Skip to contents

Graph Network

Source: R/nn_graph.R
nn_graph.Rd

Represents a neural network using a Graph that contains mostly PipeOpModules.

Usage

nn_graph(graph, shapes_in, output_map = graph$output$name, list_output = FALSE)

Arguments

graph

(Graph)
The Graph to wrap. Is not cloned.

shapes_in

(named integer)
Shape info of tensors that go into graph. Names must be graph$input$name, possibly in different order.

output_map

(character)
Which of graph's outputs to use. Must be a subset of graph$output$name.

list_output

(logical(1))
Whether output should be a list of tensors. If FALSE (default), then length(output_map) must be 1.

Value

nn_graph

Fields

  • graph :: Graph
    The graph (consisting primarily of PipeOpModules) that is wrapped by the network. When it consists only of PipeOpModules and PipeOpNOPs, the order in which the modules are called is determined once and $forward() calls them directly, which is much faster than running the graph. This is determined anew when PipeOps are added to or removed from the graph, but not when only its edges are changed, so the graph should not be modified in place. The field cannot be set to another graph, because the parameters of the modules are registered when the network is created. Create a new nn_graph instead.

  • input_map :: character()
    The names of the input arguments of the network. Unnamed arguments of $forward() are matched in this order.

  • shapes_in :: list()
    The shapes of the input tensors of the network.

  • output_map :: character()
    Which output elements of the graph are returned by the $forward() method.

  • list_output :: logical(1)
    Whether the output is a list of tensors.

  • module_list :: nn_module_list
    The list of modules in the network.

  • list_output :: logical(1)
    Whether the output is a list of tensors.

See also

Other Graph Network: ModelDescriptor(), TorchIngressToken(), as_learner_torch(), mlr_learners_torch_model, mlr_pipeops_module, mlr_pipeops_torch, mlr_pipeops_torch_ingress, mlr_pipeops_torch_ingress_categ, mlr_pipeops_torch_ingress_ltnsr, mlr_pipeops_torch_ingress_num, model_descriptor_to_learner(), model_descriptor_to_module(), model_descriptor_union(), pipeop_torch()

Examples

graph = mlr3pipelines::Graph$new()
graph$add_pipeop(po("module_1", module = nn_linear(10, 20)), clone = FALSE)
graph$add_pipeop(po("module_2", module = nn_relu()), clone = FALSE)
graph$add_pipeop(po("module_3", module = nn_linear(20, 1)), clone = FALSE)
graph$add_edge("module_1", "module_2")
graph$add_edge("module_2", "module_3")

network = nn_graph(graph, shapes_in = list(module_1.input = c(NA, 10)))

x = torch_randn(16, 10)

network(module_1.input = x)
#> torch_tensor
#> -0.2289
#>  0.1131
#> -0.1302
#> -0.1225
#>  0.0016
#> -0.3070
#> -0.4943
#> -0.0887
#> -0.2062
#> -0.1422
#>  0.1346
#>  0.2569
#> -0.0886
#>  0.0758
#> -0.3757
#> -0.2076
#> [ CPUFloatType{16,1} ][ grad_fn = <AddmmBackward0> ]