kerch 0.3.1.1.post1

Kernel Methods in PyTorch

Kerch

Kerch is a Python package meant for various kernel methods, and in particular Deep Restricted Kernel Machines. You can natively perform SVMs, LS-SVMs, (K)PCA with various kernels, automatic centering, out-of-sample, etc.

The package is built upon PyTorch and supports GPU acceleration.

• Examples
  • LS-SVM
• Installation
  • PIP
  • From source
• Acknowledgements
• Resources
• License

Examples

Training and plotting an LS-SVM

This is done by first instantiating a model, setting its dataset, searching for the hyperparameters, fitting with those parameters and plotting. The implementation can be found here.

import kerch                                        # importation is fast as the modules are only loaded when called
mdl = kerch.model.LSSVM(type="rbf",                 # kernel name
                     representation="dual")         # initiate model
mdl.set_data_prop(data=data,                        # value
                  labels=labels,                    # corresponding labels
                  proportions=[1, 0, 0])            # initiate data
mdl.hyperopt({"gamma", "sigma"},                    # define which parameters to tune
             max_evals=500,                         # define how many trials
             k=10)                                  # 10-fold cross-validation
mdl.fit()                                           # fit the optimal parameters found
kerch.plot.plot_model(mdl)                          # plot the model using the built-in method

Out-of-sample kernels with normalization and centering

The factory class alows for the fast instantiation of various implemented kernels. Centering and normalization are options to choose from and the out-of-sample will also satisfy these properties, based on statistics relative ti the sample. You can easily use numpy arrays ore even python builtins such as range(). An implementation can be found here

sample = np.sin(np.arange(0, 15) / np.pi) + .1
oos = np.sin(np.arange(15, 30) / np.pi) + .1

k = kerch.kernel.factory(type="polynomial", sample=sample, center=True, normalize=True)

k.K  # = k1.k1()
k.k1(y=oos)
k.k1(x=oos)
k.k1(x=oos, y=oos)

Installation

As for now, there are two ways to install the package.

PIP

Using pip, it suffices to run pip install kerch. Just rerun this command with the suffix --upgrade to upgrade the package to its newest version.

From source

You can also install the package directly from the GitHub repository.

git clone --recursive https://github.com/hdeplaen/kerch
cd kerch
pip install -e .

Resources

• Documentation
• E-DUALITY: ERC Adv. Grant website.
• ESAT-STADIUS: KU Leuven, Department of Electrical Engineering (ESAT), STADIUS Center for Dynamical Systems, Signal Processing and Data Analytics.

Contributors

The contributors and acknowledgements can be found in the CONRIBUTORS file.

License

Kerch has a LGPL-3.0 license, as found in the LICENSE file.