search

proteus1991 / GridDehazeNet

This repo contains the official training and testing codes for our paper: GridDehazeNet: Attention-Based Multi-Scale Network for Image Dehazing.
https://jhc.sjtu.edu.cn/~xiaohongliu/
243 stars 54 forks source link

about the loss #15

Closed Jimzhou82sub closed 4 years ago

Jimzhou82sub commented 4 years ago

In many image restoration works like super resolution, denoising and so on, when I train the network ,I found after 1or2epoch ,the loss keeps an approximately stable value, Is this normal?

proteus1991 commented 4 years ago

To my best understanding, if the network is designed well (i.e., effective), the loss would keep decreasing to achieve the local minimum, which usually needs more than 2 epochs.