YOLOR-TRT

Empowering Vision with High-Speed Object Detection

Built with the tools and technologies:

Table of Contents

• Overview
• Features
• Project Structure
  • Project Index
• Getting Started
  • Prerequisites
  • Installation
  • Usage

---

Overview

Yolor-trt is a pioneering open-source project providing an efficient, predictive personal protective equipment (PPE) detection system. This revolutionary technology leverages high-performance object detection models paired with convenient API endpoints, orchestrated services and asynchronous tasks handling. Designed for businesses, developers, and researchers, it ensures a safer workplace by accurately identifying PPE in visual content, delivering streamlined image processing, and facilitating rapid deployment.

---

Features

	Feature	Summary
⚙️	Architecture	Based on the combination of FastAPI, TensorRT and Celery for distributed PPE detection using image data sent via API calls. Task queue and message brokering are managed by RabbitMQ and Redis, ensuring a robust, high-performance system. TensorRT environment is prepared through a Dockerfile, ensuring smooth operation of the system.
🔩	Code Quality	Consistent use of Python's PEP-8 conventions ensures code readability. The code is clearly segmented into utilities, services, routes and controllers facilitating maintainability. Use of Docker and docker-compose ensures consistent operation across different environments.
📄	Documentation	The primary language used in the project is Python, with the presence of yml, py, txt, and names file types. Dependencies are managed by pip and can be found in requirements.txt. Containers are defined using Docker and orchestrated using docker-compose.yml.
🔌	Integrations	Integrated with TensorRT for efficient implementation of the deep learning model. Utilizes FastAPI for routing and API endpoints management. Deploys RabbitMQ and Redis for task queuing and message brokering respectively.
🧩	Modularity	The project is modular with distinct directories for routes, services, controllers and utility functions. Routes are managed by FastAPI routers, enhancing traceability and organization within the larger architectural framework. Utility functions are abstracted out in model_utils.py for use across the application, encouraging code reusability.
🧪	Testing	The project utilizes PyTest for running tests, as indicated by the test commands in the documentation.
⚡️	Performance	The project leverages NVIDIA's TensorRT models for efficient model inference, contributing to faster results. The use of Non-Maximum Suppression (NMS) in model_utils.py helps in predicting accurate output by eliminating redundant bounding boxes.
🛡️	Security	Use of Docker containers, which provide added security through process isolation.
📦	Dependencies	Managed by pip as defined in the requirements.txt file. The system also relies on Docker, RabbitMQ, and Redis systems.

---

Project Structure

└── yolor-trt/
    ├── PPEDetection
    │   ├── .gitignore
    │   ├── Dockerfile
    │   ├── main.py
    │   ├── requirements.txt
    │   └── src
    ├── README.md
    └── docker-compose.yml

Project Index

YOLOR-TRT/

__root__

docker-compose.yml

- The docker-compose.yml orchestrates service creation for a predictive personal protective equipment (PPE) detection system, encompassing a main PPE detection service, an associated worker service, and two additional required services: Redis and RabbitMQ - These coordinated services ensure efficient message brokering and task queuing for robust, high-performance PPE detection.

PPEDetection

main.py	- MaineCoon is a FastAPI-based application outlined in PPEDetection/main.py, offering a comprehensive routing solution within the project structure through the integration of the API router - It ensures smooth navigation and efficient functionality of various endpoints, contributing significantly to the architectural integrity of the entire codebase.
requirements.txt	- PPEDetection/requirements.txt lists the specific software dependencies required for the PPEDetection project - It enables the setup of a uniform development environment and guarantees the smooth running of the project - It includes packages necessary for computer vision tasks, web server operation, data serialization, and handling asynchronous tasks.
Dockerfile	- The Dockerfile in the PPEDetection directory sets up an environment for running a TensorRT-based application - It begins with an Nvidia TensorRT base image, installs necessary Python packages, copies project requirements, and installs them - Additionally, it updates system packages and installs utilities required for multimedia processing.

src

routes

model_route.py	- Model_route.py forms part of the routing system in the PPEDetection application - It handles image data sent via API calls for inference, initiates processing, and manages responses - It also retrieves processing results once ready, converting them to a viewable image format for return to the client.
routes.py	- Routes.py in PPEDetection/src/routes acts as a connection point, integrating model_route with the APIRouter from FastAPI - Predominantly, it assigns a specific "/model" prefix and "model" tag, enhancing the traceability and organization of the routes related to the model within the broader codebase architecture.

controllers

model_controller.py

- Model Controller in the PPEDetection project facilitates image inference tasks - Leveraging helper functions, it converts image bytes to string format, then dispatches this data for further processing - Thus, it serves as an important touchpoint for managing image parsing and initiating subsequent AI analysis tasks.

queue

worker.py	- Worker.py configures and initializes the Celery application within the Personal Protective Equipment (PPE) Detection project - It leverages environment variables for the broker and backend URI setups - Additionally, it includes task definitions from the tasks file located within the same queue module, facilitating distributed task execution across worker nodes.
tasks.py	- The 'tasks.py' within the PPEDetection/src/queue directory serves as the primary component for managing machine learning model prediction tasks - It utilizes an abstraction of the Celery's Task class to efficiently load models for the first task call, eliminating the need to reload for each subsequent task - It also provides an inference function for image data processing.

utils

asserts

coco.names

- The utility file 'coco.names' within the PPEDetection project serves as a predefined dictionary for object recognition - It lists common objects from everyday items like 'bottle' or 'clock', to animals like 'dog', and transportation mediums like 'bicycle' - It aids in identifying and labeling these objects within images or video frames for the personal protective equipment detection system.

helper

model_utils.py

- The code in 'model_utils.py' provides various utility functions geared towards image processing and detection tasks - Key functionalities include converting and resizing images, performing Non-Maximum Suppression (NMS) on prediction results, drawing bounding boxes for detection results, and handling image formats for further processing - These functions support the primary tasks of the PPEDetection project.

services

trt_loader.py

- The 'trt_loader.py' file in the PPEDetection project is responsible for managing NVIDIA's TensorRT models - It prepares, executes, and retrieves results from the inference engine - Specifically, it handles memory allocation for the engine's inputs and outputs, performs inference, and manages the transfer of data between host and device - Furthermore, it also reorganizes the model's flattened outputs back to their original shapes.

model_service.py

- The code in 'model_service.py' is responsible for coordinating the entire object detection process - It manages the loading and configuration of the YOLOR model, preprocesses images for detection, handles detection and results from the model, and conducts post-processing activities including drawing bounding boxes and scaling results back to original image size.

---

Getting Started

Prerequisites

Before getting started with yolor-trt, ensure your runtime environment meets the following requirements:

• Programming Language: Python
• Package Manager: Pip
• Container Runtime: Docker

Installation

Install yolor-trt using one of the following methods:

Build from source:

1. Clone the yolor-trt repository:

❯ git clone https://github.com/RTae/yolor-trt

2. Navigate to the project directory:

❯ cd yolor-trt

3. Install the project dependencies:

Using pip  

❯ pip install -r PPEDetection/requirements.txt

Using docker  

❯ docker build -t RTae/yolor-trt .

Usage

Run yolor-trt using the following command: Using pip  

❯ python {entrypoint}

Using docker  

❯ docker run -it {image_name}