Masking results in errors.

[simonlevine]

Hi, If we introduce zeros in the mask key of a batch, the model won't run forward. I'm wondering if there is a fix for this as this breaks node_mask_select and edge_mask_select as the E_idx contains values that are invalid for slicing the nodes tensor in the encoder. Simply removing invalid indices fixes this problem but still results in unneeded computation. I was also wondering how masking should be managed in the outputs, as masked selection results in smaller final outputs and as such losses are not computed over entire inputs. Thanks

[gaozhangyang]

Could you provide examples?