search

vandal-vpr / vg-transformers

Official Repository of "Learning Sequential Descriptors for Sequence-based Visual Place Recognition "
MIT License
39 stars 5 forks source link

Can the 6G GPU memory run the program? #7

Closed wpumain closed 1 year ago

wpumain commented 1 year ago

RTX3060 6G can the 6G GPU memory run the program?

ga1i13o commented 1 year ago

Hello, if you want to use sequence length of 5 as in our paper, you should be able to do so with a bit of caution

  1. use lightweight backbones -> resnet-18 or cct-224. You should then be able to use seqvlad
  2. if the above is not enough you can lower the train_batch_size to 2. I've found it to not affect significantly results

This should do the trick and fit into 6 Gbs; however if you want to further reduce GPU memory occupancy you can reduce the number of negative samples for the triplet ( 5 by default, nNeg in the parser) to 2 or 3.

hope this helps

wpumain commented 1 year ago

Think you for your help