feature: remove resolve_agent and use build_agent to build trained model agent

handyrl/evaluation.py

@@ -147,6 +147,12 @@ def build_agent(raw, env=None):
     elif raw.startswith('rulebase'):
         key = raw.split('-')[1] if '-' in raw else None
         return RuleBasedAgent(key)
+
+    if env is not None:
+        model = load_model(model_path, env.net())
+        agent = Agent(model)
+        return agent
+
     return None
 
 
@@ -383,14 +389,7 @@ def eval_main(args, argv):
     num_games = int(argv[1]) if len(argv) >= 2 else 100
     num_process = int(argv[2]) if len(argv) >= 3 else 1
 
-    def resolve_agent(model_path):
-        agent = build_agent(model_path, env)
-        if agent is None:
-            model = load_model(model_path, env.net())
-            agent = Agent(model)
-        return agent
-
-    main_agent = resolve_agent(model_paths[0])
+    main_agent = build_agent(model_paths[0], env)
     critic = None
 
     print('%d process, %d games' % (num_process, num_games))
@@ -399,7 +398,7 @@ def resolve_agent(model_path):
     print('seed = %d' % seed)
 
     opponent = model_paths[1] if len(model_paths) > 1 else 'random'
-    agents = [main_agent] + [resolve_agent(opponent) for _ in range(len(env.players()) - 1)]
+    agents = [main_agent] + [build_agent(opponent, env) for _ in range(len(env.players()) - 1)]
 
     evaluate_mp(env, agents, critic, env_args, {'default': {}}, num_process, num_games, seed)