Phil 5.20.20

Graduation today starting at 10:00!



  • Looks like ASRC actually wanted us to build a completely different marketing app that would be done by 3 people in 3-4 Fridays. So no sponsorship there

GPT2 Agents

  • Adding promotion – done. I have to say that I’m pretty pleased with how the parser handles it. We go from:
    62. a8=Q+ Kg1 63. Qa7 Kg2


    [62] expanded: 
    	white: F Caruana moves white pawn from a7 to a8. White pawn is promoted to white queen. Check.
    	black: Ding Liren moves black king from g2 to g1.
    [63] expanded: 
    	white: F Caruana moves white queen from a8 to a7.
    	black: Ding Liren moves black king from g1 to g2.


  • Had a bit of trouble figuring out how to deal with the end of a file with a potentially incomplete game.
  • Everything works!
  • Now I need to write games out to files, do something with the introductions, etc.
  • Also add some variability in the language later
  • Also need to handle queenside castling (O-O-O)


  • See how the 20,000 epoch run went, and start on Wasserstein Loss
  • The 20,000 epoch really wasn’t any better. rerunning the 10k and saving the model
  • The 10k version seems to be wider and fits better in the bounds. Looking at the two charts, it looks like following the loss/accuracy of the fake data make be best in this case:
  • Reading the GAN chapter in Advanced Deep Learning with Keras (github)
    • Adding a plot of plots to show the appearance of the generator over the epochs. It should help in understanding the evolution of the generator
    • Bless StackOverflow
      side = 4
      fig = plt.figure(1)
      axs = fig.subplots(side, side)
      for i in range(side*side):
          X = gen_data()
          plot_grid(axs, X.T, i, side=side)
      fig = plt.figure(2)
      plt.plot([1,2,3,4,3,2,1,2,3], label = "GREEN", color="green")
      plt.plot([1, 2, 3, 4, 5, 6, 7, 8, 9], label = "RED", color="red")
      plt.legend(loc="upper left")
  • 2:00 Meeting
    • Write up a code walkthrough proposal and send to Erik
    • set up 10:00 meeting with Vadim for tomorrow
  • 3:00 Log file meeting (T & Isaac)
    • Demo’d LMN and RB


  • Good discussion about data and paper writing
  • Tried to see if the annotated twitter data is in the DB. Some is, some is not
  • Tried to get the transformers translate example running again, including downloading the files explicitly. Still didn’t work. Submitted a ticket