rtvi-cv-customize-model

작성자: nvidia

DeepStream CV 감지 모델을 VSS Alerts Blueprint 검증(2d_cv) 모드에서 교체하는 방법 - ONNX 내보내기, 사용자 정의 bbox 파서, compose 마운트 등을 다룹니다…

npx skills add https://github.com/nvidia/skills --skill rtvi-cv-customize-model

CV Detection Model Customization — VSS Alerts Blueprint (2d_cv mode only)

The RT-CV perception container (vss-rt-cv) runs a DeepStream pipeline with a configurable primary GIE (GPU Inference Engine). By default it uses GDINO or RTDETR. This guide covers replacing it with any ONNX-format model, using YOLOv11 COCO 80 as the worked example.

This only applies to --mode verification (2d_cv). Real-time alerts mode (2d_vlm) has no CV detector.

For instance-segmentation or detection-plus-mask models, complete references/segmentation-model-contract.md before writing parser or handoff code.


When to use

Use this skill when the user wants to:

  • swap the stock vss-rt-cv detector for another ONNX model in verification mode,
  • debug a broken ONNX staging path, ghost-directory bind mount, or missing runtime TRT engine build,
  • fix a DeepStream parser load failure such as dlsym failed on the bbox parser symbol.

Do not use this skill for:

  • 2d_vlm real-time alerts mode,
  • scaffolding a brand-new standalone RTVI CV microservice (use rtvi-cv-scaffold-vss-service).

Instructions

  • Keep the answer scoped to VSS Alerts Blueprint verification mode (2d_cv) unless the user explicitly asks to compare modes.
  • Paths beginning with deploy/docker/ are relative to the VSS Blueprint repository, not the DeepStream repository. Clone or reuse a VSS checkout that is v3.2.1 or compatible, then run these commands from that repository root (see VSS Quickstart).
  • If the user asks about ONNX staging or a compose mount, explicitly say that mounting a missing file path is wrong: use the stock parent-directory mount, keep the ONNX under ${VSS_DATA_DIR}/models/yolo, and re-stage the ONNX file after any dev-profile.sh up that recreates the models directory.
  • If the user asks about parser load failures, line up parse-bbox-func-name, the extern "C" function symbol, and CHECK_CUSTOM_PARSE_FUNC_PROTOTYPE(...), then rebuild the .so with both DeepStream and CUDA include paths present.
  • If the model is detection-plus-mask or instance segmentation, route the mask-contract decision first through references/segmentation-model-contract.md before editing parser or handoff logic.

Examples

  • "Replace the VSS Alerts Blueprint verification detector with a YOLO11 ONNX model and redeploy perception-alerts."
  • "DeepStream says dlsym failed for the bbox parser after loading the custom parser library. What should I check?"
  • "I mounted the ONNX file path directly in compose before the host file existed. Is that okay?"

Architecture

NVStreamer (RTSP) → SDR (port 9010) → DeepStream (perception-alerts)
                                            └─ primary-gie (nvinfer)
                                                  ├─ ONNX → TRT engine (built once, then cached)
                                                  ├─ custom bbox parser (.so)
                                                  └─ label file (.txt)
                                            └─ Kafka → mdx-raw
                                                  └─ vss-behavior-analytics
                                                        └─ mdx-incidents
                                                              └─ vlm-as-verifier

VSS source location

This skill is documentation-only and does not ship the VSS deployment sources. Use a VSS v3.2.1 or compatible checkout; run all paths and commands from that repository root. Clone/LFS setup, stock Alerts profile files, and the YOLOv11 customization file tree are in references/vss-source-layout.md.


Adapting to a Different Model

YOLOv11 COCO 80 is the worked example throughout all steps below. The same pattern applies to any ONNX-format detector — substitute at these points:

StepWhat to change
Step 1Replace the export procedure with whatever your model's training library requires. Confirm the resulting .onnx exists on the host before continuing.
Steps 2–3Inspect your model's actual output tensor name, shape, layout, and whether NMS is applied in-graph. Do not assume it matches YOLOv11. The deepstream-dev skill (skills/deepstream-dev/) has a generation-by-generation YOLO output format table and references/nvinfer_config.md for full property reference.
Step 4Update output-blob-names to your tensor name, infer-dims to your input shape, and cluster-mode to match whether NMS is in-graph (4) or not (2).
Step 5Update the --onnx path, the input tensor name in --minShapes/--optShapes/--maxShapes (not images for non-YOLO models), and the sed commands to reference your nvinfer .txt file. Add a new if [[ $MODEL_NAME_2D == "YOURMODEL" ]] block rather than editing the YOLO block.

Step 1: Stage Your ONNX Model (Host, One-Time)

Obtain a TensorRT-compatible ONNX model and place it at ${VSS_DATA_DIR}/models/yolo/yolo11s.onnx before continuing.

See references/yolov11-onnx-export.md for the export settings and tensor layout used in this reference. Adapt the parser and nvinfer config in later steps to match your model's actual output.

After every full dev-profile.sh up, restore ownership of the recreated models/yolo directory and re-stage your ONNX file:

sudo mkdir -p "${VSS_DATA_DIR}/models/yolo"
sudo chown -R $(id -u):$(id -g) "${VSS_DATA_DIR}/models/yolo"
# Copy your ONNX model to ${VSS_DATA_DIR}/models/yolo/yolo11s.onnx

Steps 2–3: Inspect the Model Output and Build the Parser

Read and follow references/yolov11-parser.md before creating ${VSS_PROFILE_DIR}/deepstream/custom_parser/nvdsparseyolov11.cpp or editing the parser build in Dockerfiles/perception.Dockerfile.

Before writing any parser code, confirm these four things for your model:

  • Output tensor name and shape — drives output-blob-names and infer-dims in Step 4
  • Pre-NMS or post-NMS — drives cluster-mode in Step 4; see deepstream-dev Rule 13 for the generation-by-generation breakdown
  • Coordinate semanticscx/cy/w/h center-format or x1/y1/x2/y2 corner-format; the YOLOv11 example uses center-format (see references/yolov11-parser.md)
  • Parser symbol name — the extern "C" function name must match parse-bbox-func-name in Step 4 and CHECK_CUSTOM_PARSE_FUNC_PROTOTYPE(...) exactly

Confirm the model output tensor layout instead of assuming it matches YOLOv11. Keep the parser symbol identical in the exported C++ function, CHECK_CUSTOM_PARSE_FUNC_PROTOTYPE(...), and the Step 4 parse-bbox-func-name setting.

For instance segmentation or detection-plus-mask models, complete the contract decisions table in references/segmentation-model-contract.md before writing parser or handoff code.


Step 4: nvinfer Config

${VSS_PROFILE_DIR}/deepstream/configs/yolov11.txt — the primary GIE sub-config referenced by the DeepStream run config.

model-engine-file and batch-size are patched at container startup by the ds-start.sh block in Step 5. Copy the values below as-is; Step 5 will overwrite both with the correct engine path and sensor count before DeepStream reads the file.

[property]
gpu-id=0
net-scale-factor=0.0039215697906911373    # 1/255 input normalization
model-engine-file=/opt/storage/yolo11s_fp16.engine
onnx-file=/opt/storage/yolo/yolo11s.onnx
batch-size=1                             # patched at startup from NUM_SENSORS
network-mode=2                           # 0=FP32, 1=INT8, 2=FP16
network-type=0                           # 0=Detector
num-detected-classes=80
interval=0
gie-unique-id=1
output-blob-names=output0               # must match your model's output tensor name
infer-dims=3;640;640                    # C;H;W
maintain-aspect-ratio=1
parse-bbox-func-name=NvDsInferParseCustomYoloE  # must match extern "C" function name in .so
custom-lib-path=/opt/deepstream-yolo/libnvdsparseyolov11.so
labelfile-path=/opt/nvidia/deepstream/deepstream/sources/apps/sample_apps/metropolis_perception_app/mounted-configs/yolo-coco-labels.txt
cluster-mode=2                   # 2=NMS — required: Ultralytics ONNX has no NMS in-graph

[class-attrs-all]
pre-cluster-threshold=0.25
topk=300
nms-iou-threshold=0.45

Step 5: Runtime TRT Engine Build (ds-start.sh)

TRT engines are GPU-architecture-specific. Build the engine on first startup when it is missing, then reuse it on subsequent restarts. Rebuild an existing engine only when FORCE_REBUILD=true. The stock Alerts compose file bind-mounts ${VSS_DATA_DIR}/models/ at /opt/storage/, so the generated engine persists on the host.

Which ds-start.sh? Stock Alerts runs services/rtvi/rtvi-cv/ds-start.sh (via extends), not ${VSS_PROFILE_DIR}/deepstream/init-scripts/ds-start.sh. This skill edits the profile copy, then remounts it in Step 6 so those edits actually run. Full map of the three paths, remount trade-offs, and DS_CONFIG_FILE pairing: references/ds-start-entrypoint.md.

Edit ${VSS_PROFILE_DIR}/deepstream/init-scripts/ds-start.sh, then override the stock rtvi-cv ds-start.sh bind mount in the Alerts profile compose file as shown in Step 6. Without that override, profile-script edits never execute.

Critical — honor DS_CONFIG_FILE. The profile script sets CONFIG_FILE=${1:-...} and never reads DS_CONFIG_FILE. The Alerts compose command invokes ds-start.sh with no positional args, while Step 6 sets DS_CONFIG_FILE to the absolute mounted-configs run config (not the stock .../configs/... path). Replace the script's existing CONFIG_FILE=... assignment with:

# Prefer $1 when provided; otherwise use DS_CONFIG_FILE from compose (absolute path).
# Without this, YOLO_CONFIG resolves to ./yolov11.txt and the stock detector keeps running.
CONFIG_FILE="${1:-${DS_CONFIG_FILE:-/opt/nvidia/deepstream/deepstream/sources/apps/sample_apps/metropolis_perception_app/mounted-configs/run_config-api-rtdetr-protobuf.txt}}"

Then add the YOLO block (before the GDINO/RT-DETR model branches is fine):

if [[ $MODEL_NAME_2D == "YOLO" ]]; then
    YOLO_ONNX=/opt/storage/yolo/yolo11s.onnx
    YOLO_ENGINE=/opt/storage/yolo11s_fp16.engine
    YOLO_ENGINE_TMP="${YOLO_ENGINE}.building"
    FORCE_REBUILD=${FORCE_REBUILD:-false}

    if [[ ! -f "$YOLO_ONNX" ]]; then
        echo "ERROR: ONNX not found at ${YOLO_ONNX}. Stage your model there before starting."
        exit 1
    fi

    if [[ ! -s "$YOLO_ENGINE" || "${FORCE_REBUILD,,}" == "true" ]]; then
        echo "Building TensorRT engine: ${YOLO_ENGINE}"
        rm -f "$YOLO_ENGINE_TMP"
        if /usr/src/tensorrt/bin/trtexec \
            --onnx=${YOLO_ONNX} \
            --minShapes=images:1x3x640x640 \
            --optShapes=images:${NUM_SENSORS}x3x640x640 \
            --maxShapes=images:${NUM_SENSORS}x3x640x640 \
            --fp16 --saveEngine=${YOLO_ENGINE_TMP}; then
            mv "$YOLO_ENGINE_TMP" "$YOLO_ENGINE"
        else
            rm -f "$YOLO_ENGINE_TMP"
            echo "ERROR: TensorRT engine build failed; existing engine was preserved."
            exit 1
        fi
    else
        echo "Reusing cached TensorRT engine: ${YOLO_ENGINE}"
    fi

    # Patch the run config to use yolov11.txt for the primary GIE
    # (relative name is intentional: DeepStream resolves it against CONFIG_FILE's dir)
    sed -i '/^\[primary-gie\]/,/^\[/{s/config-file=.*/config-file=yolov11.txt/;}' "$CONFIG_FILE"
    # Patch engine path and batch size into yolov11.txt (same dir as CONFIG_FILE)
    YOLO_CONFIG="$(dirname "${CONFIG_FILE}")/yolov11.txt"
    sed -i "s|model-engine-file=.*|model-engine-file=${YOLO_ENGINE}|" "${YOLO_CONFIG}"
    sed -i "/^\[property\]/,/^\[/{s/^batch-size=.*/batch-size=${NUM_SENSORS}/;}" "${YOLO_CONFIG}"
fi

To add a different model, add a new if [[ $MODEL_NAME_2D == "YOURMODEL" ]] block. Update:

  • --onnx path and --saveEngine output path
  • --minShapes/--optShapes/--maxShapes input tensor name and dimensions for your model
  • The sed commands to point at your nvinfer .txt file

Step 6: Compose + Env

In the existing ${VSS_PROFILE_DIR}/compose.yml, update the perception-alerts service:

perception-alerts:
  build:
    context: $VSS_APPS_DIR/developer-profiles/dev-profile-alerts
    dockerfile: Dockerfiles/perception.Dockerfile
  volumes:
    # Keep the stock parent-directory mount. Do not replace it with a direct
    # mount of yolo11s.onnx: Docker creates a ghost directory if the host
    # file does not exist.
    - $VSS_DATA_DIR/models/:/opt/storage/
    - $VSS_APPS_DIR/developer-profiles/dev-profile-alerts/deepstream/configs/:/opt/nvidia/deepstream/deepstream/sources/apps/sample_apps/metropolis_perception_app/mounted-configs/
    # Override stock rtvi-cv ds-start.sh (extends bind). Without this,
    # edits under deepstream/init-scripts/ never run — see
    # references/ds-start-entrypoint.md.
    - $VSS_APPS_DIR/developer-profiles/dev-profile-alerts/deepstream/init-scripts/ds-start.sh:/opt/nvidia/deepstream/deepstream/sources/apps/sample_apps/metropolis_perception_app/ds-start.sh:ro
  environment:
    MODEL_NAME_2D: ${MODEL_NAME_2D}
    NUM_SENSORS: ${NUM_SENSORS}
    FORCE_REBUILD: ${FORCE_REBUILD:-false}
    # Use mounted-configs (where yolov11.txt lives), not stock .../configs/...
    DS_CONFIG_FILE: /opt/nvidia/deepstream/deepstream/sources/apps/sample_apps/metropolis_perception_app/mounted-configs/run_config-api-rtdetr-protobuf.txt

Set in ${VSS_PROFILE_DIR}/.env:

MODEL_NAME_2D="YOLO"
NUM_SENSORS=1
FORCE_REBUILD=false

Keep FORCE_REBUILD=false for normal starts. Compose's --force-recreate option recreates the container but does not rebuild the TensorRT engine unless FORCE_REBUILD=true.

Gotcha — generated.env wins at restart. Step 7's redeploy uses --env-file .../generated.env, which is produced by dev-profile.sh from .env. Editing .env alone does not change FORCE_REBUILD for that compose command. For a one-shot engine rebuild, pass the override on the shell (see Step 7). To persist the value for later dev-profile.sh runs, edit .env and re-run dev-profile.sh so it regenerates generated.env.


Step 7: Deploy

Run from the video-search-and-summarization repository root (VSS_ROOT / VSS_DEPLOY_DIR / VSS_PROFILE_DIR from VSS source location):

: "${VSS_ROOT:=$PWD}"
: "${VSS_DEPLOY_DIR:=${VSS_ROOT}/deploy/docker}"
: "${VSS_PROFILE_DIR:=${VSS_DEPLOY_DIR}/developer-profiles/dev-profile-alerts}"

1. Single-GPU host (only GPU 0): edit ${VSS_PROFILE_DIR}/.env so these keys match the required final values below. Edit the source profile env file, not generated.env (dev-profile.sh regenerates that). Stock Alerts reserves GPU 0 and puts RT-VLM / LLM / VLM on GPU 1; dev-profile.sh CLI flags cannot clear RESERVED_DEVICE_IDS or set FIXED_SHARED_DEVICE_IDS. Confirm device IDs exist with nvidia-smi --query-gpu=index --format=csv,noheader,nounits before deploying.

RESERVED_DEVICE_IDS=''
FIXED_SHARED_DEVICE_IDS='0'
RT_CV_DEVICE_ID='0'
RT_VLM_DEVICE_ID='0'
LLM_DEVICE_ID='0'
VLM_DEVICE_ID='0'

2. Deploy the official Alerts verification profile. Select the hardware profile supported by your host, as documented by the VSS Quickstart. Do not present the remaining commands as a copy-paste handoff unless execution is blocked.

"${VSS_DEPLOY_DIR}/scripts/dev-profile.sh" up \
  -p alerts \
  -m verification \
  -H <H100|L40S|RTXPRO4500BW|RTXPRO6000BW|DGX-SPARK|IGX-THOR|AGX-THOR|OTHER>

# 3. dev-profile.sh recreates the model directory. Restore ownership and
# re-stage your ONNX model.
export VSS_DATA_DIR="${VSS_DEPLOY_DIR}/data-dir"
sudo mkdir -p "${VSS_DATA_DIR}/models/yolo"
sudo chown -R $(id -u):$(id -g) "${VSS_DATA_DIR}/models/yolo"
# Copy your ONNX model to ${VSS_DATA_DIR}/models/yolo/yolo11s.onnx

# 4. Rebuild and recreate the customized perception service using the compose
# project generated by dev-profile.sh.
cd "${VSS_DEPLOY_DIR}"
docker compose \
  --env-file developer-profiles/dev-profile-alerts/generated.env \
  up -d --build --force-recreate perception-alerts

# One-shot TensorRT rebuild (after ONNX / shape / GPU / TRT changes).
# Inline FORCE_REBUILD overrides generated.env for this invocation only:
# FORCE_REBUILD=true docker compose \
#   --env-file developer-profiles/dev-profile-alerts/generated.env \
#   up -d --build --force-recreate perception-alerts

TRT engine build takes approximately 15–30 seconds on Blackwell hardware. Subsequent restarts reuse the engine persisted under ${VSS_DATA_DIR}/models/.


Common Gotchas

See references/common-gotchas.md for single-GPU reservation failures, wipe/restage after dev-profile.sh up, ghost file mounts, dlsym parser mismatches, and Redis stream contamination.

nvidia의 다른 스킬

compileiq-debug
nvidia
무언가 잘못되었을 때 사용: Search()가 멈추거나, 모든 평가가 INVALID_SCORE를 반환하거나, 점수가 개선되지 않거나, 모든 설정이 동일한 숫자를 반환하거나, ptxas 오류 등이 발생할 때
create-github-pr
nvidia
gh CLI를 사용하여 GitHub 풀 리퀘스트를 생성합니다. 사용자가 새 PR을 만들거나, 코드 리뷰를 제출하거나, 풀 리퀘스트를 열고자 할 때 사용합니다. 트리거 키워드 -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
다른 열린 이슈들을 스캔하여 주어진 PR이 함께 수정하거나 실수로 망가뜨릴 수 있는 이슈를 찾습니다. 인접 수정 기회와 모순 위험을 file:line…과 함께 출력합니다.
fhir-basics
nvidia
에이전트에게 FHIR R4 API의 작동 방식, 사용 가능한 리소스, 검색 매개변수를 사용한 쿼리 방법, 모든 응답 형식을 올바르게 파싱하는 방법을 가르칩니다…
compileiq-validate-result
nvidia
검색이 완료된 후, 속도 향상을 청구하거나 ACF를 발송하기 전에 사용합니다. dump_results CSV를 로드하고, 상위 K개 후보(단일 목표)를 추출합니다…
changelog-audit
nvidia
릴리스 전에 Warp CHANGELOG.md를 감사합니다: 누락된 항목 복구, 사용자 영향별 정렬, 항목 언어 다듬기, 줄 바꿈, (릴리스 브랜치 모드) 비교 업데이트…
maintain-dynamic-plugins
nvidia
NeMo Relay 동적 플러그인 로더, 매니페스트, Rust 네이티브 SDK, gRPC 워커 프로토콜, Python 워커 SDK, 문서, 테스트 및 릴리스 워크플로 커버리지를 유지 관리합니다.
dgx-diagnose
nvidia
일반적인 DGX Station GB300 문제 진단 — CUDA 충돌, 잘못된 GPU 타겟팅, vLLM/SGLang 컨테이너 버그, MIG 상태 문제, NVLink/Fabric Manager 오류,…