Two questions about the descriptor se_a_mask and the descriptor se_e2_a.

[ZhouXY-PKU]

Hello everyone,

I have the following two questions about the descriptor se_a_mask and the descriptor se_e2_a.

The first is that, if we use se_a_mask for a PBC system (I know DP not support that yet, just an assumption), should the results almost be the same with that if we use se_e2_a, or the accuracy of the two DP models can be significantly different (because the descriptors are not the same one)?

The second is that, as se_e2_a has no padding, should the numbers of atoms in all data frames for one mini-batch gradient descent step in training must be same , or can we use data frames with different numbers of atoms for one step of mini-batch gradient descent? (when we use se_e2_a to train, the data frames with different numbers of atoms are used in different step of mini-batch gradient descent, is that right? )

Thank you very much!

[njzjz]

FYI @dingye18 for se_a_mask

se_e2_a requires the same number of atoms, but se_atten_v2 does not. You can set batch_size to mixed:N with se_atten_v2.

[ZhouXY-PKU]

Thank you very much!