Open thauptmann opened 12 months ago
Hi!
Currently the cond_on_cat
is not well supported. You can need some additional digging into the code to add this support.
for example:
style = torch.cat([z_global, cls_emb], dim=1) if self.args.data.cond_on_cat else z_global
style_mlp
as a layer that map the dim_z_global + dim_cls_emb
to dim_z_global
in terms of the variable length issue: I think you could take a transformer like encoder and do average pooling at the end to get a 1D latent. And feed them into prior model and vae's decoder like how we use the CLIP embeddings.
Dear @ZENGXH ,
Thank you for providing the source code to your interesting work. We want to use a conditioned LION with one of our data sets to create point clouds for specific classes.
I tested it with Shapenet by setting cfg.data.cond_on_cat to true (1), but in the VAE I get the error:
AttributeError: 'tuple' object has no attribute 'transpose'
, because the points and class_label are only combined into a tuple.Our classes are embeddings based on a character sequence with variable length. How would one incorporate the embedding vector? Concatenating the embedding vector to every point or using it in a later layer? Or would it be better to use an architecture similar to the CLIP embeddings?
Greetings Tony