LSTMIX: Multivariate multistep RNN predictive framework for multidimensional mixing performance metrics.

Overview

This Python repository provides a comprehensive framework for preprocessing and augmenting multivariate and multidimensional mixing performance metrics and statistics to train Recurrent Neural Networks with different cell types (LSTM and GRU) and network architectures (Fully-Connected and Encoder-Decoder) using PyTorch. The framework includes scripts for data generation (read and preprocess), model training, hyperparameter tuning, and prediction generation via sequence rollout, and model uncertainty quantification via ensemble-based perturbations.

Repository Structure

The repository is organized into several key components:

1. datagen.py

This script is responsible for generating and preprocessing datasets for training, validation, and testing. It utilizes two additional functionalities:

• input.py: Handles the import, organization, scaling, and smoothing of raw data from DNS simulations stored in CSV files.
• Windowing class from modeltrain_LSTM.py: Used for augmenting data by windowing and packaging it into corresponding pickle (pkl) files with labels and relevant data for later use.

2. modeltrain_LSTM.py

This script contains essential classes and functionalities for model training. It includes:

• Windowing class: Implements data windowing for augmentation.
• RNN abstract classes and cell-specific child classes: Provides implementations for both fully connected and encoder-decoder architectures using either GRU or LSTM cells.
• Model training logic: Conducts the main training process, saving model states, trained datasets, and the hyperparameters used during training.

3. hyperparam_tuning.py

This script leverages Ray Tune to perform hyperparameter tuning on the RNN architectures defined in modeltrain_LSTM.py. It aims to optimize the model's performance by systematically exploring hyperparameter combinations.

4. rollout_prediction.py

This script facilitates the evaluation of the trained model. It includes functionalities to:

• Plot trained and validated datasets.
• Execute a rollout operation to predict values from the test set and compare them against the ground truth.
• Plot various metrics such as a y=x plot, Wasserstein and K-L divergence plots, and more.

Directory Structure

plaintext LSTMIX/ │  ├── Clean_CSV.py │  ├── config/ │   ├── config_paths.ini │   ├── config_sm.ini │   └── config_sv.ini │  ├── data_gen.py │  ├── figs/ │   ├── input_data/ │   ├── performance_logs/ │   ├── perturbations/ │   ├── rollouts/ │   ├── split_data/ │   ├── temporal_dist/ │   ├── temporal_EMD/ │   └── windowed/ │  ├── hyperparam_tuning.py │  ├── input_data/ │   └── inputdata.pkl │  ├── input.py ├── Load_Clean_DF.py ├── modeltrain_LSTM.py ├── perturbation.py ├── rollout_prediction.py ├── README.md ├── requirements.txt ├── tools_modeltraining.py │  ├── trained_models/ │   ├── data_sets_GRU_ED/ │   ├── data_sets_GRU_FC/ │   ├── data_sets_LSTM_ED/ │   ├── data_sets_LSTM_FC/ │   ├── GRU_ED_logs/ │   ├── GRU_ED_trained_model.pt │   ├── GRU_FC_logs/ │   ├── GRU_FC_trained_model.pt │   ├── hyperparams_GRU_ED.txt │   ├── hyperparams_GRU_FC.txt │   ├── hyperparams_LSTM_ED.txt │   ├── hyperparams_LSTM_FC.txt │   ├── LSTM_ED_logs/ │   ├── LSTM_ED_trained_model.pt │   ├── LSTM_FC_logs/ │   └── LSTM_FC_trained_model.pt │  ├── tuning/ │  ├── best_models/ │ ├── GRU_ED/ │ ├── GRU_FC/ │ ├── LSTM_ED/ │ └── LSTM_FC/ │ └── RawData/ # Not part of the repository, user data

Getting Started

To use this framework, follow these steps:

1. Clone the repository: git clone https://github.com/your-username/your-repository.git
2. Install the required dependencies: pip install -r requirements.txt
3. Execute the data generation script: python datagen.py
4. Train the LSTM model: python modeltrain_LSTM.py
5. Tune hyperparameters (optional): python hyperparam_tuning.py
6. Evaluate model performance: python rollout_prediction.py

Dependencies

• PyTorch
• Ray Tune
• NumPy
• Matplotlib
• Other dependencies specified in requirements.txt

Acknowledgments

• The developers and contributors of PyTorch, Ray Tune, and other open-source libraries used in this framework.