🐐

backend/nn.py

@@ -132,8 +132,6 @@ class NeuralNet:
     return net
 
 
-# ==================== Feature extraction ====================
-
 def extract_plan_features(plans: list, player, opponent) -> np.ndarray:
   """
   Returns (n_plans, N_FEATURES) float32 array.
@@ -143,7 +141,7 @@ def extract_plan_features(plans: list, player, opponent) -> np.ndarray:
 
   n = len(plans)
 
-  # ---- state (same for every plan) ----
+  # state (same for every plan)
   state = np.array([
     player.life    / STARTING_LIFE,
     opponent.life    / STARTING_LIFE,
@@ -155,7 +153,7 @@ def extract_plan_features(plans: list, player, opponent) -> np.ndarray:
     len(opponent.deck) / _MAX_DECK,
   ], dtype=np.float32)
 
-  # ---- current boards (same for every plan) ----
+  # current boards (same for every plan)
   my_board  = np.zeros(BOARD_SIZE * 3, dtype=np.float32)
   opp_board = np.zeros(BOARD_SIZE * 3, dtype=np.float32)
   for slot in range(BOARD_SIZE):
@@ -170,7 +168,7 @@ def extract_plan_features(plans: list, player, opponent) -> np.ndarray:
       opp_board[slot * 3 + 1] = c.defense / _MAX_DEF
       opp_board[slot * 3 + 2] = 1.0
 
-  # ---- per-plan features ----
+  # per-plan features
   plan_part = np.zeros((n, 3 + BOARD_SIZE * 3), dtype=np.float32)
   for idx, plan in enumerate(plans):
     # simulate board result
@@ -192,7 +190,7 @@ def extract_plan_features(plans: list, player, opponent) -> np.ndarray:
         plan_part[idx, 3 + slot * 3 + 1] = c.defense / _MAX_DEF
         plan_part[idx, 3 + slot * 3 + 2] = 1.0
 
-  # ---- opponent deck type one-hot (same for every plan) ----
+  # opponent deck type one-hot (same for every plan)
   opp_deck_oh = np.zeros(len(_DECK_TYPES), dtype=np.float32)
   opp_deck_oh[_DECK_TYPE_IDX.get(opponent.deck_type, 0)] = 1.0
 
@@ -204,8 +202,6 @@ def extract_plan_features(plans: list, player, opponent) -> np.ndarray:
   return np.concatenate([state_t, my_board_t, opp_board_t, plan_part, opp_deck_t], axis=1)
 
 
-# ==================== Neural player ====================
-
 class NeuralPlayer:
   """
   Wraps a NeuralNet for use in game simulation.