torch.nn.functional.l1_loss computes a criterion with the MSE, not the MAE

The L1Loss should create a criterion that measures the MAE error, but in TorchSharp it measures the MSE.

I believe that this is incorrect. See the following screenshot comparing TorchSharp and pytorch, which I believe should get identical results, but don't:

See the call to functional::mse_loss:

TorchSharp/src/Native/LibTorchSharp/THSLoss.cpp

Lines 94 to 102 in 26240d5

    
           Tensor THSNN_l1_loss(const Tensor input, const Tensor target, const int64_t reduction) 
        
           { 
        
               CATCH_RETURN_Tensor( 
        
                   auto opts = torch::nn::functional::MSELossFuncOptions(); 
        
                   ApplyReduction(opts, reduction); 
        
                   res = ResultTensor(torch::nn::functional::mse_loss(*input, *target, opts)); 
        
               ) 
        
           }

If this is indeed incorrect, I would be happy to submit a fix in a PR, but am still figuring out how to make my projects use the locally built version of TorchSharp—I'm not the biggest C# / dotnet / nuget expert.

Thank you for the report.

You are correct, the C++ implementation calls the wrong libtorch function -- mse_loss() instead of l1_loss().

I can fix it, or you can submit a PR and get yourself on our 'Contributor' board... :-)

Address issue dotnet#1359

@RensOliemans

@RensOliemans -- take a look at #1360

Fix calculation of L1Loss

Thanks, I had the same diff in RensOliemans@1f7e50b but didn't push it yet: I haven't figured out how to build the TorchSharp so that I get a .nupkg that I can test locally. Is that possible on Linux?

Does dotnet pack not work?

You can look in the azure_pipeline.yaml file for the exact CLI arguments to provide to dotnet pack

Merge pull request #1360 from NiklasGustafsson/bugs

You can look in the azure_pipeline.yaml file for the exact CLI arguments to provide to `dotnet pack`

Thanks, this works. Specifically the `/p:IncludeTorchSharpPackage=true` and `/p:IncludeLibTorchCpuPackages=true` flags for `dotnet pack`.

Closed #1359 as completed via #1360.

Thanks!

Navigation Menu

Repository navigation

`torch.nn.functional.l1_loss` computes a criterion with the MSE, not the MAE #1359

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	Tensor THSNN_l1_loss(const Tensor input, const Tensor target, const int64_t reduction)
	{
	CATCH_RETURN_Tensor(
	auto opts = torch::nn::functional::MSELossFuncOptions();
	ApplyReduction(opts, reduction);

	res = ResultTensor(torch::nn::functional::mse_loss(input, target, opts));
	)
	}

torch.nn.functional.l1_loss computes a criterion with the MSE, not the MAE #1359

Description

Activity

NiklasGustafsson commented on Jul 18, 2024

NiklasGustafsson commented on Jul 19, 2024

RensOliemans commented on Jul 19, 2024

NiklasGustafsson commented on Jul 19, 2024

NiklasGustafsson commented on Jul 19, 2024

RensOliemans commented on Jul 19, 2024

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

`torch.nn.functional.l1_loss` computes a criterion with the MSE, not the MAE #1359