Skip to content

README.md

Latest commit

 

History

History
65 lines (47 loc) · 4.47 KB

README.md

File metadata and controls

65 lines (47 loc) · 4.47 KB

Model Installer

Model installer provides users with AI models for SceneScape by:

  • downloading the configured set of models from OpenVINO Model Zoo
  • downloading/generating necessary configuration files to integrate them with SceneScape services

The models and configuration files are downloaded into a models volume that is attached to SceneScape services for both Docker and Kubernetes deployments.

Configuration

Model installer can be configured to download a specific set of models using the following parameters:

Parameter Allowed Values Format Description
models default, ocr, all Single value Specifies which set of models to download. default includes person detection, re-identification, and pose estimation models. ocr includes text detection and recognition models. all downloads both default and OCR models.
precisions FP32, FP16, INT8 Comma-separated list Model precision formats to download. Multiple precisions can be specified for the same model (e.g., FP16,FP32). The first one will be used as preferred when generating model-config.json

For Kubernetes deployment refer to the initModels section in Helm chart values, for example use --set initModels.modelType=all --set initModels.modelPrecisions=FP16,FP32 when installing the Helm chart.

For Docker deployment use MODELS and PRECISIONS environment variables when building, e.g.: make install-models MODELS=all or make install-models MODELS=all PRECISIONS="FP16,FP8".

Models Volume Folder Structure

models/
├── intel/
│   ├── model-name-1/
│   │   ├── FP16/
│   │   │   ├── model-name-1.xml (OpenVINO model)
│   │   │   ├── model-name-1.bin (OpenVINO model)
│   │   │   └── model-name-1.json (model-proc file - required only for selected models)
│   │   └── FP32/
│   │       └── ...
│   └── model-name-2/
│       └── ...
├── public/
│   └── model-name-3/
│       └── ...
└── model_configs/
    └── model_config.json (auto-generated default model configuration file)

Auto-Generation of model-config.json File

For detailed information about the file format and its usage, refer to the Model Configuration File Format documentation.

Model Classification

The function automatically assigns metadata policies and element types based on model names:

Model Pattern Metadata Policy Type Description
detection, detector, detect detectionPolicy detect Object detection models
text + detection ocrPolicy detect Text detection models
reidentification, reid reidPolicy inference Person/object re-identification
recognition, attributes, classification classificationPolicy classify Classification and attribute recognition
text + recognition ocrPolicy classify Text recognition models
pose detection3DPolicy inference Human pose estimation

Model Name Mapping

The generate_model_config.py file includes a predefined mapping for shorter, more convenient model names. The mapping is defined in the _MODEL_NAME_MAP variable.

If a model name exists in this mapping, the shortened name will be used as the key in the configuration. Otherwise, the original behavior (replacing hyphens with underscores) is used.