felixriese / susi Goto Github PK

SuSi: Python package for unsupervised, supervised and semi-supervised self-organizing maps (SOM)

Home Page: https://felixriese.github.io/susi

License: BSD 3-Clause "New" or "Revised" License

TeX 0.60% Python 98.19% JavaScript 1.21%

machine-learning data-science opensource self-organizing-map som python semi-supervised-learning supervised-learning unsupervised-learning pypi-package

susi's Introduction

SuSi: Supervised Self-organizing maps in Python

Python package for unsupervised, supervised and semi-supervised self-organizing maps (SOM)

Description

We present the SuSi package for Python. It includes a fully functional SOM for unsupervised, supervised and semi-supervised tasks:

SOMClustering: Unsupervised SOM for clustering
SOMRegressor: (Semi-)Supervised Regression SOM
SOMClassifier: (Semi-)Supervised Classification SOM

License: 3-Clause BSD license
Author: Felix M. Riese
Citation: see Citation and in the bibtex file
Documentation: Documentation
Installation: Installation guidelines
Paper: F. M. Riese, S. Keller and S. Hinz in Remote Sensing, 2020

Installation

Pip

pip3 install susi

Conda

conda install -c conda-forge susi

More information can be found in the installation guidelines.

Examples

A collection of code examples can be found in the documentation. Code examples as Jupyter Notebooks can be found here:

FAQs

How should I set the initial hyperparameters of a SOM? For more details on the hyperparameters, see in documentation/hyperparameters.
How can I optimize the hyperparameters? The SuSi hyperparameters can be optimized, for example, with scikit-learn.model_selection.GridSearchCV, since the SuSi package is developed according to several scikit-learn guidelines.

Citation

The bibtex file including both references is available in bibliography.bib.

Paper:

F. M. Riese, S. Keller and S. Hinz, "Supervised and Semi-Supervised Self-Organizing Maps for Regression and Classification Focusing on Hyperspectral Data", Remote Sensing, vol. 12, no. 1, 2020. DOI:10.3390/rs12010007

@article{riese2020supervised,
    author = {Riese, Felix~M. and Keller, Sina and Hinz, Stefan},
    title = {{Supervised and Semi-Supervised Self-Organizing Maps for
              Regression and Classification Focusing on Hyperspectral Data}},
    journal = {Remote Sensing},
    year = {2020},
    volume = {12},
    number = {1},
    article-number = {7},
    URL = {https://www.mdpi.com/2072-4292/12/1/7},
    ISSN = {2072-4292},
    DOI = {10.3390/rs12010007}
}

Code:

Felix M. Riese, "SuSi: SUpervised Self-organIzing maps in Python", Zenodo, 2019. DOI:10.5281/zenodo.2609130

@misc{riese2019susicode,
    author = {Riese, Felix~M.},
    title = {{SuSi: Supervised Self-Organizing Maps in Python}},
    year = {2019},
    DOI = {10.5281/zenodo.2609130},
    publisher = {Zenodo},
    howpublished = {\href{https://doi.org/10.5281/zenodo.2609130}{doi.org/10.5281/zenodo.2609130}}
}

License

This project is published under the 3-Clause BSD license.

susi's People

Contributors

Stargazers

Watchers

susi's Issues

Support for sparse matrix

Hi, thank you for the great package.
I just wanted to ask whether the support for sparse matrices is still on the roadmap?
Some comments in the code suggest that it was planned.

Quantization Error

Hello,

I think there might be an issue with the Quantization calculation in Susi. The error seems to be much large then when I attempt the calculation myself. I believe the line below is missing the axis being specified, (axis=1):

    quantization_errors = np.linalg.norm(
        np.subtract(weights_per_datapoint, X)
    )

Thanks,

SOMClassifier object has no attribute 'predict_proba'

If I need to plot AUC-ROC curve as in other ML techniques in SKLEARN, is there any function like predict_proba() ? or manually has to calculate? Because it returns 'SOMClassifier' object has no attribute 'predict_proba'... Is there any function for the same?

Multi-output regression for SOMRegressor?

Hi Felix,

I noticed that the fit function for SOMRegressor only allows y to be a matrix of shape = [n_samples, 1]. Is there any way to allow for multi-output regression using SOMRegressor?

Thank you!

Has SUSI quantization or topographic error?

sklearn.utils.fixes.parallel_helper is removed in the latest version of sklearn

Hi Felix, thank you for what seems to be a great SOM implementation. Seems to be because I cannot import susi. It seems that sklearn has completely removed sklearn.utils.fixes.parallel_helper.

I think that Line 501 should become: delayed(self.get_bmu)(som_array, dp)

Greetings,

Bram

Why we need to save SOMRegressor.p?

In this notebook, why we need to save som to SOMRegressor.p in the local disk when the som has been trained?
Does anybody can explain this ?
Thank you.

ValueError: Found array with dim 3. None expected <=2.

Been bashing my head against this one. I have a hyperspectral cube from drill core scanning:

box.shape
(3590, 30, 249)

som = Susi.SOMClustering()
test = som.fit(box)

ValueError: Found array with dim 3. None expected <= 2.

wrong optional type hint?

Here it says that y is Optional which makes sense since for an unsupervised SOM it isn't necessary, however the check_estimation_input function does not allow y to be None.
Am I missing something?

ImportError: cannot import name 'softmax'

ImportError Traceback (most recent call last) in () ----> 1 import susi ~\Anaconda3\lib\site-packages\susi_init_.py in () 6 import numpy as np 7 import scipy.spatial.distance as dist ----> 8 from scipy.special import softmax 9 from sklearn.base import BaseEstimator, RegressorMixin, ClassifierMixin 10 from sklearn.decomposition import PCA ImportError: cannot import name 'softmax'

Saving and loading SOM model

Hi there!

Does SUSI has a save/open dict method? If not, any ideas to do it quicky?