Open lyshuga opened 8 months ago
The training "MSE" for the diffusion head we are using is actually the MSE between predicted and actual noise, not actions. In contrast, the evaluation MSE is the mse between predicted and actual actions, so they are not comparable. See training MSE computation here.
This is a bit confusing, maybe we should proper action MSE computation to training or remove the training MSE for diffusion head alltogether.
Thanks for your reply, clear now.
Indeed, it would be great at least to add action MSE to the train logs, and detail the names. Should be easy to understand after that :)
Hi,
After doing fine-tuning, I have checked MSE plots in two tabs: " offline_metrics_red_cube/text_conditioned" and "training", and they were quite different.
I was wondering if it is the same function used in both of evaluations?