getitune-optimizing-a-model

name

getitune-optimizing-a-model

description

Optimize an exported getitune model (the Geti training library) with post-training quantization. Use when a user wants to run `OVEngine.optimize()` / `engine.optimize()` to produce an INT8 model via NNCF, understands calibration-set requirements, or needs to re-validate and run inference with a quantized model versus the original FP32/FP16 model. Covers OpenVINO NNCF post-training quantization and the accuracy/size trade-off.

Optimizing (quantizing) a model with getitune

getitune applies post-training quantization (PTQ) via NNCF to shrink an exported OpenVINO model and speed up inference. Quantization runs on an OpenVINO model (an exported .xml), producing an INT8 version.

Run everything from library/.

Workflow

from getitune.engine import create_engine

# Load an exported OpenVINO model, then quantize it
ov_engine = create_engine(
    model="/path/to/exported_model.xml",
    data="/path/to/dataset",
)
ov_engine.optimize()                 # INT8 post-training quantization via NNCF
int8_metrics = ov_engine.test()      # validate the quantized model
predictions = ov_engine.predict()    # run inference with the quantized model

1. Start from an exported OpenVINO model (.xml). If you only have a checkpoint, export it first with the getitune-exporting-a-model skill.
   • Done when: create_engine(model="....xml", data=...) builds an OVEngine.
2. Provide a calibration dataset. Calibration images are taken automatically from the training subset; 200-500 images is the recommended calibration size.
   • Done when: optimize() runs without a "not enough calibration data" issue.
3. Run optimize(). This replaces the engine's model in place with the INT8 version.
   • Done when: the call completes and subsequent test()/predict() use INT8.
4. Re-validate accuracy with test() and compare against the FP32/FP16 baseline; a small accuracy drop is expected in exchange for size/latency.
   • Done when: the INT8 metric is within your acceptable tolerance of baseline.

Comparing against the original model

After optimize() the engine holds the INT8 model. To re-check the original FP32/FP16 model, either pass the original .xml path directly to .test() / .predict(), or create the engine again from the original .xml.

Notes

• Quantization is OpenVINO/NNCF-based and applies to exported IR models — it is not a training-time step.
• Only OpenVINO IR (.xml) is supported. An ONNX model must be converted to OpenVINO IR first before it can be optimized.
• In the Geti application this is exposed as the quantize job (application/backend/app/execution/quantization/); library optimize() is the same capability without the job/queue wrapper.

Verify

# from library/
just lint
just test-unit -- -k optimize      # when you touched optimization code

Related skills

• getitune-exporting-a-model — produce the OpenVINO .xml to quantize.
• getitune-running-inference — run inference with the quantized model.