DAC-SDC-2024 GPU Track 3rd Place

This repository contains the code for the 3rd place solution in the DAC-SDC-2024 GPU Track.

Functionality

You can find everything you need in GPU-Track/SKKU_IRIS_scripts

GPU Server: Training & Model export

Prerequisites

This project heavily based on Ultralytics. Our training and export codes are build on it. You need to install ultralytics

pip install ultralytics

or just use shell script file install.sh for conda environment settings.

bash install.sh

generate_ultralytics_data.py

This file generate the instance segmentation data as the form of ultralytics. In this challenge, our strategy is treating the bounding box as the instance segmentation polygon. Therefore, this script converts all bounding boxes into instance segmentation polygons.

train.py

This script trains a YOLOv8 segmentation model on a GPU.

export.py

This script exports a pretrained YOLOv8 segmentation pytorch model to ONNX model

Jetson Nano: Build engines & Inference

iris_utils_v6.py

This file contains all the pre-/post-processing and TensorRT engine execution behaviors.

• Pre-process:

  • Resize the source image to fit the model input size.
  • Convert data types.
  • Copy data from host (CPU) to device (CUDA).

• Post-process:

  • Execute non-maximum suppression and operations for mask generation using the TensorRT engine (IRIS_post.trt).
  • Crop masks.
  • Upscale masks.
  • Apply thresholds to masks.
  • Rescale masks and bounding boxes.

• TensorRTEngine:

  • Run the neural network using the TensorRT engine (IRIS.trt).

build_engines.py

This script generate TensorRT engines (*.trt, *_post.trt) based on ONNX model file.

IRIS.trt

This is the TensorRT engine file for the neural network model, based on a custom YOLOv8n-seg model.

IRIS_post.trt

This TensorRT engine file implements parts of the post-processing, including:

• Non-maximum suppression.
• Sigmoid activation and matrix multiplication for mask generation.

Usage

To run the inference code on Jetson Nano, follow these steps:

1. Prerequisites:

   • Ensure you have the required dependencies installed, including TensorRT, pycuda, numpy, opencv-python, and tqdm.
   • We recommend you to use JetPack SDK 4.6.1
   • You can install rest of them using pip:

     pip install pycuda tqdm

2. Setup:

   • Place your images in a directory named images located in the parent directory of the project. Ensure the images are in .jpg format.

3. Running the Code:

   • You can run the script by providing the path to the TensorRT engine file (IRIS.trt). If no engine file path is provided, it defaults to IRIS.trt in the current directory.

     python iris_utils_v6.py [path_to_engine_file]

   • Example:

     python iris_utils_v6.py IRIS.trt

     or you can run the script using dac_sdc.ipynb notebook

4. Output:

   • The script processes the images and generates a result dictionary containing the processed output for each image. The processing includes pre-processing the images, running the TensorRT engine, and post-processing the outputs.

5. Profiling:

   • The script includes time profiling for different stages (pre-processing, engine execution, and post-processing) to help analyze performance.
   • For this functionality, you need to install line_profiler

    pip install line_profiler

   • Then you can run the script with an environment variable LINE_PROFILE=1

    LINE_PROFILE=1 python iris_utils_v6.py IRIS.trt