splinter21

1、“We trained the large model from smaller ones and used a smaller learning rate for dcn.” Do you mean this(for example): step 1>5front-10back with DCN+TSA，lr=1e-4，(model S(hallow)). step 2>5front-40back with DCN+TSA，lr(DCN)=5e-5(e.g.),lr_others=1e-4....

> The purpose is the same, but the algorithm is different. I tried subpixel convolution before the ESPCN paper was published([github log](https://github.com/nagadomi/waifu2x/blob/v0.12/lib/DepthExpand2x.lua)) but eventually I chose [deconvolution](https://github.com/torch/nn/blob/master/doc/convolution.md#spatialfullconvolution). Why did you...

有没有实验数据表格？r_blur-scale-PSNR 以及有没有考虑加jpeg噪声

楼主一提这个问题，我都不敢训了 你说的table4是指fsrcnn吗

look at the ablation study part in the paper  sum gets better performance than prod, maybe you can modify prod to sum.

Will some preprocess be helpful? For example, make the LR(input) of dataset more jpeg artifacts for training，and the model will learn to reduce these blocking artifacts?

@HuangBiaoVivo I think the author means that you should prepare the LR-GT image pair, and LR should be made by bicubic first -> weixin-jpeg, not weixin-jpeg first -> download it...

I noticed that you trained the denoising model with 4 datasets, but trained the upsacler with only SYNLA-Dataset, why?

> I have to find where I put the script to originally generate the dataset, but the dataset generation is very slow compared to the speed of training. It is...

> > I noticed that you trained the denoising model with 4 datasets, but trained the upsacler with only SYNLA-Dataset, why? > > The downscaling operation mainly affects salient features...