Update chess-gpt-eval-contrastive/main.py

chess-gpt-eval-contrastive/main.py

@@ -625,6 +625,29 @@ def play_games(
             games_saved += 1
         else:
             print("Duplicate game; not saved.")
+
+        if update_activations:
+            if player_one_resignation or player_one_failed_to_find_legal_move:
+                player_one.update_activations("lost")
+                player_two.update_activations("won")
+            elif player_two_resignation or player_two_failed_to_find_legal_move:
+                player_one.update_activations("won")
+                player_two.update_activations("lost")
+            else:
+                if board.result() == "1-0":
+                    player_one.update_activations("won")
+                    player_two.update_activations("lost")
+                elif board.result() == "0-1":
+                    player_one.update_activations("lost")
+                    player_two.update_activations("won")
+                else:  # Draw
+                    player_one.update_activations("draw")
+                    player_two.update_activations("draw")
+        
+        if games_saved % contrastive_activation_save_interval == 0:
+            player_one.save_activations(activations_path)
+            player_two.save_activations(activations_path)
+            
     if isinstance(player_one, StockfishPlayer):
         player_one.close()
     if isinstance(player_two, StockfishPlayer):
@@ -658,11 +681,12 @@ move_num_in_gamestate = False
 book_opening = True
 random_opening = True
 random_opening_moves = 10
+contrastive_activation_save_interval = 10
+activations_path="activations.pkl"
+update_activations = True # False = use them
 if __name__ == "__main__":
     for nanogpt_player in player_ones:
-        i = 0
-        for rm in [25]: #range(5, 25, 5):
-        #for i in [0]: # [3] #range(11):
+        for i in [0]: # [3] #range(11):
             num_games = 500
             # player_one = GPTPlayer(model="gpt-3.5-turbo-instruct")
             # player_one = LocalLlamaPlayer(model_name="meta-llama/Llama-2-7b-hf")
@@ -675,6 +699,8 @@ if __name__ == "__main__":
             #player_one_recording_name = f"xformer_rdm_{rm}"
             player_one = MambaPlayer(model_name=player_one_recording_name, move_num_in_gamestate=move_num_in_gamestate)
             player_one_recording_name = f"random_mamba_start/mamba_rdmstart_{rm}"
+            if not update_activations:
+                player_one.apply_contrastive_activations()
 
             #player_two = StockfishPlayer(skill_level=i, play_time=0.1)
             player_two = LC0PLayer(skill=i)
@@ -685,5 +711,5 @@ if __name__ == "__main__":
             #print(f"\n\nSTARTING GAMES AGAINST STOCKFISH LEVEL {i}\n\n")
             print(f"\n\nSTARTING GAMES AGAINST LC0 LEVEL {i}\n\n")
 
-            play_games(player_one, player_two, num_games, book_opening=book_opening, random_opening=random_opening, random_opening_moves=random_opening_moves, random_move_start=rm)
+            play_games(player_one, player_two, num_games, book_opening=book_opening, random_opening=random_opening, random_opening_moves=random_opening_moves)