process them in parallel and combine the results.

The custom node in OVMS simplifies linking deep learning models into complete pipeline. Custom node can be used to implement all operations on the data which cannot be handled by the neural network model. It is represented by a C++ dynamic library implementing OVMS API defined in custom_node_interface.h.

Super-Resolution Pipeline Workflow

Figure1 shows the super-resolution pipeline in a flowchart, where we use "demultiply_counter=3" without loss of generality. The whole pipeline starts with input data from the Request node via gRPC calls. Batched input data with 5D shape(3,1,3,270,480) is split into a single batch by the DAG demultiplexer. Each single batch of data is fed into a custom node for image preprocessing. The two outputs of the custom node serve as inputs for model A inference. In the end, all inference results are gathered as output C, which will be sent by the Response node to the client via gRPC calls.

Figure 1: Super-Resolution Pipeline Workflow in OpenVINO™ Model Server

Here is an example configuration for the super-resolution pipeline deployed with OVMS.

    "model_config_list": [
            "config": {
                "name": "single-image-super-resolution",
                "base_path": "/models/super_resolution_model_preprocessed/",
                "nireq": 1,
                "plugin_config": {
                    "NUM_STREAMS": "1",
                    "INFERENCE_PRECISION_HINT": "bf16",
                    "AFFINITY": "CORE"
    "custom_node_library_config_list": [
            "name": "sr_preprocess",
            "base_path": "/models/"
    "pipeline_config_list": [
            "name": "super_resolution",
            "inputs": [
            "demultiply_count": -1,
            "nodes": [
                    "name": "sr_preprocess_node",
                    "library_name": "sr_preprocess",
                    "type": "custom",
                    "params": {
                        "image_preprocess_method": "SimpleResize",
                        "debug": "false",
                    "inputs": [
                            "image": {
                                "node_name": "request",
                                "data_item": "data"
                    "outputs": [
                            "data_item": "out1",
                            "alias": "output1"
                            "data_item": "out2",
                            "alias": "output2"
                    "name": "super_resolution_node",
                    "model_name": "single-image-super-resolution",
                    "type": "DL model",
                    "inputs": [
                            "0": {
                                "node_name": "sr_preprocess_node",
                                "data_item": "output1"
                            "1": {
                                "node_name": "sr_preprocess_node",
                                "data_item": "output2"
                    "outputs": [
                            "data_item": "129",
                            "alias": "super_resolution_output"
            "outputs": [
                    "129": {
                        "node_name": "super_resolution_node",
                        "data_item": "super_resolution_output"

“pipeline_config_list” contains super-resolution pipeline information, data enter from the “request” node, flow to “sr_preprocess_node” for image preprocessing, generated two outputs will serve as inputs in “super_resolution_node” for inference, gathered inference results will be returned by “response” node.

  • "demultiply_count": acceptable input data batch size when Demultiplexing in DAG feature enabled, “demultiply_count” with value -1 means OVMS can accept dynamic batch input data.

“model_config_list”: contains the basic configuration for super-resolution deep learning model and OpenVINOTM CPU plugin configuration.

  • "nireq": set number of infer requests used in OVMS server for deep learning model
  • "NUM_STREAMS": set number of streams used in the CPU plugin
  • "INFERENCE_PRECISION_HINT": option to select preferred inference precision in CPU plugin. We can set "INFERENCE_PRECISION_HINT":bf16 on the Xeon platform that supports BF16 precision, such as the 4th Gen Intel® Xeon® Scalable processor (formerly codenamed Sapphire Rapids). Otherwise, we should set "INFERENCE_PRECISION_HINT":f32 as the default value.

“custom_node_library_config_list”: contains the name and path of the custom node dynamic library

Image Preprocessing with libvips in Custom Node

In this blog, we use a single-image-super-resolution model from Open Model Zoo for the super-resolution pipeline. The model requires two inputs according to the model specification. The first input is the original image (shape [1,3,270,480]). The second input is a 4x resized image with bicubic interpolation (shape [1,3,1080,1920]). Both input images expected color space is BGR. Therefore, image preprocessing for input image is required.

Figure2: Custom Node for Image Preprocessing in the Super-Resolution Pipeline

Figure2 shows the custom node designed for image preprocessing in the super-resolution pipeline. The custom node takes the original input image as input data. At first, input data is assigned to output 1 without modification. Besides, the input data is resized 4x with bicubic interpolation and assigned as output 2. The two outputs are passed to the model node for inference. For image processing in the custom node, we utilize libvips – an open-source image processing library that is designed to be fast and efficient with low memory usage. Please see the detailed custom node implementation in super_resolution_nhwc.cpp.

Although libvips is very sufficient for image processing operations with less memory, libvips does not provide functionality for layout (NCHW->NHWC) and color space (RGB->BGR) conversion, which is required by the super-resolution model as inputs. Instead, we can integrate layout and color space conversion into models using OpenVINOTM Preprocessing API.

Integrate Preprocessing with OpenVINOTM Preprocessing API

OpenVINOTM Preprocessing API allows adding custom preprocessing steps into the execution graph of OpenVINOTM models.

Here is a sample code to integrate layout (NCHW-> NHWC) and color space (BRG->RGB) conversion into the super-resolution model with OpenVINOTM Preprocessing API.

from openvino.runtime import Core, Layout, Type, serialize
from openvino.preprocess import ColorFormat, PrePostProcessor

core = Core()
input_tensor_name_1 = "0"

model_path = "./super_resolution/1/single-image-super-resolution-1032.xml"
model = core.read_model(model_path)
ppp = PrePostProcessor(model)
# Input 1
model =

In the code snippet above, we first load the original model and initialize the PrePostProcessor object with the original model. Then we modify the model's 1st input element type to “uint8”, change the color format from the default “BGR” to “RGB”, and set the layout from “NCHW” to “NHWC”. In the end, we build a new model and serialize it on the disk. The whole model preprocessing can be done offline, please find details in

Build Model Server Docker Image for Super-Resolution Pipeline

Build OVMS docker image with custom node

git clone -b super_resolution_demo
cd model_server
IMAGE_TAG_SUFFIX=-sr make docker_build

Copy compiled custom nodes library to the “models” directory

cp src/custom_nodes/lib/ubuntu/  models/

Setup client environment

cd models
pip install -r requirement.txt
sudo apt-get install libvips libjpeg-turbo8-dev

Integrate preprocessing with OpenVINOTM Preprocessing API


The resulting model will be saved in the “super_resolution_model_preprocessed/1” directory.

Super-Resolution Pipeline Demo

Start the OpenVINOTM Model Server with docker binding with 8 cores

docker run --cpuset-cpus 0-7 --rm -v ${PWD}:/models -p 9001:9001 \
openvino/model_server:latest-sr \
--config_path /models/image_super_resolution_config_nhwc.json \
--port 9001

Run client with command line

python --grpc_port 9001  --input_images_dir images \
--model_name super_resolution --height 270 --width 480 \
--batch_size 1 -niter 100

Figure 3 shows the original input image (shape 270x480).

Figure 3: Original low resolution Input Image(270x480)

Figure 4 shows the resized image (shape 1080x1920) after image preprocessing in the custom node.

Figure 4: Bicubic Interpolation Resized Image (1080x1920)

Figure 5 shows the inference result of the super-resolution model (shape1080x1920).  

Figure 5: Super-Resolution Model Inference Result (1080x1920)


In this blog, we demonstrate an end-to-end super-resolution pipeline deployment with OpenVINOTM Model Server. The whole pipeline takes dynamic batched images (RGB, NHWC) as input, demultiplexing into single batch data, preprocess with a custom node, runs an inference with a super-resolution model, send gathered inference results to the client in the end.

This blog provides following examples that utilize OpenVINOTM Model Server and OpenVINOTM features: 

  • Enable OVMS DAG demultiplexing feature
  • Provide custom node for image preprocessing using libvips
  • Provide sample code for integrating preprocessing into the model with OpenVINOTM Preprocessing API.
  • Support super-resolution end-to-end pipeline with image preprocessing and model inference with OVMS DAG scheduler

Remote Tensor API Sample

January 28, 2023

This AI pipeline implements zero-copy between SYCL and OpenVINO through the Remote Tensor API of the GPU Plugin.

  1. Introduction

The development of SYCL simplifies the use of OpenCL, which can fully exploit the computing power of GPU in the pipeline. Meanwhile, SYCL has more flexibility to do customized pre- and post-processing of OpenVINO. To further optimize the pipeline, developers can use GPU Plugin to avoid the memory copy overhead between SYCL and OpenVINO. The GPU plugin provides the ov::RemoteContext and ov::RemoteTensor interfaces for video memory sharing and interoperability with existing native APIs, such as OpenCL, Microsoft DirectX, or VAAPI. For details, please refer to the online documentation of OpenVINO.

Based on the pseudocode of the online documentation, here we provide a simple pipeline sample with Remote Tensor API. Because in the rapid iteration of oneAPI, sometimes customers need quick verification so that this sample can be used for testing. OneAPI also provides a real-world, end-to-end example, which optimizes PointPillars for lidar object detection.

  1. Components

SYCL preprocessing is based on the Sepia Filter sample, which demonstrates how to convert a color image to a Sepia tone image, a monochromatic image with a distinctive Brown Gray color. The sample program works by offloading the compute-intensive conversion of each pixel to Sepia tone using SYCL*-compliant code for CPU and GPU.

OpenVINO inferencing is based on the OpenVINO classification sample, the input from SYCL filtered image in the device will be sent into OpenVINO as a remote tensor without a memory copy.

Remote Tensor API: Create RemoteContext from SYCL pre-processing’s native handle. After model compiling, do memory sharing between the application and GPU plugin with from cl::Buffer to remote tensor.

auto cl_queue = get_native<backend::opencl>(sycl_queue);
auto remote_context = ov::intel_gpu::ocl::ClContext(core, cl_queue);
ov::CompiledModel compiled_model = core.compile_model(model, remote_context);
auto cl_buffers = get_native<backend::opencl>(image_buf_out);
auto remote_tensor = remote_context.create_tensor(ov::element::u8, {batch, input_height, input_width, 3}, cl_buffers);
infer_request.set_tensor(input_tensor_name, remote_tensor);
  1. Build Sample on Linux

Download the source code from the link. Prepare the model and images.

To run the sample, you need to specify a model and image:

Use pre-trained models from the Open Model Zoo. The models can be downloaded using the Model Downloader. Use images from the media files collection.

mkdir build
cd build
cmake ..

Run on Intel NUC Core 11 iGPU with OpenVINO 2022.2 and oneAPI 2022.3.

./intel64/hello_nv12_input_classification_oneAPI../model/FP32/alexnet.xml ../image/dog512.bmp GPU 2

Sample Output:

  Loaded image with a width of 512, a height of 512 and 3 channels
[ INFO ] OpenVINO Runtime version ......... 2022.2.0
[ INFO ] Build ........... 2022.2.0-7713-af16ea1d79a-releases/2022/2
[ INFO ] 
[ INFO ] Loading model files: ../model/FP32/alexnet.xml
Running on Intel(R) Iris(R) Xe Graphics [0x9a49]
---Load model - 1126ms
---Create an infer request - 0.515ms
Use remote tensor API and set_tensor
No 1. do inf: 
---sycl buffer  - 0.005ms
---sycl filter total time - 93.342ms
---kernel time: 0.051822 milliseconds
---Set tensor - 0.138ms
---Run infer req - 4.725ms
---get tensor - 0.062ms

Top 3 results:

Image ../image/dog512.bmp

classid probability label
------- ----------- -----
176     0.4579958   "Saluki, gazelle hound"
212     0.2727944   "English setter"
169     0.0513433   "borzoi, Russian wolfhound"

---sum of inference - 98.268ms
No 2. do inf: 
---sycl buffer  - 0.001ms
---sycl filter total time - 0.582ms
---kernel time: 0.05177 milliseconds
---Set tensor - 0.086ms
---Run infer req - 4.162ms
---get tensor - 0.049ms

Top 3 results:

Image ../image/dog512.bmp

classid probability label
------- ----------- -----
176     0.4579958   "Saluki, gazelle hound"
212     0.2727944   "English setter"
169     0.0513433   "borzoi, Russian wolfhound"

---sum of inference - 4.881ms 

Warning: With the updating of OpenVINO and oneAPI, different versions may cause problems with the tools in the common directory or the new SYCL header name. Please use the same version or debug following the corresponding release instructions.


Accelerate Inference of Sparse Transformer Models with OpenVINO™ and 4th Gen Intel® Xeon® Scalable Processors

Authors: Alexander Kozlov, Vui Seng Chua, Yujie Pan, Rajesh Poornachandran, Sreekanth Yalachigere, Dmitry Gorokhov, Nilesh Jain, Ravi Iyer, Yury Gorbachev


When it comes to the inference of overparametrized Deep Neural Networks, perhaps, weight pruning is one of the most popular and promising techniques that is used to reduce model footprint, decrease the memory throughput required for inference, and finally improve performance. Since Language Models (LMs) are highly overparametrized and contain lots of MatMul operations with weights it looks natural to prune the redundant weights and benefit from sparsity at inference time. There are several types of pruning methods available:

  • Fine-grained pruning (single weights).
  • Coarse pruning: group-level pruning (groups of weights), vector pruning (rows in weights matrices), and filter pruning (filters in ConvNets).

Contemporary Language Models are basically represented by Transformer-based architectures. Using coarse pruning methods for such models is problematic because of the many connections between the layers. This trait means that, first, not every pruning type is applicable to such models and, second, pruning of some dimension in one layer requires adjustments in the rest of the layers connected to it.

Fine-grained sparsity does not have such a constraint and can be applied to each layer independently. However, it requires special support on the HW and inference SW level to get real performance improvements from weight sparsity. There are two main approaches that help to leverage from weight sparsity at inference:

  • Skip multiplication and addition for zero weights in dot products of weights and activations. This usually results in a special instruction set that implements such logic.
  • Weights compression/decompression to reduce the memory throughput. Compression is performed at the model load/compilation stage while decompression happens on the fly right before the computation when weights are in the cache. Such a method can be implemented on the HW or SW level.

In this blog post, we focus on the SW weight decompression method and showcase the end-to-end workflow from model optimization to deployment with OpenVINO.

Sparsity support in OpenVINO

Starting from OpenVINO 2022.3release, OpenVINO runtime contains a feature that enables weights compression/decompression that can lead to performance improvement on the 4thGen Intel® Xeon® Scalable Processors. However, there are some prerequisites that should be considered to enable this feature during the model deployment:

  • Currently, this feature is available only to MatMul operations with weights (Fully-connected layers). So currently, there is no support for sparse Convolutional layers or other operations.
  • MatMul layers should contain a high level of weights sparsity, for example, 80% or higher which is achievable, especially for large Transformer models trained on simple tasks such as Text Classification.
  • The deployment scenario should be memory-bound. For example, this prerequisite is applicable to cloud deployment when there are multiple containers running inference of the same model in parallel and competing for the same RAM and CPU resources.

The first two prerequisites assume that the model is pruned using special optimization methods designed to introduce sparsity in weight matrices. It is worth noting that pruning methods require model fine-tuning on the target dataset in order to reduce accuracy degradation caused by zeroing out weights within the model. It assumes the availability of the HW capable of DL model training. Nowadays, many frameworks and libraries offer such methods. For example, PyTorch provides some capabilities for NN pruning. There are also resources that offer pre-trained sparse models that can be used as a starting point, for example, SparseZoo from Neural Magic.

OpenVINO also provides instruments for DL model pruning implemented in Neural Network Compression Framework (NNCF) that is aimed specifically for model optimization and offers different optimization options: from post-training optimization to deep compression when stacking several optimization methods. NNCF is also integrated into Hugging Face Optimum library which is designed to optimize NLP models from Hugging Face Hub.

Using only sparsity is not so beneficial compared to another popular optimization method such as bit quantization which can guarantee better performance-accuracy trade-offs after optimization in the general case. However, the good thing about sparsity is that it can be stacked with 8-bit quantization so that the performance improvements of one method reinforce the optimization effect of another one leading to a higher cumulative speedup when applying both. Considering this, OpenVINO runtime provides an acceleration feature for sparse and 8-bit quantized models. The runtime flow is shown in the scheme below:

Below, we demonstrate two end-to-end workflows:

  • Pruning and 8-bit quantization of the floating-point BERT model using Hugging Face Optimum and NNCF as an optimization backend.
  • Quantization of sparse BERT model pruned with 3rd party optimization solution.

Both workflows end up with inference using OpenVINO API where we show how to turn on a runtime option that allows leveraging from sparse weights.

Pruning and 8-bit quantization with Hugging Face Optimum and NNCF

This flow assumes that there is a Transformer model coming from the Hugging Face Transformers library that is fine-tuned for a downstream task. In this example, we will consider the text classification problem, in particular the SST2 dataset from the GLUE benchmark, and the BERT-base model fine-tuned for it. To do the optimization, we used an Optimum-Intel library which contains the optimization capabilities based on the NNCF framework and is designed for inference with OpenVINO. You can find the exact characteristics and steps to reproduce the result in this model card on the Hugging Face Hub. The model is 80% sparse and 8-bit quantized.

To run a pre-optimized model you can use the following code from this notebook:

from pathlib import Path

from import OVModelForSequenceClassification
from transformers import AutoTokenizer, pipeline
from huggingface_hub import hf_hub_download

model_id = "OpenVINO/bert-base-uncased-sst2-int8-unstructured80"
ov_model = OVModelForSequenceClassification.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)

Quantization of already pruned model

In case if you deal with already pruned model, you can use Post-Training Quantization from the Optimum-Intel library to make it 8-bit quantized as well. The code snippet below shows how to quantize the sparse BERT model optimized for MNLI dataset using Neural Magic SW solution. This model is publicly available so that we download it using Optimum API and quantize on fly using calibration data from MNLI dataset. The code snippet below shows how to do that.

from functools import partial
from pathlib import Path

from datasets import load_dataset
from transformers import AutoModelForSequenceClassification, AutoTokenizer

from import OVQuantizer
from import OVConfig

model_id = "neuralmagic/oBERT-12-downstream-pruned-unstructured-90-mnli"
quantized_sparse_dir = Path("bert_90_sparse_quantized")

# Instantiate model and tokenizer in PyTorch and load them from the HF Hub
torch_model = AutoModelForSequenceClassification.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)

def preprocess_function(examples, tokenizer):
    Define a function that tokenizes the data and returns it in the format expected by the model.
    :param: examples: a dictionary containing the input data which are the items from caliration dataset.
            tokenizer: a tokenizer object that is used to tokenize the text data.
            the data that can be fed directly to the model.
    return tokenizer(
        examples["premise"], examples["hypothesis"], padding="max_length", max_length=128, truncation=True

# Create quantization config (default) and OVQuantizer
# OVConfig is a wrapper class on top of NNCF config. 
# Use "compression" field to control quantization parameters
# For more information about the parameters refer to NNCF GitHub documentatioin
quantization_config = OVConfig()
quantizer = OVQuantizer.from_pretrained(torch_model, feature="sequence-classification")

# Instantiate a dataset and convert it to calibration dataset using HF API
# The latter one produces a model input
dataset = load_dataset("glue", "mnli")
calibration_dataset = quantizer.get_calibration_dataset(
    preprocess_function=partial(preprocess_function, tokenizer=tokenizer),
# Apply static quantization and export the resulting quantized model to OpenVINO IR format
    quantization_config=quantization_config, calibration_dataset=calibration_dataset, save_directory=quantized_sparse_dir

Enabling sparsity optimization inOpenVINO Runtime and 4th Gen Intel® Xeon® Scalable Processors

Once you get ready with the sparse quantized model you can use the latest advances of the OpenVINO runtime to speed up such models. The model compression feature is enabled in the runtime at the model compilation step using a special option called: “CPU_SPARSE_WEIGHTS_DECOMPRESSION_RATE”. Its value controls the minimum sparsity rate that MatMul operation should have to be optimized at inference time. This property is passed to the compile_model API as it is shown below:

from openvino.runtime import Core

core = Core()
model = core.read_model(model="path_to_model_xml")
# MatMul layers with higher sparsity rate than 80% are optimized
compiled_model = core.compile_model(model=model, device_name="CPU", config=configuration)

An important note is that a high sparsity rate is required to see the performance benefit from this feature. And we note again that this feature is available only on the 4th Gen Intel® Xeon® Scalable Processors and it is basically for throughput-oriented scenarios. To simulate such a scenario, you can use the benchmark_app application supplied with OpenVINO distribution and limit the number of resources available for inference. Below we show the performance difference between the two runs sparsity optimization in the runtime:

  • Benchmarking without sparsity optimization:
# Dump benchmarking config for dense inference
with open("perf_config.json", "w") as outfile:
            "CPU": {"NUM_STREAMS": 4, "INFERENCE_NUM_THREADS": 4}
benchmark_app -m bert_90_sparse_quantized/openvino_model.xml -shape "input_ids[1,16],attention_mask[1,16],token_type_ids[1,16]" -load_config perf_config.json
  • Benchmarking when sparsity optimization is enabled:
# Dump benchmarking config for sparse inference
with open("perf_config_sparse.json", "w") as outfile:
benchmark_app -m bert_90_sparse_quantized/openvino_model.xml -shape "input_ids[1,16],attention_mask[1,16],token_type_ids[1,16]" -load_config perf_config_sparse.json

Performance Results

We performed a benchmarking of our sparse and 8-bit quantized BERT model on 4th Gen Intel® Xeon® Scalable Processors with various settings. We ran two series of experiments where we vary the number of parallel threads and streams available for the asynchronous inference in the first experiments and we investigate how the sequence length impact the relative speedup in the second series of experiments.

The table below shows relative speedup for various combinations of number of streams and threads and at the fixed sequence length after enabling sparsity acceleration in the OpenVINO runtime.

Based on this, we can conclude that one can expect significant performance improvement with any number of streams/threads larger than one. The optimal performance is achieved at eight streams/threads. However, we would like to note that this is model specific and depends on the model architecture and sparsity distribution.

The chart below also shows the relationship between the possible acceleration and the sequence length.

As you can see the benefit from sparsity is decreasing with the growth of the sequence length processed by the model. This effect can be explained by the fact that for larger sequence lengths the size of the weights is no longer a performance bottleneck and weight compression does not have so much impact on the inference time. It means that such a weight sparsity acceleration feature does not suit well for large text processing tasks but could be very helpful for Question Answering, Sequence Classification, and similar tasks.


  • Jupyter notebook with the performance benchmarks.
  • Model card for sparse and quantized BERT-base model