Mode collapse with a large dataset

[Poofy1]

First off I would just like to say this most likely is not a bug. I am mostly just asking for advice because I'm not very familiar with the program's config and this page didn't explain the settings in much depth.

I have a dataset of about 65000 images and I expected the snapshot images to be very diverse but they are all very similar. This started becoming noticeable around the 200kimg mark. I believe this is mode collapse. I am using the default settings currently. I have not really tried anything because I don't know what to change. Perhaps increasing batch_size or ada_kimg will help prevent it?

My main questions:
Is there a config setting that I can change that will help prevent mode collapse from happening? Do you have any other advice to prevent mode collapse?

Desktop:

• OS: Windows 10
• PyTorch version 1.7.1
• CUDA toolkit version CUDA 11.1
• NVIDIA driver version: 466.11
• RTX 3070
• Anaconda enviroment

[abstractdonut]

Have you continued training your model since posting this? If so, do recent snapshots show more diversity?

Does your dataset have a lot of variation? Are all images cropped and centered and do they contain a single class of object from a single point of view, or are there multiple types of objects depicted?

[Poofy1]

As I trained more it became more diverse, so it has seemed to fix itself. Thanks :)

[katie-cathy-hunt]

@Poofy1 Hi! After how many kimgs did the snapshot become more diverse?