Rtvi Cv Customize Model

[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

# 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}}"

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

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

MODEL_NAME_2D="YOLO"
NUM_SENSORS=1
FORCE_REBUILD=false

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

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'

"${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