About pre-train
ztt0821 opened this issue · comments
Hi, may I ask we need to pre-train only for the RGB image or both grey and RGB images all need pre-train μ first by using L2 Loss, then pre-train σ_μ and σ_n by using the Pre-trained loss ?
Hi,
According to our experience, it is suggested to pre-train μ first by using L2 Loss on both gray and RGB images.
As for gray images, after pre-training μ, the whole framework can be trained end-to-end by using the final loss. (There is no need to pre-train σ_μ and σ_n specifically. That means it only needs two stage to train DBSN.)
As for RGB images, it is also suggested to pre-train σ_μ and σ_n specifically. (Because for each position in RGB image, we operate a 3x3 matrix, and pre-training phases can benefit to convergence. That means it needs three stage to train DBSN.)