tags: CS449

Project Meeting

Agenda

  • Set our weekly goals and work distribution
  • Review the training code of conditional DCGAN
  • Understand the differences between CDCGAN and the ordinary GAN
  • How did they generate fake data (add noise)?

Weekly goals

Week 1:

Reminders: DB set up google colab with code
GAN

    • Look through code + Figure out how it works
      • Data
      • Generator
      • Discriminator
  • Figure out potential areas for optimization
  • Start training?
  • Next meeting: 5/4

Week 2

    • Answer all remaining questions about code
    • Get code to run on all systems
      • Run both train and get_level (generate data), make sure the codes run as it is expected
    • Make a list of parameters to modify (and try to optimize) link here
      • Think about what effects we will get with various changes (var in demo_lse.py)
      • Play around with conditional channel: change length and selected channel
      • Reach out to Zach and our mentor to see what we can modify

Week 3:

  • What to ask Cameron:
    • How to fix cuda issue – we got it done
    • How to upload existing code to colab
    • How to convert output of GAN to RL?
    • What parameters are we going to change and how to quantitatively assess the error
    • Setting up RL environment and what challenges we might face
    • Evaluate output and optimize(?)
      • Figure out how to pipe GAN output to RL
    • Decide what information/data should we save during/after training? Loss and images(3-5)
    • Each person do five experiments

Week 4:

Week 5:

What to write in our proposal update:

  • Fixing FID (answering his question 1) (Yi-Ping)
  • Add messing with the fitness function
  • Update on our understanding of the code we are using to generate levels
  • Explain the paramters we messed with to generate levels (answer his question 2)

Work distribution

Select a repo