Deep Learning For Computer Vision

Optimize deep learning solutions across multiple Intel® platforms—CPU, GPU, FPGA, and VPU—and accelerate deep neural network workloads.

Intel® Deep Learning Deployment Toolkit

This toolkit allows developers to deploy pretrained deep learning models through a high-level C++ or Python* inference engine API integrated with application logic. It supports multiple Intel® platforms and is included in the Intel® Distribution of OpenVINO™ toolkit.

This toolkit comprises the following two components:

Model Optimizer

This Python*-based command line tool imports trained models from popular deep learning frameworks such as Caffe*, TensorFlow*, and Apache MXNet*, and Open Neural Network Exchange (ONNX*).

  • Runs on multiple operating systems : Windows*, Linux*, and macOS*
  • Perform analysis and adjustments for optimal execution on endpoint target devices using static, trained models
  • Serialize and adjust the model into an intermediate representation (IR) format from Intel
  • Support over 100 public models for Caffe, TensorFlow, MXNet, and ONNX.

Standard frameworks are not required when generating IR files for models consisting of standard layers. When processing custom layers in original models, the Model Optimizer provides a flexible mechanism of extensions.

Inference Engine

This engine uses a common API to deliver inference solutions on the platform of your choice: CPU, GPU, VPU, or FPGA.

  • Run different layers on different targets (for example, a GPU and selected layers on a CPU)
  • Implement custom layers on a CPU while running the remaining topology on a GPU—without having to rewrite the custom layers
  • Optimize workloads (computational graph analysis, scheduling, and model compression) for target hardware with an embedded-friendly scoring solution
  • Take advantage of new asynchronous execution to improve frame-rate performance while limiting wasted cycles
  • Use a convenient C++ or Python API to work on IR files and optimize inference

Inference Support

In addition to supporting processors with and without integrated graphics, this toolkit enables optimal performance on other hardware accelerators from Intel, such as:

  • Intel® Programmable Acceleration Card
  • Intel® Movidius™ Vision Processing Unit (VPU)
  • Intel® Vision Accelerator Design
workflow for training models

Deep Learning Workbench

This web-based graphical environment that allows users to visualize a simulation of the performance of deep learning models and datasets on various Intel® architecture configurations (CPU, GPU, VPU). It provides key performance metrics such as latency, throughput, and performance counters for each layer of the selected neural network. This tool includes simple configuration for many inference experiments to detect optimal performance settings.

  • Run single versus multiple inferences.
  • Calibrate to reduce precision of certain model layers from FP32 to Int8.
  • Automatically determine the optimized algorithm based on convolution layer parameters and hardware configuration with the Winograd Algorithmic Tuner.
  • Run experiments on known data sets and determine accuracy of the model after parameter tuning or calibration using the accuracy checker.

Deep Learning Workbench Developer Guide

Discover the Capabilities

Hardware Acceleration

Harness the performance of Intel®-based accelerators: CPUs, GPUs, FPGAs, VPUs, and IPUs.