DeepStream Eval & Finetune

Purpose

Deploy an object detector, measure its deployed mAP and performance on the user's KPI data, fine-tune it, redeploy it, and produce a before/after report. Both accuracy legs must run through the real DeepStream pipeline. Report measured results without a Go/No-Go verdict.

This skill orchestrates deepstream-import-vision-model for deployment and the TAO Skill Bank router tao-launch-workflow (target tao-finetune-huggingface-model) for single-run fine-tuning. HF AutoML uses this skill's NGC harness. Read the applicable direct reference before each stage.

Prerequisites