search

bennyguo / instant-nsr-pl

Neural Surface reconstruction based on Instant-NGP. Efficient and customizable boilerplate for your research projects. Train NeuS in 10min!
MIT License
857 stars 84 forks source link

how to save the trained model and the hashmap used in Hashencoding #50

Closed xuxumiao777 closed 1 year ago

xuxumiao777 commented 1 year ago

Hi, bennyguo I hopes to find a simple way to save the trained model and the hashmap. Do you know how to do that?

bennyguo commented 1 year ago

Hi! The checkpointing saving is implemented via the PyTorch-Lightning Checkpointing callback: https://github.com/bennyguo/instant-nsr-pl/blob/2d8970ddf2cf405e99d09652560ed62e4d1aa7a5/launch.py#L71-L74 You could use all the parameters here in the checkpoint section of the config files to customize the checkpoint saving behavior (now we just save the checkpoint after the last training step). You can find the checkpoint in the ckpt directory in your trial output directory, and you may use torch.load to inspect the state dict stored in it (including the hash tables of course) for further processing.

xuxumiao777 commented 1 year ago

Thank you, your answer helps me.