ResNet101 PyTorch segmentation example#

The example presents an inference scenario using DALI and ResNet101.

ResNet101 is a segmentation model. Together with DALI on board they form the following scenario:

  1. Preprocessing - DALI-based typical ResNet preprocessing. Instead of images the input data is a video. Includes GPU decoding (using NVDEC), resize and normalization.

  2. Inference - the model returns the probabilities of a given class in every pixel.

  3. Postprocessing - DALI takes the original image and the probabilities and extracts a particular class.

Every step mentioned above is executed by the Triton server. Triton client is used only for reading the test data from disk and handling the result.

Setting up the model#

To set up the model, please run install_dependencies.sh inside your Triton docker container. This script installs PyTorch, which is a part of this example.

Running the example#

To run the example please follow the list:

  1. Clone the repository:

$ git clone https://github.com/triton-inference-server/dali_backend.git

  1. Download the latest Triton release:

$ docker pull nvcr.io/nvidia/tritonserver:23.05-py3

  1. Run the tritonserver container (remember to set the MODEL_PATH properly):

$ MODEL_REPO="//dali_backend/docs/examples/segmentation" && docker run -it --rm --gpus all --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 -p8000:8000 -p8001:8001 -p8002:8002 --privileged -v $MODEL_REPO:/segmentation nvcr.io/nvidia/tritonserver:23.05-py3 bash

  1. Inside the container, install required dependencies

$ cd /segmentation
$ bash install_dependencies.sh

  1. Run the tritonserver:

$ tritonserver --model-repository /segmentation/model_repository

  1. In a separate window run the client script:

$ cd //dali_backend/qa/L0_segmentation
$ python client.py
On this page