Bugs of training in Windows 10 and the solutions

[Zerui-Yu]

In fact deploying this project on windows 10 is quite troublesome, my GPU is 4090, there are a lot of bugs , including:

1. AttributeError: module 'distutils' has no attribute '_msvccompiler'
   Solution: lower the setuptools
2. [WinError 1455] The page file is too small to complete the operation (页面文件太小，无法完成操作)
   Solution: First, Advanced System Settings --> Increase the virtual memory (e.g., 20GB), second, restart your computer and find the following code in dtu.py, colmap.py, blender.py :

def general_loader(self, dataset, batch_size):
    sampler = None
    return DataLoader(
        dataset, 
        num_workers=os.cpu_count(),
        batch_size=batch_size,
        pin_memory=True,
        sampler=sampler
    )

Change os.cpu_count() to a suitable value, maybe 4 or 6? Just have a try.
3. Loss=NaN after 15000 iters
Solution: modify os.cpu_count() to 0, but this will dramatically reduce the training speed to about 5 iters/s after 15000 iters.

All in all, this works.

[zf-123]

Hi, are your results on the dtu dataset consistent with the paper?

[Zerui-Yu]

Hi, are your results on the dtu dataset consistent with the paper?

I tried my best, but I still didn't run the project correctly, so I didn't bother to verify. If you want a better reconstuction result, try PGSR

[zf-123]

Hi, are your results on the dtu dataset consistent with the paper?

I tried my best, but I still didn't run the project correctly, so I didn't bother to verify. If you want a better reconstuction result, try PGSR

Thanks