TODO for v1.0.0 (Late February 2024 Update)

[mantasu]

Features and Improvements for v1.0.0

Data

• Update scripts to process dataset-by-dataset instead of task-by-task
• Find or add more datasets for classification, detection, and segmentation
• Write processing scripts, update preprocess.py CLI
• Upload the datasets as a single big one to Kaggle

Architecture

• Add prediction interface to handle predictions and conversion + saving to different formats
• Update base model to be extendable with glasses classifier, detector, and segmenter children
• Change class-based model instantiation to arg-based, create default models for each specification
• Require task for CLI, add size alias, detect device when None

Features

• Add detection dataset, tiny detection model, and binary detection wrapper
• Identify the best hyperparameters and set them as default for each wrapper
• Train models of different sizes and kinds for different tasks (36 in total)

Documentation

• Add a demo file and a collab notebook for quick examples
• Crop only the eye area in the tables, add more example images to showcase features
• Add the results for each model + their sizes, document training parameters
• Document data download and preprocessing steps
• Add type hints and documentation for the new classes/functions
• Add dataset + model architecture references