Performance Degradation in Training

[BheeshmSharma]

Thanks for sharing the code.
However, I encountered an issue. When I perform training using a single sample (as per the provided code), the results are decent after 300 steps. However, when I scale this up to the full dataset (WebVid10M), I notice that the qualitative performance starts to degrade.
I’m not sure where things might be going wrong. Could you suggest any training-related details that might not be covered in the repo but are crucial for training on the full dataset?

[PoliteApps]

Hi @BheeshmSharma did you find out what was the issue? I'm planning to start working with this repo as well.