Tong Chen
Tong Chen
I have caught the same problem. How do you solve this? I have googled it and they said might due to the structure problem.
Have you reproduced this code successfully? I think it might miss a lot of file and command.
> > clip_caps[i] = clip_text.mean(0) ~~~~~~~~~^^^ RuntimeError: The expanded size of the tensor (59136) must match the existing size (768) at non-singleton dimension 0. Target sizes: [59136]. Tensor sizes: [768]...
yeah same issue
> 作者您好，我在lolv2-real数据集上训练遇到了两个问题 1.在训练早期的时候，比如eopch=10或者20时，进行验证，网络输出结果可能会是大部分区域都是黑色的图片(甚至整个图片像素值全是0) 2.我在lolv2-real数据集上训练了几次（也修改了batchsize），都复现不出论文中那种效果，每个指标都比论文中的差（有些指标还差挺多的） 是的我在LOLv2 real的dataset上也差了很多
RTX 3090 AND Ada 6000, batchsize=8, all of the other parameters haven't been changed. On the LOL-v1 dataset, we trained for 1000 epochs and achieved PSNR/SSIM/LPIPS: 27.8865/0.8760/0.0839 with 'use_GT_mean' with...