How do DeePMD-kit behave if Attention layer = 0 ?

[ald3hyde]

Dear all,

I try to build the model using the attention descriptor (se_attn_v2). I want to compress the model to be quickly calculated. However, as you show in the DeePMD-kit v2 paper, when we use se_attn type descriptor, the embedding matrix is embedded through attention layers depending on the number of attention layers (attn_layer). If attn_layer = 0, the embedding matrix equals to the matrix when we use "se_e2_a"?

[wanghan-iapcm]

almost the same, the only difference is that se_atten uses type embedding while se_a uses different embedding nets for different atom types.

[ald3hyde]

Thank you, Dr. Wang, for replaying my question.