--- title: "Devel: Customised model components with the bru_mapper system" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{Devel: Customised model components with the bru_mapper system} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} header-includes: - \newcommand{\bm}[1]{\boldsymbol{#1}} --- ```{r, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>", eval = FALSE ) ``` (Vignette under construction!) ## Mapper system introduction ```{r setup, eval = TRUE, echo=FALSE, include=FALSE} library(inlabru) ``` Each inlabru latent model component generates an _effect_, given the latent _state_ vector. The purpose of the mapper system is to define the link from state vectors to effect vectors. For an ordinary model component, named $c$, this link can be represented as a matrix-vector product $$ \eta_c(u_c) = A_c(\text{input}) u_c, $$ where $u_c$ is the latent state vector, $\eta_c(u_c)$ is the resulting component effect vector, and $A_c(\text{input})$ is a _component model matrix_. This matrix depends on the component inputs (covariates, index information, etc) from `main`, `group`, `replicate`, and `weights` in the component definition. The wider scope of component definitions is discussed in the [component vignette](https://inlabru-org.github.io/inlabru/articles/component.html). Here, we will focus on the mapper system itself, how the basic building blocks are used to construct the built-in mappers, and finally how to construct new mappers. The in-package documentation for the mapper methods is contained in four parts: ```{r, eval=FALSE} ?bru_mapper # Mapper constructors ?bru_mapper_generics # Generic and default methods ?bru_mapper_methods # Specialised mapper methods ?bru_get_mapper # Mapper extraction methods ``` Regular users normally at most need some of the methods from `?bru_mapper` and sometimes `?bru_mapper_generics`. The methods in `?bru_mapper_methods` provides more details and allows the user to query mappers and to use them outside the context of inlabru model definitions. The `bru_mapper_define()` and `bru_get_mapper()` methods are needed for those implementing their own mapper class. ## Mappers The main purpose of each mapper class is to allow evaluating a component effect, from given `input` and latent `state`, by calling ```{r, eval=FALSE} ibm_eval(mapper, input, state) ``` for the `mapper` associated with the component definition. ### Basic mappers Basic mappers take covariate vectors or matrices as input and numeric vectors as state. Constructors: * `bru_mapper_const()` * `bru_mapper_linear()` * `bru_mapper_index(n)` * `bru_mapper_factor(values, factor_mapping)` * `bru_mapper_matrix()` * `bru_mapper_harmonics(order, scaling, intercept, interval)` #### const The `const` mapper defines a mapping from an empty state vector. This is used to define offset components, $\eta(u)_j = z_j$, where $z$ is a fixed input vector, indexed by $j$. #### linear The `linear` mapper defines a mapper from a length 1 state vector. This is used to define ordinary "fixed" covariate effects, linear in both the covariate and in the state variable; $\eta(u)_j = z_j u$, where $u$ is the state, and $z$ is the input covariate vector. #### index The `index(n)` mapper defines a direct mapping from a length `n` state variable. This is used for structured and unstructured "random effect" component effects; $\eta(u)_j = u_{z_j}$, where $u$ is the state vector, and $z$ is the input index vector. #### factor The `factor(values, factor_mapping)` mapper defines a direct mapping between a state vector of length equal to the number of factor levels of `values` (for `factor_mapping = "full"`), or one less (for `factor_mapping = "contrast"`). Each factor level is represented as a dummy 0/1 variable, or equivalently, the input is used as a label index into the state vector; $\eta(u)_j = u_{z_j}$. In the `"contrast"` case, the first factor level is removed from the model, and the corresponding effect defined to be zero. #### matrix The `matrix` mapper defines a direct matrix multiplication between a pre-computed model matrix $A_\text{input}$ with the state vector; $\eta(u) = A_\text{input} u$. This is used e.g. for linear model formula input, that is converted to a component model matrix before handing it over to the mapper system. #### harmonics The `harmonics(order, scaling, intercept, interval)` mapper defines a sum of weighted harmonics on a real interval $(L, U)$, each frequency additionally scaled by factors determined by `scale`; $\eta(u_j) = \sum_{k=1}^{1+2p} A_{j,k} u_k$, where $A_{j,k}=s_k$ for $k=1$, $A_{j,2k}=s_{k+1}\cos[2\pi k (z_j - L)/(U-L)]$ for $k=1,2,\dots,p$, and $A_{j,2k+1}=s_{k+1}\sin[2\pi k (z_j - L)/(U-L)]$ for $k=1,2,\dots,p$, where $p=$`order`. When `intercept` is `TRUE`, the state vector has length $1+2p$. When `intercept` is `FALSE`, the first, constant, column is removed, and the state vector has length $2p$. ### Transformation mappers Transformation mappers are mappers that would normally be combined with other mappers, as steps in a sequence of transformations, or as individual transformation mappers. * `bru_mapper_scale()` * `bru_mapper_marginal(qfun, pfun, ...)` * `bru_mapper_aggregate(rescale)` * `bru_mapper_logsumexp(rescale)` #### scale The `scale` mapper multiplies the `input` vector with the `state` vector. This is used for INLA models to implement the `weights` component scaling, as the final stage of a `pipe` mapper, see below. ```{r, eval=FALSE} mapper <- bru_mapper_scale() ibm_eval(mapper, input = ..., state = ... ) ``` #### marginal The `marginal` mapper transforms the `state` vector from standardised Gaussian marginal distribution to the distribution defined by a quantile function. This can be used to define components with latent marginal `N(0,1)` distribution but non-Gaussian effect. ```{r, eval=FALSE} mapper <- bru_mapper_marginal( qfun = ..., pfun = ..., dfun = ..., ..., inverse = ... ) ibm_eval(mapper, input = NULL, # If a list is given, it overrides the parameter specification state = ... ) # Examples: mapper <- bru_mapper_marginal(qfun = qexp, rate = 1 / 8) mapper <- bru_mapper_marginal( qfun = qexp, pfun = pexp, dfun = dexp, rate = 1 / 8 ) ``` #### aggregate The `aggregate` mapper aggregates the `state` vector elements by blockwise summation after scaling the elements by weights. This can be used for summation, integration, and averaging (for `rescale=TRUE`). ```{r, eval=FALSE} mapper <- bru_mapper_aggregate(rescale = ...) ibm_eval(mapper, input = list(block = ..., weights = ...), state = ... ) ``` #### logsumexp The `logsumexp` mapper aggregates the `exp(state)` vector elements by blockwise summation after scaling the elements by weights, and takes the logarithm. The implementation takes care to avoid numerical overflow. This can be used for summation, integration, and averaging (for `rescale=TRUE`). ```{r, eval=FALSE} mapper <- bru_mapper_logsumexp(rescale = ...) ibm_eval(mapper, input = list(block = ..., weights = ...), state = ... ) ``` ### Compound mappers Compound mappers define collections or chains of mappings, and can take various forms of input. The state vector is normally a numeric vector, but can in some cases be a list of vectors. * `bru_mapper_collect(mappers, hidden)` * `bru_mapper_multi(mappers)` * `bru_mapper_pipe(mappers)` * `bru_mapper_repeat(mapper, n_rep)` #### collect The `collect` mapper defines a compound mapper where the Jacobian is block-diagonal, and each block is defined by a separate sub-mapper. When `hidden = TRUE`, it can be used for INLA models where the user-visible part of the latent state is only a part of the internal state, such as for `"bym"` models. ```{r, eval=FALSE} mapper <- bru_mapper_collect(list(name1 = ..., name2 = ..., ...), hidden = FALSE ) ibm_eval(mapper, input = list(name1 = ..., name2 = ..., ...), state = ... ) # If hidden = TRUE, the inla_f=TRUE argument "hides" all but the first mapper: ibm_eval(mapper, input = name1_input, state = ..., inla_f = TRUE ) ``` #### multi The `multi` mapper defines a compound mapper where the Jacobian is the row-wise Kronecker product between the sub-mapper Jacobians. This can be used for INLA models to construct the internal mapper for `main`, `group`, and `replicate`, but also for user-defined models where a single `rgeneric` or `cgeneric` model is defined on e.g. a space-time product domain. ```{r, eval=FALSE} mapper <- bru_mapper_multi(list(name1 = ..., name2 = ..., ...)) ibm_eval(mapper, input = list(name1 = ..., name2 = ..., ...), state = ... ) ``` #### pipe Pipe mappers chain multiple mappers into a sequence, where the evaluation result of each mapper is given as the state vector for the next mapper. ```{r, eval=FALSE} mapper <- bru_mapper_pipe(list(name1 = ..., name2 = ..., ...)) ibm_eval(mapper, input = list(name1 = ..., name2 = ..., ...), state = ... ) ``` #### repeat Repeat mappers take a single mapper and repeats it, by defining the Jacobian as a Kronecker product of a row vector of ones, corresponding to a state vector of length `nrep * ibm_n(mapper, ...)`. The `ibm_` method inputs are passed on to the sub-mapper methods. ```{r, eval=FALSE} mapper <- bru_mapper_repeat(mapper = ..., n_rep = ...) ibm_eval(mapper, input = ..., state = ... ) ``` #### The core model component mapper All model component mappers are currently defined as a `pipe` mapper containing a `multi` mapper followed by a `scale` mapper: ```{r, eval=FALSE} mapper <- bru_mapper_pipe( list( mapper = bru_mapper_multi(list(main = ..., group = ..., replicate = ...)), scale = bru_mapper_scale() ) ) ibm_eval(mapper, input = list( mapper = list(main = ..., group = ..., replicate = ...), scale = ... ), state = ... ) ``` ### Object mappers For object with predefined mapper conversion methods, use `bru_mapper(object)` to obtain a suitable mapper object. Current predefined object mappers are defined for objects of these classes: * `fm_mesh_2d` and `inla.mesh`; Use `bru_mapper(object)` or `bru_mapper_fmesher(object)` * `fm_mesh_1d` and `inla.mesh.1d`; Use `bru_mapper(object, indexed)` or, if `indexed=TRUE`, `bru_mapper_fmesher(object)`. * Any object supporting the `fmesher::fm_dof()` and `fmesher::fm_basis()` methods; Use `bru_mapper_fmesher(object)`. If `fmesher` in the future gains a common base class for such objects, a corresponding `bru_mapper()` conversion method will be added. More details are given below. ### Model object mappers For inla model objects with predefined mappers, use `bru_get_mapper(object)` Current predefined model object mappers are defined for models of these classes: * `inla.spde`; this includes `inla.spde2.matern()` and `inla.spde2.pcmatern()` models. The conversion method calls `bru_mapper()` in the appropriate way for the mesh type used for the model * `inla.rgeneric`; This relies on the rgeneric definition including a `"mapper"` callback argument. A less invasive method, that works for both rgeneric and cgeneric models, is to define a `bru_get_mapper()` method for the class, which needs to include a unique class identifier, e.g. add `class(object) <- c("my_unique_model_class", class(object))` and define ```{r, eval=FALSE} bru_get_mapper.my_unique_model_class <- function(model, ...) { ... } ``` and register it in the namespace. ### Special mappers * `bru_mapper_taylor()` ## Mapper methods ```{r, eval=FALSE} ibm_n(mapper, inla_f, ...) ibm_n_output(mapper, input, ...) ibm_values(mapper, inla_f, ...) ibm_jacobian(mapper, input, state, ...) ibm_eval(mapper, input, state, ...) ibm_names(mapper, ...) ibm_inla_subset(mapper, ...) ``` An example where the `inla_f` argument matters is the `bru_mapper_collect` class, when the `hidden=TRUE` argument is used to indicate that only the first mapper should be used for the `INLA::f()` inputs, e.g. for `"bym2"` models. For `inla_f=FALSE` (the default), the `ibm_n` and `ibm_values` methods return the total number of latent variables, but with `inla_f=TRUE` we get the values needed by `INLA::f()` instead: ```{r eval=TRUE} mapper <- bru_mapper_collect( list( a = bru_mapper_index(3), b = bru_mapper_index(2) ), hidden = TRUE ) ibm_n(mapper) ibm_values(mapper) ibm_n(mapper, inla_f = TRUE) ibm_values(mapper, inla_f = TRUE) ``` ```{r eval=TRUE} ibm_n(mapper, multi = TRUE) ibm_values(mapper, multi = TRUE) ibm_n(mapper, inla_f = TRUE, multi = TRUE) ibm_values(mapper, inla_f = TRUE, multi = TRUE) ``` For the `bru_mapper_multi` class, the `multi` argument provides access to the inner layers of the multi-mapper: ```{r eval=TRUE} mapper <- bru_mapper_multi(list( a = bru_mapper_index(3), b = bru_mapper_index(2) )) ibm_n(mapper) ibm_n(mapper, multi = TRUE) ibm_values(mapper) ibm_values(mapper, multi = TRUE) ``` ```{r eval=TRUE} ibm_n(mapper, inla_f = TRUE) ibm_n(mapper, multi = TRUE, inla_f = TRUE) ibm_values(mapper, inla_f = TRUE) ibm_values(mapper, multi = TRUE, inla_f = TRUE) ``` The default `ibm_inla_subset` method determines the inla subset by comparing the full values information from `ibm_values(mapper, inla_f = FALSE)` to the inla specific values information from `ibm_values(mapper, inla_f = TRUE)`, to determine the logical vector identifying the inla values subset. Custom mappers should normally not need to specialise this method. ## Mappers for inla.mesh objects For component models referenced by a character label (e.g. `"iid"`, `"bym2"`, `"rw2"` etc), inlabru will construct default mappers, that in most cases replicate the default `INLA::f()` behaviour. For the `fm_mesh_1d`, `fm_mesh_2d`, `inla.mesh` and `inla.mesh.1d` classes, default mappers can be constructed by the pre-defined `bru_mapper` S3 methods. For 2d meshes (`fm_mesh_2d`, `inla.mesh`), ```{r, eval=FALSE} bru_mapper(mesh) ``` For 1d meshes (`fm_mesh_1d`, `inla.mesh.1d`), ```{r, eval=FALSE} # If ibm_values() should return mesh$loc (e.g. for "rw2" models # with degree=1 meshes) bru_mapper(mesh, indexed = FALSE) # If ibm_values() should return seq_along(mesh$loc) (e.g. for # inla.spde2.pcmatern() models) bru_mapper(mesh, indexed = TRUE) ``` ## Customised mappers A mapper object should store enough information in order for the `ibm_*` methods to work. The simplest case of a customised mapper is to just attached a new class label to the front of the S3 `class()` information of an existing mapper, to obtain a class to override some of the standard `ibm_*` method implementations. More commonly, one would instead start with a basic `list()`, that might _contain_ an existing mapper object, and then add methods that know how to use that information. In all these cases, the `bru_mapper_define()` method should be used to properly set the class information: ```{r, eval=FALSE} bru_mapper_define(mapper, new_class) ``` Implementations can avoid having to define `ibm_n` and `ibm_values` methods, by instead computing and storing `n`, `n_inla`, `values`, and `values_inla` in the mapper object during construction: ```{r, eval=FALSE} bru_mapper_define( mapper = list(n = 10, values = 1:10), new_class = "my_bru_mapper_class_name" ) ``` The default `ibm_n` and `ibm_values` methods checks if these values are available, and return the appropriate values depending on the `inla_f` argument. When `*_inla` values are requested but not available, the methods fall back to the non-inla versions. If the needed information is not found, the default methods give an error message. Note: Before version 2.6.0, `ibm_amatrix` was used instead of `ibm_jacobian`. Version 2.6.0 still supports this, by having the default `ibm_jacobian` method call `ibm_amatrix`, but will give a deprecation message later, and it may be removed in a future version. ## Example Let's build a mapper class `bru_mapper_p_quadratic` for a model component that takes input covariates with values $a_{ij}$, $i=1,\dots,m$ and $j=1,\dots,p$, in a `matrix` or `data.frame` and evaluates a full quadratic expression $$ \eta_i = x_0 + \sum_{j=1}^p a_{ij} x_j + \frac{1}{2}\sum_{j=1}^p\sum_{k=1}^j \gamma_{j,k} a_{ij}a_{ik} x_{j,k}, $$ where the latent component vector is $\bm{x}=[x_0,x_1,\dots,x_p,x_{1,1},\dots,x_{p,1},x_{2,2}\dots,x_{p,p}]$, and $$ \gamma_{j,k} = \begin{cases} 1, & j = k,\\ 2, & j > k. \end{cases} $$ We start with the constructor method. Like the `bru_mapper_matrix`, we require the user to supply a vector of covariate labels, and store them as a character vector. We also include a `min_degree` parameter to control if an intercept (`min_degree <= 0`) and linear terms (`min_degree <= 1`) should be included in the model. ```{r, eval=TRUE} bru_mapper_p_quadratic <- function(labels, min_degree = 0, ...) { if (is.factor(labels)) { mapper <- list( labels = levels(labels), min_degree = min_degree ) } else { mapper <- list( labels = as.character(labels), min_degree = min_degree ) } bru_mapper_define(mapper, new_class = "bru_mapper_p_quadratic") } ``` The `ibm_n` method can compute the value of $n$ from $n=1+p+\frac{p(p+1)}{2}$: ```{r, eval=TRUE} ibm_n.bru_mapper_p_quadratic <- function(mapper, ...) { p <- length(mapper$labels) (mapper$min_degree <= 0) + (mapper$min_degree <= 1) * p + p * (p + 1) / 2 } ``` For `ibm_values`, the default method is sufficient, returning the vector $[1,\dots,n]$ with $n$ obtained by `ibm_n(mapper)`. However, for clearer result naming, we can use a character vector, at least for INLA `f()` models that allow it (we could have an option argument to the `bru_mapper_p_quadratic` constructor to control this): ```{r, eval=TRUE} ibm_values.bru_mapper_p_quadratic <- function(mapper, ...) { p <- length(mapper$labels) n <- ibm_n(mapper) jk <- expand.grid(seq_len(p), seq_len(p)) jk <- jk[jk[, 2] <= jk[, 1], , drop = FALSE] c( if (mapper$min_degree <= 0) "Intercept" else NULL, if (mapper$min_degree <= 1) mapper$labels else NULL, paste0(mapper$labels[jk[, 1]], ":", mapper$labels[jk[, 2]]) ) } ``` While the approach to `ibm_n` and `ibm_values` above works, it is wasteful to recompute the `n` and `values` information for each call. Instead we can use that the default method will check if `n` or `values` are stored in the mapper object, and then return them. If we change the `.` in the method definitions to `_`, we can end the constructor method like this, making subsequent method calls faster: ```{r, eval=FALSE} bru_mapper_p_quadratic <- function(labels, min_degree = 0, ...) { ... mapper <- bru_mapper_define(mapper, new_class = "bru_mapper_p_quadratic") mapper$n <- ibm_n_bru_mapper_p_quadratic(mapper) mapper$values <- ibm_values_bru_mapper_p_quadratic(mapper) mapper } ``` Note that the order matters, since our `ibm_values_bru_mapper_p_quadratic` function calls `ibm_n(mapper)`. We can now define the main part of the mapper interface that computes the model matrix linking the latent variables to the component effect. It's required that `NULL` input should return a $0$-by-$n$ matrix. ```{r, eval=TRUE} ibm_jacobian.bru_mapper_p_quadratic <- function(mapper, input, ...) { if (is.null(input)) { return(Matrix::Matrix(0, 0, ibm_n(mapper))) } p <- length(mapper$labels) n <- ibm_n(mapper) N <- NROW(input) A <- list() in_ <- as(input, "Matrix") idx <- 0 if (mapper$min_degree <= 0) { idx <- idx + 1 A[[idx]] <- Matrix::Matrix(1, N) } if (mapper$min_degree <= 1) { idx <- idx + 1 A[[idx]] <- in_ } for (k in seq_len(p)) { idx <- idx + 1 A[[idx]] <- in_[, seq(k, p, by = 1), drop = FALSE] * in_[, k] A[[idx]][, k] <- A[[idx]][, k] / 2 } A <- do.call(cbind, A) colnames(A) <- as.character(ibm_values(mapper)) A } ``` If the output of a mapper is of a different size than `NROW(input)`, the mapper should define a `ibm_n_output(mapper, input, ...)` method to return the output size for a given input.