full and remaining unmasked latent embeddings

[ZhaoLiruii]

Hello, I would like to know how this sentence in the paper is achieved: "In MDT, both the full and remaining unmasked latent embeddings are fed into the diffusion model during training."? I can't find this part in the code

[ZhaoLiruii]

@gasvn

[gasvn]

MDT/masked_diffusion/train_util.py

Line 227 in 4b8bcbb

def forward_backward(self, batch, cond):

Check this function.