QUICKSTART.md

Quickstart

You can follow these steps to quickly recreate the models covered in the tutorial.

Step 1 - Train a PyTorch model on the CIFAR10 dataset

Execute the following command on a machine with an NVIDIA GPU.

python3 train.py model_bn --checkpoint_path=data/model_bn.pth

Step 2 - Export the trained model to ONNX

Execute the following command on a machine with an NVIDIA GPU.

python3 export.py model_bn data/model_bn.onnx --checkpoint_path=data/model_bn.pth

Tip: Once exported to ONNX, the models can be profiled using the trtexec tool as described in TUTORIAL.md

Step 3 - Build the TensorRT engine

Execute the following command on a machine with an NVIDIA GPU. To use the DLA, you must call this on a machine with a DLA, like Jetson Orin.

python3 build.py data/model_bn.onnx --output=data/model_bn.engine --int8 --dla_core=0 --gpu_fallback --batch_size=32

Step 4 - Evaluate the model on the CIFAR10 test dataset

Execute the following command on a machine with an NVIDIA GPU. You must call this on the same machine that you called build.py.

python3 eval.py data/model_bn.engine --batch_size=32