In DavisVaughan/nodegraph: Infrastructure of node based matrix graphs

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-",
  out.width = "100%"
)

nodegraph

The goal of nodegraph is to provide infrastructure for delaying matrix and array operations.

Installation

devtools::install_github("DavisVaughan/nodegraph")

Example

nodegraph "remembers" the operations that you perform on your matrix/array, but doesn't actually perform the computation. It performs all of the computations at once when you call compute(), computing any necessary dependencies and storing them as it goes.

library(nodegraph)

Create a delay_array to get started.

mat <- matrix(1:10)

delay_mat <- as_delay_array(mat)
delay_mat

It looks similar to a matrix, but under the hood it's pretty different. Let's try and do something with it.

delay_mat2 <- as_delay_array(matrix(1:10))

res <- delay_mat + delay_mat2
res

The output of the + operation knows the shape, but not the actual result. This is because it has not yet been computed.

You can add more operations, and they get chained together. R matrices are promoted to delay_arrays before the computation takes place.

res2 <- res / matrix(5, nrow = 10)
res2

To actually perform the computation, call compute(). It will compute any child dependencies and itself. Along the way, it will set the values so it doesn't have to calculate them again.

compute(res2)

Look, it remembered that res2 has been computed!

res2

Look, res (a dependency of res2) has been computed now too!

res

Extensibility

nodegraph is being designed with extensibility in mind. If you create a new class that inherits from delay_array, then you can define methods for the following 2 functions to get the laziness you see here:

• compute_engine() - Given a set of known children (arguments), and an operation (like +), this defines how your engine computes the result of that operation.

• compute_dim_engine() - Given a set of arguments for an operation (i.e. x and y in the expression x+y), this defines what the dimension of the output should be. At first glance, you might think this is the same no matter what backend you use, but some backends allow for broadcasting which can change the dimensions substantially. The default method uses R's strict dimensionality rules.

Eager execution

If you want to enable eager execution (for debugging or whatever reason), you can do so with set_computation_type("eager") which will result in the computation being performed immediately:

delay_mat + delay_mat2

set_computation_type("eager")

delay_mat + delay_mat2

set_computation_type("lazy")

delay_mat + delay_mat2

Plots

You can plot a lazy matrix and it's dependency chain with plot(). Currently it's pretty ugly. I can't show it here because DiagrammeR requires this to be an HTML document.

new_res <- delay_mat + delay_mat2 / delay_mat
plot(new_res)

DavisVaughan/nodegraph documentation built on May 21, 2019, 10:15 a.m.