import warnings from pettingzoo.utils.conversions import aec_to_parallel_wrapper, parallel_to_aec_wrapper from pettingzoo.utils.wrappers import BaseWrapper from .api_test import missing_attr_warning def parallel_api_test(par_env, num_cycles=10): if not hasattr(par_env, 'possible_agents'): warnings.warn(missing_attr_warning.format(name='possible_agents')) assert not isinstance(par_env.unwrapped, aec_to_parallel_wrapper) assert not isinstance(par_env.unwrapped, parallel_to_aec_wrapper) assert not isinstance(par_env.unwrapped, BaseWrapper) MAX_RESETS = 2 for n_resets in range(MAX_RESETS): obs = par_env.reset() assert isinstance(obs, dict) assert set(obs.keys()) == (set(par_env.agents)) done = {agent: False for agent in par_env.agents} live_agents = set(par_env.agents[:]) has_finished = set() for i in range(num_cycles): actions = {agent: par_env.action_space(agent).sample() for agent in par_env.agents if agent in done and not done[agent]} obs, rew, done, info = par_env.step(actions) for agent in par_env.agents: assert agent not in has_finished, "agent cannot be revived once done" if agent not in live_agents: live_agents.add(agent) assert isinstance(obs, dict) assert isinstance(rew, dict) assert isinstance(done, dict) assert isinstance(info, dict) agents_set = set(live_agents) keys = 'observation reward done info'.split() vals = [obs, rew, done, info] for k, v in zip(keys, vals): if set(v.keys()) == agents_set: continue warnings.warn('Agent was given: {} but was done last turn'.format(k)) if hasattr(par_env, 'possible_agents'): assert set(par_env.agents).issubset(set(par_env.possible_agents)), "possible_agents defined but does not contain all agents" has_finished |= {agent for agent, d in done.items() if d} if not par_env.agents and has_finished != set(par_env.possible_agents): warnings.warn('No agents present but not all possible_agents are done') elif not par_env.agents: warnings.warn('No agents present') for agent in par_env.agents: assert par_env.observation_space(agent) is par_env.observation_space(agent), "observation_space should return the exact same space object (not a copy) for an agent. Consider decorating your observation_space(self, agent) method with @functools.lru_cache(maxsize=None)" assert par_env.action_space(agent) is par_env.action_space(agent), "action_space should return the exact same space object (not a copy) for an agent (ensures that action space seeding works as expected). Consider decorating your action_space(self, agent) method with @functools.lru_cache(maxsize=None)" for agent, d in done.items(): if d: live_agents.remove(agent) assert set(par_env.agents) == live_agents