Training fails without alpha_mask

Hello, 
I encountered the same problem discussed in #1038 , and I tried training after removing the alpha_mask block in train.py as mentioned in the previous issue.
However, the estimated training time is now 5 hours, which I do not find reasonable (it used to take less than 10 minutes with the same nerf synthetic dataset). 

Has anyone faced the same problem? 