search

tamarott / SinGAN

Official pytorch implementation of the paper: "SinGAN: Learning a Generative Model from a Single Natural Image"
https://tamarott.github.io/SinGAN.htm
Other
3.29k stars 608 forks source link

The indicators in the paper cannot be obtained #133

Open Howardfive opened 3 years ago

Howardfive commented 3 years ago

I've been studying your work recently, and it's been an excellent one.When I was using the SR module, I could not reproduce the indicators in your paper (PSNR/SSIM).The picture I used was 33039_LR and did not modify any parameters in the code you provided.So I would like to ask about your parameter configuration and hardware configuration.Thank you very much.

tamarott commented 3 years ago

The reported numbers are average over the full BSD100 dataset (not per a single result). You can find all the results in the "Downloads" folder.

Howardfive commented 3 years ago

Thank you very much for your answer.I have recalcated the PSNR and SSIM indexes with the generated pictures(BSD100) you provided, but the accuracy is not as high as in your paper, which is about 1.94 lower.I want to ask you where I am wrong.

import cv2 import skimage.measure import compare_psnr,compare_ssim psnr = 0.0 ssim = 0.0 filename = os.listdir('/home/SinGAN-master/SinGAN-master/Downloads/SR_BSD100/SR_BSD100_100/') for name in filename: img = cv2.imread('/home/SinGAN-master/SinGAN-master/Downloads/SR_BSD100/SR_BSD100_100/' + name) gt = cv2.imread('/home/dataset/BSDS100/BSDS100_GTmod') # the cropped image(e.g. [3,321,481]→[3,320,480]) psnr1 = compare_psnr(gt,img,255) ssim1 = compare_ssim(gt,img,multichannel=True) psnr += psnr1 ssim += ssim1 avg_psnr = psnr /100 avg_ssim = ssim /100 print("psnr",avg_psnr) print("ssim",avg_ssim)

Thank you very much.