TommyZihao/Train_Custom_Dataset
Train Custom Dataset
This project is a computer vision training pipeline and image classification framework. It provides a workflow for preparing custom image datasets and fine-tuning pre-trained neural networks to recognize user-defined categories.
The system includes a model interpretability toolkit that generates saliency maps to highlight influential image regions and uses dimensionality reduction to project high-dimensional semantic features into 2D or 3D visualizations.
The framework covers the full lifecycle of model development, including dataset preparation with proportional class splitting, performance evaluation via confusion matrices and ROC curves, and a real-time inference pipeline for processing static images, video files, and live camera streams.
Features
- Custom Vision Training - Provides a complete pipeline for fine-tuning computer vision models on specialized custom datasets.
- Image Classifier Fine-Tuning - Provides a full pipeline to train image classifiers on custom datasets using pre-trained weights and fine-tuning.
- Performance Metrics - Calculates precision and recall and generates confusion matrices and ROC curves to evaluate model accuracy.
- Dataset Stratification - Organizes raw image files into training and testing sets while maintaining the original proportional representation of each category.
- Image Classification - Implements a framework to train and deploy models that assign labels to images based on visual content.
- Classification Training - Implements a complete workflow for training models to assign labels to images in custom datasets.
- Model Performance Evaluators - Quantifies the accuracy and reliability of image classification models by comparing predictions against ground truth labels.
- Training and Testing Splits - Provides utilities to split image datasets into training and testing sets while maintaining proportional class distribution.
- Vision Dataset Preparation - Provides tools for organizing raw image data and analyzing class distributions to prepare datasets for training.
- Data Collections & Datasets - Includes workflows to collect images and organize them into structured datasets for classification analysis.
- Real-Time Model Inference on Frames - Provides a deployment layer for running model inference on static images, video files, and live camera streams.
- Inference Pipelines - Implements an inference pipeline that processes static images, video files, and live camera streams for real-time output.
- High-Dimensional Feature Visualization - Visualizes high-dimensional internal layer activations using dimensionality reduction to identify semantic clusters in the data.
- Model Evaluation Metrics - Includes tools for measuring model performance and quality using confusion matrices and ROC curves.
- Model Interpretability - Offers interpretability tools to explain model decisions by visualizing the influential regions of an image.
- Model Interpretability Toolkits - Ships a toolkit for visualizing internal representations of neural networks via saliency maps and feature projections.
- Pixel Saliency Maps - Generates pixel-level saliency maps to highlight the image regions most influential to the final classification result.
- Real-Time Vision Classifications - Performs real-time image classification and object recognition on live camera streams and video files.
- Model Inference Deployment - Enables the execution of trained models to perform real-time predictions on various visual input sources.
