RT-DETRv3 Tutorial

This tutorial will guide you through using trtutils with RT-DETRv3 models. We will cover:

1. Downloading ONNX weights from RT-DETRv3

2. Building a TensorRT engine

3. Running inference with the engine

4. Advanced features and optimizations

Downloading ONNX Weights

RT-DETRv3 models can be automatically downloaded and converted to ONNX format using the trtutils CLI:

# Download and convert RT-DETRv3 models to ONNX
# Available models: rtdetrv3_r18, rtdetrv3_r34, rtdetrv3_r50, rtdetrv3_r101
$ python3 -m trtutils download --model rtdetrv3_r18 --output rtdetrv3_r18.onnx --imgsz 640 --opset 16

# For other RT-DETRv3 variants
$ python3 -m trtutils download --model rtdetrv3_r50 --output rtdetrv3_r50.onnx --imgsz 640 --opset 16

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 image:1,3,640,640 \
    --shape im_shape:1,2 \
    --shape scale_factor: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="rtdetrv3_r18.onnx",
    output="rtdetrv3_r18.engine",
    fp16=True,
    shapes=[("image", (1, 3, 640, 640)), ("im_shape", (1, 2)), ("scale_factor", (1, 2))],
)

# For Jetson devices with DLA support
build_engine(
    onnx="rtdetrv3_r18.onnx",
    output="rtdetrv3_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=[("image", (1, 3, 640, 640)), ("im_shape", (1, 2)), ("scale_factor", (1, 2))],
)