Add ppo_sp

docs/features.md

@@ -18,6 +18,20 @@
 ## Global
 - lp: 2, max 65535 to 2 bytes
 - oppo_lp: 2, max 65535 to 2 bytes
+- n_my_decks: 1, int
+- n_my_extras:
+- n_my_hands:
+- n_my_graves:
+- n_my_removes:
+- n_my_monsters:
+- n_my_spell_traps:
+- n_op_decks:
+- n_op_extras:
+- n_op_hands:
+- n_op_graves:
+- n_op_removes:
+- n_op_monsters:
+- n_op_spell_traps:
 - turn: 1, int, trunc to 8
 - phase: 1, int, one-hot (10)
 - is_first: 1, int, 0: False, 1: True

scripts/eval.py

@@ -43,6 +43,8 @@ class Args:
     """the maximum number of options"""
     n_history_actions: int = 16
     """the number of history actions to use"""
+    num_embeddings: Optional[int] = None
+    """the number of embeddings of the agent"""
 
     player: int = -1
     """the player to play as, -1 means random, 0 is the first player, 1 is the second player"""
@@ -138,6 +140,8 @@ if __name__ == "__main__":
 
     if args.agent:
         # count lines of code_list
+        embedding_shape = args.num_embeddings
+        if embedding_shape is None:
             with open(args.code_list_file, "r") as f:
                 code_list = f.readlines()
                 embedding_shape = len(code_list)

ygoai/rl/agent.py

@@ -105,6 +105,7 @@ class Encoder(nn.Module):
         self.a_option_embed = nn.Embedding(6, c // divisor // 2)
         self.a_number_embed = nn.Embedding(13, c // divisor // 2)
         self.a_place_embed = nn.Embedding(31, c // divisor // 2)
+        # TODO: maybe same embedding as attribute_embed
         self.a_attrib_embed = nn.Embedding(10, c // divisor // 2)
         self.a_feat_norm = nn.LayerNorm(c, elementwise_affine=affine)