The loss doesn't decrease

[hugo-xie]

Hello, I'm trying to train InteractDiffusion using the HOI dataset according to the readme, but I've noticed that the loss isn't decreasing. Instead, it's fluctuating. Is this normal?

[jiuntian]

Yes it's normal

[hugo-xie]

Yes it's normal

So does the loss remain in a fluctuating state from the beginning to the end of training? How do we then choose the final model checkpoint

[jiuntian]

we did evaluation per 100k