The result is very different from that in the paper

Question

The result is very different from that in the paper

TingMAC opened this issue 6 years ago · comments

I downloaded the pre_trained model and the nature image, and run the command python demo.py --dataroot ./facades/nat_new4 --valDataroot ./facades/nat_new4 --netG ./demo_model/netG_epoch_8.pth, but the obtained results in result_cvpr18 are different from those in the paper. There is obvious haze in the results. If you have time, please tell me how can I solve the problem?

He Zhang · Answer 1 · Fri Jan 18 2019 06:02:24 GMT+0800 (China Standard Time)

If you strictly follow the readme.md and do exactly everything same, you should reproduce the results exactly same as in my github and also the paper. It won't be huge difference if you following everything and use the same image.

Make sure you are using net.train() rather than net.eval().

If you want to reproduce the table results, please using the demo.py to generate the target image and then do the comparison using SSIM.

Let me know if you have further concerns.

TingMAC · Answer 2 · Mon Jan 21 2019 10:17:14 GMT+0800 (China Standard Time)

Thank you. I have reproduce the results.

TingMAC · Answer 3 · Mon Jan 21 2019 17:06:00 GMT+0800 (China Standard Time)

Another question. If the image is not 256256 or 512512, there will be an error message：‘’invalid argument 0: Sizes of tensors must match except in dimension 1. Got 3 and 2 in dimension 2 at /opt/conda/conda-bld/pytorch_1525796793591/work/torch/lib/THC/generic/THCTensorMath.cu:111‘’.
Is there a way to run the code with input hazy image of any arbitrary size？