search

CAIIVS / chuchichaestli

Where you find all the state-of-the-art cooking utensils (salt, pepper, gradient descent... the usual).
GNU General Public License v3.0
3 stars 0 forks source link

fwp default timestep to none #3

Closed mstadelmann closed 5 months ago

mstadelmann commented 6 months ago

@bil-y not sure if that is all there is; if that is the best way to fix it, I did not check what the downstream functions are doing with t_emb. You're certainly a lot faster evaluating this than me.

mstadelmann commented 6 months ago

Another question: there are two parameters that seem currently unused (class_embed_type and num_class_embeds).

I assume this is still work in progress? Will num_class_embeds control the number of time embedding dimensions?

bil-y commented 6 months ago

Another question: there are two parameters that seem currently unused (class_embed_type and num_class_embeds).

I assume this is still work in progress? Will num_class_embeds control the number of time embedding dimensions?

These parameters are for class-conditioned generation. I haven't removed them yet, but I don't think we are interested in implementing this functionality.

bil-y commented 6 months ago

@bil-y not sure if that is all there is; if that is the best way to fix it, I did not check what the downstream functions are doing with t_emb. You're certainly a lot faster evaluating this than me.

Looks good to me, but I will check the downstream functions.

mstadelmann commented 5 months ago

Another question: there are two parameters that seem currently unused (class_embed_type and num_class_embeds). I assume this is still work in progress? Will num_class_embeds control the number of time embedding dimensions?

These parameters are for class-conditioned generation. I haven't removed them yet, but I don't think we are interested in implementing this functionality.

Do you want me to remove them here, or should we leave it as is?

bil-y commented 5 months ago

Another question: there are two parameters that seem currently unused (class_embed_type and num_class_embeds). I assume this is still work in progress? Will num_class_embeds control the number of time embedding dimensions?

These parameters are for class-conditioned generation. I haven't removed them yet, but I don't think we are interested in implementing this functionality.

Do you want me to remove them here, or should we leave it as is?

I'll clean that up separately, together with the add_attention flag.