RT-DETRv1 Tutorial

This tutorial will guide you through using trtutils with RT-DETRv1 models. We will cover:

1. Downloading ONNX weights from RT-DETRv1

2. Building a TensorRT engine

3. Running inference with the engine

4. Advanced features and optimizations

Downloading ONNX Weights

RT-DETRv1 models can be automatically downloaded and converted to ONNX format using the trtutils CLI:

# Download and convert RT-DETRv1 models to ONNX
# Available models: rtdetrv1_r18, rtdetrv1_r50, rtdetrv1_r101
$ python3 -m trtutils download --model rtdetrv1_r18 --output rtdetrv1_r18.onnx --imgsz 640 --opset 17

# For other RT-DETRv1 variants
$ python3 -m trtutils download --model rtdetrv1_r50 --output rtdetrv1_r50.onnx --imgsz 640 --opset 17

Building TensorRT Engine

Once you have the ONNX weights, you can build a TensorRT engine using trtutils:

# Note we can build directly from the ONNX weights
# Note: Need to specify the input shapes (namely the batch dimension is
# left dynamic)
python3 -m trtutils build \
    --onnx $ONNX_PATH \
    --output $OUTPUT_PATH \
    --fp16 \
    --shape images:1,3,640,640 \
    --shape orig_image_size:1,2

Alternatively, if you want to export the engine using the Python API:

from trtutils.builder import build_engine

# Build the engine with FP16 precision
build_engine(
    onnx="rtdetrv1_r18.onnx",
    output="rtdetrv1_r18.engine",
    fp16=True,
    shapes=[("images", (1, 3, 640, 640)), ("orig_image_size", (1, 2))],
)

# For Jetson devices with DLA support
build_engine(
    onnx="rtdetrv1_r18.onnx",
    output="rtdetrv1_r18_dla.engine",
    int8=True,  # Orin series optimize for int8
    fp16=True,  # Can use fp16 on Xavier series
    dla_core=0,  # Use DLA core 0
    shapes=[("images", (1, 3, 640, 640)), ("orig_image_size", (1, 2))],
)