The trtexec tool is a command-line wrapper included as part of the TensorRT samples.
TAO 5.0.0 exposes the trtexec tool in the TAO Deploy container (or task group when run via launcher)
for deploying the model with an x86-based CPU and discrete GPUs. To run trtexec on other platforms, such as
Jetson devices, or with versions of TensorRT that are not used by default in the TAO containers, you can follow the
official TensorRT documentation on how to get trtexec.
This section describes how to generate a TensorRT engine using trtexec, which allows you to deploy TAO-trained
models on TensorRT, Triton, and Deepstream.
To generate an .onnx file for ActionRecognitionNet, refer to the ActionRecognitionNet documentation.
ActionRecognitionNet currently does not support INT8 calibration.
# convert 2D RGB model with input sequence length is 32 and input size is 224x224
trtexec --onnx=/path/to/model.onnx \
--maxShapes=input_rgb:16x3x96x224x224 \
--minShapes=input_rgb:1x3x96x224x224 \
--optShapes=input_rgb:4x3x96x224x224 \
--fp16 \
--saveEngine=/path/to/save/trt/model.engine
# convert 3D RGB model with input sequence length is 32 and input size is 224x224:
trtexec --onnx=/path/to/model.onnx \
--maxShapes=input_rgb:16x3x32x224x224 \
--minShapes=input_rgb:1x3x32x224x224 \
--optShapes=input_rgb:4x3x32x224x224 \
--fp16 \
--saveEngine=/path/to/save/trt/model.engine
# convert 2D optical-flow model with input sequence length is 32 and input size is 224x224
trtexec --onnx=/path/to/model.onnx \
--maxShapes=input_of:16x3x64x224x224 \
--minShapes=input_of:1x3x64x224x224 \
--optShapes=input_of:4x3x64x224x224 \
--fp16 \
--saveEngine=/path/to/save/trt/model.engine
# convert 3D optical-flow model with input sequence length is 32 and input size is 224x224:
trtexec --onnx=/path/to/model.onnx \
--maxShapes=input_of:16x2x32x224x224 \
--minShapes=input_of:1x2x32x224x224 \
--optShapes=input_of:4x2x32x224x224 \
--fp16 \
--saveEngine=/path/to/save/trt/model.engine