vincentfung13/MINE

minimum hardward requirements

emjay73 opened this issue · 5 comments

emjay73 commented

Thank you for your nice work!

What if I want to run your code, do I need 48 V100 GPUs as you mentioned in the paper?

What are the minimum requirements to run this code?

Thanks in advance.

vincentfung13 commented

Hi, thank you for your interest in our work. The 48 V100 GPUs setting was for training the large RealEstate10K dataset. For running the LLFF training provided in the current code and the demo, a couple of GPUs at most should do.

lijx10 commented

It takes only 1 GPU, e.g., 2080ti, to run network inference and the demo we provided. You may need to tune the batch size and/or input image size if there is out-of-memory problem at inference.

emjay73 commented

Thanks,
Just for information, can you briefly note the training time required for each dataset? (with number of GPUs)

vincentfung13 commented
  • For the LLFF dataset, it takes around a few hours with 2 GPUs;
  • For the KITTI dataset, it takes approximately 1 day with 4 GPUs;
  • For the Flowers Light Fields dataset, it takes 2 days with 4 GPUs.

Note that these numbers depend on the specific hyperparams you choose for your training (such as input size and the number of planes), and you might need to tune it a little bit yourself by looking at the validation results.

emjay73 commented

OK, That seems doable now. 🤣
I'll check the hyperparameters that you mentioned later.
Thank you!