search

jacquelinelala / GFN

Gated Fusion Network for Joint Image Deblurring and Super-Resolution(BMVC 2018 Oral)
http://xinyizhang.tech/bmvc2018/
141 stars 39 forks source link

question about training details --how to train GFN to reach the PSNR 27.74+ on LR-GOPRO 4× #24

Closed LIUHUACHENG closed 2 years ago

LIUHUACHENG commented 3 years ago

hello ,Dear author ,Xinyi Zhang: I meet some problems during training the GFN model.
I git clone the code and use the README methods to generate training and testing data,use the Default configuration parameters of the model.After complete three stages training to get the stage3-55epoch model,use the completed model to test on testsets,only get about 24 PSNR on on LR-GOPRO 4. But I use the provided pre-training model to test on testset,got 27.81 PSNR,are there some training details or tricks need to add ,if possible ,can you provide more training details?

jacquelinelala commented 3 years ago

Hi, you may check the intermediate results of the first two steps. It may occur gradient explosion. By the way, there is no trick in our method. The released model is trained by the released code.

LIUHUACHENG commented 3 years ago

Ok, thank your reply,I try it again.

cyyycc commented 1 year ago

Ok, thank your reply,I try it again.

I met the same problem , hava you solved this problem?