Does the minimal code of Cdpam loss provided is a trained model or not trained?
ericwudayi opened this issue · comments
Hi, Does the minimal code of Cdpam loss provided is a trained model or not trained?:
I have tried this:
For denoising: if modelA is a denoising model then
import cdpam
loss_fn = cdpam.CDPAM()
wav_syn = modelA(wav_noise)
loss = loss_fn.forward(wav_true, wav_syn)
loss.backward() --> update modelA's parameter
I expect that modelA would learn something. Am I right?
Hi, Yes this is correct. However, we think that this loss should be applied in conjunction with other losses like L1 / multiresolution STFT etc. I am planning to post the paper on arxiv soon, but the idea for now is for this loss to work well along with other conventional losses to remove perceptible artifacts not captured by other losses. This is especially true at the beginning of training (like the denoiser that you mentioned). Hope this clarifies things!
Thank you very much!