Question about hyperparameters #287
Description
I've quite a lot of questions about this model. I've successfully trained latent diffusion on AFHQ dataset. But i'm having a hard time understanding many hyperparameters in yaml files.
In Autoencoder yaml:
embed_dim: Why we are using embeddings in an autoencoder?n_embed: What is this?double_z: What is the purpose of this? I've noticed that it's True for KL autoencoder & False for VQ autoencoder. Why?ch: I know it means channels. But how does this changes model architecture?ch_mult: How does this work?lossconfig.target: This is set totaming.modules.losses.vqperceptual.VQLPIPSWithDiscriminatoris it using a discriminator (like in a GAN)? Why an autoencoder needs an Discriminator?lossconfig.params.disc_weight: Is that related to discriminator in VQLPIPSWithDiscriminator & how does it influences it?lossconfig.params.codebook_weight: What is a codebook weight in a VQ autoencoder?
In Latent Diffusion yaml:
first_stage_key: What is this? In every yaml file it's set toimage.num_timesteps_cond: What does this do? In every file it's set to1.log_every_t: How does this work?
I would be grateful for any form of assistance. Thank you!