questions

[Armandpl]

things im confused about, quickly jotted down

• why use Independant distrib?
• is using batch_first w/ gru slower?
• understand layer norm
• i don't get why they call it log probs for the recon loss? they take the mse of the symlog and call that a 'Distribution' and take its 'log_prob'

[Armandpl]

• [ ] zero out the model state when getting is_first?
• [ ] storing the first obs w/ action=0?
• [ ] why not use the provided weight innit in pytorch?

[Armandpl]

• understand how nn init works, what is xavier, why are the nn implem doing the init by hand