Merge remote-tracking branch 'origin/master'

Othello/experiments/trainBaselinePlayerVsBest.py

@@ -25,7 +25,7 @@ def reset(self):
         return self
 
     def run(self, lr, silent=False):
-        self.player1 = self.pretrained_player if self.pretrained_player else LargeFCBaselinePlayer(lr=lr)
+        self.player1 = self.pretrained_player if self.pretrained_player else FCBaselinePlayer(lr=lr)
 
         # Player 2 has the same start conditions as Player 1 but does not train
         self.player2 = self.player1.copy(shared_weights=False)
@@ -47,7 +47,6 @@ def run(self, lr, silent=False):
 
             results, losses = self.simulation.run_simulations(games_per_evaluation)
             self.add_loss(np.mean(losses))
-            self.add_results(("Losses", np.mean(losses)))
             self.add_results(("Best", np.mean(results)))
 
             # evaluate

Othello/experiments/trainBaselinePlayerVsSelf.py

@@ -25,7 +25,7 @@ def reset(self):
         return self
 
     def run(self, lr, silent=False):
-        self.player1 = self.pretrained_player if self.pretrained_player else LargeFCBaselinePlayer(lr=lr)
+        self.player1 = self.pretrained_player if self.pretrained_player else FCBaselinePlayer(lr=lr)
 
         games_per_evaluation = self.games // self.evaluations
         start_time = datetime.now()
@@ -44,7 +44,7 @@ def run(self, lr, silent=False):
 
             results, losses = self.simulation.run_simulations(games_per_evaluation)
             self.add_loss(np.mean(losses))
-            self.add_results(("Losses", np.mean(losses)))
+            self.add_results(("Self", np.mean(results)))
 
             # evaluate
             if episode*games_per_evaluation % 1000 == 0:
@@ -54,8 +54,8 @@ def run(self, lr, silent=False):
 
                 if not silent and Printer.print_episode(episode*games_per_evaluation, self.games, datetime.now() - start_time):
                     self.plot_and_save(
-                        "%s vs BEST" % (self.player1.__str__() + (" milestones" if MILESTONES else "")),
-                        "Train %s vs Best version of self\nGames: %s Evaluations: %s\nTime: %s"
+                        "%s vs SELF" % (self.player1.__str__() + (" milestones" if MILESTONES else "")),
+                        "Train %s vs Self\nGames: %s Evaluations: %s\nTime: %s"
                         % (self.player1, episode*games_per_evaluation, self.evaluations, config.time_diff(start_time)))
 
             # If x/5th of training is completed, save milestone

Othello/experiments/trainBaselinePlayerVsTraditionalOpponent.py

@@ -45,7 +45,7 @@ def run(self, lr, silent=False):
 
             results, losses = self.simulation.run_simulations(games_per_evaluation)
             self.add_loss(np.mean(losses))
-            self.add_results(("Losses", np.mean(losses)))
+            self.add_results(("Training Results", np.mean(results)))
 
             # evaluate
             self.player1.strategy.train, self.player1.strategy.model.training = False, False  # eval mode

TicTacToe/experiments/reinforced/trainBaselinePlayerVsBest.py

@@ -40,7 +40,6 @@ def run(self, lr, weight_decay, silent=False):
             self.simulation = TicTacToe([self.player1, self.player2])
             results, losses = self.simulation.run_simulations(games_per_evaluation)
             self.add_loss(np.mean(losses))
-            self.add_results(("Loss", np.mean(losses)))
             self.add_results(("Best", np.mean(results)))
 
             # evaluate

TicTacToe/experiments/reinforced/trainBaselinePlayerVsSelf.py

@@ -39,7 +39,6 @@ def run(self, lr, weight_decay, silent=False):
 
             results, losses = self.simulation.run_simulations(games_per_evaluation)
             self.add_loss(np.mean(losses))
-            self.add_results(("Loss", np.mean(losses)))
             self.add_results(("Self", np.mean(results)))
 
             # evaluate
@@ -50,8 +49,8 @@ def run(self, lr, weight_decay, silent=False):
 
                 if not silent and Printer.print_episode(episode*games_per_evaluation, self.games, datetime.now() - start_time):
                     self.plot_and_save(
-                        "%s vs BEST" % (self.player1),
-                        "Train %s vs Best version of self\nGames: %s Evaluations: %s\nTime: %s"
+                        "%s vs SELF" % (self.player1),
+                        "Train %s vs Self version of self\nGames: %s Evaluations: %s\nTime: %s"
                         % (self.player1, episode*games_per_evaluation, self.evaluations, config.time_diff(start_time)))
 
         self.final_score, self.final_results, self.results_overview = evaluate_against_base_players(self.player1, silent=False)

TicTacToe/experiments/reinforced/trainBaselinePlayerVsTraditionalOpponent.py

@@ -46,7 +46,7 @@ def run(self, lr, weight_decay, silent=False):
 
             results, losses = self.simulation.run_simulations(games_per_evaluation)
             self.add_loss(np.mean(losses))
-            self.add_results(("Loss", np.mean(losses)))
+            self.add_results(("Training Results", np.mean(results)))
 
             # evaluate
             self.player1.strategy.train, self.player1.strategy.model.training = False, False  # eval mode

abstractClasses.py

@@ -330,7 +330,7 @@ def forward(self, input, legal_moves_map):
             x = input
             # set illegal moves to zero, softmax, set illegal moves to zero again
             # x = input * legal_moves_map
-            x = self.softmax(x)
+            x = self.softmax(x).exp()
             x = x * legal_moves_map
 
             return x