finish modifying how t gets fed to both conditional and unconditional mode

[JdbermeoUZH]

DoD:

• [ ] Launch cddpm_only TTA with no class-free-guidance

  • [x] for small DDPM (cpt 6)
  • [x] for small DDPM (cpt 10)
  • [x] for large DDPM (cpt 8)

• [ ] Launch cddpm_only TTA with class-free-guidance (w=30)

  • [x] for small DDPM (cpt 6)
  • [x] for large DDPM (cpt 8)

• [ ] Launch cddpm_only TTA with class-free-guidance (w=100)

  • [x] for small DDPM (cpt 6)
  • [x] for small DDPM (cpt 10)
  • [x] for large DDPM (cpt 8)

• [ ] Launch cddpm_only TTA job with implicit classifier guidance

  • [x] for small DDPM (cpt 6)
  • [x] for large DDPM (cpt 8)

[JdbermeoUZH]

Idea is to:

• Get a sense of how much faster is the smaller than the bigger one
• How do the different forms of guidance push it towards the better ground truth
• Do they differ significantly in the things they converge to

[JdbermeoUZH]

Check loss functions