Haodong Yao

I have the same issue. It takes around 36 hrs to download bedroom training image alone.

我猜测题主想问的是，如果我们想要优化状态量R的话，应该是计算d(Rp) / d(fi), 为什么可以用d(Rp) / d(delta(fi))代替？也就是为什么变成了对扰动的小量求导而不是对状态量本身求导。 我的理解是，这里做了一个叫“lifting”的操作，也就是将误差项转变成了关于delta(fi)的方程, 所以这时候我们需要对delta(fi)求导以对新的方程进行最小化。 具体这样做的好处，还想请教一下@gaoxiang12

> I also got the same 2 errors you got, don't remove the line : torch.set_default_tensor_type('torch.cuda.FloatTensor') > ,instead remove this line: "batch = next(raysRGB_iter).to(device)" (for depth also), and use the...

Same here. Please let me know if you have resolved it. Thanks.

> I uploaded the zip files I found on my computer. I hope there are the good ones. > https://drive.google.com/drive/folders/1wUJB1ZOO7fgqIh3sqGYXD9IF37nasRVk?usp=sharing Thanks for sharing! Could you please also provide val_perturbed.zip? Thanks...

> I find a file named `val_perturbed.zip` I hope it contents the correct files. I'm uploading it now Thank you! Really appreciate it.