This page contains the main concepts of the OpenVINO™ DL Workbench.
Project is a combination of an OpenVINO IR model, a dataset, and a target device you use to perform experiments in the DL Workbench. You can work with multiple projects during one session. When creating a project, you can:
Intermediate Representation (IR) is the OpenVINO™ format of pretrained model representation with two files:
Original model is your custom pretrained model in one of the supported formats: OpenVINO™ IR, Caffe*, MXNet*, ONNX*, or TensorFlow*.
Model task is the problem the model was trained to solve:
Model precision is the type of variables used to store model weights:
Dataset is a set of testing and validation images with annotations. Dataset format depends on a task a model is trained to perform.
Target is a system on which you want to perform an inference of your model. This can be your local workstation or a remote system.
Local target is the machine on which you run the DL Workbench.
Remote target is a machine in your local network to which you connect your local workstation to gather performance data remotely.
Device is a hardware accelerator on which a model is executed, for example, Intel® Movidius™ Neural Compute Stick 2 (NCS2).
Profiling is analysis of neural network performance to explore areas where optimization can be applied.
Throughput is the number of images processed in a given amount of time. Measured in frames per second (FPS).
Latency is the time required to complete a unit of work, for example, time required to perform an inference for a single image. Measured in milliseconds.
Inference Engine (IE) is a set of C++ libraries providing a common API to deliver inference solutions on the platform of your choice: CPU, GPU, or VPU. OpenVINO™ Inference Engine is used inside the DL Workbench to infer models.
Accuracy is the quality of predictions made by a neural network. Different use cases measure quality differently, so an accuracy metric depends on a particular model task.
Calibration is the process of lowering the precision of a model from FP32 to INT8. Calibration accelerates the performance of certain models on hardware that supports INT8 precision. An INT8 model takes up less memory footprint and speeds up inference time at the cost of a small reduction in accuracy. See INT8 Calibration for details.
Deployment is the process of putting your model into a real-life application. DL Workbench enables you to download a deployment package with your model optimized for particular devices, including required libraries, API, and scripts. See Build Your Application with Deployment Package for details.
Deployment target is the target on which you plan to run a product application, so you perform calibration and tune runtime hyperparameters for this particular target.