Classification Demo on Jetson Orin Nano

Load pre-trained EfficientNet model, and make inferences using different engines.

Startup [Already done]

Preparing environment for running EfficientNet in Jetson-Orin Requirements are listed below

pytorch 1.13.0 torchvision 0.14.1 pycuda 2024.1 efficientnet_pytorch 0.7.1 numpy 1.20.3 onnx 1.16.0 tensorrt 8.5.2 (SDK Manager will install it) cv2 4.5.4 (SDK Manager will install it)

Installation of Pytorch and Torchvision are from here

DO NOT USE 'sudo' command

• Since we build torchvision from source, if you execute startup.sh with sudo, torchvision will be accessible only with sudo command

sh startup.sh

TODO

• Install required python packages

  $ pip3 install -r requirements.txt

• Open the jupyter notebook file (demo.ipynb), run each cell, see what happens.

• classification.py will do the similar thing at once.

  $ python3 classification.py --engine pytorch
  $ python3 classification.py --engine tensorrt_fp32
  $ python3 classification.py --engine tensorrt_fp16

  • Use -h option to see the usage

What happens in "Classification.py"

• For the given model version (e.g., b0) and inference engine type (e.g., pytorch), build the EfficientNet model

  • For more details about Efficientnet, please check this out.
  • If you select TensorRT engine, it will automatically build ONNX & TensorRT model from the original PyTorch model (it takes about 10 minutes)

• Then, open the input image and apply preprocessing

  • Resize to 224 x 224
  • Center crop
  • Convert numpy array (np.ndarray) to torch.Tensor
  • Normalize input: substract mean and divide by std. per channel

• Make a prediction for the input image using the model

• Print inference time

• Overlay predictions on image and save it