Create a Measure for a Generic Torch Task — msr

Short form for constructing a MeasureTorch. See the Custom Learning Problems article for how to create and use such measures.

Usage

msr_torch(
  id,
  fun,
  minimize = NA,
  range = c(-Inf, Inf),
  predict_type = "response",
  properties = character(),
  label = NA_character_,
  obs_loss = NULL
)

Arguments

id: (character(1))
The id of the measure.
fun: (function())
The scoring function. It receives whichever of the arguments truth, response, prob, se, lazy_tensor, prediction, task, learner, train_set and weights it declares, and must return a single number. Asking for anything else is an error. An argument that the prediction does not have – weights on a task without a weights_measure column, or prob on a response-only prediction – is not passed at all, so a default declared for it is what the function sees. weights are also withheld when the measure's $use_weights is set to "ignore", so give the argument a default if the measure should score with and without weights, and none if it cannot do without them – asking for weights that do not exist is an error, not a score.
minimize: (logical(1))
Whether a smaller score is better. NA (default) means the direction is unknown.
range: (numeric(2))
The range of possible scores.
predict_type: (character(1))
The predict type the measure requires: "response" (default), "prob", "se" or "lazy_tensor". A measure asking for one it did not declare here still receives it, if the prediction has it.
properties: (character())
Properties of the measure, see Measure. The "requires_task", "requires_learner", "requires_train_set" and "weights" properties are added automatically when fun declares the corresponding argument. "requires_model" is not: a learner argument only says that the learner object is needed, and mlr3 hands that over even when the model was not stored – after a resample() with store_models = FALSE, learner$network is then NULL and a measure reading it scores whatever an empty model gives. Pass properties = "requires_model" yourself whenever the measure reaches for the trained network, so that mlr3 refuses to score instead.
label: (character(1))
The label of the measure.
obs_loss: (function() or NULL)
The per-observation loss. Declared like fun, except that train_set is not available here, and if specified adds the "obs_loss" property. It must return one number per observation: a multi-target loss reduces over the targets (rowMeans()), not over the observations (mean()), and returning a single number is an error rather than a column of that number repeated.

Value

MeasureTorch

Examples

m = msr_torch("hamming", function(truth, response) mean(as.matrix(truth) != response))
m$properties
#> character(0)

# with a per-observation loss
m = msr_torch("mse", function(truth, response) mean((truth - response)^2),
  obs_loss = function(truth, response) (truth - response)^2)
m$properties
#> [1] "obs_loss"