Export and Optimize Geti Model#

Overview#

This guide starts by downloading the trained YOLOX PyTorch weights from Intel Geti and the COCO dataset used during training. A workspace is then set up and the Training Extensions repository is cloned, which provides the conversion script. After installing the required Python and Rust dependencies, the export_and_optimize.py script is run to convert the model to OpenVINO IR format — producing a full-precision FP32 model and an INT8 post-training quantized model optimized for Intel hardware.

Prerequisites#

Before you begin, ensure you have the following:

A trained model exported from Intel Geti as a PyTorch weights file (.pth)

Note: Image is for illustration purposes only.
A COCO-format dataset (.zip) used during training (required for post-training optimization)

Note: Images are for illustration purposes only.
Git installed
Internet access to download dependencies

Step 1: Set Up the Workspace#

Create the working directory structure:

mkdir generate_model
cd generate_model

mkdir model
mkdir coco_dataset
mkdir output

Directory	Purpose
`model/`	Stores the downloaded PyTorch weights file
`coco_dataset/`	Stores the COCO dataset used for optimization
`output/`	Stores the exported and optimized model files

Step 2: Add Model Weights and Dataset#

Copy and Extract the PyTorch Model#

Place the downloaded Pytorch_model.zip file into the model/ directory and extract it:

# Copy Pytorch_model.zip into the model directory, then unzip
cp /path/to/Pytorch_model.zip model/
cd model
unzip Pytorch_model.zip
cd ..

After extraction, the model/ directory should contain a weights.pth file.

Copy and Extract the COCO Dataset#

Place the downloaded COCO dataset archive into the coco_dataset/ directory and extract it:

# Copy the COCO dataset zip into the coco_dataset directory, then unzip
cp /path/to/<coco_dataset>.zip coco_dataset/
cd coco_dataset
unzip <coco_dataset>.zip
cd ..

After extraction, the coco_dataset/ directory should follow the standard COCO layout:

coco_dataset/
├── annotations/
└── images/

Step 3: Clone the Training Extensions Repository#

git clone https://github.com/open-edge-platform/training_extensions.git

Step 4: Install Dependencies#

Install `uv` (Python Package Manager)#

curl -LsSf https://astral.sh/uv/install.sh | sh
source $HOME/.local/bin/env

Install Rust Toolchain (required by some dependencies)#

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y

Step 5: Set Up the Python Environment#

Navigate to the library directory within the cloned repository and check out the required branch:

cd training_extensions/library
git checkout kp/test_yolox

Create and activate a virtual environment, then sync all dependencies:

uv venv
source .venv/bin/activate
source "$HOME/.cargo/env"
uv sync

Step 6: Export and Optimize the Model#

Run the export_and_optimize.py script with the appropriate paths and model configuration:

python export_and_optimize.py \
    --weights        /path/to/model/weights.pth \
    --source_dataset /path/to/coco_dataset \
    --output_dir     /path/to/output \
    --model_name     yolox_tiny

Arguments#

Argument	Required	Description
`--weights`	Yes	Path to the PyTorch weights file (`.pth`)
`--source_dataset`	Yes	Path to the COCO dataset directory
`--output_dir`	Yes	Directory where exported and optimized model files are saved
`--model_name`	Yes	Model variant to use. Supported values: `yolox_tiny`, `yolox_s`, `yolox_l`, `yolox_x` (default: `yolox_tiny`)

Example with Absolute Paths#

Assuming the workspace is located at ~/generate_model:

python export_and_optimize.py \
    --weights        ~/generate_model/model/weights.pth \
    --source_dataset ~/generate_model/coco_dataset \
    --output_dir     ~/generate_model/output \
    --model_name     yolox_tiny

Output#

After the script completes, the output/ directory will contain the exported and optimized model files ready for deployment in the Worker Safety Gear Detection pipeline:

output/
├── otx-workspace/
│   ├── exported_model.xml    # FP32 – full-precision exported model
│   └── optimized_model.xml   # INT8 – post-training quantized model

File	Precision	Description
`exported_model.xml`	FP32	Full-precision model exported directly from the PyTorch weights
`optimized_model.xml`	INT8	Post-training quantized model optimized using the COCO dataset

Both files can be used directly with the OpenVINO inference engine. The INT8 model (optimized_model.xml) offers faster inference with reduced memory footprint, while the FP32 model (exported_model.xml) retains full numerical precision.

Generated Model Output

Note: Image is for illustration purposes only.

Troubleshooting#

Issue	Resolution
`uv: command not found`	Re-run `source $HOME/.local/bin/env` or open a new terminal session
Rust compilation errors	Ensure `source "$HOME/.cargo/env"` was run after the Rust installation
Dataset not found	Verify the COCO dataset was extracted and the `annotations/` folder exists
Incorrect model output	Confirm `--model_name` matches the architecture used during Geti training