nn_loss is not exported

[tdhock]

Hi! First of all, thanks very much for maintaining torch, which I find very useful in my machine learning research.

I'm trying to define a nn_module based on my own custom loss function.

I looked in nn-loss.R and adapted the code therein to get

my_mse_loss <- torch::nn_module(
  "my_mse_loss",
  inherit = torch::nn_loss,
  initialize = function(reduction = "mean") {
    super$initialize(reduction = reduction)
  },
  forward = function(input, target) {
    torch::nnf_mse_loss(input, target, reduction = self$reduction)
  }
)

but when running the code above I get the error below

> my_mse_loss <- torch::nn_module(
+ "my_mse_loss",
+ inherit = torch::nn_loss,
+ initialize = function(reduction = "mean") {
+ super$initialize(reduction = reduction)
+ },
+ forward = function(input, target) {
+ torch::nnf_mse_loss(input, target, reduction = self$reduction)
+ }
+ )
Erreur : 'nn_loss' n’est un object exporté depuis 'namespace:torch'

I expected that the code above should work without error, so the user could be able to define their own loss modules.
The error above indicates that nn_loss is not exported.
Can nn_loss be exported please?
If not, what is the recommended method to define our own loss modules?

A work-around is below, inherit=nn_mse_loss (which is exported, but are we supposed to inherit from nn_mse_loss?)

my_mse_loss <- torch::nn_module(
  "my_mse_loss",
  inherit = torch::nn_mse_loss,
  initialize = function(reduction = "mean") {
    super$initialize(reduction = reduction)
  },
  forward = function(input, target) {
    torch::nnf_mse_loss(input, target, reduction = self$reduction)
  }
)

[dfalbel]

@tdhock I don't think you need to inheirt from nn_loss, just a bare nn_module seems fine to me:

my_mse_loss <- torch::nn_module(
  "my_mse_loss",
  initialize = function(reduction = "mean") {
    super$initialize(reduction = reduction)
  },
  forward = function(input, target) {
    torch::nnf_mse_loss(input, target, reduction = self$reduction)
  }
)

There's nothing really special about the nn_loss module.

[tdhock]

Thanks!
if that is the case, why does the current definition of nn_mse_loss have inherit=nn_loss?
According to your advice, the inherit could be removed and it still should work fine, right?

[dfalbel]

Yes, the only difference is that it doesn't need self$reduction <- reduction and instead uses self$initialize(reduction).
Honestly, there's no clear win on the approach we used here versus just doing:

my_mse_loss <- torch::nn_module(
  "my_mse_loss",
  initialize = function(reduction = "mean") {
   self$reduction <- reduction
  },
  forward = function(input, target) {
    torch::nnf_mse_loss(input, target, reduction = self$reduction)
  }
)

torch/R/nn-loss.R

Lines 374 to 384 in abd2a09

	nn_mse_loss <- nn_module(
	"nn_mse_loss",
	inherit = nn_loss,
	initialize = function(reduction = "mean") {
	super$initialize(reduction = reduction)
	},
	forward = function(input, target) {
	nnf_mse_loss(input, target, reduction = self$reduction)
	}
	)

[tdhock]

ok great thanks