pytorch-zssr Poor inference result

Poor inference result

Open jaegerstar opened this issue 7 years ago • 8 comments

trafficstars

I set "lincoln.png" in the example folder as input and run the code. It turns out that the result is hardly different from the input LR. I took a brief look at the code but nothing seems wrong. I wonder if you had the same problem.

Jan 11 '18 14:01 jaegerstar

Hi,

Can you please try playing with the parameters: -- lr 0.0001 (larger than default) or -- lr 0.000001 (smaller than default) and also trying running for more iterations: --num_batches 20000 (instead of 15000 by default).

The result is sensitive to these parameters. The default params I set there should work ok on the kennedy picture.

Thanks.

Jan 11 '18 14:01 jacobgil

Also realized now you were running on the lincoln image after x2 resizing.

The LR lincoln image I produced the result with is this one: http://www.wisdom.weizmann.ac.il/~vision/zssr/real/img_002_SRF_2_LR.png I'd suggest running on that one first.

Jan 11 '18 17:01 jacobgil

sorry for the late reply. But afer tweaking those parameters ,I still have no luck. Can you share the parameters that you set, or the process that you get those result ? Much appreciated.

Jan 14 '18 08:01 jaegerstar

I run into the same issue as @jaegerstar . The low_res.png and zssr.png differs in size but looks exactly the same.