Support for different batch sizes

[amrutadharmapurikar]

There is a batch_size argument in the DataLoaders in litegs.training.star, is there any plan to support larger batch sizes while training and testing?

[liaoKM]

The batch_size argument is a legacy parameter retained from previous implementations. Currently, LiteGS maintains high GPU utilization even with 1 batch in typical scenarios, and increasing the batch size shows no performance benefits. We keep this interface because future research might explore how consecutive frames affect reconstruction speed and accuracy. Batchsize support would naturally emerge as a byproduct of that investigation.
While we do have plans to support larger batch sizes, it's currently not a high-priority task. Would you mind sharing the specific scenario driving your need for batch size adjustments (e.g., are you observing low GPU utilization with LiteGS in certain applications)? If this proves to be a common requirement or presents an interesting technical challenge, we would reconsider its prioritization.

[amrutadharmapurikar]

I just wanted to use a higher batch size to replicate the training process of MVGS, but that's alright.