DL Workbench Key Concepts¶
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:
Upload your own model or download one from the Intel® [Open Model Zoo](omz_models_group_intel) directly in the DL Workbench.
Upload your own dataset or generate one in the DL Workbench.
Select your local workstation as a target or connect to a remote machine.
Intermediate Representation (IR)¶
Intermediate Representation (IR) is the OpenVINO™ format of pretrained model representation with two files:
XML file describing the network topology
BIN file containing weights and biases
Inference Engine operates with models in the IR format. You can convert your model of a supported format into the IR format with the Model Optimizer directly in the DL Workbench.
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:
Classification is a task of defining a category for an image.
Object detection is a task of identifying instances of objects of a certain class and highlighting them with bounding boxes.
Instance segmentation is a task of identifying instances of objects of a certain class and highlighting them with masks.
Semantic segmentation is a task of detecting all object types in an image at the pixel level and highlighting them as regions.
Image Inpainting is a task of reconstructing missing areas in an image.
Super-Resolution is a task of enhancing the image resolution.
Style Transfer is a task of adopting the appearance of another image.
Facial Landmark Detection is a task of detecting key landmarks of a face, such as eyes, nose, and mouth.
Face Recognition is a task of identifying a person based on an image of their face.
Model precision is the type of variables used to store model weights:
FP32 - 32-bit floating point numbers
FP16 - 16-bit floating point numbers
INT8 - 8-bit integer numbers
Usually a model has a mix of precisions. When converting a model, you have to choose between FP32 and FP16. FP16 is supported by all OpenVINO™ plugins while keeping good accuracy and speed gains over FP32.
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)¶
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.