reduce batch size or number of iterations when using smaller images?

[malfonsoarquimea]

In order to train a dynerf in a reasonable amount of time on my hardware (nvidia 3090) I assume that I should reduce the image size and reduce either the number of iterations or the batch size (number of rays per iteration). I can fit about 2500 rays on my GPU, but I can accumulate gradients to make a larger effective batch size if needed. So, tu put it clear, I would want to train in about 4 days using your dataset with only one 3090, so I reduced the images to 333x250 pixels and then I can:

• reduce batch size to 2500 -3000 and maintain the number of iterations at 300k +250k +100k as explained in the supplementary material
• Reduce the number of iterations to approx 35k + 30k +12k but maintain the batch size at approx 24k

Can you provide some insight on what would work better? Thanks in advance!

[zhaoyang-lv]

We have tried before to train on images with reduced resolution, similar to the size you have. We did not get as sharp results as those trained on high resolution.

I could not predict whether it can train on a single GPU though. We never did such experiments. Even for reduced resolution results, we used at least 4 gpus. Significantly reducing the batch size might cause the network hard to converge at high resolution, which might be different trained at a smaller resolution.

[malfonsoarquimea]

thanks for your response. I will try to experiment with and will let you know the results