Autoencoder is identity function on atom coordinates? Equivalence to EDM

[guanjq]

Hi Minkai,

Thank you for sharing this work! When I analyze the sampling results of GeoLDM, I found the latent variable z_x is almost equal to the decoded atom positions. Below are molecules I reconstructed with decoded atom pos and atom type (left) and z_x and decoded atom type (right) respectively. They are almost same.

z_x + recon atom type	recon atom pos + recon atom type

A further analysis on the reconstruction results of the auto encoder in GeoLDM indicates that both encoder and decoder are almost identity functions on atom coordinates. If so, can I consider GeoLDM is actually equivalent to 3D space diffusion (i.e. EDM) since #latent variables is equal to #atoms and both encoder and decoder are identity functions on atom coordinates, except that there is an auto-encoder part on atom types?

If this is correct, I’m also wondering how did you train the autoencoder in your published version. I can understand the training will lead to identity functions with the reconstruction loss only, but you mentioned in the repo that the encoder is remained untrained. If so, why is the encoder not a random mapping but a identity function instead?

Thanks!