Commits · master · Jan-Frederik Konopka / CSGNet · GitLab

Snippets Groups Projects

Dec 25, 2019
- add requirements.txt · ef8e95ac
  Jan Konopka authored 5 years ago
  
  ef8e95ac
- delete some redundant files · d6d58f5c
  Jan Konopka authored 5 years ago
  
  d6d58f5c
Dec 21, 2019
- cleanup: move powell's method into own file · 1ee0ad00
  Jan-Frederik Konopka authored 5 years ago
  
  1ee0ad00
Dec 20, 2019
- ! add powell's method via copy-paste + visualize rasters as pngs · 4fa9c98a
  Jan-Frederik Konopka authored 5 years ago
  
  4fa9c98a
Nov 16, 2019
- ! official parallelization test (nb changed playout while loop) · 2387367f
  Jan-Frederik Konopka authored 5 years ago
  
  2387367f
Nov 15, 2019
- ! official exploration rate test + minor fixups · 1d9b443b
  Jan-Frederik Konopka authored 5 years ago
  
  1d9b443b
Nov 10, 2019
- allow supervised pretraining for value network + do stop_selection over... · 84b4a670
  Jan-Frederik Konopka authored 5 years ago
  84b4a670
Nov 06, 2019
- make mcts work better and as intended + mcts try-operations-stop-trick + beam search implementation · 0e026238
  Jan-Frederik Konopka authored 5 years ago
  
  0e026238
Oct 31, 2019
- add test mode and printing · d348eaa3
  Jan-Frederik Konopka authored 5 years ago
  
  d348eaa3
Oct 27, 2019
- allow lr drop (was incorrectly implemented) + file with old seemingly somewhat... · c4bf9b57
  Jan-Frederik Konopka authored 5 years ago
  c4bf9b57
Oct 22, 2019
- debugging policy gradients, found matrix multiplication and detach() issues,... · 4d6058ba
  Jan-Frederik Konopka authored 5 years ago
  4d6058ba
Oct 17, 2019
- corrected implementation of policy gradients RL, however it does not work · f342d058
  Jan-Frederik Konopka authored 5 years ago
  
  f342d058
Oct 16, 2019
- (wrong) implementation of policy gradients RL (for compatibility with pipeline) · c65b0518
  Jan-Frederik Konopka authored 5 years ago
  
  c65b0518
Oct 14, 2019
- attempt GPU parallelization (separate files) + allow MCTS testing · 48cb0409
  Jan-Frederik Konopka authored 5 years ago
  
  48cb0409
Oct 04, 2019
- (re)implement keep_root speedup for MCTS (+ failed neural network worker approach) · bc62d12c
  Jan-Frederik Konopka authored 5 years ago
  
  bc62d12c
Sep 26, 2019
- allow loading of h5 datasets (i.e. the CAD dataset) + remove double inversion of input · 457ff0f1
  Jan-Frederik Konopka authored 5 years ago
  
  457ff0f1
Sep 25, 2019
- communication queue + code quality improvements (partly from reverted nn_worker changes) · 53f4e738
  Jan-Frederik Konopka authored 5 years ago
  
  53f4e738
Sep 21, 2019
- Revert "(failed) NN Worker: parallelize MCTS with separate worker thread to... · 516e605d
  Jan-Frederik Konopka authored 5 years ago
  516e605d
- Revert "re last commit: actually it works with .cuda() if we always do... · f9ac5ba4
  Jan-Frederik Konopka authored 5 years ago
  f9ac5ba4
Sep 19, 2019
- re last commit: actually it works with .cuda() if we always do .clone(), this... · 0c5bfb26
  Jan-Frederik Konopka authored 5 years ago
  0c5bfb26
- (failed) NN Worker: parallelize MCTS with separate worker thread to hold all... · ef203e0a
  Jan-Frederik Konopka authored 5 years ago
  ef203e0a
Sep 17, 2019
- reset point: pre: bundle NN evals of ALL workers in a thread · 8f86367d
  Jan-Frederik Konopka authored 5 years ago
  
  8f86367d
Sep 15, 2019
- parallelize MCTS execution by allowing different games to be played in... · 79e61a79
  Jan-Frederik Konopka authored 5 years ago
  79e61a79
Sep 11, 2019
- fix a bug where we would discard leaves_nn instead of evaluating them · e097d2e7
  Jan-Frederik Konopka authored 5 years ago
  
  e097d2e7
- speed up RL by doing smarter backprop + separate RL into functions for... · 78102d31
  Jan-Frederik Konopka authored 5 years ago
  78102d31
- tree parallelization in 1 thread. required as precursor for proper... · 21100096
  Jan-Frederik Konopka authored 5 years ago
  21100096
Sep 10, 2019
- (more pre-parallelization changes) rm pure MCTS option (was defunct anyway),... · 4d8d468d
  Jan-Frederik Konopka authored 5 years ago
  4d8d468d
- reset point for RL parallelization (18m to converge) · d01f45c5
  Jan-Frederik Konopka authored 5 years ago
  
  d01f45c5
Sep 09, 2019
- fix really stupid i accidentally introduced during speedups · dd7cf7a8
  Jan-Frederik Konopka authored 5 years ago
  
  dd7cf7a8
- more attempted minor speedups · b272d49d
  Jan-Frederik Konopka authored 5 years ago
  
  b272d49d
- remove maximax and speed up some other minor stuff (10-20% speedup) · d2c79179
  Jan-Frederik Konopka authored 5 years ago
  
  d2c79179
- store "after"_hidden_state instead of "before"_hidden_state in nodes, i hoped... · 801f2fb4
  Jan-Frederik Konopka authored 5 years ago
  801f2fb4
- cleanup + reset point for RL parallelization (63s for 320+100t) · 68d66c7a
  Jan-Frederik Konopka authored 5 years ago
  
  68d66c7a
Sep 03, 2019
- okay, i am like 3/4 thru with parallelizing now, this is hard and not really... · 7fb98c51
  Jan-Frederik Konopka authored 5 years ago
  7fb98c51
- parallelize training by using proper training mini-batches (also in forward... · 2bdf8435
  Jan-Frederik Konopka authored 5 years ago
  2bdf8435
Sep 02, 2019
- do halfway (reset point): allow parallelization of (supervised) training by... · df595a93
  Jan-Frederik Konopka authored 5 years ago
  df595a93
Aug 30, 2019
- use CSG from CSGNet + combined shape-controlling parameter shape_actions (code... · 022c2dc7
  Jan-Frederik Konopka authored 5 years ago
  022c2dc7
- HUGE BUGFIX: RNN, during MCTS, when root only, was always fed -1 instead of its move · b71accfa
  Jan-Frederik Konopka authored 5 years ago
  
  b71accfa
Aug 28, 2019
- allow multi-stage pipeline (idea: SL+RL) + clean up code · 0ddc0359
  Jan-Frederik Konopka authored 5 years ago
  
  0ddc0359
Aug 22, 2019
- load from file + chamfer distance + invert input + minor changes, higher... · 772167cf
  Jan-Frederik Konopka authored 5 years ago
  772167cf