Very Very different PSNR between the original and the latest version

[731875679]

I have followed this program since Aug, 2023. (Many thanks to the relevant workers!) Recently, I noticed the program updated a lot. Therefore, I also updated the codes in my computer. However, when I trained the same directory, which is created by running COLMAP with a series of figures, I got a really very different PSNR after same times of training. With the original version, only use the loss function like belows,

Ll1 = l1_loss(image, gt_image)
loss = (1.0 - opt.lambda_dssim) * Ll1 + opt.lambda_dssim * (1.0 - ssim(image, gt_image))

the PSNR is aroud 20 or 30. However, with the latest version, loss funtion like,

Ll1 = l1_loss(image, gt_image)
if FUSED_SSIM_AVAILABLE:
ssim_value = fused_ssim(image.unsqueeze(0), gt_image.unsqueeze(0))
else:
ssim_value = ssim(image, gt_image)

loss = (1.0 - opt.lambda_dssim) * Ll1 + opt.lambda_dssim * (1.0 - ssim_value)

I got PSNR is aroud 7 or 8, less a lot. I have examined they have both the same training_report function, which print the PSNR. However, observed with my naked eyes that the difference between the figures was very tiny. High distance in PSNR number, low in figures. I want to know WHY?

Really appreciate!!!

[AsherJingkongChen]

Please update your submodules by git submodule update --init --recursive --remote

See #1050

[731875679]

Please update your submodules by git submodule update --init --recursive --remote

See #1050

THANKS A LOT! I will try.

[731875679]

Please update your submodules by git submodule update --init --recursive --remote

See #1050

everything gonna be fine!!! you are right!!! really appreciate!!!