sdg_pallet_model

NVIDIA's sdg_pallet_model wrapper

Deploy ModelKit Contents Security Report ModelKit Diff Model Card

SDG Pallet Model

This repository contains code for performing optimized TensorRT inference with a pre-trained pallet detection model that was trained using synthetic data with NVIDIA Omniverse Replicator.
The model takes as input a monocular RGB image, and outputs the pallet box estimates. The box esimates are defined for each pallet side face. So a single pallet may have multiple box estimates.

If you have any questions, please feel free to reach out by opening an issue!

Instructions

Step 1 - Install dependencies

Assumes you've already set up your system with OpenCV, PyTorch and numpy.

Install einops for some utility functions.

pip3 install einops

Install torch2trt. This is used for the TRTModule class which simplifies engine inference.

git clone https://github.com/NVIDIA-AI-IOT/torch2trt
cd torch2trt
python3 setup.py develop

Step 2 - Download the ONNX model

Download the pallet model ONNX file.

Model	Notes	Links
pallet_model_v1_all	Trained for wood and other pallets (metal, plastic).	onnx
pallet_model_v1_wood	Trained only for wood pallets.	onnx

Step 3 - Build the TensorRT engine

Option 1 (recommended) - Build the FP16 engine

To build the FP16 engine, call the following:

./build_trt_fp16.sh

Option 2 - Build the INT8 engine

The INT8 model instructions do not yet include calibration. Please only use this model for throughput profiling. The accuracy is likely to vary from FP32/FP16 models. However, once calibration is included, this may become the recommended option given the improved throughput results.

To build the INT8 engine, call the following:

./build_trt_int8.sh

We hope to provide instructions for using the Deep Learning Accelerator (DLA) on Jetson AGX Orin, and INT8 calibration soon.

Step 3 - Profile the engine

To profile the engine with the trtexec tool, call the following:

./profile_engine.sh

Here are the results for a model inference at 256x256 resolution, profiled on Jetson AGX Orin.

Precision	Throughput (FPS)
FP16	465
INT8	710

Notes:

Called jetson_clocks before running
Using MAXN power mode by calling sudo nvpmodel -m0
Batch size 1
--useSpinWait flag enabled to stabilize timings
--useCudaGraph flag enabled to use CUDA graph optimizations. Cuda graph isn't yet used in the predict function.

Step 4 - Run inference on an example image.

python3 predict.py   --output=

For more options

python3 predict.py --help

Next Steps

Try modifying the predict.py code to visualize inference on a live camera feed.