You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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
The text was updated successfully, but these errors were encountered:
thauptmann
changed the title
Conditioning on Categoriwa
Conditioning on Category
Aug 29, 2023
style=z_global# torch.cat([z_global, cls_emb], dim=1) if self.args.data.cond_on_cat else z_global
into style = torch.cat([z_global, cls_emb], dim=1) if self.args.data.cond_on_cat else z_global
set style_mlp as a layer that map the dim_z_global + dim_cls_emb to dim_z_global
change the global prior to take conditional input
the data loader may also require changes
there may be other change needed
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
The text was updated successfully, but these errors were encountered: