Possibly exploding object size when defining a `nn_module` in the `$initialize()` method.

[sebffischer]

Because the default of parent_env for nn_module is parent.frame(), defining an nn_module in an environment that is not a special environment (such as the global environment or a package environment) can lead to exploding object sizes when serializing.

The special environments (global, pkg env) are no problem, because their serialization is handled special anyway.

library(torch)

nn_module1 = nn_module("nn_module1",
  initialize = function(data) {
    data
    self$layer = nn_module("layer",
      initialize = function() {
        self$linear = nn_linear(10, 1)
      },
      forward = function(x) {
        self$linear(x)
      }
    )()
  },
  forward = function(x) {
    self$layer(x)
  }
)

pryr::object_size(nn_module1(rnorm(1L)))
#> 499.06 kB
pryr::object_size(nn_module1(rnorm(10000000L)))
#> 80.50 MB


net = nn_module1(rnorm(1))
net$layer$forward |> environment() |> parent.env() |> parent.env() |> with(data)
#> [1] 1.308416

^{Created on 2025-05-28 with reprex v2.1.1}

I also think this is not a crazy edge case, but something that can actually happen and some big object (like training data) can cause issues here.

I am not really sure what to do about this.