ISED: Intrinsic smooth embedding learning framework

ISED is a Python library that provides temporal smoothness embedding learning through a unique approach for dimensionality reduction of high-dimensional dynamical systems in order to attain smoothned trajectories.

Version update

v0.1.0 --> 0.2.0

Features

• Preprocess the high-dimensional time series for subsequences: Includes the attainment of optimal subsequence length and dimensionality, we offer buffering/appending/sliding window approaches to make subsequences.
• Embedding learning (ISED core): Learn and transform unseen data into low-dimensional smooth trajectories. It requires both time series and subsequences.

Installation

To use ISED, clone the repository and ensure that the required dependencies are installed. Dependencies include NumPy, SciPy, Scikit-learn, and Matplotlib, among others.

# Clone this repository
git clone https://github.com/LeonBai/ISED_learner.git

# Install required packages
pip install -r requirements.txt

Requirements

• Python 3.6+
• NumPy >= 1.18.0
• TensorFlow >= 2.0.0
• SciPy >= 1.4.0
• scikit-learn >= 0.22.0
• scikit-dimension >= 0.3.4
• scikit-image >= 0.16.0
• GPy >= 1.9.0

Notebook (Test_run.ipynb)

The notebook Test_run.ipynb provides a structured example to test the ISED model workflow:

• Loading Functions: Import all relevant functions from the ISED.py script.
• Loading Data: Simulation data is loaded, and a preprocessing method is applied.
• Determine Latent Dimension: The latent dimension (latent_dim) is set for the analysis.
• Training and Evaluation: Train the ISEDModel and analyze the embedding and decoded dynamics.

Usage

Below is a simple code snippet demonstrating how to use the ISED model with pre-processing method:

from ISED_learner import ISED
from sklearn import preprocessing

# Load simulation data file
data_file = './ISED/data/Simulation_data/Xs.pkl'  ## Or your self-defined data path, currently accepting both .pkl and .npy files

# Preprocess data, normalization, and subsequencing with opted subsequence methods (choose from sliding window, buffering and appending methods)
## New in 0.2.0 version 

dp = utils.DataProcessor(
        data_source=data_file,
        alpha=1,
        method='buffering',
        window_length=None,
    )
X_train, X_test, X_norm = dp.process(
    length_mode='quarter',
    x_norm_type='standard', 
    id_method=None
)

latent_dim = 2

model = ISED_learner.ISEDModel(
    input_dim=X_train.shape[-1],
    seq_length=X_train.shape[1],
    latent_dim=latent_dim,
    batch_size=50,
    epochs=500,    # each .fit() runs this many epochs
    encoder_layers=[(50, 'relu'), (latent_dim, 'relu')],
    rnn_layers=[(20, 'gru'), (latent_dim, 'gru')],
    decoder_layers=[(latent_dim, 'relu'), (30, 'relu')],
    optimizer='adam',
    use_early_stopping=True,
    loss_weights={
        'MI_loss': 1.0,
        'GSM_loss': 1.0,
        'LSM_loss': 1.0,
        'time_loss': 0.0, ### This part loss is not included in manuscript. Can set to 1 for extra embedding performance up, 
    },
    verbose=0
) 

# Train ISEDModel on train data

model.fit(X_train[:500],X_norm[:500])


# Attain trajectories on test data
length = 320

from sklearn import preprocessing
from sklearn.decomposition import PCA

z_y = model.transform(X_test[:352])
encoded_data =  preprocessing.MinMaxScaler().fit_transform(z_y)

License

This project is licensed under the MIT License.

Author

• Wenjun Bai (wjbai@atr.jp)

For more details, please visit the GitHub Repository.

Contributing

Contributions are welcome! Please submit a pull request or file an issue to help improve ISED.