TODO.md

Documentation

• Pydocs
• Getting Started Guide

Datasets

• Choose a demonstration dataset for images
• Choose a demonstration dataset for text
• Automatic dataset balancing
• Stratified validation splits
• Collect and prepare more datasets
• Allow for manifest files instead of directories

Benchmarking

• Create a benchmarking script
• Enable fine-tuning for faster training (Just simply loading pretrained weights will not fit for downstream task with different labels)
• (Optional) If datasets other than text and image are chosen, an appropriate model for this task should be defined and implemented for training.

Inference

• Set up cloud storage to upload pre-trained weights and enable download command
• Define API for use as a library

Models

• Upgrade ResNet architecture to something more modern
• Upgrade RoBERTa architecture to something more modern
• Pre-train models at different sizes
• Pre-train models on new datasets

Export

• Add tools to export to other formats (e.g. ONNX)

Publishing

• Build a live demo
• Announce on social media