TLViz — Visualising and analysing component models

TLViz is a Python package for visualising component-based decomposition models like PARAFAC and PCA.

Documentation

The documentation is available on the TensorLy website and includes

• A primer on tensors, tensor factorisations and the notation we use
• An example gallery
• The API reference

Dependencies

TLViz supports Python 3.8 or above (it may also work with Python 3.6 and 3.7, though that is not officially supported).

Installation requires matplotlib, numpy, pandas, scipy, statsmodels and xarray.

Installation

To install the latest stable release of TLViz and its dependencies, run:

pip install tensorly-viz

There is also functionality to create improved QQ-plots with Pingoiun. However, this is disabled by default due to the restrictive GPL lisence. To enable this possibility, you must manually install Pingoiun.

To install the latest development version of TLViz, you can either clone this repo or run

pip install git+https://github.com/marieroald/tlviz.git

Some extra dependencies are needed to run the examples, tests or build the documentation. For more information about installing these dependencies, see the installation guide.

Example

import tlviz
import matplotlib.pyplot as plt
from tensorly.decomposition import parafac

def fit_parafac(dataset, num_components, num_inits):
    model_candidates = [
        parafac(dataset.data, num_components, init="random", random_state=i)
        for i in range(num_inits)
    ]
    model = tlviz.multimodel_evaluation.get_model_with_lowest_error(
        model_candidates, dataset
    )
    return tlviz.postprocessing.postprocess(model, dataset)

data = tlviz.data.load_aminoacids()
cp_tensor = fit_parafac(data, 3, num_inits=3)
tlviz.visualisation.components_plot(cp_tensor)
plt.show()

Loading Aminoacids dataset from:
Bro, R, PARAFAC: Tutorial and applications, Chemometrics and Intelligent Laboratory Systems, 1997, 38, 149-171

This example uses TensorLy to fit five three-component PARAFAC models to the data. Then it uses TLViz to do the following steps:

1. Select the model that gave the lowest reconstruction error.
2. Normalise the component vectors, storing their magnitude in a separate weight-vector.
3. Permute the components in descending weight (i.e. signal strength) order.
4. Flip the components so they point in a logical direction compared to the data.
5. Convert the factor matrices into Pandas DataFrames with logical indices.
6. Plot the components using matplotlib.

All these steps are described in the API documentation with references to the literature.

Testing

The test suite requires an additional set of dependencies. To install these, run

pip install tlviz[test]

or

pip install -e .[test]

inside your local copy of the TLViz repository.

The tests can be run by calling pytest with no additional arguments. All doctests are ran by default and a coverage summary will be printed on the screen. To generate a coverage report, run coverage html.

Contributing

Contributions are welcome to TLViz, see the contribution guidelines.