mir-group / nequip

NequIP is a code for building E(3)-equivariant interatomic potentials
https://www.nature.com/articles/s41467-022-29939-5
MIT License
611 stars 135 forks source link

The difference between two validation metrics in file log❓ [QUESTION] #336

Closed whu-dft closed 1 year ago

whu-dft commented 1 year ago

Hi, Nequip is a very nice work! I'm trying to reproduce the result using the official codes currently. I found there are two validation metrics in the log file. The difference between them is quit large (see below):

image

What is the difference between the parts enclosed by the red box in the picture?

Linux-cpp-lisp commented 1 year ago

Hi @whu-dft ,

Thanks for your interest in our work!

The upper metrics are averages across a single minibatch (see the "batch" column). The later ones are averages over the entire epoch (note the column is now epoch and not batch). Note also that nequip correctly does the whole epoch averages as over all samples in the epoch, and not as a mean of means, which is incorrect.

There is only one log line during the epoch because you have fewer batches than the log frequency, so you only get one line.