I have a regression task (image reconstruction) and the benchmark was done using L1Loss (Pytorch). The input images were normalised to [-1, 1] so that tanh activation could be used. When I tried switching the loss to perceptual SSIM loss from piqa, I get an error, since the input can be negative due to the normalisation.

A possible solution is to change normalisation to [0, 1] range and use it together with sigmoid activation, but I am wondering if there is a way around it, i.e. is there a way to use SSIM loss together with tanh?

  • $\begingroup$ did you find any solution? $\endgroup$ Commented Aug 21, 2021 at 12:46


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.