PyTorch Recurrence Matrix / Plot Operator

PyTorch operator to calculate unthresholded Recurrence Matrix / Plot from Tensors on CPU/GPU.

A recurrence plot (RP) is an advanced technique of nonlinear data analysis. It is a visualisation (or a graph) of a square matrix, in which the matrix elements correspond to those times at which a state of a dynamical system recurs (columns and rows correspond then to a certain pair of times). Techniqually, the RP reveals all the times when the phase space trajectory of the dynamical system visits roughly the same area in the phase space.

Source: recurrence-plot.tk

Although a CPU implementation is available, this is just for debugging, because it is non optimized. If CUDA is not present, the CPU version will be compiled.

Requirements

• CUDA
• pytorch
• CUDA device (GPU)

Build & Install

The pytorch operator can be directly build and installed using the system python or within a docker container

Build and install using system python

python setup.py install

Build wheel in docker container

docker build --tag pytorch_recurrence_build .
docker run --rm -v /tmp/out:/out pytorch_recurrence_build

Wheel file can be found in /tmp/out

Install wheel

pip install /tmp/out/*.whl

Usage

Define Recurrence layer

import torch
from recurrence_matrix import recurrence_matrix


class RecurrenceFunction(torch.autograd.Function):

    @staticmethod
    def forward(ctx, x):
        return recurrence_matrix(x)


class RecurrenceLayer(torch.nn.Module):

    @staticmethod
    def forward(x):
        return RecurrenceFunction.apply(x)

Add layer to existing image consuming model

model = torch.nn.Sequential(RecurrenceLayer(), image_model)

Resize image using MaxPooling

image_size = 224
resize_factor = int(num_inputs/image_size)
if resize_factor < 1:
    resize_factor = 1

model = torch.nn.Sequential(
    RecurrenceLayer(),
    torch.nn.Conv2d(1, network_channels, padding=3, kernel_size=7, bias=False),
    torch.nn.MaxPool2d(resize_factor),
    image_model
)

License

MIT licensed as found in the LICENSE file.