PyTorch convert function for op 'l1_loss' not implemented

[hex41434]

Hi,

I encountered an error while attempting to convert a PyTorch model to Core ML using the CoreMLTools convert function. The model utilizes the l1_loss operation, and during the conversion process, I received the following error:

RuntimeError: PyTorch convert function for op 'l1_loss' not implemented.

Request:

I am requesting either:

• Implementation support for the l1_loss operation in the CoreMLTools conversion process, or
• Guidance on recommended practices or alternative methods for handling l1_loss during model conversion.

Thank you for your attention to this issue!

[YifanShenSZ]

Hi @hex41434, could you please elaborate your use case a bit? Concretely, why would a loss function present in inference?

As a temporary workaround, you may explicitly decompose your torch.nn.L1Loss layer with torch.mean(torch.abs(y - x))

The long term solution will be to add an op translation. Here is an example for Frobenius norm