2019-07-26

End of era



The era of test40 has finished, that training of test40 has been stopped.

Instead we've just started to train test60. What's new there (relative to test40):

  • Residual tower size is 320x24 (24 blocks, 320 filters).
  • Policy head is AlphaZero-style (8×8×73 instead of plain 1858 vector).
  • Value head is WDL.
  • Instead of using fixed nodes for every move while training, KLD threshold is used instead (think less in obvious cases, think more in complicated cases, but in average it's still about the same as it was).
  • [disabled after net 60021 due to issues] Illegal moves are masked from policy head, so that neural network doesn't have to learn detecting legal moves itself.
  • Instead of training value head just on game outcome {-1; 0; 1}, a fraction of tree's Q value after the move is added.
  • Using different FPU at root vs for the rest of the tree.
As it was written in earlier post, it's not going to be a "clean run". If we discover improvements while test60 is being trained, it's likely that they will be incorporated in the middle of the training process.

Let's see how it goes.

12 comments:

  1. I feel that there is still some potential for t40

    ReplyDelete
  2. Any chance of an official "unclean" T40 extension? Though I guess the Jh nets do that to some extent

    ReplyDelete
    Replies
    1. There's a chance, and if it will be done it will actually be by jhorthos. But running 4 runs at the same time would split contribution forces too much.

      Delete
  3. Very good. All glory to KDL! all glory to Q learning! Lc0 will just go full berserk

    ReplyDelete
  4. Go leela! Only thing slowing me down is optimalization of opencl version on AMD GPUs.. It would be great to see improvement there! (as it is even the first evaluation is not finished as it hangs somewhere in the middle of it)

    ReplyDelete
  5. I feel too much experimentation is harmful to the community engagement. We should always have a state of the art network being trained (until convergence) and one experiment.

    I'd try adding value function delta (like TD learning) to the objective function, not add future value function. Just my $.02.

    ReplyDelete
  6. maelic, see if adding -parallism=16 to the client command line doesn't help a bit there. As it is it's trying to start 32 games at once, which, depending on your GPU might take it forever before you see it start to print to the console. Changing the number of parallel threads doesn't add or subtract anything in the long run, unless you've got a top of the top of the line GPU anyway.

    ReplyDelete
    Replies
    1. Hi - thank you for the tip, sadly I am out from the said computer for now for at least 14 days.. The GPU in question is AMD Radeon RX 5700, not top top tier GPU but strong one for sure. I suspect there might be an issue with driver-opencl compatibility as it is a very new card but in the end I am no expert so I am just not sure why it is so slow..

      Delete
  7. Now is a great time to contribute your GPUs! :gpu: :fire:
    Read here for info on how to contribute: https://discordapp.com/channels/425419482568196106/427066771627966466/604323576807555073
    Read more details about it in #dev-log and/or https://blog.lczero.org/2019/07/end-of-era.html
    Note: T60 is a larger net at 320x24b so game generation will be slower than normal.

    ReplyDelete
  8. Great work on the T40. I have been testing T40 for months. And it is currently the strongest engine I have tested. Hardware is 2950x for the AB engines, and RTX 2080 ti for Lc0. LR = 1.14. Waiting to see if T60 can surpass T40.

    ReplyDelete
    Replies
    1. I take it T60 will "obviously" surpass T40. Unless there is some tradeoff here with the bigger net running slower on the same hardware. From time to time I was logging onto the Google cloud and run the free GPUs with the scripts provided, will they need update for T60? I havent tried in a while but they looked like they could use a bit of maintenance, minor errors and warnings were popping up as the ecosystem was evolving.

      Delete

Note: Only a member of this blog may post a comment.