Issues with Transferring GIGA to Custom Dataset and Training Performance on Smaller Scales

[Fanqyu]

Hi,

Thanks for your excellent work!

I’ve been working on transferring GIGA to a new custom dataset, but I’ve encountered a few issues. The inputs strictly follow the required format, but I noticed that you didn’t use the data augmentation from VGN. Based on GIGA’s requirements, I adapted the data augmentation code accordingly. However, even after applying it as expected, the training still seems slower than anticipated.

Additionally, I’m curious about GIGA’s performance on smaller-scale datasets. Specifically, I’m working with a dataset where both the Pile and Packed categories contain 5000 scenes each. After experimentation, I found that VGN can handle this scale.

If you’ve already experimented with or have any insights regarding these issues, I would greatly appreciate your guidance.

Thanks a lot!