class Actor(object):
def __init__(self, config):
self.config = config
self.envs = []
for _ in range(config['env_num']):
env = gym.make(config['env_name'])
env = wrap_deepmind(env, dim=config['env_dim'], obs_format='NCHW')
self.envs.append(env)
self.vector_env = VectorEnv(self.envs)
self.obs_batch = self.vector_env.reset()
obs_shape = env.observation_space.shape
act_dim = env.action_space.n
model = AtariModel(act_dim)
algorithm = parl.algorithms.IMPALA(
model,
sample_batch_steps=self.config['sample_batch_steps'],
gamma=self.config['gamma'],
vf_loss_coeff=self.config['vf_loss_coeff'],
clip_rho_threshold=self.config['clip_rho_threshold'],
clip_pg_rho_threshold=self.config['clip_pg_rho_threshold'])
self.agent = AtariAgent(algorithm, obs_shape, act_dim)
def sample(self):
env_sample_data = {}
for env_id in range(self.config['env_num']):
env_sample_data[env_id] = defaultdict(list)
for i in range(self.config['sample_batch_steps']):
actions, behaviour_logits = self.agent.sample(
np.stack(self.obs_batch))
next_obs_batch, reward_batch, done_batch, info_batch = \
self.vector_env.step(actions)
for env_id in range(self.config['env_num']):
env_sample_data[env_id]['obs'].append(self.obs_batch[env_id])
env_sample_data[env_id]['actions'].append(actions[env_id])
env_sample_data[env_id]['behaviour_logits'].append(
behaviour_logits[env_id])
env_sample_data[env_id]['rewards'].append(reward_batch[env_id])
env_sample_data[env_id]['dones'].append(done_batch[env_id])
self.obs_batch = next_obs_batch
# Merge data of envs
sample_data = defaultdict(list)
for env_id in range(self.config['env_num']):
for data_name in [
'obs', 'actions', 'behaviour_logits', 'rewards', 'dones'
]:
sample_data[data_name].extend(
env_sample_data[env_id][data_name])
# size of sample_data: env_num * sample_batch_steps
for key in sample_data:
sample_data[key] = np.stack(sample_data[key])
return sample_data
def get_metrics(self):
metrics = defaultdict(list)
for env in self.envs:
monitor = get_wrapper_by_cls(env, MonitorEnv)
if monitor is not None:
for episode_rewards, episode_steps in monitor.next_episode_results(
):
metrics['episode_rewards'].append(episode_rewards)
metrics['episode_steps'].append(episode_steps)
return metrics
def set_weights(self, weights):
self.agent.set_weights(weights)
class Actor(object):
def __init__(self, config):
self.config = config
self.envs = []
for _ in range(config['env_num']):
env = gym.make(config['env_name'])
env.seed(ENV_SEED)
env = MonitorEnv(env)
env = ClipRewardEnv(env)
env = StateStack(env, k=4)
self.envs.append(env)
# env = gym.make(config['env_name'])
# obs_shape = env.observation_space.shape
self.vector_env = VectorEnv(self.envs)
self.obs_batch = self.vector_env.reset()
obs_dim = self.envs[0].observation_space.shape
act_dim = self.envs[0].action_space.shape[0]
max_action = float(self.envs[0].action_space.high[0])
# obs_shape = env.observation_space.shape
# act_dim = env.action_space.n
model = MujocoModel(act_dim)
algorithm = DVtrace(
model,
max_action,
sample_batch_steps=self.config['sample_batch_steps'],
gamma=self.config['gamma'],
vf_loss_coeff=self.config['vf_loss_coeff'],
clip_rho_threshold=self.config['clip_rho_threshold'],
clip_pg_rho_threshold=self.config['clip_pg_rho_threshold'])
self.agent = AtariAgent(algorithm, obs_dim, act_dim)
def sample(self):
env_sample_data = {}
for env_id in range(self.config['env_num']):
env_sample_data[env_id] = defaultdict(list)
for i in range(self.config['sample_batch_steps']):
actions, mean, std = self.agent.sample(self.obs_batch)
next_obs_batch, reward_batch, done_batch, info_batch = self.vector_env.step(actions)
for env_id in range(self.config['env_num']):
env_sample_data[env_id]['obs'].append(self.obs_batch[env_id])
env_sample_data[env_id]['actions'].append(actions[env_id])
env_sample_data[env_id]['mean'].append(mean[env_id])
env_sample_data[env_id]['std'].append(std[env_id])
env_sample_data[env_id]['rewards'].append(reward_batch[env_id])
env_sample_data[env_id]['dones'].append(done_batch[env_id])
self.obs_batch = next_obs_batch
# Merge data of envs
sample_data = defaultdict(list)
for env_id in range(self.config['env_num']):
for data_name in [
'obs', 'actions', 'mean', 'std', 'rewards', 'dones'
]:
sample_data[data_name].extend(
env_sample_data[env_id][data_name])
# size of sample_data: env_num * sample_batch_steps
for key in sample_data:
sample_data[key] = np.stack(sample_data[key])
return sample_data
def get_metrics(self):
metrics = defaultdict(list)
for env in self.envs:
monitor = get_wrapper_by_cls(env, MonitorEnv)
if monitor is not None:
for episode_rewards, episode_steps in monitor.next_episode_results(
):
metrics['episode_rewards'].append(episode_rewards)
metrics['episode_steps'].append(episode_steps)
return metrics
def set_weights(self, weights):
self.agent.set_weights(weights)
class Learner(object):
def __init__(self, config):
self.config = config
self.sample_data_queue = queue.Queue()
self.batch_buffer = defaultdict(list)
#=========== Create Agent ==========
env = gym.make(config['env_name'])
env = wrap_deepmind(env, dim=config['env_dim'], obs_format='NCHW')
obs_shape = env.observation_space.shape
act_dim = env.action_space.n
self.config['obs_shape'] = obs_shape
self.config['act_dim'] = act_dim
model = AtariModel(act_dim)
algorithm = parl.algorithms.A3C(model,
vf_loss_coeff=config['vf_loss_coeff'])
self.agent = AtariAgent(
algorithm,
obs_shape=self.config['obs_shape'],
predict_thread_num=self.config['predict_thread_num'],
learn_data_provider=self.learn_data_provider)
if machine_info.is_gpu_available():
assert get_gpu_count() == 1, 'Only support training in single GPU,\
Please set environment variable: `export CUDA_VISIBLE_DEVICES=[GPU_ID_YOU_WANT_TO_USE]` .'
else:
cpu_num = os.environ.get('CPU_NUM')
assert cpu_num is not None and cpu_num == '1', 'Only support training in single CPU,\
Please set environment variable: `export CPU_NUM=1`.'
#========== Learner ==========
self.lr, self.entropy_coeff = None, None
self.lr_scheduler = PiecewiseScheduler(config['lr_scheduler'])
self.entropy_coeff_scheduler = PiecewiseScheduler(
config['entropy_coeff_scheduler'])
self.total_loss_stat = WindowStat(100)
self.pi_loss_stat = WindowStat(100)
self.vf_loss_stat = WindowStat(100)
self.entropy_stat = WindowStat(100)
self.learn_time_stat = TimeStat(100)
self.start_time = None
# learn thread
self.learn_thread = threading.Thread(target=self.run_learn)
self.learn_thread.setDaemon(True)
self.learn_thread.start()
self.predict_input_queue = queue.Queue()
# predict thread
self.predict_threads = []
for i in six.moves.range(self.config['predict_thread_num']):
predict_thread = threading.Thread(target=self.run_predict,
args=(i, ))
predict_thread.setDaemon(True)
predict_thread.start()
self.predict_threads.append(predict_thread)
#========== Remote Simulator ===========
self.remote_count = 0
self.remote_metrics_queue = queue.Queue()
self.sample_total_steps = 0
self.create_actors()
def learn_data_provider(self):
""" Data generator for fluid.layers.py_reader
"""
B = self.config['train_batch_size']
while True:
sample_data = self.sample_data_queue.get()
self.sample_total_steps += len(sample_data['obs'])
for key in sample_data:
self.batch_buffer[key].extend(sample_data[key])
if len(self.batch_buffer['obs']) >= B:
batch = {}
for key in self.batch_buffer:
batch[key] = np.array(self.batch_buffer[key][:B])
obs_np = batch['obs'].astype('float32')
actions_np = batch['actions'].astype('int64')
advantages_np = batch['advantages'].astype('float32')
target_values_np = batch['target_values'].astype('float32')
self.lr = self.lr_scheduler.step()
self.lr = np.array(self.lr, dtype='float32')
self.entropy_coeff = self.entropy_coeff_scheduler.step()
self.entropy_coeff = np.array(self.entropy_coeff,
dtype='float32')
yield [
obs_np, actions_np, advantages_np, target_values_np,
self.lr, self.entropy_coeff
]
for key in self.batch_buffer:
self.batch_buffer[key] = self.batch_buffer[key][B:]
def run_predict(self, thread_id):
""" predict thread
"""
batch_ident = []
batch_obs = []
while True:
ident, obs = self.predict_input_queue.get()
batch_ident.append(ident)
batch_obs.append(obs)
while len(batch_obs) < self.config['max_predict_batch_size']:
try:
ident, obs = self.predict_input_queue.get_nowait()
batch_ident.append(ident)
batch_obs.append(obs)
except queue.Empty:
break
if batch_obs:
batch_obs = np.array(batch_obs)
actions, values = self.agent.sample(batch_obs, thread_id)
for i, ident in enumerate(batch_ident):
self.predict_output_queues[ident].put(
(actions[i], values[i]))
batch_ident = []
batch_obs = []
def run_learn(self):
""" Learn loop
"""
while True:
with self.learn_time_stat:
total_loss, pi_loss, vf_loss, entropy = self.agent.learn()
self.total_loss_stat.add(total_loss)
self.pi_loss_stat.add(pi_loss)
self.vf_loss_stat.add(vf_loss)
self.entropy_stat.add(entropy)
def create_actors(self):
""" Connect to the cluster and start sampling of the remote actor.
"""
parl.connect(self.config['master_address'])
logger.info('Waiting for {} remote actors to connect.'.format(
self.config['actor_num']))
ident = 0
self.predict_output_queues = []
for i in six.moves.range(self.config['actor_num']):
self.remote_count += 1
logger.info('Remote simulator count: {}'.format(self.remote_count))
if self.start_time is None:
self.start_time = time.time()
q = queue.Queue()
self.predict_output_queues.append(q)
remote_thread = threading.Thread(target=self.run_remote_sample,
args=(ident, ))
remote_thread.setDaemon(True)
remote_thread.start()
ident += 1
def run_remote_sample(self, ident):
""" Interacts with remote simulator.
"""
remote_actor = Actor(self.config)
mem = defaultdict(list)
obs = remote_actor.reset()
while True:
self.predict_input_queue.put((ident, obs))
action, value = self.predict_output_queues[ident].get()
next_obs, reward, done = remote_actor.step(action)
mem['obs'].append(obs)
mem['actions'].append(action)
mem['rewards'].append(reward)
mem['values'].append(value)
if done:
next_value = 0
advantages = calc_gae(mem['rewards'], mem['values'],
next_value, self.config['gamma'],
self.config['lambda'])
target_values = advantages + mem['values']
self.sample_data_queue.put({
'obs': mem['obs'],
'actions': mem['actions'],
'advantages': advantages,
'target_values': target_values
})
mem = defaultdict(list)
next_obs = remote_actor.reset()
elif len(mem['obs']) == self.config['t_max'] + 1:
next_value = mem['values'][-1]
advantages = calc_gae(mem['rewards'][:-1], mem['values'][:-1],
next_value, self.config['gamma'],
self.config['lambda'])
target_values = advantages + mem['values'][:-1]
self.sample_data_queue.put({
'obs': mem['obs'][:-1],
'actions': mem['actions'][:-1],
'advantages': advantages,
'target_values': target_values
})
for key in mem:
mem[key] = [mem[key][-1]]
obs = next_obs
if done:
metrics = remote_actor.get_metrics()
if metrics:
self.remote_metrics_queue.put(metrics)
def log_metrics(self):
""" Log metrics of learner and simulators
"""
if self.start_time is None:
return
metrics = []
while True:
try:
metric = self.remote_metrics_queue.get_nowait()
metrics.append(metric)
except queue.Empty:
break
episode_rewards, episode_steps = [], []
for x in metrics:
episode_rewards.extend(x['episode_rewards'])
episode_steps.extend(x['episode_steps'])
max_episode_rewards, mean_episode_rewards, min_episode_rewards, \
max_episode_steps, mean_episode_steps, min_episode_steps =\
None, None, None, None, None, None
if episode_rewards:
mean_episode_rewards = np.mean(np.array(episode_rewards).flatten())
max_episode_rewards = np.max(np.array(episode_rewards).flatten())
min_episode_rewards = np.min(np.array(episode_rewards).flatten())
mean_episode_steps = np.mean(np.array(episode_steps).flatten())
max_episode_steps = np.max(np.array(episode_steps).flatten())
min_episode_steps = np.min(np.array(episode_steps).flatten())
metric = {
'Sample steps': self.sample_total_steps,
'max_episode_rewards': max_episode_rewards,
'mean_episode_rewards': mean_episode_rewards,
'min_episode_rewards': min_episode_rewards,
'max_episode_steps': max_episode_steps,
'mean_episode_steps': mean_episode_steps,
'min_episode_steps': min_episode_steps,
'total_loss': self.total_loss_stat.mean,
'pi_loss': self.pi_loss_stat.mean,
'vf_loss': self.vf_loss_stat.mean,
'entropy': self.entropy_stat.mean,
'learn_time_s': self.learn_time_stat.mean,
'elapsed_time_s': int(time.time() - self.start_time),
'lr': self.lr,
'entropy_coeff': self.entropy_coeff,
}
for key, value in metric.items():
if value is not None:
summary.add_scalar(key, value, self.sample_total_steps)
logger.info(metric)
class Actor(object):
def __init__(self, config):
self.config = config
self.envs = []
for _ in range(config['env_num']):
env = gym.make(config['env_name'])
env = wrap_deepmind(env, dim=config['env_dim'], obs_format='NCHW')
self.envs.append(env)
self.vector_env = VectorEnv(self.envs)
self.obs_batch = self.vector_env.reset()
obs_shape = env.observation_space.shape
act_dim = env.action_space.n
self.config['obs_shape'] = obs_shape
self.config['act_dim'] = act_dim
model = AtariModel(act_dim)
algorithm = parl.algorithms.A3C(model,
vf_loss_coeff=config['vf_loss_coeff'])
self.agent = AtariAgent(algorithm, config)
def sample(self):
sample_data = defaultdict(list)
env_sample_data = {}
for env_id in range(self.config['env_num']):
env_sample_data[env_id] = defaultdict(list)
for i in range(self.config['sample_batch_steps']):
actions_batch, values_batch = self.agent.sample(
np.stack(self.obs_batch))
next_obs_batch, reward_batch, done_batch, info_batch = \
self.vector_env.step(actions_batch)
for env_id in range(self.config['env_num']):
env_sample_data[env_id]['obs'].append(self.obs_batch[env_id])
env_sample_data[env_id]['actions'].append(
actions_batch[env_id])
env_sample_data[env_id]['rewards'].append(reward_batch[env_id])
env_sample_data[env_id]['dones'].append(done_batch[env_id])
env_sample_data[env_id]['values'].append(values_batch[env_id])
# Calculate advantages when the episode is done or reach max sample steps.
if done_batch[
env_id] or i == self.config['sample_batch_steps'] - 1:
next_value = 0
if not done_batch[env_id]:
next_obs = np.expand_dims(next_obs_batch[env_id], 0)
next_value = self.agent.value(next_obs)
values = env_sample_data[env_id]['values']
rewards = env_sample_data[env_id]['rewards']
advantages = calc_gae(rewards, values, next_value,
self.config['gamma'],
self.config['lambda'])
target_values = advantages + values
sample_data['obs'].extend(env_sample_data[env_id]['obs'])
sample_data['actions'].extend(
env_sample_data[env_id]['actions'])
sample_data['advantages'].extend(advantages)
sample_data['target_values'].extend(target_values)
env_sample_data[env_id] = defaultdict(list)
self.obs_batch = next_obs_batch
# size of sample_data: env_num * sample_batch_steps
for key in sample_data:
sample_data[key] = np.stack(sample_data[key])
return sample_data
def get_metrics(self):
metrics = defaultdict(list)
for env in self.envs:
monitor = get_wrapper_by_cls(env, MonitorEnv)
if monitor is not None:
for episode_rewards, episode_steps in monitor.next_episode_results(
):
metrics['episode_rewards'].append(episode_rewards)
metrics['episode_steps'].append(episode_steps)
return metrics
def set_weights(self, params):
self.agent.set_weights(params)