shape_helpers: Helpers for Shape Inference

In mlr3torch: Deep Learning with 'mlr3'

Helpers for Shape Inference

Description

Helpers for writing the private$.shapes_out() method of a PipeOpTorch. They all propagate unknown (NA) dimensions, see the "Shape Inference" section of PipeOpTorch for the shape conventions.

Usage

shape_to_str(x)

broadcast_shapes(shapes, id)

resolve_dim(dim, shape, insert = FALSE)

Arguments

x	(integer() | list() of integer() | NULL) The shape(s) to format. NULL stands for an unknown shape.
shapes	(list() of integer()) The input shapes, all with the same number of dimensions.
id	(character(1)) The id of the PipeOp, which the error message names.
dim	(integer()) The dimension(s) the operator addresses. Negative ones count back from the last dimension: -1 is the last dimension, -2 the one before it, and so on.
shape	(integer()) The input shape.
insert	(logical(1)) Whether a dimension is inserted rather than addressed, as by nn_unsqueeze(): there is then one more position than the shape has dimensions, -1 appending a new last one.

Details

• broadcast_shapes() applies the broadcasting rules of torch, generalized to shapes that may contain NA. Per dimension a known size that is not 1 wins; if all known sizes are 1 and some input is unknown, the result is unknown, because the unknown one may be greater than 1 and would then determine the size. The shapes must already have the same number of dimensions: shorter ones are not left-padded with 1s, because the first dimension is the batch dimension.

• resolve_dim() resolves dimension indices that count from the end, as in torch, to positive ones. Indices that are out of range stay out of range, so that assert_dim_in_range() reports them.

• shape_to_str() formats a shape, or a list() of them, for an error message.

Value

broadcast_shapes() returns an integer() shape, resolve_dim() an integer() of the same length as dim, and shape_to_str() a character(1).

See Also

Other Shape Inference: infer_shapes(), reshape_output_shape()

Examples

broadcast_shapes(list(c(NA, 1), c(NA, 6)), id = "nn_merge_sum")
resolve_dim(-1, c(NA, 3, 8))
shape_to_str(c(NA, 3, 8))

mlr3torch documentation built on Oct. 30, 2026, 5:10 p.m.