SUAVE v0.1.2a1 - Alpha Release

To install the latest version (with pip):

pip install --upgrade scikit-learn

Release Highlights

1. Improved Flexibility in the fit Method

Added freeze_vae and predictor_fine_tuning parameters to the fit method. These allow training or fine-tuning of SUAVE on new downstream tasks by freezing the VAE and focusing on task-specific predictors.

2. Resolved animate_monitor Inconsistencies

Fixed an issue where the animate_monitor visualization of task losses did not match the validation evaluation results.

3. Normalized Evaluation Metrics

All evaluation results are now normalized by sample size, ensuring consistency across different batch sizes and datasets.

4. Enhanced eval_loss with Deterministic Mode

Added a deterministic parameter to the eval_loss method. By default, predictions for the same sample remain consistent across evaluations.

5. New Reconstruction Evaluation Methods

Introduced two new methods eval_recon_loss and eval_recon_loss_with_pred: Evaluates reconstruction error without requiring labels.

SUAVE v0.1.1a1 - Alpha Release

SUAVE 0.1.1a1 - Alpha Release

SUAVE (Supervised and Unified Analysis of Variational Embeddings) is a Python library that combines Variational Autoencoder (VAE) with Multi-Task Learning (MTL) for dimensionality reduction and predictive modeling. This release (still in Alpha) focuses on improving training flexibility, early stopping mechanisms, and code maintainability, paving the way for future enhancements and more robust workflows.

Key Features in This Release

• Predictor Fine-Tuning Support:
  Introduced a predictor_fine_tuning option to separate the VAE training phase from the predictor training phase. With this feature, users can freeze the VAE (once its training has converged or stopped) and continue to fine-tune the downstream predictors for improved performance on specific tasks.

• Refined Early Stopping Logic:
  Enhanced the early stopping mechanisms for both the VAE and the multi-task prediction heads. It is now possible to terminate VAE training early while continuing to optimize specific prediction tasks. This refined approach is particularly useful in sequential task modeling scenarios, allowing for focused refinement of the downstream heads after stabilizing the latent representations.

Known Limitations

• Model Stability: Alpha version may exhibit instability with large datasets or complex architectures.
• Documentation: Current API documentation is limited, but examples are available in the repository.

Upcoming Features

• Support for Regression Tasks:
  • Extending predictive capabilities to include regression in addition to classification tasks.
• Multimodal Learning:
  • Adding support for multimodal inputs such as image, text, wave, and graph data.
• Advanced Latent Space Analysis Tools:
  • Tools for better visualization and interpretation of the latent space.
  • Feature importance analysis to enhance interpretability of latent representations.

Getting Started

Install the library via PyPI:

pip install suave-ml --upgrade