ShunChengWu / 3DSSG

Other
136 stars 21 forks source link

Some issues with training on multiple Gpus #24

Closed zkaiWu closed 2 years ago

zkaiWu commented 2 years ago

Sorry to bother you again There are som error when I used more than one gpu to train sgpn, but that is fine when I just use one gpu. 68feaf90f79678f47b291d5416c03e9

ShunChengWu commented 2 years ago

the current implementation only supports batch size 1. This is related to the GNN we used. If you want to train with multiple GPUs, you may need to change the data loader to Pytorch Geometric loader.

zkaiWu commented 2 years ago

So if I what to train SGPN, how many GPU's memory I need. 12G is not enough, I what to ask if 32G is ok?

ShunChengWu commented 2 years ago

training with multiple GPU is not possible for the current code. If you want to train with SGPN you should set the maximum edge to 80. https://github.com/ShunChengWu/3DSSG/blob/a143511eb7024cbb347019de734e0e495659fdcd/config_example.json#L72