def multi_agent_match(policy_indices, max_num_episodes=int(1e9), max_num_frames=1e10):
log.debug('Starting eval process with policies %r', policy_indices)
for i, rival in enumerate(RIVALS):
rival.policy_index = policy_indices[i]
curr_dir = os.path.dirname(os.path.abspath(__file__))
evaluation_filename = join(curr_dir,
f'eval_{"vs".join([str(pi) for pi in policy_indices])}.txt')
with open(evaluation_filename, 'w') as fobj:
fobj.write('start\n')
common_config = RIVALS[0].cfg
render_action_repeat = common_config.render_action_repeat if common_config.render_action_repeat is not None else common_config.env_frameskip
if render_action_repeat is None:
log.warning('Not using action repeat!')
render_action_repeat = 1
log.debug('Using action repeat %d during evaluation', render_action_repeat)
common_config.env_frameskip = 1 # for evaluation
common_config.num_envs = 1
common_config.timelimit = 4.0 # for faster evaluation
def make_env_func(env_config):
return create_env(ENV_NAME, cfg=common_config, env_config=env_config)
env = make_env_func(AttrDict({'worker_index': 0, 'vector_index': 0}))
env.seed(0)
is_multiagent = is_multiagent_env(env)
if not is_multiagent:
env = MultiAgentWrapper(env)
else:
assert env.num_agents == len(RIVALS)
device = torch.device('cuda')
for rival in RIVALS:
rival.actor_critic = create_actor_critic(rival.cfg, env.observation_space, env.action_space)
rival.actor_critic.model_to_device(device)
policy_id = rival.policy_index
checkpoints = LearnerWorker.get_checkpoints(
LearnerWorker.checkpoint_dir(rival.cfg, policy_id))
checkpoint_dict = LearnerWorker.load_checkpoint(checkpoints, device)
rival.actor_critic.load_state_dict(checkpoint_dict['model'])
episode_rewards = []
num_frames = 0
last_render_start = time.time()
def max_frames_reached(frames):
return max_num_frames is not None and frames > max_num_frames
wins = [0 for _ in RIVALS]
ties = 0
frag_differences = []
with torch.no_grad():
for _ in range(max_num_episodes):
obs = env.reset()
obs_dict_torch = dict()
done = [False] * len(obs)
for rival in RIVALS:
rival.rnn_states = torch.zeros([1, rival.cfg.hidden_size],
dtype=torch.float32,
device=device)
episode_reward = 0
prev_frame = time.time()
while True:
actions = []
for i, obs_dict in enumerate(obs):
for key, x in obs_dict.items():
obs_dict_torch[key] = torch.from_numpy(x).to(device).float().view(
1, *x.shape)
rival = RIVALS[i]
policy_outputs = rival.actor_critic(obs_dict_torch, rival.rnn_states)
rival.rnn_states = policy_outputs.rnn_states
actions.append(policy_outputs.actions[0].cpu().numpy())
for _ in range(render_action_repeat):
if not NO_RENDER:
target_delay = 1.0 / FPS if FPS > 0 else 0
current_delay = time.time() - last_render_start
time_wait = target_delay - current_delay
if time_wait > 0:
# log.info('Wait time %.3f', time_wait)
time.sleep(time_wait)
last_render_start = time.time()
env.render()
obs, rew, done, infos = env.step(actions)
if all(done):
log.debug('Finished episode!')
frag_diff = infos[0]['PLAYER1_FRAGCOUNT'] - infos[0]['PLAYER2_FRAGCOUNT']
if frag_diff > 0:
wins[0] += 1
elif frag_diff < 0:
wins[1] += 1
else:
ties += 1
frag_differences.append(frag_diff)
avg_frag_diff = np.mean(frag_differences)
report = f'wins: {wins}, ties: {ties}, avg_frag_diff: {avg_frag_diff}'
with open(evaluation_filename, 'a') as fobj:
fobj.write(report + '\n')
# log.info('%d:%d', infos[0]['PLAYER1_FRAGCOUNT'], infos[0]['PLAYER2_FRAGCOUNT'])
episode_reward += np.mean(rew)
num_frames += 1
if num_frames % 100 == 0:
log.debug('%.1f', render_action_repeat / (time.time() - prev_frame))
prev_frame = time.time()
if all(done):
log.info('Episode finished at %d frames', num_frames)
break
if all(done) or max_frames_reached(num_frames):
break
if not NO_RENDER:
env.render()
time.sleep(0.01)
episode_rewards.append(episode_reward)
last_episodes = episode_rewards[-100:]
avg_reward = sum(last_episodes) / len(last_episodes)
log.info(
'Episode reward: %f, avg reward for %d episodes: %f',
episode_reward,
len(last_episodes),
avg_reward,
)
if max_frames_reached(num_frames):
break
env.close()
def enjoy(cfg, max_num_frames=1e9):
cfg = load_from_checkpoint(cfg)
render_action_repeat = cfg.render_action_repeat if cfg.render_action_repeat is not None else cfg.env_frameskip
if render_action_repeat is None:
log.warning('Not using action repeat!')
render_action_repeat = 1
log.debug('Using action repeat %d during evaluation', render_action_repeat)
cfg.env_frameskip = 1 # for evaluation
cfg.num_envs = 1
if cfg.record_to:
tstamp = datetime.datetime.now().strftime('%Y_%m_%d__%H_%M_%S')
cfg.record_to = join(cfg.record_to, f'{cfg.experiment}', tstamp)
if not os.path.isdir(cfg.record_to):
os.makedirs(cfg.record_to)
else:
cfg.record_to = None
def make_env_func(env_config):
return create_env(cfg.env, cfg=cfg, env_config=env_config)
env = make_env_func(AttrDict({'worker_index': 0, 'vector_index': 0}))
# env.seed(0)
is_multiagent = is_multiagent_env(env)
if not is_multiagent:
env = MultiAgentWrapper(env)
if hasattr(env.unwrapped, 'reset_on_init'):
# reset call ruins the demo recording for VizDoom
env.unwrapped.reset_on_init = False
actor_critic = create_actor_critic(cfg, env.observation_space,
env.action_space)
device = torch.device('cpu' if cfg.device == 'cpu' else 'cuda')
actor_critic.model_to_device(device)
policy_id = cfg.policy_index
checkpoints = LearnerWorker.get_checkpoints(
LearnerWorker.checkpoint_dir(cfg, policy_id))
checkpoint_dict = LearnerWorker.load_checkpoint(checkpoints, device)
actor_critic.load_state_dict(checkpoint_dict['model'])
episode_rewards = [deque([], maxlen=100) for _ in range(env.num_agents)]
true_rewards = [deque([], maxlen=100) for _ in range(env.num_agents)]
num_frames = 0
last_render_start = time.time()
def max_frames_reached(frames):
return max_num_frames is not None and frames > max_num_frames
obs = env.reset()
rnn_states = torch.zeros(
[env.num_agents, get_hidden_size(cfg)],
dtype=torch.float32,
device=device)
episode_reward = np.zeros(env.num_agents)
finished_episode = [False] * env.num_agents
with torch.no_grad():
while not max_frames_reached(num_frames):
obs_torch = AttrDict(transform_dict_observations(obs))
for key, x in obs_torch.items():
obs_torch[key] = torch.from_numpy(x).to(device).float()
policy_outputs = actor_critic(obs_torch,
rnn_states,
with_action_distribution=True)
# sample actions from the distribution by default
actions = policy_outputs.actions
action_distribution = policy_outputs.action_distribution
if isinstance(action_distribution, ContinuousActionDistribution):
if not cfg.continuous_actions_sample: # TODO: add similar option for discrete actions
actions = action_distribution.means
actions = actions.cpu().numpy()
rnn_states = policy_outputs.rnn_states
for _ in range(render_action_repeat):
if not cfg.no_render:
target_delay = 1.0 / cfg.fps if cfg.fps > 0 else 0
current_delay = time.time() - last_render_start
time_wait = target_delay - current_delay
if time_wait > 0:
# log.info('Wait time %.3f', time_wait)
time.sleep(time_wait)
last_render_start = time.time()
env.render()
obs, rew, done, infos = env.step(actions)
episode_reward += rew
num_frames += 1
for agent_i, done_flag in enumerate(done):
if done_flag:
finished_episode[agent_i] = True
episode_rewards[agent_i].append(
episode_reward[agent_i])
true_rewards[agent_i].append(infos[agent_i].get(
'true_reward', math.nan))
log.info(
'Episode finished for agent %d at %d frames. Reward: %.3f, true_reward: %.3f',
agent_i, num_frames, episode_reward[agent_i],
true_rewards[agent_i][-1])
rnn_states[agent_i] = torch.zeros(
[get_hidden_size(cfg)],
dtype=torch.float32,
device=device)
episode_reward[agent_i] = 0
# if episode terminated synchronously for all agents, pause a bit before starting a new one
if all(done):
if not cfg.no_render:
env.render()
time.sleep(0.05)
if all(finished_episode):
finished_episode = [False] * env.num_agents
avg_episode_rewards_str, avg_true_reward_str = '', ''
for agent_i in range(env.num_agents):
avg_rew = np.mean(episode_rewards[agent_i])
avg_true_rew = np.mean(true_rewards[agent_i])
if not np.isnan(avg_rew):
if avg_episode_rewards_str:
avg_episode_rewards_str += ', '
avg_episode_rewards_str += f'#{agent_i}: {avg_rew:.3f}'
if not np.isnan(avg_true_rew):
if avg_true_reward_str:
avg_true_reward_str += ', '
avg_true_reward_str += f'#{agent_i}: {avg_true_rew:.3f}'
log.info('Avg episode rewards: %s, true rewards: %s',
avg_episode_rewards_str, avg_true_reward_str)
log.info(
'Avg episode reward: %.3f, avg true_reward: %.3f',
np.mean([
np.mean(episode_rewards[i])
for i in range(env.num_agents)
]),
np.mean([
np.mean(true_rewards[i])
for i in range(env.num_agents)
]))
# VizDoom multiplayer stuff
# for player in [1, 2, 3, 4, 5, 6, 7, 8]:
# key = f'PLAYER{player}_FRAGCOUNT'
# if key in infos[0]:
# log.debug('Score for player %d: %r', player, infos[0][key])
env.close()
return ExperimentStatus.SUCCESS, np.mean(episode_rewards)
def enjoy(cfg, max_num_episodes=1000000, max_num_frames=1e9):
cfg = load_from_checkpoint(cfg)
render_action_repeat = cfg.render_action_repeat if cfg.render_action_repeat is not None else cfg.env_frameskip
if render_action_repeat is None:
log.warning('Not using action repeat!')
render_action_repeat = 1
log.debug('Using action repeat %d during evaluation', render_action_repeat)
cfg.env_frameskip = 1 # for evaluation
cfg.num_envs = 1
if cfg.record_to:
tstamp = datetime.datetime.now().strftime('%Y_%m_%d__%H_%M_%S')
cfg.record_to = join(cfg.record_to, f'{cfg.experiment}', tstamp)
if not os.path.isdir(cfg.record_to):
os.makedirs(cfg.record_to)
else:
cfg.record_to = None
def make_env_func(env_config):
return create_env(cfg.env, cfg=cfg, env_config=env_config)
env = make_env_func(AttrDict({'worker_index': 0, 'vector_index': 0}))
env.seed(0)
is_multiagent = hasattr(env, 'num_agents') and env.num_agents > 1
if not is_multiagent:
env = MultiAgentWrapper(env)
if hasattr(env.unwrapped, 'reset_on_init'):
# reset call ruins the demo recording for VizDoom
env.unwrapped.reset_on_init = False
actor_critic = create_actor_critic(cfg, env.observation_space, env.action_space)
device = torch.device('cpu' if cfg.device == 'cpu' else 'cuda')
actor_critic.model_to_device(device)
policy_id = cfg.policy_index
checkpoints = LearnerWorker.get_checkpoints(LearnerWorker.checkpoint_dir(cfg, policy_id))
checkpoint_dict = LearnerWorker.load_checkpoint(checkpoints, device)
actor_critic.load_state_dict(checkpoint_dict['model'])
episode_rewards = []
true_rewards = deque([], maxlen=100)
num_frames = 0
last_render_start = time.time()
def max_frames_reached(frames):
return max_num_frames is not None and frames > max_num_frames
obs = env.reset()
with torch.no_grad():
for _ in range(max_num_episodes):
done = [False] * len(obs)
rnn_states = torch.zeros([env.num_agents, get_hidden_size(cfg)], dtype=torch.float32, device=device)
episode_reward = 0
while True:
obs_torch = AttrDict(transform_dict_observations(obs))
for key, x in obs_torch.items():
obs_torch[key] = torch.from_numpy(x).to(device).float()
policy_outputs = actor_critic(obs_torch, rnn_states, with_action_distribution=True)
action_distribution = policy_outputs.action_distribution
# sample actions from the distribution by default
actions = policy_outputs.actions
if isinstance(action_distribution, ContinuousActionDistribution):
if not cfg.continuous_actions_sample: # TODO: add similar option for discrete actions
actions = action_distribution.means
actions = actions.cpu().numpy()
rnn_states = policy_outputs.rnn_states
for _ in range(render_action_repeat):
if not cfg.no_render:
target_delay = 1.0 / cfg.fps if cfg.fps > 0 else 0
current_delay = time.time() - last_render_start
time_wait = target_delay - current_delay
if time_wait > 0:
# log.info('Wait time %.3f', time_wait)
time.sleep(time_wait)
last_render_start = time.time()
env.render()
obs, rew, done, infos = env.step(actions)
episode_reward += np.mean(rew)
num_frames += 1
if all(done):
true_rewards.append(infos[0].get('true_reward', math.nan))
log.info('Episode finished at %d frames', num_frames)
if not math.isnan(np.mean(true_rewards)):
log.info('true rew %.3f avg true rew %.3f', true_rewards[-1], np.mean(true_rewards))
# VizDoom multiplayer stuff
# for player in [1, 2, 3, 4, 5, 6, 7, 8]:
# key = f'PLAYER{player}_FRAGCOUNT'
# if key in infos[0]:
# log.debug('Score for player %d: %r', player, infos[0][key])
break
if all(done) or max_frames_reached(num_frames):
break
if not cfg.no_render:
env.render()
time.sleep(0.01)
episode_rewards.append(episode_reward)
last_episodes = episode_rewards[-100:]
avg_reward = sum(last_episodes) / len(last_episodes)
log.info(
'Episode reward: %f, avg reward for %d episodes: %f', episode_reward, len(last_episodes), avg_reward,
)
if max_frames_reached(num_frames):
break
env.close()
return ExperimentStatus.SUCCESS, np.mean(episode_rewards)