CKConv: Continuous Kernel Convolution For Sequential Data

This repository contains the source code accompanying the paper:

CKConv: Continuous Kernel Convolution For Sequential Data [Slides]
David W. Romero, Anna Kuzina, Erik J. Bekkers, Jakub M. Tomczak & Mark Hoogendoorn.

Fitting Convolutional Kernels via MLPs

In this folder, you can find the code used in our experiments to reconstruct convolutional kernels via a (Sine) MLP.

Functionality:

Several functions can be found in functions.py. Some functions may receive some parameters as input. For example, a Gaussian might be parameterized via its mean and its variance. This parameters are controlled via argparse and follow the dictionary of parameters found in config.py.

Other parameters in config.py are used to parameterized the MLP used in the experiment. For instance, define the nonlinearity used.

An example is, for instance:

python fit_function.py --config.function SinusChirp --config.min -15.0 --config.max 0.0 --config.no_samples 1000 --config.optim Adam --config.lr 1e-4 --config.no_iterations 20000 --config.seed 0 --config.device cuda --config.kernelnet_norm_type LayerNorm --config.kernelnet_activation_function Sine --config.kernelnet_no_hidden 32 --config.padding 0 --config.kernelnet_omega_0 1000.0

This call will fit a SinusShirp defined between -15 and 0 sampled at 1000 positions. The network used to approximate that function is given by a 3-layer MLP with 32 hidden units, Sine nonlinearities and omega_0 = 1000.