Download and Convert Models#

The recommended way to obtain models for Deep Learning Streamer is to use the standalone conversion scripts in scripts/download_models. They download models from their original sources, convert them to OpenVINO IR, and save the resulting files in the requested output directory.

For the complete setup, command reference, supported options, and examples, see the Model Conversion Scripts README.

Available Scripts#

  • download_hf_models.py converts supported Hugging Face models with optimum-cli. It also provides custom conversion paths for selected models, including CLIP and RT-DETR.

  • download_ultralytics_models.py converts Ultralytics models, including YOLO detection, segmentation, pose, OBB, classification, and YOLOE models. It supports model names, local .pt files, and Hugging Face repositories.

  • download_timm_models.py converts supported TIMM image-classification models hosted on Hugging Face.

  • download_other_models.sh downloads and converts selected helper models that are not handled by the other scripts, including yolox-tiny, yolox_s, and yolov7.

Reproducible Downloads#

Model references can include an @... suffix to pin the source version:

  • Hugging Face and TIMM use repo_id@revision, typically a commit SHA.

  • Ultralytics uses model.pt@tag, where tag is an ultralytics/assets GitHub release tag.

  • download_other_models.sh uses sources and tool versions defined in the script and does not support per-model version suffixes.

Without a pinned revision, the Hugging Face, TIMM, and Ultralytics scripts may resolve the latest available model at runtime.

Model Usage#

After conversion, use the generated OpenVINO .xml file with the appropriate Deep Learning Streamer inference element. The matching .bin file must remain in the same directory. See the Supported Models table and the GStreamer samples for model-specific elements and pipeline examples.