taylanates24/false_positive_elimination

This is a classification model and a post-processing method of any object detection model for false positive elimination.

false_positive_elimination

Introduction

Welcome to the False Fositive Elimination project! This repository hosts a machine learning application aimed at classifying false positives using a classification algorithm built on the timm library and utilizing PyTorch Lightning. This project is particularly useful for improving the accuracy of object detection models by effectively distinguishing false positives.

Project Structure

false_positive_elimination/
├── docker/
│   └── Dockerfile
├── LICENSE
├── README.md
├── classifier.py
├── model.py
├── get_optim.py
├── inference.py
├── convert_tensorrt.py
├── convert_onnx.py
├── train.py
├── train.yaml
├── data/
│   ├── augmentations.py
│   ├── dataset.py
└── datasets

Getting Started

Prerequisites

Before you begin, ensure you have Docker installed on your machine. If you don't have Docker installed, you can download it from Docker's official website

Installation with Docker

Clone the Repository:

git clone https://github.com/taylanates24/false_positive_classification.git
cd false_positive_classification

Build the Docker Image:

Build an image from a Dockerfile in the project directory. This will install all necessary dependencies.

docker build -t false_positive_elimination -f docker/Dockerfile .

Run the container

Start a container from the image you've just built.

docker run -v $(pwd):/workspace -it --gpus all --rm --ipc host false_positive_elimination:latest

Training

1. Change the train.yaml file with appropriate variables.
   • Choose the number of epochs, learning_rate, validation frequency, optimizer, learning rate scheduler, number of GPUs, batch size, augmentations and their parameters.
2. Run the following code to train the code.

python3 train.py --train_cfg train.yaml

Choosing The Right Confidence Threshold

Coming Soon! The model eventually gives a probability (a softmax output.) We will draw a Precision-Recall curve to choose the right confidence threshold so that you can have a Precision of 1. That means you do not eliminate any true positives!

About the Model

In this project, I use the timm library of the Hugging Face. In this way, you can choose a model among a lot of pre-trained and deployed Hugging Face models by:

available_models = timm.list_models()

And you can use them by:

model = timm.create_model('tf_efficientnet_lite2', pretrained=pretrained, num_classes=num_classes)

I have chosen the EfficientNet model because its accuracy and latency are perfect.