Question on Epochs and training Iterations

nickgkan / 3d_diffuser_actor

Code for the paper "3D Diffuser Actor: Policy Diffusion with 3D Scene Representations"

https://3d-diffuser-actor.github.io/

MIT License

159 stars 16 forks source link

Question on Epochs and training Iterations #36

Closed pawanw17 closed 3 days ago

pawanw17 commented 4 weeks ago

Hi, thanks for the great work!

I had a doubt regarding training iterations and epochs, The paper mentions your training setup with 1.6e4 epochs (for peract). However the code uses training iterations, is there a formula to go from iterations to epochs or are they used interchangeably? this will help me tune in for better time estimations.

twke18 commented 1 week ago

Hi,

Thanks for your interest! Sorry for the confusion about the training iterations and epochs. We estimate the epoch based on the formula of total_iteration / total_num_demonstration * num_demonstration_per_batch, which is 6e5 / 1800 * 48 = 1.6e4 on PerAct. We observe the reported epoch on CALVIN may not be correct, which will be updated in the paper.