def init_demo_buffer(self, demoDataFile, update_stats=True): #function that initializes the demo buffer
demoData = np.load(demoDataFile) #load the demonstration data from data file
info_keys = [key.replace('info_', '') for key in self.input_dims.keys() if key.startswith('info_')]
info_values = [np.empty((self.T - 1, 1, self.input_dims['info_' + key]), np.float32) for key in info_keys]
demo_data_obs = demoData['obs']
demo_data_acs = demoData['acs']
demo_data_info = demoData['info']
for epsd in range(self.num_demo): # we initialize the whole demo buffer at the start of the training
obs, acts, goals, achieved_goals = [], [] ,[] ,[]
i = 0
for transition in range(self.T - 1):
obs.append([demo_data_obs[epsd][transition].get('observation')])
acts.append([demo_data_acs[epsd][transition]])
goals.append([demo_data_obs[epsd][transition].get('desired_goal')])
achieved_goals.append([demo_data_obs[epsd][transition].get('achieved_goal')])
for idx, key in enumerate(info_keys):
info_values[idx][transition, i] = demo_data_info[epsd][transition][key]
obs.append([demo_data_obs[epsd][self.T - 1].get('observation')])
achieved_goals.append([demo_data_obs[epsd][self.T - 1].get('achieved_goal')])
episode = dict(o=obs,
u=acts,
g=goals,
ag=achieved_goals)
for key, value in zip(info_keys, info_values):
episode['info_{}'.format(key)] = value
episode = convert_episode_to_batch_major(episode)
global DEMO_BUFFER
DEMO_BUFFER.store_episode(episode) # create the observation dict and append them into the demonstration buffer
logger.debug("Demo buffer size currently ", DEMO_BUFFER.get_current_size()) #print out the demonstration buffer size
if update_stats:
# add transitions to normalizer to normalize the demo data as well
episode['o_2'] = episode['o'][:, 1:, :]
episode['ag_2'] = episode['ag'][:, 1:, :]
num_normalizing_transitions = transitions_in_episode_batch(episode)
transitions = self.sample_transitions(episode, num_normalizing_transitions)
o, g, ag = transitions['o'], transitions['g'], transitions['ag']
transitions['o'], transitions['g'] = self._preprocess_og(o, ag, g)
# No need to preprocess the o_2 and g_2 since this is only used for stats
self.o_stats.update(transitions['o'])
self.g_stats.update(transitions['g'])
self.o_stats.recompute_stats()
self.g_stats.recompute_stats()
episode.clear()
logger.info("Demo buffer size: ", DEMO_BUFFER.get_current_size()) #print out the demonstration buffer size
def display_var_info(vars):
from baselines import logger
count_params = 0
for v in vars:
name = v.name
if "/Adam" in name or "beta1_power" in name or "beta2_power" in name: continue
count_params += np.prod(v.shape.as_list())
if "/b:" in name: continue # Wx+b, bias is not interesting to look at => count params, but not print
logger.info(" %s%s%s" % (name, " "*(55-len(name)), str(v.shape)))
logger.info("Total model parameters: %0.1f million" % (count_params*1e-6))
def get_target_updates(vars, target_vars, tau):
logger.info('setting up target updates ...')
soft_updates = []
init_updates = []
assert len(vars) == len(target_vars)
for var, target_var in zip(vars, target_vars):
logger.info(' {} <- {}'.format(target_var.name, var.name))
init_updates.append(tf.assign(target_var, var))
soft_updates.append(tf.assign(target_var, (1. - tau) * target_var + tau * var))
assert len(init_updates) == len(vars)
assert len(soft_updates) == len(vars)
return tf.group(*init_updates), tf.group(*soft_updates)
def get_perturbed_actor_updates(actor, perturbed_actor, param_noise_stddev):
assert len(actor.vars) == len(perturbed_actor.vars)
assert len(actor.perturbable_vars) == len(perturbed_actor.perturbable_vars)
updates = []
for var, perturbed_var in zip(actor.vars, perturbed_actor.vars):
if var in actor.perturbable_vars:
logger.info(' {} <- {} + noise'.format(perturbed_var.name, var.name))
updates.append(tf.assign(perturbed_var, var + tf.random_normal(tf.shape(var), mean=0., stddev=param_noise_stddev)))
else:
logger.info(' {} <- {}'.format(perturbed_var.name, var.name))
updates.append(tf.assign(perturbed_var, var))
assert len(updates) == len(actor.vars)
return tf.group(*updates)
def step_wait(self):
obs, rews, dones, infos = self.venv.step_wait()
self.step_id += 1
if self.recording:
self.video_recorder.capture_frame()
self.recorded_frames += 1
if self.recorded_frames > self.video_length:
logger.info("Saving video to ", self.video_recorder.path)
self.close_video_recorder()
elif self._video_enabled():
self.start_video_recorder()
return obs, rews, dones, infos
def setup_param_noise(self, normalized_obs0):
assert self.param_noise is not None
# Configure perturbed actor.
param_noise_actor = copy(self.actor)
param_noise_actor.name = 'param_noise_actor'
self.perturbed_actor_tf = param_noise_actor(normalized_obs0)
logger.info('setting up param noise')
self.perturb_policy_ops = get_perturbed_actor_updates(self.actor, param_noise_actor, self.param_noise_stddev)
# Configure separate copy for stddev adoption.
adaptive_param_noise_actor = copy(self.actor)
adaptive_param_noise_actor.name = 'adaptive_param_noise_actor'
adaptive_actor_tf = adaptive_param_noise_actor(normalized_obs0)
self.perturb_adaptive_policy_ops = get_perturbed_actor_updates(self.actor, adaptive_param_noise_actor, self.param_noise_stddev)
self.adaptive_policy_distance = tf.sqrt(tf.reduce_mean(tf.square(self.actor_tf - adaptive_actor_tf)))
def setup_critic_optimizer(self):
logger.info('setting up critic optimizer')
normalized_critic_target_tf = tf.clip_by_value(normalize(self.critic_target, self.ret_rms), self.return_range[0], self.return_range[1])
self.critic_loss = tf.reduce_mean(tf.square(self.normalized_critic_tf - normalized_critic_target_tf))
if self.critic_l2_reg > 0.:
critic_reg_vars = [var for var in self.critic.trainable_vars if 'kernel' in var.name and 'output' not in var.name]
for var in critic_reg_vars:
logger.info(' regularizing: {}'.format(var.name))
logger.info(' applying l2 regularization with {}'.format(self.critic_l2_reg))
critic_reg = tc.layers.apply_regularization(
tc.layers.l2_regularizer(self.critic_l2_reg),
weights_list=critic_reg_vars
)
self.critic_loss += critic_reg
critic_shapes = [var.get_shape().as_list() for var in self.critic.trainable_vars]
critic_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in critic_shapes])
logger.info(' critic shapes: {}'.format(critic_shapes))
logger.info(' critic params: {}'.format(critic_nb_params))
self.critic_grads = U.flatgrad(self.critic_loss, self.critic.trainable_vars, clip_norm=self.clip_norm)
self.critic_optimizer = MpiAdam(var_list=self.critic.trainable_vars,
beta1=0.9, beta2=0.999, epsilon=1e-08)
def train(policy, rollout_worker, evaluator,
n_epochs, n_test_rollouts, n_cycles, n_batches, policy_save_interval,
save_policies, **kwargs):
rank = MPI.COMM_WORLD.Get_rank()
latest_policy_path = os.path.join(logger.get_dir(), 'policy_latest.pkl')
best_policy_path = os.path.join(logger.get_dir(), 'policy_best.pkl')
periodic_policy_path = os.path.join(logger.get_dir(), 'policy_{}.pkl')
logger.info("Training...")
best_success_rate = -1
for epoch in range(n_epochs):
# train
rollout_worker.clear_history()
for _ in range(n_cycles):
episode = rollout_worker.generate_rollouts()
policy.store_episode(episode)
for _ in range(n_batches):
policy.train()
policy.update_target_net()
# test
evaluator.clear_history()
for _ in range(n_test_rollouts):
evaluator.generate_rollouts()
# record logs
logger.record_tabular('epoch', epoch)
for key, val in evaluator.logs('test'):
logger.record_tabular(key, mpi_average(val))
for key, val in rollout_worker.logs('train'):
logger.record_tabular(key, mpi_average(val))
for key, val in policy.logs():
logger.record_tabular(key, mpi_average(val))
if rank == 0:
logger.dump_tabular()
# save the policy if it's better than the previous ones
success_rate = mpi_average(evaluator.current_success_rate())
if rank == 0 and success_rate >= best_success_rate and save_policies:
best_success_rate = success_rate
logger.info('New best success rate: {}. Saving policy to {} ...'.format(best_success_rate, best_policy_path))
evaluator.save_policy(best_policy_path)
evaluator.save_policy(latest_policy_path)
if rank == 0 and policy_save_interval > 0 and epoch % policy_save_interval == 0 and save_policies:
policy_path = periodic_policy_path.format(epoch)
logger.info('Saving periodic policy to {} ...'.format(policy_path))
evaluator.save_policy(policy_path)
# make sure that different threads have different seeds
local_uniform = np.random.uniform(size=(1,))
root_uniform = local_uniform.copy()
MPI.COMM_WORLD.Bcast(root_uniform, root=0)
if rank != 0:
assert local_uniform[0] != root_uniform[0]
def setup_actor_optimizer(self):
logger.info('setting up actor optimizer')
self.actor_loss = -tf.reduce_mean(self.critic_with_actor_tf)
actor_shapes = [var.get_shape().as_list() for var in self.actor.trainable_vars]
actor_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in actor_shapes])
logger.info(' actor shapes: {}'.format(actor_shapes))
logger.info(' actor params: {}'.format(actor_nb_params))
self.actor_grads = U.flatgrad(self.actor_loss, self.actor.trainable_vars, clip_norm=self.clip_norm)
self.actor_optimizer = MpiAdam(var_list=self.actor.trainable_vars,
beta1=0.9, beta2=0.999, epsilon=1e-08)
def _create_network(self, reuse=False):
logger.info("Creating a DDPG agent with action space %d x %s..." % (self.dimu, self.max_u))
self.sess = tf.get_default_session()
if self.sess is None:
self.sess = tf.InteractiveSession()
# running averages
with tf.variable_scope('o_stats') as vs:
if reuse:
vs.reuse_variables()
self.o_stats = Normalizer(self.dimo, self.norm_eps, self.norm_clip, sess=self.sess)
with tf.variable_scope('g_stats') as vs:
if reuse:
vs.reuse_variables()
self.g_stats = Normalizer(self.dimg, self.norm_eps, self.norm_clip, sess=self.sess)
# mini-batch sampling.
batch = self.staging_tf.get()
batch_tf = OrderedDict([(key, batch[i])
for i, key in enumerate(self.stage_shapes.keys())])
batch_tf['r'] = tf.reshape(batch_tf['r'], [-1, 1])
# networks
with tf.variable_scope('main') as vs:
if reuse:
vs.reuse_variables()
self.main = self.create_actor_critic(batch_tf, net_type='main', **self.__dict__)
vs.reuse_variables()
with tf.variable_scope('target') as vs:
if reuse:
vs.reuse_variables()
target_batch_tf = batch_tf.copy()
target_batch_tf['o'] = batch_tf['o_2']
target_batch_tf['g'] = batch_tf['g_2']
self.target = self.create_actor_critic(
target_batch_tf, net_type='target', **self.__dict__)
vs.reuse_variables()
assert len(self._vars("main")) == len(self._vars("target"))
# loss functions
target_Q_pi_tf = self.target.Q_pi_tf
clip_range = (-self.clip_return, 0. if self.clip_pos_returns else np.inf)
target_tf = tf.clip_by_value(batch_tf['r'] + self.gamma * target_Q_pi_tf, *clip_range)
self.Q_loss_tf = tf.reduce_mean(tf.square(tf.stop_gradient(target_tf) - self.main.Q_tf))
self.pi_loss_tf = -tf.reduce_mean(self.main.Q_pi_tf)
self.pi_loss_tf += self.action_l2 * tf.reduce_mean(tf.square(self.main.pi_tf / self.max_u))
Q_grads_tf = tf.gradients(self.Q_loss_tf, self._vars('main/Q'))
pi_grads_tf = tf.gradients(self.pi_loss_tf, self._vars('main/pi'))
assert len(self._vars('main/Q')) == len(Q_grads_tf)
assert len(self._vars('main/pi')) == len(pi_grads_tf)
self.Q_grads_vars_tf = zip(Q_grads_tf, self._vars('main/Q'))
self.pi_grads_vars_tf = zip(pi_grads_tf, self._vars('main/pi'))
self.Q_grad_tf = flatten_grads(grads=Q_grads_tf, var_list=self._vars('main/Q'))
self.pi_grad_tf = flatten_grads(grads=pi_grads_tf, var_list=self._vars('main/pi'))
# optimizers
self.Q_adam = MpiAdam(self._vars('main/Q'), scale_grad_by_procs=False)
self.pi_adam = MpiAdam(self._vars('main/pi'), scale_grad_by_procs=False)
# polyak averaging
self.main_vars = self._vars('main/Q') + self._vars('main/pi')
self.target_vars = self._vars('target/Q') + self._vars('target/pi')
self.stats_vars = self._global_vars('o_stats') + self._global_vars('g_stats')
self.init_target_net_op = list(
map(lambda v: v[0].assign(v[1]), zip(self.target_vars, self.main_vars)))
self.update_target_net_op = list(
map(lambda v: v[0].assign(self.polyak * v[0] + (1. - self.polyak) * v[1]), zip(self.target_vars, self.main_vars)))
# initialize all variables
tf.variables_initializer(self._global_vars('')).run()
self._sync_optimizers()
self._init_target_net()
def run_baselines(env, seed, log_dir):
"""
Create baselines model and training.
Replace the ddpg and its training with the algorithm you want to run.
:param env: Environment of the task.
:param seed: Random seed for the trail.
:param log_dir: Log dir path.
:return
"""
rank = MPI.COMM_WORLD.Get_rank()
seed = seed + 1000000 * rank
set_global_seeds(seed)
env.seed(seed)
# Set up logger for baselines
configure(dir=log_dir)
baselines_logger.info('rank {}: seed={}, logdir={}'.format(
rank, seed, baselines_logger.get_dir()))
# Set up params for baselines ddpg
nb_actions = env.action_space.shape[-1]
layer_norm = False
action_noise = OrnsteinUhlenbeckActionNoise(
mu=np.zeros(nb_actions),
sigma=float(params["sigma"]) * np.ones(nb_actions))
memory = Memory(
limit=params["replay_buffer_size"],
action_shape=env.action_space.shape,
observation_shape=env.observation_space.shape)
critic = Critic(layer_norm=layer_norm)
actor = Actor(nb_actions, layer_norm=layer_norm)
training.train(
env=env,
eval_env=None,
param_noise=None,
action_noise=action_noise,
actor=actor,
critic=critic,
memory=memory,
nb_epochs=params["n_epochs"],
nb_epoch_cycles=params["n_epoch_cycles"],
render_eval=False,
reward_scale=1.,
render=False,
normalize_returns=False,
normalize_observations=False,
critic_l2_reg=0,
actor_lr=params["policy_lr"],
critic_lr=params["qf_lr"],
popart=False,
gamma=params["discount"],
clip_norm=None,
nb_train_steps=params["n_train_steps"],
nb_rollout_steps=params["n_rollout_steps"],
nb_eval_steps=100,
batch_size=64)
return osp.join(log_dir, "progress.csv")
def _create_network(self, reuse=False):
logger.info("Creating a DDPG agent with action space %d x %s..." % (self.dimu, self.max_u))
self.sess = tf.get_default_session()
if self.sess is None:
self.sess = tf.InteractiveSession()
# running averages
with tf.variable_scope('o_stats') as vs:
if reuse:
vs.reuse_variables()
self.o_stats = Normalizer(self.dimo, self.norm_eps, self.norm_clip, sess=self.sess)
with tf.variable_scope('g_stats') as vs:
if reuse:
vs.reuse_variables()
self.g_stats = Normalizer(self.dimg, self.norm_eps, self.norm_clip, sess=self.sess)
# mini-batch sampling.
batch = self.staging_tf.get()
batch_tf = OrderedDict([(key, batch[i])
for i, key in enumerate(self.stage_shapes.keys())])
batch_tf['r'] = tf.reshape(batch_tf['r'], [-1, 1])
# networks
with tf.variable_scope('main') as vs:
if reuse:
vs.reuse_variables()
self.main = self.create_actor_critic(batch_tf, net_type='main', **self.__dict__)
vs.reuse_variables()
with tf.variable_scope('target') as vs:
if reuse:
vs.reuse_variables()
target_batch_tf = batch_tf.copy()
target_batch_tf['o'] = batch_tf['o_2']
target_batch_tf['g'] = batch_tf['g_2']
self.target = self.create_actor_critic(
target_batch_tf, net_type='target', **self.__dict__)
vs.reuse_variables()
assert len(self._vars("main")) == len(self._vars("target"))
# loss functions
target_Q_pi_tf = self.target.Q_pi_tf
clip_range = (-self.clip_return, 0. if self.clip_pos_returns else np.inf)
target_tf = tf.clip_by_value(batch_tf['r'] + self.gamma * target_Q_pi_tf, *clip_range)
self.Q_loss_tf = tf.reduce_mean(tf.square(tf.stop_gradient(target_tf) - self.main.Q_tf))
self.pi_loss_tf = -tf.reduce_mean(self.main.Q_pi_tf)
self.pi_loss_tf += self.action_l2 * tf.reduce_mean(tf.square(self.main.pi_tf / self.max_u))
Q_grads_tf = tf.gradients(self.Q_loss_tf, self._vars('main/Q'))
pi_grads_tf = tf.gradients(self.pi_loss_tf, self._vars('main/pi'))
assert len(self._vars('main/Q')) == len(Q_grads_tf)
assert len(self._vars('main/pi')) == len(pi_grads_tf)
self.Q_grads_vars_tf = zip(Q_grads_tf, self._vars('main/Q'))
self.pi_grads_vars_tf = zip(pi_grads_tf, self._vars('main/pi'))
self.Q_grad_tf = flatten_grads(grads=Q_grads_tf, var_list=self._vars('main/Q'))
self.pi_grad_tf = flatten_grads(grads=pi_grads_tf, var_list=self._vars('main/pi'))
# optimizers
self.Q_adam = MpiAdam(self._vars('main/Q'), scale_grad_by_procs=False)
self.pi_adam = MpiAdam(self._vars('main/pi'), scale_grad_by_procs=False)
# polyak averaging
self.main_vars = self._vars('main/Q') + self._vars('main/pi')
self.target_vars = self._vars('target/Q') + self._vars('target/pi')
self.stats_vars = self._global_vars('o_stats') + self._global_vars('g_stats')
self.init_target_net_op = list(
map(lambda v: v[0].assign(v[1]), zip(self.target_vars, self.main_vars)))
self.update_target_net_op = list(
map(lambda v: v[0].assign(self.polyak * v[0] + (1. - self.polyak) * v[1]), zip(self.target_vars, self.main_vars)))
# initialize all variables
tf.variables_initializer(self._global_vars('')).run()
self._sync_optimizers()
self._init_target_net()
def learn(env,
seed=None,
total_timesteps=None,
nb_epochs=None, # with default settings, perform 1M steps total
nb_epoch_cycles=20,
nb_rollout_steps=100,
LOG_STD_MIN=0.1,
LOG_STD_MAX=0.9,
reward_scale=1.0,
render=False,
render_eval=False,
normalize_returns=False,
normalize_observations=True,
actor_lr=1e-4,
critic_lr=1e-3,
gamma=0.99,
nb_train_steps=50, # per epoch cycle and MPI worker,
nb_eval_steps=100,
nb_save_epochs=None,
batch_size=64, # per MPI worker
tau=0.01,
action_range=(-250.0, 250.0),
observation_range=(-5.0, 5.0),
eval_env=None,
load_path=None,
save_dir=None,
param_noise_adaption_interval=50,
**network_kwargs):
set_global_seeds(seed)
if total_timesteps is not None:
assert nb_epochs is None
nb_epochs = int(total_timesteps) // (nb_epoch_cycles * nb_rollout_steps)
else:
nb_epochs = 500
if MPI is not None:
rank = MPI.COMM_WORLD.Get_rank()
else:
rank = 0
memory = Memory(limit=int(1e6))
network_spec = [
{
'layer_type': 'dense',
'units': int (256),
'activation': 'relu',
'nodes_in': ['main'],
'nodes_out': ['main']
},
{
'layer_type': 'dense',
'units': int (128),
'activation': 'relu',
'nodes_in': ['main'],
'nodes_out': ['main']
}
]
vnetwork_spec = [
{
'layer_type': 'concat',
'nodes_in': ['action_movement', 'observation_self'],
'nodes_out': ['main']
},
{
'layer_type': 'dense',
'units': int (256),
'activation': 'relu',
'nodes_in': ['main'],
'nodes_out': ['main']
},
{
'layer_type': 'dense',
'units': int (128),
'activation': 'relu',
'nodes_in': ['main'],
'nodes_out': ['main']
}
]
vnetwork_spec2 = [
{
'layer_type': 'concat',
'nodes_in': ['observation_self'],
'nodes_out': ['main']
},
{
'layer_type': 'dense',
'units': int (256),
'activation': 'relu',
'nodes_in': ['main'],
'nodes_out': ['main']
},
{
'layer_type': 'dense',
'units': int (128),
'activation': 'relu',
'nodes_in': ['main'],
'nodes_out': ['main']
}
]
ob_space = env.observation_space
ac_space = env.action_space
network = SacPolicy(scope="sac", ob_space=ob_space, ac_space=ac_space, network_spec=network_spec,
v_network_spec=vnetwork_spec, v_network_spec2=vnetwork_spec2,
stochastic=False, reuse=False, build_act=True,
trainable_vars=None, not_trainable_vars=None,
gaussian_fixed_var=False, weight_decay=0.0, ema_beta=0.99999,
normalize_observations=normalize_observations, normalize_returns=normalize_returns,
observation_range=observation_range,
LOG_STD_MIN=LOG_STD_MIN, LOG_STD_MAX=LOG_STD_MAX)
target_network = SacPolicy(scope="sac_target", ob_space=ob_space, ac_space=ac_space, network_spec=network_spec,
v_network_spec=vnetwork_spec, v_network_spec2=vnetwork_spec2,
stochastic=False, reuse=False, build_act=True,
trainable_vars=None, not_trainable_vars=None,
gaussian_fixed_var=False, weight_decay=0.0, ema_beta=0.99999,
normalize_observations=normalize_observations, normalize_returns=normalize_returns,
observation_range=observation_range,
LOG_STD_MIN=LOG_STD_MIN, LOG_STD_MAX=LOG_STD_MAX)
# obs0
obs0 = dict()
for k, v in ob_space.spaces.items():
obs0[k] = network.phs[k]
# obs1
obs1 = dict()
for k, v in ob_space.spaces.items():
obs1[k] = target_network.phs[k]
# actions
acts = dict()
for k, v in ac_space.spaces.items():
acts[k] = network.phs[k]
self.dones = tf.placeholder(tf.float32, shape=(None, 1), name='dones')
self.rewards = tf.placeholder(tf.float32, shape=(None, None, 1), name='rewards')
# min double-q
min_q_pi = tf.minimum(network.q1_pi, network.q2_pi)
eval_episode_rewards_history = deque(maxlen=100)
episode_rewards_history = deque(maxlen=100)
sess = U.get_session()
saver = functools.partial(save_variables, sess=sess)
loader = functools.partial(load_variables, sess=sess)
if load_path != None:
loader(load_path)
# Prepare everything.
agent.initialize(sess)
sess.graph.finalize()
agent.reset()
obs = env.reset()
if eval_env is not None:
eval_obs = eval_env.reset()
nenvs = env.num_envs
n_agents = obs['observation_self'].shape[0]
episode_reward = np.zeros((nenvs, n_agents), dtype = np.float32) #vector
episode_step = np.zeros(nenvs, dtype = int) # vector
episodes = 0 #scalar
t = 0 # scalar
epoch = 0
start_time = time.time()
epoch_episode_rewards = []
epoch_episode_steps = []
epoch_actions = []
epoch_qs = []
epoch_episodes = 0
for epoch in range(nb_epochs):
for cycle in range(nb_epoch_cycles):
# Perform rollouts.
if nenvs > 1:
# if simulating multiple envs in parallel, impossible to reset agent at the end of the episode in each
# of the environments, so resetting here instead
agent.reset()
for t_rollout in range(nb_rollout_steps):
# Predict next action.
action, q, _, _ = agent.step(obs, apply_noise=True, compute_Q=True)
# Execute next action.
if rank == 0 and render:
env.render()
# max_action is of dimension A, whereas action is dimension (nenvs, A) - the multiplication gets broadcasted to the batch
for k, v in action.items():
action[k] *= max_action
nenvs_actions = []
for i in range(nenvs):
nenv_action = {'action_movement' : action['action_movement'][i*n_agents:(i + 1)*n_agents]}
nenvs_actions.append(nenv_action)
new_obs, r, done, info = env.step(nenvs_actions) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
# note these outputs are batched from vecenv
t += 1
if rank == 0 and render:
env.render()
episode_reward += r
episode_step += 1
# Book-keeping.
epoch_actions.append(action)
epoch_qs.append(q)
agent.store_transition(obs, action, r, new_obs, done) #the batched data will be unrolled in memory.py's append.
obs = new_obs
for d in range(len(done)):
if done[d]:
# Episode done.
epoch_episode_rewards.append(episode_reward[d])
episode_rewards_history.append(episode_reward[d])
epoch_episode_steps.append(episode_step[d])
episode_reward[d] = 0.
episode_step[d] = 0
epoch_episodes += 1
episodes += 1
if nenvs == 1:
agent.reset()
# Train.
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
for t_train in range(nb_train_steps):
# Adapt param noise, if necessary.
if memory.nb_entries >= batch_size and t_train % param_noise_adaption_interval == 0:
distance = agent.adapt_param_noise()
epoch_adaptive_distances.append(distance)
cl, al = agent.train()
epoch_critic_losses.append(cl)
epoch_actor_losses.append(al)
agent.update_target_net()
# Evaluate.
eval_episode_rewards = []
eval_qs = []
if eval_env is not None:
nenvs_eval = eval_obs.shape[0]
eval_episode_reward = np.zeros(nenvs_eval, dtype = np.float32)
for t_rollout in range(nb_eval_steps):
eval_action, eval_q, _, _ = agent.step(eval_obs, apply_noise=False, compute_Q=True)
eval_obs, eval_r, eval_done, eval_info = eval_env.step(max_action * eval_action) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
if render_eval:
eval_env.render()
eval_episode_reward += eval_r
eval_qs.append(eval_q)
for d in range(len(eval_done)):
if eval_done[d]:
eval_episode_rewards.append(eval_episode_reward[d])
eval_episode_rewards_history.append(eval_episode_reward[d])
eval_episode_reward[d] = 0.0
if MPI is not None:
mpi_size = MPI.COMM_WORLD.Get_size()
else:
mpi_size = 1
# Log stats.
# XXX shouldn't call np.mean on variable length lists
duration = time.time() - start_time
stats = agent.get_stats()
combined_stats = stats.copy()
combined_stats['rollout/return'] = np.mean(epoch_episode_rewards)
combined_stats['rollout/return_std'] = np.std(epoch_episode_rewards)
combined_stats['rollout/return_history'] = np.mean(episode_rewards_history)
combined_stats['rollout/return_history_std'] = np.std(episode_rewards_history)
combined_stats['rollout/episode_steps'] = np.mean(epoch_episode_steps)
combined_stats['rollout/Q_mean'] = np.mean(epoch_qs)
combined_stats['train/loss_actor'] = np.mean(epoch_actor_losses)
combined_stats['train/loss_critic'] = np.mean(epoch_critic_losses)
combined_stats['train/param_noise_distance'] = np.mean(epoch_adaptive_distances)
combined_stats['total/duration'] = duration
combined_stats['total/steps_per_second'] = float(t) / float(duration)
combined_stats['total/episodes'] = episodes
combined_stats['rollout/episodes'] = epoch_episodes
# Evaluation statistics.
if eval_env is not None:
combined_stats['eval/return'] = eval_episode_rewards
combined_stats['eval/return_history'] = np.mean(eval_episode_rewards_history)
combined_stats['eval/Q'] = eval_qs
combined_stats['eval/episodes'] = len(eval_episode_rewards)
def as_scalar(x):
if isinstance(x, np.ndarray):
assert x.size == 1
return x[0]
elif np.isscalar(x):
return x
else:
raise ValueError('expected scalar, got %s'%x)
combined_stats_sums = np.array([ np.array(x).flatten()[0] for x in combined_stats.values()])
if MPI is not None:
combined_stats_sums = MPI.COMM_WORLD.allreduce(combined_stats_sums)
combined_stats = {k : v / mpi_size for (k,v) in zip(combined_stats.keys(), combined_stats_sums)}
# Total statistics.
combined_stats['total/epochs'] = epoch + 1
combined_stats['total/steps'] = t
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
if rank == 0:
logger.dump_tabular()
logger.info('')
logdir = logger.get_dir()
if rank == 0 and logdir:
if hasattr(env, 'get_state'):
with open(os.path.join(logdir, 'env_state.pkl'), 'wb') as f:
pickle.dump(env.get_state(), f)
if eval_env and hasattr(eval_env, 'get_state'):
with open(os.path.join(logdir, 'eval_env_state.pkl'), 'wb') as f:
pickle.dump(eval_env.get_state(), f)
if nb_save_epochs != None and (epoch + 1) % nb_save_epochs == 0:
if save_dir == None:
checkdir = osp.join(logger.get_dir(), 'checkpoints')
else:
checkdir = osp.join(save_dir, 'checkpoints')
os.makedirs(checkdir, exist_ok=True)
savepath = osp.join(checkdir, '%.5i'%epoch)
print('Saving to', savepath)
saver(savepath)
return agent
def run(env_id, seed, noise_type, num_cpu, layer_norm, logdir, gym_monitor, evaluation, bind_to_core, **kwargs):
kwargs['logdir'] = logdir
whoami = mpi_fork(num_cpu, bind_to_core=bind_to_core)
if whoami == 'parent':
sys.exit(0)
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
# Write to temp directory for all non-master workers.
actual_dir = None
Logger.CURRENT.close()
Logger.CURRENT = Logger(dir=mkdtemp(), output_formats=[])
logger.set_level(logger.DISABLED)
# Create envs.
if rank == 0:
env = gym.make(env_id)
if gym_monitor and logdir:
env = gym.wrappers.Monitor(env, os.path.join(logdir, 'gym_train'), force=True)
env = SimpleMonitor(env)
if evaluation:
eval_env = gym.make(env_id)
if gym_monitor and logdir:
eval_env = gym.wrappers.Monitor(eval_env, os.path.join(logdir, 'gym_eval'), force=True)
eval_env = SimpleMonitor(eval_env)
else:
eval_env = None
else:
env = gym.make(env_id)
if evaluation:
eval_env = gym.make(env_id)
else:
eval_env = None
# Parse noise_type
action_noise = None
param_noise = None
nb_actions = env.action_space.shape[-1]
for current_noise_type in noise_type.split(','):
current_noise_type = current_noise_type.strip()
if current_noise_type == 'none':
pass
elif 'adaptive-param' in current_noise_type:
_, stddev = current_noise_type.split('_')
param_noise = AdaptiveParamNoiseSpec(initial_stddev=float(stddev), desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = NormalActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
else:
raise RuntimeError('unknown noise type "{}"'.format(current_noise_type))
# Configure components.
memory = Memory(limit=int(1e6), action_shape=env.action_space.shape, observation_shape=env.observation_space.shape)
critic = Critic(layer_norm=layer_norm)
actor = Actor(nb_actions, layer_norm=layer_norm)
# Seed everything to make things reproducible.
seed = seed + 1000000 * rank
logger.info('rank {}: seed={}, logdir={}'.format(rank, seed, logger.get_dir()))
tf.reset_default_graph()
set_global_seeds(seed)
env.seed(seed)
if eval_env is not None:
eval_env.seed(seed)
# Disable logging for rank != 0 to avoid noise.
if rank == 0:
start_time = time.time()
training.train(env=env, eval_env=eval_env, param_noise=param_noise,
action_noise=action_noise, actor=actor, critic=critic, memory=memory, **kwargs)
env.close()
if eval_env is not None:
eval_env.close()
Logger.CURRENT.close()
if rank == 0:
logger.info('total runtime: {}s'.format(time.time() - start_time))
def train(policy, rollout_worker, evaluator, n_epochs, n_test_rollouts,
n_cycles, n_batches, policy_save_interval, save_policies, num_cpu,
dump_buffer, w_potential, w_linear, w_rotational, rank_method,
clip_energy, **kwargs):
rank = MPI.COMM_WORLD.Get_rank()
latest_policy_path = os.path.join(logger.get_dir(), 'policy_latest.pkl')
best_policy_path = os.path.join(logger.get_dir(), 'policy_best.pkl')
periodic_policy_path = os.path.join(logger.get_dir(), 'policy_{}.pkl')
logger.info("Training...")
best_success_rate = -1
t = 1
for epoch in range(n_epochs):
# train
rollout_worker.clear_history()
for cycle in range(n_cycles):
episode = rollout_worker.generate_rollouts()
policy.store_episode(episode, dump_buffer, w_potential, w_linear,
w_rotational, rank_method, clip_energy)
for batch in range(n_batches):
t = ((epoch * n_cycles * n_batches) +
(cycle * n_batches) + batch) * num_cpu
policy.train(t, dump_buffer)
policy.update_target_net()
# test
evaluator.clear_history()
for _ in range(n_test_rollouts):
evaluator.generate_rollouts()
# record logs
logger.record_tabular('epoch', epoch)
for key, val in evaluator.logs('test'):
logger.record_tabular(key, mpi_average(val))
for key, val in rollout_worker.logs('train'):
logger.record_tabular(key, mpi_average(val))
for key, val in policy.logs():
logger.record_tabular(key, mpi_average(val))
if rank == 0:
logger.dump_tabular()
if dump_buffer:
policy.dump_buffer(epoch)
# save the policy if it's better than the previous ones
success_rate = mpi_average(evaluator.current_success_rate())
if rank == 0 and success_rate >= best_success_rate and save_policies:
best_success_rate = success_rate
logger.info(
'New best success rate: {}. Saving policy to {} ...'.format(
best_success_rate, best_policy_path))
evaluator.save_policy(best_policy_path)
evaluator.save_policy(latest_policy_path)
if rank == 0 and policy_save_interval > 0 and epoch % policy_save_interval == 0 and save_policies:
policy_path = periodic_policy_path.format(epoch)
logger.info('Saving periodic policy to {} ...'.format(policy_path))
evaluator.save_policy(policy_path)
# make sure that different threads have different seeds
local_uniform = np.random.uniform(size=(1, ))
root_uniform = local_uniform.copy()
MPI.COMM_WORLD.Bcast(root_uniform, root=0)
if rank != 0:
assert local_uniform[0] != root_uniform[0]
def train(env, nb_epochs, nb_epoch_cycles, render_eval, reward_scale, render, param_noise, actor, critic,
normalize_returns, normalize_observations, critic_l2_reg, actor_lr, critic_lr, action_noise,
popart, gamma, clip_norm, nb_train_steps, nb_rollout_steps, nb_eval_steps, batch_size, memory,
tau=0.01, eval_env=None, param_noise_adaption_interval=50):
rank = MPI.COMM_WORLD.Get_rank()
assert (np.abs(env.action_space.low) == env.action_space.high).all() # we assume symmetric actions.
max_action = env.action_space.high
logger.info('scaling actions by {} before executing in env'.format(max_action))
agent = DDPG(actor, critic, memory, env.observation_space.shape, env.action_space.shape,
gamma=gamma, tau=tau, normalize_returns=normalize_returns, normalize_observations=normalize_observations,
batch_size=batch_size, action_noise=action_noise, param_noise=param_noise, critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr, critic_lr=critic_lr, enable_popart=popart, clip_norm=clip_norm,
reward_scale=reward_scale)
logger.info('Using agent with the following configuration:')
logger.info(str(agent.__dict__.items()))
# Set up logging stuff only for a single worker.
if rank == 0:
saver = tf.train.Saver()
else:
saver = None
step = 0
episode = 0
eval_episode_rewards_history = deque(maxlen=100)
episode_rewards_history = deque(maxlen=100)
with U.single_threaded_session() as sess:
# Prepare everything.
agent.initialize(sess)
sess.graph.finalize()
agent.reset()
obs = env.reset()
if eval_env is not None:
eval_obs = eval_env.reset()
done = False
episode_reward = 0.
episode_step = 0
episodes = 0
t = 0
epoch = 0
start_time = time.time()
epoch_episode_rewards = []
epoch_episode_steps = []
epoch_episode_eval_rewards = []
epoch_episode_eval_steps = []
epoch_start_time = time.time()
epoch_actions = []
epoch_qs = []
epoch_episodes = 0
for epoch in range(nb_epochs):
for cycle in range(nb_epoch_cycles):
# Perform rollouts.
for t_rollout in range(nb_rollout_steps):
# Predict next action.
action, q = agent.pi(obs, apply_noise=True, compute_Q=True)
assert action.shape == env.action_space.shape
# Execute next action.
if rank == 0 and render:
env.render()
assert max_action.shape == action.shape
new_obs, r, done, info = env.step(max_action * action) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
t += 1
if rank == 0 and render:
env.render()
episode_reward += r
episode_step += 1
# Book-keeping.
epoch_actions.append(action)
epoch_qs.append(q)
agent.store_transition(obs, action, r, new_obs, done)
obs = new_obs
if done:
# Episode done.
epoch_episode_rewards.append(episode_reward)
episode_rewards_history.append(episode_reward)
epoch_episode_steps.append(episode_step)
episode_reward = 0.
episode_step = 0
epoch_episodes += 1
episodes += 1
agent.reset()
obs = env.reset()
# Train.
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
for t_train in range(nb_train_steps):
# Adapt param noise, if necessary.
if memory.nb_entries >= batch_size and t % param_noise_adaption_interval == 0:
distance = agent.adapt_param_noise()
epoch_adaptive_distances.append(distance)
cl, al = agent.train()
epoch_critic_losses.append(cl)
epoch_actor_losses.append(al)
agent.update_target_net()
# Evaluate.
eval_episode_rewards = []
eval_qs = []
if eval_env is not None:
eval_episode_reward = 0.
for t_rollout in range(nb_eval_steps):
eval_action, eval_q = agent.pi(eval_obs, apply_noise=False, compute_Q=True)
eval_obs, eval_r, eval_done, eval_info = eval_env.step(max_action * eval_action) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
if render_eval:
eval_env.render()
eval_episode_reward += eval_r
eval_qs.append(eval_q)
if eval_done:
eval_obs = eval_env.reset()
eval_episode_rewards.append(eval_episode_reward)
eval_episode_rewards_history.append(eval_episode_reward)
eval_episode_reward = 0.
mpi_size = MPI.COMM_WORLD.Get_size()
# Log stats.
# XXX shouldn't call np.mean on variable length lists
duration = time.time() - start_time
stats = agent.get_stats()
combined_stats = stats.copy()
combined_stats['rollout/return'] = np.mean(epoch_episode_rewards)
combined_stats['rollout/return_history'] = np.mean(episode_rewards_history)
combined_stats['rollout/episode_steps'] = np.mean(epoch_episode_steps)
combined_stats['rollout/actions_mean'] = np.mean(epoch_actions)
combined_stats['rollout/Q_mean'] = np.mean(epoch_qs)
combined_stats['train/loss_actor'] = np.mean(epoch_actor_losses)
combined_stats['train/loss_critic'] = np.mean(epoch_critic_losses)
combined_stats['train/param_noise_distance'] = np.mean(epoch_adaptive_distances)
combined_stats['total/duration'] = duration
combined_stats['total/steps_per_second'] = float(t) / float(duration)
combined_stats['total/episodes'] = episodes
combined_stats['rollout/episodes'] = epoch_episodes
combined_stats['rollout/actions_std'] = np.std(epoch_actions)
# Evaluation statistics.
if eval_env is not None:
combined_stats['eval/return'] = eval_episode_rewards
combined_stats['eval/return_history'] = np.mean(eval_episode_rewards_history)
combined_stats['eval/Q'] = eval_qs
combined_stats['eval/episodes'] = len(eval_episode_rewards)
def as_scalar(x):
if isinstance(x, np.ndarray):
assert x.size == 1
return x[0]
elif np.isscalar(x):
return x
else:
raise ValueError('expected scalar, got %s'%x)
combined_stats_sums = MPI.COMM_WORLD.allreduce(np.array([as_scalar(x) for x in combined_stats.values()]))
combined_stats = {k : v / mpi_size for (k,v) in zip(combined_stats.keys(), combined_stats_sums)}
# Total statistics.
combined_stats['total/epochs'] = epoch + 1
combined_stats['total/steps'] = t
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
logger.dump_tabular()
logger.info('')
logdir = logger.get_dir()
if rank == 0 and logdir:
if hasattr(env, 'get_state'):
with open(os.path.join(logdir, 'env_state.pkl'), 'wb') as f:
pickle.dump(env.get_state(), f)
if eval_env and hasattr(eval_env, 'get_state'):
with open(os.path.join(logdir, 'eval_env_state.pkl'), 'wb') as f:
pickle.dump(eval_env.get_state(), f)
def define_dynamics_prediction_rew(self, convfeat, rep_size, enlargement):
# Dynamics loss with random features.
# Random target network.
for ph in self.ph_ob.values():
if len(ph.shape.as_list()) == 5: # B,T,H,W,C
logger.info(
f"CnnTarget: using '{ph.name}' shape {ph.shape} as image input"
)
xr = ph[:, 1:]
xr = tf.cast(xr, tf.float32)
xr = tf.reshape(xr, (-1, *ph.shape.as_list()[-3:]))[:, :, :,
-1:]
xr = tf.clip_by_value((xr - self.ph_mean) / self.ph_std, -5.0,
5.0)
xr = tf.nn.leaky_relu(
conv(
xr,
"c1r",
nf=convfeat * 1,
rf=8,
stride=4,
init_scale=np.sqrt(2),
))
xr = tf.nn.leaky_relu(
conv(
xr,
"c2r",
nf=convfeat * 2 * 1,
rf=4,
stride=2,
init_scale=np.sqrt(2),
))
xr = tf.nn.leaky_relu(
conv(
xr,
"c3r",
nf=convfeat * 2 * 1,
rf=3,
stride=1,
init_scale=np.sqrt(2),
))
rgbr = [to2d(xr)]
X_r = fc(rgbr[0], "fc1r", nh=rep_size, init_scale=np.sqrt(2))
# Predictor network.
ac_one_hot = tf.one_hot(self.ph_ac, self.ac_space.n, axis=2)
assert ac_one_hot.get_shape().ndims == 3
assert ac_one_hot.get_shape().as_list() == [
None,
None,
self.ac_space.n,
], ac_one_hot.get_shape().as_list()
ac_one_hot = tf.reshape(ac_one_hot, (-1, self.ac_space.n))
def cond(x):
return tf.concat([x, ac_one_hot], 1)
for ph in self.ph_ob.values():
if len(ph.shape.as_list()) == 5: # B,T,H,W,C
logger.info(
f"CnnTarget: using '{ph.name}' shape {ph.shape} as image input"
)
xrp = ph[:, :-1]
xrp = tf.cast(xrp, tf.float32)
xrp = tf.reshape(xrp, (-1, *ph.shape.as_list()[-3:]))
# ph_mean, ph_std are 84x84x1, so we subtract the average of the last channel from all channels. Is this ok?
xrp = tf.clip_by_value((xrp - self.ph_mean) / self.ph_std,
-5.0, 5.0)
xrp = tf.nn.leaky_relu(
conv(
xrp,
"c1rp_pred",
nf=convfeat,
rf=8,
stride=4,
init_scale=np.sqrt(2),
))
xrp = tf.nn.leaky_relu(
conv(
xrp,
"c2rp_pred",
nf=convfeat * 2,
rf=4,
stride=2,
init_scale=np.sqrt(2),
))
xrp = tf.nn.leaky_relu(
conv(
xrp,
"c3rp_pred",
nf=convfeat * 2,
rf=3,
stride=1,
init_scale=np.sqrt(2),
))
rgbrp = to2d(xrp)
X_r_hat = tf.nn.relu(
fc(
cond(rgbrp),
"fc1r_hat1_pred",
nh=256 * enlargement,
init_scale=np.sqrt(2),
))
X_r_hat = tf.nn.relu(
fc(
cond(X_r_hat),
"fc1r_hat2_pred",
nh=256 * enlargement,
init_scale=np.sqrt(2),
))
X_r_hat = fc(cond(X_r_hat),
"fc1r_hat3_pred",
nh=rep_size,
init_scale=np.sqrt(2))
self.feat_var = tf.reduce_mean(tf.nn.moments(X_r, axes=[0])[1])
self.max_feat = tf.reduce_max(tf.abs(X_r))
self.int_rew = tf.reduce_mean(
tf.square(tf.stop_gradient(X_r) - X_r_hat),
axis=-1,
keep_dims=True)
self.int_rew = tf.reshape(self.int_rew,
(self.sy_nenvs, self.sy_nsteps - 1))
noisy_targets = tf.stop_gradient(X_r)
self.aux_loss = tf.reduce_mean(tf.square(noisy_targets - X_r_hat), -1)
mask = tf.random_uniform(shape=tf.shape(self.aux_loss),
minval=0.0,
maxval=1.0,
dtype=tf.float32)
mask = tf.cast(mask < self.proportion_of_exp_used_for_predictor_update,
tf.float32)
self.aux_loss = tf.reduce_sum(mask * self.aux_loss) / tf.maximum(
tf.reduce_sum(mask), 1.0)
def apply_policy(
ph_ob,
reuse,
scope,
hidsize,
memsize,
extrahid,
sy_nenvs,
sy_nsteps,
pdparamsize,
additional_inputs=None,
):
meta_rl = False
data_format = "NHWC"
ph = ph_ob
assert len(ph.shape.as_list()) == 5 # B,T,H,W,C
logger.info(
f"CnnPolicy: using '{ph.name}' shape {ph.shape} as image input")
X = tf.cast(ph, tf.float32) / 255.0
X = tf.reshape(X, (-1, *ph.shape.as_list()[-3:]))
activ = tf.nn.relu
yes_gpu = any(get_available_gpus())
with tf.variable_scope(
scope,
reuse=reuse), tf.device("/gpu:0" if yes_gpu else "/cpu:0"):
X = activ(
conv(
X,
"c1",
nf=32,
rf=8,
stride=4,
init_scale=np.sqrt(2),
data_format=data_format,
))
X = activ(
conv(
X,
"c2",
nf=64,
rf=4,
stride=2,
init_scale=np.sqrt(2),
data_format=data_format,
))
X = activ(
conv(
X,
"c3",
nf=64,
rf=4,
stride=1,
init_scale=np.sqrt(2),
data_format=data_format,
))
X = to2d(X)
mix_other_observations = [X]
if ('prev_acs' in additional_inputs) and ('prev_rew'
in additional_inputs):
# Cast numpy arrays to tf tensors
prev_acs = tf.cast(additional_inputs['prev_acs'], tf.float32)
prev_rew = tf.cast(additional_inputs['prev_rew'], tf.float32)
# Flatten out time dimension
prev_acs = tf.reshape(prev_acs,
(-1, *prev_acs.shape.as_list()[2:]))
prev_rew = tf.reshape(prev_rew,
(-1, *prev_rew.shape.as_list()[2:]))
# Add to 2D features going to FC layers
mix_other_observations.extend([prev_acs, prev_rew])
X = tf.concat(mix_other_observations, axis=1)
X = activ(fc(X, "fc1", nh=hidsize, init_scale=np.sqrt(2)))
additional_size = 448
X = activ(
fc(X,
"fc_additional",
nh=additional_size,
init_scale=np.sqrt(2)))
snext = tf.zeros((sy_nenvs, memsize))
mix_timeout = [X]
Xtout = tf.concat(mix_timeout, axis=1)
if extrahid:
Xtout = X + activ(
fc(Xtout, "fc2val", nh=additional_size, init_scale=0.1))
X = X + activ(
fc(X, "fc2act", nh=additional_size, init_scale=0.1))
pdparam = fc(X, "pd", nh=pdparamsize, init_scale=0.01)
vpred_int = fc(Xtout, "vf_int", nh=1, init_scale=0.01)
vpred_ext = fc(Xtout, "vf_ext", nh=1, init_scale=0.01)
pdparam = tf.reshape(pdparam, (sy_nenvs, sy_nsteps, pdparamsize))
vpred_int = tf.reshape(vpred_int, (sy_nenvs, sy_nsteps))
vpred_ext = tf.reshape(vpred_ext, (sy_nenvs, sy_nsteps))
return pdparam, vpred_int, vpred_ext, snext
def learn(*,
network,
env,
total_timesteps,
seed=None,
nsteps=2048,
ent_coef=0.0,
lr=3e-4,
vf_coef=0.5,
max_grad_norm=0.5,
gamma=0.99,
lam=0.95,
log_interval=10,
nminibatches=4,
noptepochs=4,
cliprange=0.2,
save_interval=0,
load_path=None,
callback=None,
**network_kwargs):
'''
Learn policy using PPO algorithm (https://arxiv.org/abs/1707.06347)
Parameters:
----------
network: policy network architecture. Either string (mlp, lstm, lnlstm, cnn_lstm, cnn, cnn_small, conv_only - see baselines.common/models.py for full list)
specifying the standard network architecture, or a function that takes tensorflow tensor as input and returns
tuple (output_tensor, extra_feed) where output tensor is the last network layer output, extra_feed is None for feed-forward
neural nets, and extra_feed is a dictionary describing how to feed state into the network for recurrent neural nets.
See baselines.common/policies.py/lstm for more details on using recurrent nets in policies
env: baselines.common.vec_env.VecEnv environment. Needs to be vectorized for parallel environment simulation.
The environments produced by gym.make can be wrapped using baselines.common.vec_env.DummyVecEnv class.
nsteps: int number of steps of the vectorized environment per update (i.e. batch size is nsteps * nenv where
nenv is number of environment copies simulated in parallel)
total_timesteps: int number of timesteps (i.e. number of actions taken in the environment)
ent_coef: float policy entropy coefficient in the optimization objective
lr: float or function learning rate, constant or a schedule function [0,1] -> R+ where 1 is beginning of the
training and 0 is the end of the training.
vf_coef: float value function loss coefficient in the optimization objective
max_grad_norm: float or None gradient norm clipping coefficient
gamma: float discounting factor
lam: float advantage estimation discounting factor (lambda in the paper)
log_interval: int number of timesteps between logging events
nminibatches: int number of training minibatches per update
noptepochs: int number of training epochs per update
cliprange: float or function clipping range, constant or schedule function [0,1] -> R+ where 1 is beginning of the training
and 0 is the end of the training
save_interval: int number of timesteps between saving events
load_path: str path to load the model from
**network_kwargs: keyword arguments to the policy / network builder. See baselines.common/policies.py/build_policy and arguments to a particular type of network
For instance, 'mlp' network architecture has arguments num_hidden and num_layers.
'''
set_global_seeds(seed)
if isinstance(lr, float):
lr = constfn(lr)
else:
assert callable(lr)
if isinstance(cliprange, float):
cliprange = constfn(cliprange)
else:
assert callable(cliprange)
total_timesteps = int(total_timesteps)
policy = build_policy(env, network, **network_kwargs)
nenvs = env.num_envs
ob_space = env.observation_space
ac_space = env.action_space
nbatch = nenvs * nsteps
nbatch_train = nbatch // nminibatches
make_model = lambda: Model(policy=policy,
ob_space=ob_space,
ac_space=ac_space,
nbatch_act=nenvs,
nbatch_train=nbatch_train,
nsteps=nsteps,
ent_coef=ent_coef,
vf_coef=vf_coef,
max_grad_norm=max_grad_norm)
# if save_interval and logger.get_dir():
# import cloudpickle
# with open(osp.join(logger.get_dir(), 'make_model.pkl'), 'wb') as fh:
# fh.write(cloudpickle.dumps(make_model))
model = make_model()
if load_path is not None:
model.load(load_path)
runner = Runner(env=env, model=model, nsteps=nsteps, gamma=gamma, lam=lam)
epinfobuf = deque(maxlen=100)
tfirststart = time.time()
nupdates = total_timesteps // nbatch
for update in range(1, nupdates + 1):
if callback is not None:
if callback(locals(), globals()):
break
assert nbatch % nminibatches == 0
nbatch_train = nbatch // nminibatches
tstart = time.time()
frac = 1.0 - (update - 1.0) / nupdates
cliprangenow = cliprange(frac)
obs, returns, masks, actions, values, neglogpacs, states, epinfos = runner.run(
) # pylint: disable=E0632
epinfobuf.extend(epinfos)
mblrvals = []
mblossvals = []
if states is None: # nonrecurrent version
inds = np.arange(nbatch)
for iteration in range(noptepochs):
np.random.shuffle(inds)
lrnow = lr(frac, iteration)
logger.info('lrnow={} (frac={},iteration={})'.format(
lrnow, frac, iteration))
mblrvals.append(lrnow)
for start in range(0, nbatch, nbatch_train):
end = start + nbatch_train
mbinds = inds[start:end]
slices = (arr[mbinds]
for arr in (obs, returns, masks, actions, values,
neglogpacs))
mblossvals.append(model.train(lrnow, cliprangenow,
*slices))
else: # recurrent version
assert nenvs % nminibatches == 0
envsperbatch = nenvs // nminibatches
envinds = np.arange(nenvs)
flatinds = np.arange(nenvs * nsteps).reshape(nenvs, nsteps)
envsperbatch = nbatch_train // nsteps
for iteration in range(noptepochs):
np.random.shuffle(envinds)
lrnow = lr(frac, iteration)
for start in range(0, nenvs, envsperbatch):
end = start + envsperbatch
mbenvinds = envinds[start:end]
mbflatinds = flatinds[mbenvinds].ravel()
slices = (arr[mbflatinds]
for arr in (obs, returns, masks, actions, values,
neglogpacs))
mbstates = states[mbenvinds]
mblossvals.append(
model.train(lrnow, cliprangenow, *slices, mbstates))
lossvals = np.mean(mblossvals, axis=0)
lrvals = np.mean(mblrvals)
tnow = time.time()
fps = int(nbatch / (tnow - tstart))
if update % log_interval == 0 or update == 1:
ev = explained_variance(values, returns)
logger.logkv("serial_timesteps", update * nsteps)
logger.logkv("nupdates", update)
logger.logkv("total_timesteps", update * nbatch)
logger.logkv("fps", fps)
logger.logkv("explained_variance", float(ev))
logger.logkv('eprewmean',
safemean([epinfo['r'] for epinfo in epinfobuf]))
logger.logkv('eplenmean',
safemean([epinfo['l'] for epinfo in epinfobuf]))
logger.logkv('time_elapsed', tnow - tfirststart)
logger.logkv('lr_mean', lrvals)
logger.logkv('epscoremean',
safemean([epinfo['s'] for epinfo in epinfobuf]))
for (lossval, lossname) in zip(lossvals, model.loss_names):
logger.logkv(lossname, lossval)
if MPI.COMM_WORLD.Get_rank() == 0:
logger.dumpkvs()
if save_interval and (update % save_interval == 0 or update == 1
or update == nupdates) and logger.get_dir(
) and MPI.COMM_WORLD.Get_rank() == 0:
checkdir = osp.join(logger.get_dir(), 'checkpoints')
os.makedirs(checkdir, exist_ok=True)
savepath = osp.join(checkdir, '%.5i' % update)
print('Saving to', savepath)
model.save(savepath)
env.close()
return model
def call(self, perform, save_networks, use_expert, expert, on_policy):
runner, model, buffer, steps = self.runner, self.model, self.buffer, self.steps
expert_buffer = self.expert_buffer
if on_policy:
enc_obs, obs, actions, rewards, mus, dones, masks = runner.run()
runner.myrun()
# if self.flag>0:
# print(self.flag,'=================================')
# print(enc_obs, obs, actions, rewards, mus, dones, masks)
# self.flag = self.flag -1
self.episode_stats.feed(rewards, dones)
if buffer is not None and not perform:
buffer.put(enc_obs, actions, rewards, mus, dones, masks)
else:
# get obs, actions, rewards, mus, dones from buffer.
obs, actions, rewards, mus, dones, masks = buffer.get()
#enc_obs, obs, actions, rewards, mus, dones, masks = runner.run()
if not perform:
# reshape stuff correctly
obs = obs.reshape(runner.batch_ob_shape)
actions = actions.reshape([runner.nbatch])
rewards = rewards.reshape([runner.nbatch])
mus = mus.reshape([runner.nbatch, runner.nact])
dones = dones.reshape([runner.nbatch])
masks = masks.reshape([runner.batch_ob_shape[0]])
if not use_expert:
names_ops, values_ops = model.train(obs, actions, rewards,
dones, mus,
model.initial_state, masks,
steps)
else:
expert_obs, expert_actions, expert_rewards, expert_mus, expert_dones, expert_masks = expert.get(
)
expert_obs = expert_obs.reshape(runner.batch_ob_shape)
expert_actions = expert_actions.reshape([runner.nbatch])
expert_rewards = expert_rewards.reshape([runner.nbatch])
expert_mus = expert_mus.reshape([runner.nbatch, runner.nact])
expert_dones = expert_dones.reshape([runner.nbatch])
expert_masks = expert_masks.reshape([runner.batch_ob_shape[0]])
names_ops, values_ops = model.expert_train(
obs, actions, rewards, dones, mus, model.initial_state,
masks, steps, expert_obs, expert_actions, expert_rewards,
expert_mus, expert_dones, expert_masks)
if on_policy and (int(steps / runner.nbatch) % self.log_interval
== 0):
logger.record_tabular("total_timesteps", steps)
logger.record_tabular("fps",
int(steps / (time.time() - self.tstart)))
# IMP: In EpisodicLife env, during training, we get done=True at each loss of life, not just at the terminal state.
# Thus, this is mean until end of life, not end of episode.
# For true episode rewards, see the monitor files in the log folder.
logger.record_tabular("mean_episode_length",
self.episode_stats.mean_length())
logger.record_tabular("mean_episode_reward",
self.episode_stats.mean_reward())
for name, val in zip(names_ops, values_ops):
logger.record_tabular(name, float(val))
logger.dump_tabular()
if save_networks and (int(steps / runner.nbatch) %
self.log_interval * 10 == 0):
model.save(int(steps))
else: #if perform
expert_buffer.append(
[enc_obs, actions, rewards, mus, dones, masks])
if len(expert_buffer) > 0 and len(expert_buffer) % 100 == 0:
expert_dir = os.path.join('./expert') + '/'
if not os.path.exists(expert_dir):
os.makedirs(expert_dir)
pwritefile = open(os.path.join(expert_dir, 'expert_test.pkl'),
'wb')
pickle.dump(expert_buffer, pwritefile, -1)
pwritefile.close()
logger.info('Successfully Saved the Expert Data')
obs = obs.reshape(runner.batch_ob_shape)
actions = actions.reshape([runner.nbatch])
rewards = rewards.reshape([runner.nbatch])
mus = mus.reshape([runner.nbatch, runner.nact])
dones = dones.reshape([runner.nbatch])
masks = masks.reshape([runner.batch_ob_shape[0]])
if on_policy and (int(steps / runner.nbatch) % self.log_interval
== 0):
logger.record_tabular("total_timesteps", steps)
logger.record_tabular("fps",
int(steps / (time.time() - self.tstart)))
# IMP: In EpisodicLife env, during training, we get done=True at each loss of life, not just at the terminal state.
# Thus, this is mean until end of life, not end of episode.
# For true episode rewards, see the monitor files in the log folder.
logger.record_tabular("mean_episode_length",
self.episode_stats.mean_length())
logger.record_tabular("mean_episode_reward",
self.episode_stats.mean_reward())
logger.dump_tabular()
def _create_network(self, reuse=False):
logger.info("Creating a DDPG agent with action space %d x %s..." % (self.dimu, self.max_u))
self.sess = tf_util.get_session()
# running averages
with tf.variable_scope('o_stats') as vs:
if reuse:
vs.reuse_variables()
self.o_stats = Normalizer(self.dimo, self.norm_eps, self.norm_clip, sess=self.sess)
with tf.variable_scope('g_stats') as vs:
if reuse:
vs.reuse_variables()
self.g_stats = Normalizer(self.dimg, self.norm_eps, self.norm_clip, sess=self.sess)
# mini-batch sampling.
batch = self.staging_tf.get()
batch_tf = OrderedDict([(key, batch[i])
for i, key in enumerate(self.stage_shapes.keys())])
batch_tf['r'] = tf.reshape(batch_tf['r'], [-1, 1])
#choose only the demo buffer samples
mask = np.concatenate((np.zeros(self.batch_size - self.demo_batch_size), np.ones(self.demo_batch_size)), axis = 0)
# networks
with tf.variable_scope('main') as vs:
if reuse:
vs.reuse_variables()
self.main = self.create_actor_critic(batch_tf, net_type='main', **self.__dict__)
vs.reuse_variables()
with tf.variable_scope('target') as vs:
if reuse:
vs.reuse_variables()
target_batch_tf = batch_tf.copy()
target_batch_tf['o'] = batch_tf['o_2']
target_batch_tf['g'] = batch_tf['g_2']
self.target = self.create_actor_critic(
target_batch_tf, net_type='target', **self.__dict__)
vs.reuse_variables()
assert len(self._vars("main")) == len(self._vars("target"))
# loss functions
target_Q_pi_tf = self.target.Q_pi_tf
clip_range = (-self.clip_return, 0. if self.clip_pos_returns else np.inf)
target_tf = tf.clip_by_value(batch_tf['r'] + self.gamma * target_Q_pi_tf, *clip_range)
self.Q_loss_tf = tf.reduce_mean(tf.square(tf.stop_gradient(target_tf) - self.main.Q_tf))
if self.bc_loss ==1 and self.q_filter == 1 : # train with demonstrations and use bc_loss and q_filter both
maskMain = tf.reshape(tf.boolean_mask(self.main.Q_tf > self.main.Q_pi_tf, mask), [-1]) #where is the demonstrator action better than actor action according to the critic? choose those samples only
#define the cloning loss on the actor's actions only on the samples which adhere to the above masks
self.cloning_loss_tf = tf.reduce_sum(tf.square(tf.boolean_mask(tf.boolean_mask((self.main.pi_tf), mask), maskMain, axis=0) - tf.boolean_mask(tf.boolean_mask((batch_tf['u']), mask), maskMain, axis=0)))
self.pi_loss_tf = -self.prm_loss_weight * tf.reduce_mean(self.main.Q_pi_tf) #primary loss scaled by it's respective weight prm_loss_weight
self.pi_loss_tf += self.prm_loss_weight * self.action_l2 * tf.reduce_mean(tf.square(self.main.pi_tf / self.max_u)) #L2 loss on action values scaled by the same weight prm_loss_weight
self.pi_loss_tf += self.aux_loss_weight * self.cloning_loss_tf #adding the cloning loss to the actor loss as an auxilliary loss scaled by its weight aux_loss_weight
elif self.bc_loss == 1 and self.q_filter == 0: # train with demonstrations without q_filter
self.cloning_loss_tf = tf.reduce_sum(tf.square(tf.boolean_mask((self.main.pi_tf), mask) - tf.boolean_mask((batch_tf['u']), mask)))
self.pi_loss_tf = -self.prm_loss_weight * tf.reduce_mean(self.main.Q_pi_tf)
self.pi_loss_tf += self.prm_loss_weight * self.action_l2 * tf.reduce_mean(tf.square(self.main.pi_tf / self.max_u))
self.pi_loss_tf += self.aux_loss_weight * self.cloning_loss_tf
else: #If not training with demonstrations
self.pi_loss_tf = -tf.reduce_mean(self.main.Q_pi_tf)
self.pi_loss_tf += self.action_l2 * tf.reduce_mean(tf.square(self.main.pi_tf / self.max_u))
Q_grads_tf = tf.gradients(self.Q_loss_tf, self._vars('main/Q'))
pi_grads_tf = tf.gradients(self.pi_loss_tf, self._vars('main/pi'))
assert len(self._vars('main/Q')) == len(Q_grads_tf)
assert len(self._vars('main/pi')) == len(pi_grads_tf)
self.Q_grads_vars_tf = zip(Q_grads_tf, self._vars('main/Q'))
self.pi_grads_vars_tf = zip(pi_grads_tf, self._vars('main/pi'))
self.Q_grad_tf = flatten_grads(grads=Q_grads_tf, var_list=self._vars('main/Q'))
self.pi_grad_tf = flatten_grads(grads=pi_grads_tf, var_list=self._vars('main/pi'))
# optimizers
self.Q_adam = MpiAdam(self._vars('main/Q'), scale_grad_by_procs=False)
self.pi_adam = MpiAdam(self._vars('main/pi'), scale_grad_by_procs=False)
# polyak averaging
self.main_vars = self._vars('main/Q') + self._vars('main/pi')
self.target_vars = self._vars('target/Q') + self._vars('target/pi')
self.stats_vars = self._global_vars('o_stats') + self._global_vars('g_stats')
self.init_target_net_op = list(
map(lambda v: v[0].assign(v[1]), zip(self.target_vars, self.main_vars)))
self.update_target_net_op = list(
map(lambda v: v[0].assign(self.polyak * v[0] + (1. - self.polyak) * v[1]), zip(self.target_vars, self.main_vars)))
# initialize all variables
tf.variables_initializer(self._global_vars('')).run()
self._sync_optimizers()
self._init_target_net()
def learn(network, env,
seed=None,
total_timesteps=None,
nb_epochs=None, # with default settings, perform 1M steps total
nb_epoch_cycles=20,
nb_rollout_steps=100,
reward_scale=1.0,
render=False,
render_eval=False,
noise_type='adaptive-param_0.2',
normalize_returns=False,
normalize_observations=True,
critic_l2_reg=1e-2,
actor_lr=1e-4,
critic_lr=1e-3,
popart=False,
gamma=0.99,
clip_norm=None,
nb_train_steps=50, # per epoch cycle and MPI worker,
nb_eval_steps=100,
batch_size=64, # per MPI worker
tau=0.01,
eval_env=None,
param_noise_adaption_interval=50,
**network_kwargs):
set_global_seeds(seed)
if total_timesteps is not None:
assert nb_epochs is None
nb_epochs = int(total_timesteps) // (nb_epoch_cycles * nb_rollout_steps)
else:
nb_epochs = 500
if MPI is not None:
rank = MPI.COMM_WORLD.Get_rank()
else:
rank = 0
nb_actions = env.action_space.shape[-1]
assert (np.abs(env.action_space.low) == env.action_space.high).all() # we assume symmetric actions.
memory = Memory(limit=int(1e6), action_shape=env.action_space.shape, observation_shape=env.observation_space.shape)
critic = Critic(network=network, **network_kwargs)
actor = Actor(nb_actions, network=network, **network_kwargs)
action_noise = None
param_noise = None
if noise_type is not None:
for current_noise_type in noise_type.split(','):
current_noise_type = current_noise_type.strip()
if current_noise_type == 'none':
pass
elif 'adaptive-param' in current_noise_type:
_, stddev = current_noise_type.split('_')
param_noise = AdaptiveParamNoiseSpec(initial_stddev=float(stddev), desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = NormalActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
else:
raise RuntimeError('unknown noise type "{}"'.format(current_noise_type))
max_action = env.action_space.high
logger.info('scaling actions by {} before executing in env'.format(max_action))
agent = DDPG(actor, critic, memory, env.observation_space.shape, env.action_space.shape,
gamma=gamma, tau=tau, normalize_returns=normalize_returns, normalize_observations=normalize_observations,
batch_size=batch_size, action_noise=action_noise, param_noise=param_noise, critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr, critic_lr=critic_lr, enable_popart=popart, clip_norm=clip_norm,
reward_scale=reward_scale)
logger.info('Using agent with the following configuration:')
logger.info(str(agent.__dict__.items()))
eval_episode_rewards_history = deque(maxlen=100)
episode_rewards_history = deque(maxlen=100)
sess = U.get_session()
# Prepare everything.
agent.initialize(sess)
sess.graph.finalize()
agent.reset()
obs = env.reset()
if eval_env is not None:
eval_obs = eval_env.reset()
nenvs = obs.shape[0]
episode_reward = np.zeros(nenvs, dtype = np.float32) #vector
episode_step = np.zeros(nenvs, dtype = int) # vector
episodes = 0 #scalar
t = 0 # scalar
epoch = 0
start_time = time.time()
epoch_episode_rewards = []
epoch_episode_steps = []
epoch_actions = []
epoch_qs = []
epoch_episodes = 0
for epoch in range(nb_epochs):
for cycle in range(nb_epoch_cycles):
# Perform rollouts.
if nenvs > 1:
# if simulating multiple envs in parallel, impossible to reset agent at the end of the episode in each
# of the environments, so resetting here instead
agent.reset()
for t_rollout in range(nb_rollout_steps):
# Predict next action.
action, q, _, _ = agent.step(obs, apply_noise=True, compute_Q=True)
# Execute next action.
if rank == 0 and render:
env.render()
# max_action is of dimension A, whereas action is dimension (nenvs, A) - the multiplication gets broadcasted to the batch
new_obs, r, done, info = env.step(max_action * action) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
# note these outputs are batched from vecenv
t += 1
if rank == 0 and render:
env.render()
episode_reward += r
episode_step += 1
# Book-keeping.
epoch_actions.append(action)
epoch_qs.append(q)
agent.store_transition(obs, action, r, new_obs, done) #the batched data will be unrolled in memory.py's append.
obs = new_obs
for d in range(len(done)):
if done[d]:
# Episode done.
epoch_episode_rewards.append(episode_reward[d])
episode_rewards_history.append(episode_reward[d])
epoch_episode_steps.append(episode_step[d])
episode_reward[d] = 0.
episode_step[d] = 0
epoch_episodes += 1
episodes += 1
if nenvs == 1:
agent.reset()
# Train.
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
for t_train in range(nb_train_steps):
# Adapt param noise, if necessary.
if memory.nb_entries >= batch_size and t_train % param_noise_adaption_interval == 0:
distance = agent.adapt_param_noise()
epoch_adaptive_distances.append(distance)
cl, al = agent.train()
epoch_critic_losses.append(cl)
epoch_actor_losses.append(al)
agent.update_target_net()
# Evaluate.
eval_episode_rewards = []
eval_qs = []
if eval_env is not None:
nenvs_eval = eval_obs.shape[0]
eval_episode_reward = np.zeros(nenvs_eval, dtype = np.float32)
for t_rollout in range(nb_eval_steps):
eval_action, eval_q, _, _ = agent.step(eval_obs, apply_noise=False, compute_Q=True)
eval_obs, eval_r, eval_done, eval_info = eval_env.step(max_action * eval_action) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
if render_eval:
eval_env.render()
eval_episode_reward += eval_r
eval_qs.append(eval_q)
for d in range(len(eval_done)):
if eval_done[d]:
eval_episode_rewards.append(eval_episode_reward[d])
eval_episode_rewards_history.append(eval_episode_reward[d])
eval_episode_reward[d] = 0.0
if MPI is not None:
mpi_size = MPI.COMM_WORLD.Get_size()
else:
mpi_size = 1
# Log stats.
# XXX shouldn't call np.mean on variable length lists
duration = time.time() - start_time
stats = agent.get_stats()
combined_stats = stats.copy()
combined_stats['rollout/return'] = np.mean(epoch_episode_rewards)
combined_stats['rollout/return_history'] = np.mean(episode_rewards_history)
combined_stats['rollout/episode_steps'] = np.mean(epoch_episode_steps)
combined_stats['rollout/actions_mean'] = np.mean(epoch_actions)
combined_stats['rollout/Q_mean'] = np.mean(epoch_qs)
combined_stats['train/loss_actor'] = np.mean(epoch_actor_losses)
combined_stats['train/loss_critic'] = np.mean(epoch_critic_losses)
combined_stats['train/param_noise_distance'] = np.mean(epoch_adaptive_distances)
combined_stats['total/duration'] = duration
combined_stats['total/steps_per_second'] = float(t) / float(duration)
combined_stats['total/episodes'] = episodes
combined_stats['rollout/episodes'] = epoch_episodes
combined_stats['rollout/actions_std'] = np.std(epoch_actions)
# Evaluation statistics.
if eval_env is not None:
combined_stats['eval/return'] = eval_episode_rewards
combined_stats['eval/return_history'] = np.mean(eval_episode_rewards_history)
combined_stats['eval/Q'] = eval_qs
combined_stats['eval/episodes'] = len(eval_episode_rewards)
def as_scalar(x):
if isinstance(x, np.ndarray):
assert x.size == 1
return x[0]
elif np.isscalar(x):
return x
else:
raise ValueError('expected scalar, got %s'%x)
combined_stats_sums = np.array([ np.array(x).flatten()[0] for x in combined_stats.values()])
if MPI is not None:
combined_stats_sums = MPI.COMM_WORLD.allreduce(combined_stats_sums)
combined_stats = {k : v / mpi_size for (k,v) in zip(combined_stats.keys(), combined_stats_sums)}
# Total statistics.
combined_stats['total/epochs'] = epoch + 1
combined_stats['total/steps'] = t
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
if rank == 0:
logger.dump_tabular()
logger.info('')
logdir = logger.get_dir()
if rank == 0 and logdir:
if hasattr(env, 'get_state'):
with open(os.path.join(logdir, 'env_state.pkl'), 'wb') as f:
pickle.dump(env.get_state(), f)
if eval_env and hasattr(eval_env, 'get_state'):
with open(os.path.join(logdir, 'eval_env_state.pkl'), 'wb') as f:
pickle.dump(eval_env.get_state(), f)
return agent
def learn(*,
network,
env,
env_name,
total_timesteps,
eval_env=None,
seed=None,
nsteps=2048,
ent_coef=0.0,
lr=3e-4,
vf_coef=0.5,
max_grad_norm=0.5,
gamma=0.99,
lam=0.95,
log_interval=10,
nminibatches=4,
noptepochs=4,
cliprange=0.2,
save_interval=0,
load_path=None,
model_fn=None,
update_fn=None,
init_fn=None,
mpi_rank_weight=1,
comm=None,
data_aug='normal',
**network_kwargs):
'''
Learn policy using PPO algorithm (https://arxiv.org/abs/1707.06347)
Parameters:
----------
network: policy network architecture. Either string (mlp, lstm, lnlstm, cnn_lstm, cnn, cnn_small, conv_only - see baselines.common/models.py for full list)
specifying the standard network architecture, or a function that takes tensorflow tensor as input and returns
tuple (output_tensor, extra_feed) where output tensor is the last network layer output, extra_feed is None for feed-forward
neural nets, and extra_feed is a dictionary describing how to feed state into the network for recurrent neural nets.
See common/models.py/lstm for more details on using recurrent nets in policies
env: baselines.common.vec_env.VecEnv environment. Needs to be vectorized for parallel environment simulation.
The environments produced by gym.make can be wrapped using baselines.common.vec_env.DummyVecEnv class.
nsteps: int number of steps of the vectorized environment per update (i.e. batch size is nsteps * nenv where
nenv is number of environment copies simulated in parallel)
total_timesteps: int number of timesteps (i.e. number of actions taken in the environment)
ent_coef: float policy entropy coefficient in the optimization objective
lr: float or function learning rate, constant or a schedule function [0,1] -> R+ where 1 is beginning of the
training and 0 is the end of the training.
vf_coef: float value function loss coefficient in the optimization objective
max_grad_norm: float or None gradient norm clipping coefficient
gamma: float discounting factor
lam: float advantage estimation discounting factor (lambda in the paper)
log_interval: int number of timesteps between logging events
nminibatches: int number of training minibatches per update. For recurrent policies,
should be smaller or equal than number of environments run in parallel.
noptepochs: int number of training epochs per update
cliprange: float or function clipping range, constant or schedule function [0,1] -> R+ where 1 is beginning of the training
and 0 is the end of the training
save_interval: int number of timesteps between saving events
load_path: str path to load the model from
**network_kwargs: keyword arguments to the policy / network builder. See baselines.common/policies.py/build_policy and arguments to a particular type of network
For instance, 'mlp' network architecture has arguments num_hidden and num_layers.
'''
set_global_seeds(seed)
if isinstance(lr, float): lr = constfn(lr)
else: assert callable(lr)
if isinstance(cliprange, float): cliprange = constfn(cliprange)
else: assert callable(cliprange)
total_timesteps = int(total_timesteps)
policy = build_policy(env, network, data_aug=data_aug, **network_kwargs)
# Get the nb of env
nenvs = env.num_envs
# Get state_space and action_space
ob_space = env.observation_space
ac_space = env.action_space
# Calculate the batch_size
nbatch = nenvs * nsteps
nbatch_train = nbatch // nminibatches
is_mpi_root = (MPI is None or MPI.COMM_WORLD.Get_rank() == 0)
# Instantiate the model object (that creates act_model and train_model)
if model_fn is None:
from baselines.ppo2.model import Model
model_fn = Model
model = model_fn(policy=policy,
ob_space=ob_space,
ac_space=ac_space,
nbatch_act=nenvs,
nbatch_train=nbatch_train,
nsteps=nsteps,
ent_coef=ent_coef,
vf_coef=vf_coef,
max_grad_norm=max_grad_norm,
comm=comm,
mpi_rank_weight=mpi_rank_weight)
if load_path is not None:
model.load(load_path)
# Instantiate the runner object
runner = Runner(env=env,
model=model,
nsteps=nsteps,
gamma=gamma,
lam=lam,
data_aug=data_aug)
if eval_env is not None:
eval_runner = Runner(env=eval_env,
model=model,
nsteps=nsteps,
gamma=gamma,
lam=lam,
data_aug=data_aug,
eval_flag=True)
epinfobuf = deque(maxlen=100)
if eval_env is not None:
eval_epinfobuf = deque(maxlen=100)
if init_fn is not None:
init_fn()
# Start total timer
tfirststart = time.perf_counter()
# random filter
sess = tf.get_default_session()
rand_processes = [v for v in tf.global_variables() if 'randcnn' in v.name]
if len(rand_processes) > 0:
init_process = tf.variables_initializer(rand_processes)
else:
init_process = None
nupdates = total_timesteps // nbatch
traj_dict = {'obs': [], 'acts': [], 'rews': [], 'dones': [], 'infos': []}
traj_count = 0
for update in range(1, nupdates + 1):
assert nbatch % nminibatches == 0
# Start timer
tstart = time.perf_counter()
frac = 1.0 - (update - 1.0) / nupdates
# Calculate the learning rate
lrnow = lr(frac)
# Calculate the cliprange
cliprangenow = cliprange(frac)
if update % log_interval == 0 and is_mpi_root:
logger.info('Stepping environment...')
# randomize parameters
if len(rand_processes) > 0:
sess.run(init_process)
logger.info('Randomizing parameters...')
# Get minibatch
obs, returns, masks, actions, values, neglogpacs, states, epinfos = runner.run(
) #pylint: disable=E0632
traj_dict['obs'].append(obs)
traj_dict['acts'].append(actions)
traj_dict['rews'].append(returns)
traj_dict['dones'].append(masks)
traj_dict['infos'].append(epinfos)
if eval_env is not None:
eval_obs, eval_returns, eval_masks, eval_actions, eval_values, eval_neglogpacs, eval_states, eval_epinfos = eval_runner.run(
) #pylint: disable=E0632
if update % log_interval == 0 and is_mpi_root: logger.info('Done.')
epinfobuf.extend(epinfos)
if eval_env is not None:
eval_epinfobuf.extend(eval_epinfos)
# Here what we're going to do is for each minibatch calculate the loss and append it.
mblossvals = []
if states is None: # nonrecurrent version
# Index of each element of batch_size
# Create the indices array
inds = np.arange(nbatch)
for _ in range(noptepochs):
# Randomize the indexes
np.random.shuffle(inds)
# 0 to batch_size with batch_train_size step
for start in range(0, nbatch, nbatch_train):
end = start + nbatch_train
mbinds = inds[start:end]
slices = (arr[mbinds]
for arr in (obs, returns, masks, actions, values,
neglogpacs))
if load_path is not None:
mblossvals.append(
model.train(lrnow, cliprangenow, *slices))
else: # recurrent version
assert nenvs % nminibatches == 0
envsperbatch = nenvs // nminibatches
envinds = np.arange(nenvs)
flatinds = np.arange(nenvs * nsteps).reshape(nenvs, nsteps)
for _ in range(noptepochs):
np.random.shuffle(envinds)
for start in range(0, nenvs, envsperbatch):
end = start + envsperbatch
mbenvinds = envinds[start:end]
mbflatinds = flatinds[mbenvinds].ravel()
slices = (arr[mbflatinds]
for arr in (obs, returns, masks, actions, values,
neglogpacs))
mbstates = states[mbenvinds]
if load_path is not None:
mblossvals.append(
model.train(lrnow, cliprangenow, *slices,
mbstates))
# Feedforward --> get losses --> update
lossvals = np.mean(mblossvals, axis=0)
# End timer
tnow = time.perf_counter()
# Calculate the fps (frame per second)
fps = int(nbatch / (tnow - tstart))
if update_fn is not None:
update_fn(update)
if update % log_interval == 0 or update == 1:
# Calculates if value function is a good predicator of the returns (ev > 1)
# or if it's just worse than predicting nothing (ev =< 0)
ev = explained_variance(values, returns)
logger.logkv("misc/serial_timesteps", update * nsteps)
logger.logkv("misc/nupdates", update)
logger.logkv("misc/total_timesteps", update * nbatch)
logger.logkv("fps", fps)
logger.logkv("misc/explained_variance", float(ev))
logger.logkv('eprewmean',
safemean([epinfo['r'] for epinfo in epinfobuf]))
logger.logkv('eplenmean',
safemean([epinfo['l'] for epinfo in epinfobuf]))
if eval_env is not None:
logger.logkv(
'eval_eprewmean',
safemean([epinfo['r'] for epinfo in eval_epinfobuf]))
logger.logkv(
'eval_eplenmean',
safemean([epinfo['l'] for epinfo in eval_epinfobuf]))
logger.logkv('misc/time_elapsed', tnow - tfirststart)
for (lossval, lossname) in zip(lossvals, model.loss_names):
logger.logkv('loss/' + lossname, lossval)
logger.dumpkvs()
if save_interval and (update % save_interval == 0 or update
== 1) and logger.get_dir() and is_mpi_root:
if load_path is not None:
video_out_path = osp.join(
logger.get_dir(),
f'video_{env_name}_traj_{traj_count}.mp4')
video_writer = TensorFrameWriter(video_out_path,
make_grid=False,
adjust_axis=False)
obs = obs
for image, done in zip(obs, masks):
if not done:
video_writer.add_tensor(image)
else:
video_writer.close()
breakpoint()
traj_count += 1
video_out_path = osp.join(
logger.get_dir(),
f'video_{env_name}_traj_{traj_count}.mp4')
video_writer = TensorFrameWriter(video_out_path,
make_grid=False,
adjust_axis=False)
else:
checkdir = osp.join(logger.get_dir(), 'checkpoints')
os.makedirs(checkdir, exist_ok=True)
savepath = osp.join(checkdir, '%.5i' % update)
print('Saving to', savepath)
model.save(savepath)
break
all_dones = traj_dict['dones'].copy()
all_dones = np.concatenate(all_dones, axis=0)
traj_ends, = np.nonzero(all_dones)
real_n_traj = len(traj_ends)
trajectories = []
traj_starts = np.concatenate(([0], traj_ends[:-1]), axis=0)
traj_start_end = np.stack((traj_starts, traj_ends), axis=1)
for start_idx, end_idx in traj_start_end:
trajectories.append(
{k: v[start_idx:end_idx]
for k, v in traj_dict.items()})
n_traj = len(trajectories)
import pickle
traj_filename = osp.join(logger.get_dir(), f'demo_{env_name}.pickle')
with open(traj_filename, 'wb') as handle:
pickle.dump(traj_dict, handle, protocol=pickle.HIGHEST_PROTOCOL)
video_writer.close()
return model
def __init__(
self,
*,
scope,
ob_space,
ac_space,
stochpol_fn,
nsteps,
nepochs=4,
nminibatches=1,
gamma=0.99,
gamma_ext=0.99,
lam=0.95,
ent_coef=0,
cliprange=0.2,
max_grad_norm=1.0,
vf_coef=1.0,
lr=30e-5,
adam_hps=None,
testing=False,
comm=None,
comm_train=None,
use_news=False,
update_ob_stats_every_step=True,
int_coeff=None,
ext_coeff=None,
):
self.lr = lr
self.ext_coeff = ext_coeff
self.int_coeff = int_coeff
self.use_news = use_news
self.update_ob_stats_every_step = update_ob_stats_every_step
self.abs_scope = (tf.get_variable_scope().name + '/' +
scope).lstrip('/')
self.testing = testing
self.comm_log = MPI.COMM_SELF
if comm is not None and comm.Get_size() > 1:
self.comm_log = comm
assert not testing or comm.Get_rank(
) != 0, "Worker number zero can't be testing"
if comm_train is not None:
self.comm_train, self.comm_train_size = comm_train, comm_train.Get_size(
)
else:
self.comm_train, self.comm_train_size = self.comm_log, self.comm_log.Get_size(
)
self.is_log_leader = self.comm_log.Get_rank() == 0
self.is_train_leader = self.comm_train.Get_rank() == 0
with tf.variable_scope(scope):
self.best_ret = -np.inf
self.local_best_ret = -np.inf
self.rooms = []
self.local_rooms = []
self.scores = []
self.ob_space = ob_space
self.ac_space = ac_space
self.stochpol = stochpol_fn()
self.nepochs = nepochs
self.cliprange = cliprange
self.nsteps = nsteps
self.nminibatches = nminibatches
self.gamma = gamma
self.gamma_ext = gamma_ext
self.lam = lam
self.adam_hps = adam_hps or dict()
self.ph_adv = tf.placeholder(tf.float32, [None, None])
self.ph_ret_int = tf.placeholder(tf.float32, [None, None])
self.ph_ret_ext = tf.placeholder(tf.float32, [None, None])
self.ph_oldnlp = tf.placeholder(tf.float32, [None, None])
self.ph_oldvpred = tf.placeholder(tf.float32, [None, None])
self.ph_lr = tf.placeholder(tf.float32, [])
self.ph_lr_pred = tf.placeholder(tf.float32, [])
self.ph_cliprange = tf.placeholder(tf.float32, [])
#Define loss.
neglogpac = self.stochpol.pd_opt.neglogp(self.stochpol.ph_ac)
entropy = tf.reduce_mean(self.stochpol.pd_opt.entropy())
vf_loss_int = (0.5 * vf_coef) * tf.reduce_mean(
tf.square(self.stochpol.vpred_int_opt - self.ph_ret_int))
vf_loss_ext = (0.5 * vf_coef) * tf.reduce_mean(
tf.square(self.stochpol.vpred_ext_opt - self.ph_ret_ext))
vf_loss = vf_loss_int + vf_loss_ext
ratio = tf.exp(self.ph_oldnlp - neglogpac) # p_new / p_old
negadv = -self.ph_adv
pg_losses1 = negadv * ratio
pg_losses2 = negadv * tf.clip_by_value(
ratio, 1.0 - self.ph_cliprange, 1.0 + self.ph_cliprange)
pg_loss = tf.reduce_mean(tf.maximum(pg_losses1, pg_losses2))
ent_loss = (-ent_coef) * entropy
approxkl = .5 * tf.reduce_mean(
tf.square(neglogpac - self.ph_oldnlp))
maxkl = .5 * tf.reduce_max(tf.square(neglogpac - self.ph_oldnlp))
clipfrac = tf.reduce_mean(
tf.to_float(tf.greater(tf.abs(ratio - 1.0),
self.ph_cliprange)))
loss = pg_loss + ent_loss + vf_loss + self.stochpol.aux_loss
#Create optimizer.
params = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope=self.abs_scope)
logger.info("PPO: using MpiAdamOptimizer connected to %i peers" %
self.comm_train_size)
trainer = MpiAdamOptimizer(self.comm_train,
learning_rate=self.ph_lr,
**self.adam_hps)
grads_and_vars = trainer.compute_gradients(loss, params)
grads, vars = zip(*grads_and_vars)
if max_grad_norm:
_, _grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
global_grad_norm = tf.global_norm(grads)
grads_and_vars = list(zip(grads, vars))
self._train = trainer.apply_gradients(grads_and_vars)
#Quantities for reporting.
self._losses = [
loss, pg_loss, vf_loss, entropy, clipfrac, approxkl, maxkl,
self.stochpol.aux_loss, self.stochpol.feat_var,
self.stochpol.max_feat, global_grad_norm
]
self.loss_names = [
'tot', 'pg', 'vf', 'ent', 'clipfrac', 'approxkl', 'maxkl',
"auxloss", "featvar", "maxfeat", "gradnorm"
]
self.I = None
self.disable_policy_update = None
allvars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=self.abs_scope)
if self.is_log_leader:
tf_util.display_var_info(allvars)
tf.get_default_session().run(tf.variables_initializer(allvars))
sync_from_root(tf.get_default_session(),
allvars) #Syncs initialization across mpi workers.
self.t0 = time.time()
self.global_tcount = 0
def save(training_step):
logger.info(
'Saved network with {} training steps'.format(training_step))
self.saver.save(self.sess,
self.cpk_dir + 'demo',
global_step=training_step)
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = NormalActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
else:
raise RuntimeError('unknown noise type "{}"'.format(current_noise_type))
# Configure components of DDPG
memory = Memory(limit=int(1e6), action_shape=env.action_space.shape, observation_shape=env.observation_space.shape)
critic = Critic(layer_norm=args.layer_norm)
actor = Actor(nb_actions, layer_norm=args.layer_norm)
# Seed everything to make things reproducible.
seed = args.seed + 1000000 * rank
logger.info('rank {}: seed={}, logdir={}'.format(rank, seed, logger.get_dir()))
# tf.reset_default_graph()
set_global_seeds(seed)
env.seed(seed)
if eval_env is not None:
eval_env.seed(seed)
# Disable logging for rank != 0 to avoid noise.
if rank == 0:
start_time = time.time()
# Derive the different numbers for the training process
num_timesteps = args.num_timesteps
nb_rollout_steps = args.nb_rollout_steps
nb_epoch_cycles = args.nb_epoch_cycles
nb_epochs = num_timesteps//(nb_rollout_steps*nb_epoch_cycles)
def train(env,
nb_epochs,
nb_epoch_cycles,
render_eval,
reward_scale,
render,
param_noise,
actor,
critic,
normalize_returns,
normalize_observations,
critic_l2_reg,
actor_lr,
critic_lr,
action_noise,
popart,
gamma,
clip_norm,
nb_train_steps,
nb_rollout_steps,
nb_eval_steps,
batch_size,
memory,
tau=0.01,
eval_env=None,
param_noise_adaption_interval=50):
rank = MPI.COMM_WORLD.Get_rank()
assert (np.abs(env.action_space.low) == env.action_space.high
).all() # we assume symmetric actions.
max_action = env.action_space.high
logger.info(
'scaling actions by {} before executing in env'.format(max_action))
agent = DDPG(actor,
critic,
memory,
env.observation_space.shape,
env.action_space.shape,
gamma=gamma,
tau=tau,
normalize_returns=normalize_returns,
normalize_observations=normalize_observations,
batch_size=batch_size,
action_noise=action_noise,
param_noise=param_noise,
critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr,
critic_lr=critic_lr,
enable_popart=popart,
clip_norm=clip_norm,
reward_scale=reward_scale)
logger.info('Using agent with the following configuration:')
logger.info(str(agent.__dict__.items()))
# Set up logging stuff only for a single worker.
if rank == 0:
saver = tf.train.Saver()
else:
saver = None
step = 0
episode = 0
eval_episode_rewards_history = deque(maxlen=100)
episode_rewards_history = deque(maxlen=100)
tf_config = tf.ConfigProto(inter_op_parallelism_threads=1,
intra_op_parallelism_threads=1)
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
# Prepare everything.
agent.initialize(sess)
sess.graph.finalize()
agent.reset()
obs = env.reset()
if eval_env is not None:
eval_obs = eval_env.reset()
done = False
episode_reward = 0.
episode_step = 0
episodes = 0
t = 0
epoch = 0
start_time = time.time()
epoch_episode_rewards = []
epoch_episode_steps = []
epoch_episode_eval_rewards = []
epoch_episode_eval_steps = []
epoch_start_time = time.time()
epoch_actions = []
epoch_qs = []
epoch_episodes = 0
for epoch in range(nb_epochs):
for cycle in range(nb_epoch_cycles):
# Perform rollouts.
for t_rollout in range(nb_rollout_steps):
# Predict next action.
action, q = agent.pi(obs, apply_noise=True, compute_Q=True)
assert action.shape == env.action_space.shape
# Execute next action.
if rank == 0 and render:
env.render()
assert max_action.shape == action.shape
new_obs, r, done, info = env.step(
max_action * action
) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
t += 1
if rank == 0 and render:
env.render()
episode_reward += r
episode_step += 1
# Book-keeping.
epoch_actions.append(action)
epoch_qs.append(q)
agent.store_transition(obs, action, r, new_obs, done)
obs = new_obs
if done:
# Episode done.
epoch_episode_rewards.append(episode_reward)
episode_rewards_history.append(episode_reward)
epoch_episode_steps.append(episode_step)
episode_reward = 0.
episode_step = 0
epoch_episodes += 1
episodes += 1
agent.reset()
obs = env.reset()
# Train.
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
for t_train in range(nb_train_steps):
# Adapt param noise, if necessary.
if memory.nb_entries >= batch_size and t % param_noise_adaption_interval == 0:
distance = agent.adapt_param_noise()
epoch_adaptive_distances.append(distance)
cl, al = agent.train()
epoch_critic_losses.append(cl)
epoch_actor_losses.append(al)
agent.update_target_net()
# Evaluate.
eval_episode_rewards = []
eval_qs = []
if eval_env is not None:
eval_episode_reward = 0.
for t_rollout in range(nb_eval_steps):
eval_action, eval_q = agent.pi(eval_obs,
apply_noise=False,
compute_Q=True)
eval_obs, eval_r, eval_done, eval_info = eval_env.step(
max_action * eval_action
) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
if render_eval:
eval_env.render()
eval_episode_reward += eval_r
eval_qs.append(eval_q)
if eval_done:
eval_obs = eval_env.reset()
eval_episode_rewards.append(eval_episode_reward)
eval_episode_rewards_history.append(
eval_episode_reward)
eval_episode_reward = 0.
# Log stats.
epoch_train_duration = time.time() - epoch_start_time
duration = time.time() - start_time
stats = agent.get_stats()
combined_stats = {}
for key in sorted(stats.keys()):
combined_stats[key] = mpi_mean(stats[key])
# Rollout statistics.
combined_stats['rollout/return'] = mpi_mean(epoch_episode_rewards)
combined_stats['rollout/return_history'] = mpi_mean(
np.mean(episode_rewards_history))
combined_stats['rollout/episode_steps'] = mpi_mean(
epoch_episode_steps)
combined_stats['rollout/episodes'] = mpi_sum(epoch_episodes)
combined_stats['rollout/actions_mean'] = mpi_mean(epoch_actions)
combined_stats['rollout/actions_std'] = mpi_std(epoch_actions)
combined_stats['rollout/Q_mean'] = mpi_mean(epoch_qs)
# Train statistics.
combined_stats['train/loss_actor'] = mpi_mean(epoch_actor_losses)
combined_stats['train/loss_critic'] = mpi_mean(epoch_critic_losses)
combined_stats['train/param_noise_distance'] = mpi_mean(
epoch_adaptive_distances)
# Evaluation statistics.
if eval_env is not None:
combined_stats['eval/return'] = mpi_mean(eval_episode_rewards)
combined_stats['eval/return_history'] = mpi_mean(
np.mean(eval_episode_rewards_history))
combined_stats['eval/Q'] = mpi_mean(eval_qs)
combined_stats['eval/episodes'] = mpi_mean(
len(eval_episode_rewards))
# Total statistics.
combined_stats['total/duration'] = mpi_mean(duration)
combined_stats['total/steps_per_second'] = mpi_mean(
float(t) / float(duration))
combined_stats['total/episodes'] = mpi_mean(episodes)
combined_stats['total/epochs'] = epoch + 1
combined_stats['total/steps'] = t
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
logger.dump_tabular()
logger.info('')
logdir = logger.get_dir()
if rank == 0 and logdir:
if hasattr(env, 'get_state'):
with open(os.path.join(logdir, 'env_state.pkl'),
'wb') as f:
pickle.dump(env.get_state(), f)
if eval_env and hasattr(eval_env, 'get_state'):
with open(os.path.join(logdir, 'eval_env_state.pkl'),
'wb') as f:
pickle.dump(eval_env.get_state(), f)
def log_params(params, logger=logger):
for key in sorted(params.keys()):
logger.info('{}: {}'.format(key, params[key]))
def run(seed, noise_type, layer_norm, **kwargs):
"""Configure things."""
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0: logger.set_level(logger.DISABLED)
"""Create Simulation envs."""
env = PegintoHoles()
"""Create True envs"""
# env = Env_robot_control()
"""Parse noise_type"""
action_noise = None
param_noise = None
nb_actions = env.action_dim
for current_noise_type in noise_type.split(','):
current_noise_type = current_noise_type.strip()
if current_noise_type == 'none':
pass
elif 'adaptive-param' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(
mu=np.zeros(nb_actions),
sigma=float(0.2) * np.ones(nb_actions))
param_noise = AdaptiveParamNoiseSpec(
initial_stddev=float(stddev),
desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = NormalActionNoise(mu=np.zeros(nb_actions),
sigma=float(stddev) *
np.ones(nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(
mu=np.zeros(nb_actions),
sigma=float(stddev) * np.ones(nb_actions))
else:
raise RuntimeError(
'unknown noise type "{}"'.format(current_noise_type))
"""Configure components."""
memory = Memory(limit=int(1e6),
action_shape=env.action_dim,
observation_shape=env.state_dim)
critic = Critic(layer_norm=layer_norm)
actor = Actor(nb_actions, layer_norm=layer_norm)
"""Seed everything to make things reproducible."""
seed = seed + 1000000 * rank
logger.info('rank {}: seed={}, logdir={}'.format(rank, seed,
logger.get_dir()))
tf.reset_default_graph()
set_global_seeds(seed)
"""Disable logging to avoid noise."""
start_time = time.time()
"""Train the model"""
training.train(env=env,
param_noise=param_noise,
action_noise=action_noise,
actor=actor,
critic=critic,
memory=memory,
**kwargs)
"""Eval the result"""
logger.info('total runtime: {}s'.format(time.time() - start_time))
def learn(*,
network,
env,
total_timesteps,
seed=None,
eval_env=None,
replay_strategy='future',
policy_save_interval=5,
clip_return=True,
demo_file=None,
override_params=None,
load_path=None,
save_path=None,
**kwargs):
override_params = override_params or {}
if MPI is not None:
rank = MPI.COMM_WORLD.Get_rank()
num_cpu = MPI.COMM_WORLD.Get_size()
# Seed everything.
rank_seed = seed + 1000000 * rank if seed is not None else None
set_global_seeds(rank_seed)
# Prepare params.
params = config.DEFAULT_PARAMS
env_name = env.spec.id
params['env_name'] = env_name
params['replay_strategy'] = replay_strategy
if env_name in config.DEFAULT_ENV_PARAMS:
params.update(config.DEFAULT_ENV_PARAMS[env_name]
) # merge env-specific parameters in
params.update(
**override_params) # makes it possible to override any parameter
with open(os.path.join(logger.get_dir(), 'params.json'), 'w') as f:
json.dump(params, f)
params = config.prepare_params(params)
params['rollout_batch_size'] = env.num_envs
if demo_file is not None:
params['bc_loss'] = 1
params.update(kwargs)
config.log_params(params, logger=logger)
if num_cpu == 1:
logger.warn()
logger.warn('*** Warning ***')
logger.warn(
'You are running HER with just a single MPI worker. This will work, but the '
+
'experiments that we report in Plappert et al. (2018, https://arxiv.org/abs/1802.09464) '
+
'were obtained with --num_cpu 19. This makes a significant difference and if you '
+
'are looking to reproduce those results, be aware of this. Please also refer to '
+
'https://github.com/openai/baselines/issues/314 for further details.'
)
logger.warn('****************')
logger.warn()
dims = config.configure_dims(params)
policy = config.configure_ddpg(dims=dims,
params=params,
clip_return=clip_return)
if load_path is not None:
tf_util.load_variables(load_path)
rollout_params = {
'exploit': False,
'use_target_net': False,
'use_demo_states': True,
'compute_Q': False,
'T': params['T'],
}
eval_params = {
'exploit': True,
'use_target_net': params['test_with_polyak'],
'use_demo_states': False,
'compute_Q': True,
'T': params['T'],
}
for name in [
'T', 'rollout_batch_size', 'gamma', 'noise_eps', 'random_eps'
]:
rollout_params[name] = params[name]
eval_params[name] = params[name]
eval_env = eval_env or env
rollout_worker = RolloutWorker(env,
policy,
dims,
logger,
monitor=True,
**rollout_params)
evaluator = RolloutWorker(eval_env, policy, dims, logger, **eval_params)
n_cycles = params['n_cycles']
n_epochs = total_timesteps // n_cycles // rollout_worker.T // rollout_worker.rollout_batch_size
logger.info("n_epochs:", n_epochs)
return train(save_path=save_path,
policy=policy,
rollout_worker=rollout_worker,
evaluator=evaluator,
n_epochs=n_epochs,
n_test_rollouts=params['n_test_rollouts'],
n_cycles=params['n_cycles'],
n_batches=params['n_batches'],
policy_save_interval=policy_save_interval,
demo_file=demo_file)
def define_self_prediction_rew(self, convfeat, rep_size, enlargement):
logger.info(
"Using RND BONUS ****************************************************"
)
# RND bonus.
# Random target network.
for ph in self.ph_ob.values():
if len(ph.shape.as_list()) == 5: # B,T,H,W,C
logger.info(
f"CnnTarget: using '{ph.name}' shape {ph.shape} as image input"
)
xr = ph[:, 1:]
xr = tf.cast(xr, tf.float32)
xr = tf.reshape(xr, (-1, *ph.shape.as_list()[-3:]))[:, :, :,
-1:]
xr = tf.clip_by_value((xr - self.ph_mean) / self.ph_std, -5.0,
5.0)
xr = tf.nn.leaky_relu(
conv(
xr,
"c1r",
nf=convfeat * 1,
rf=8,
stride=4,
init_scale=np.sqrt(2),
))
xr = tf.nn.leaky_relu(
conv(
xr,
"c2r",
nf=convfeat * 2 * 1,
rf=4,
stride=2,
init_scale=np.sqrt(2),
))
xr = tf.nn.leaky_relu(
conv(
xr,
"c3r",
nf=convfeat * 2 * 1,
rf=3,
stride=1,
init_scale=np.sqrt(2),
))
rgbr = [to2d(xr)]
X_r = fc(rgbr[0], "fc1r", nh=rep_size, init_scale=np.sqrt(2))
# Predictor network.
for ph in self.ph_ob.values():
if len(ph.shape.as_list()) == 5: # B,T,H,W,C
logger.info(
f"CnnTarget: using '{ph.name}' shape {ph.shape} as image input"
)
xrp = ph[:, 1:]
xrp = tf.cast(xrp, tf.float32)
xrp = tf.reshape(xrp, (-1, *ph.shape.as_list()[-3:]))[:, :, :,
-1:]
xrp = tf.clip_by_value((xrp - self.ph_mean) / self.ph_std,
-5.0, 5.0)
xrp = tf.nn.leaky_relu(
conv(
xrp,
"c1rp_pred",
nf=convfeat,
rf=8,
stride=4,
init_scale=np.sqrt(2),
))
xrp = tf.nn.leaky_relu(
conv(
xrp,
"c2rp_pred",
nf=convfeat * 2,
rf=4,
stride=2,
init_scale=np.sqrt(2),
))
xrp = tf.nn.leaky_relu(
conv(
xrp,
"c3rp_pred",
nf=convfeat * 2,
rf=3,
stride=1,
init_scale=np.sqrt(2),
))
rgbrp = to2d(xrp)
X_r_hat = tf.nn.relu(
fc(
rgbrp,
"fc1r_hat1_pred",
nh=256 * enlargement,
init_scale=np.sqrt(2),
))
X_r_hat = tf.nn.relu(
fc(
X_r_hat,
"fc1r_hat2_pred",
nh=256 * enlargement,
init_scale=np.sqrt(2),
))
X_r_hat = fc(X_r_hat,
"fc1r_hat3_pred",
nh=rep_size,
init_scale=np.sqrt(2))
self.feat_var = tf.reduce_mean(tf.nn.moments(X_r, axes=[0])[1])
self.max_feat = tf.reduce_max(tf.abs(X_r))
self.int_rew = tf.reduce_mean(
tf.square(tf.stop_gradient(X_r) - X_r_hat),
axis=-1,
keep_dims=True)
self.int_rew = tf.reshape(self.int_rew,
(self.sy_nenvs, self.sy_nsteps - 1))
targets = tf.stop_gradient(X_r)
self.aux_loss = tf.reduce_mean(tf.square(targets - X_r_hat), -1)
mask = tf.random_uniform(shape=tf.shape(self.aux_loss),
minval=0.0,
maxval=1.0,
dtype=tf.float32)
mask = tf.cast(mask < self.proportion_of_exp_used_for_predictor_update,
tf.float32)
self.aux_loss = tf.reduce_sum(mask * self.aux_loss) / tf.maximum(
tf.reduce_sum(mask), 1.0)
def __init__(self, *, sess, policy, ob_space, ac_space, nbatch_act, nbatch_train,
nsteps, ent_coef, vf_coef, max_grad_norm):
# sess = tf.get_default_session()
train_model = policy(sess, ob_space, ac_space, nbatch_train, nsteps)
norm_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
act_model = policy(sess, ob_space, ac_space, nbatch_act, 1)
A = train_model.pdtype.sample_placeholder([None])
ADV = tf.placeholder(tf.float32, [None])
R = tf.placeholder(tf.float32, [None])
OLDNEGLOGPAC = tf.placeholder(tf.float32, [None])
OLDVPRED = tf.placeholder(tf.float32, [None])
LR = tf.placeholder(tf.float32, [])
CLIPRANGE = tf.placeholder(tf.float32, [])
neglogpac = train_model.pd.neglogp(A)
entropy = tf.reduce_mean(train_model.pd.entropy())
vpred = train_model.vf
vpredclipped = OLDVPRED + tf.clip_by_value(train_model.vf - OLDVPRED, - CLIPRANGE, CLIPRANGE)
vf_losses1 = tf.square(vpred - R)
vf_losses2 = tf.square(vpredclipped - R)
vf_loss = .5 * tf.reduce_mean(tf.maximum(vf_losses1, vf_losses2))
ratio = tf.exp(OLDNEGLOGPAC - neglogpac)
pg_losses = -ADV * ratio
pg_losses2 = -ADV * tf.clip_by_value(ratio, 1.0 - CLIPRANGE, 1.0 + CLIPRANGE)
pg_loss = tf.reduce_mean(tf.maximum(pg_losses, pg_losses2))
approxkl = .5 * tf.reduce_mean(tf.square(neglogpac - OLDNEGLOGPAC))
clipfrac = tf.reduce_mean(tf.to_float(tf.greater(tf.abs(ratio - 1.0), CLIPRANGE)))
params = tf.trainable_variables()
weight_params = [v for v in params if '/b' not in v.name]
total_num_params = 0
for p in params:
shape = p.get_shape().as_list()
num_params = np.prod(shape)
# mpi_print('param', p, num_params)
total_num_params += num_params
logger.info('total num params:', total_num_params)
l2_loss = tf.reduce_sum([tf.nn.l2_loss(v) for v in weight_params])
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef + l2_loss * L2_WEIGHT
trainer = tf.train.AdamOptimizer(learning_rate=LR, epsilon=1e-5)
grads_and_var = trainer.compute_gradients(loss, params)
grads, var = zip(*grads_and_var)
if max_grad_norm is not None:
grads, _grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads_and_var = list(zip(grads, var))
_train = trainer.apply_gradients(grads_and_var)
def train(lr, cliprange, obs, returns, masks, actions, values, neglogpacs, states=None):
advs = returns - values
adv_mean = np.mean(advs, axis=0, keepdims=True)
adv_std = np.std(advs, axis=0, keepdims=True)
advs = (advs - adv_mean) / (adv_std + 1e-8)
td_map = {train_model.X:obs, A:actions, ADV:advs, R:returns, LR:lr,
CLIPRANGE:cliprange, OLDNEGLOGPAC:neglogpacs, OLDVPRED:values}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
return sess.run(
[pg_loss, vf_loss, entropy, approxkl, clipfrac, l2_loss, _train],
td_map
)[:-1]
self.loss_names = ['policy_loss', 'value_loss', 'policy_entropy', 'approxkl', 'clipfrac', 'l2_loss']
def save(save_path):
ps = sess.run(params)
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
sess.run(restores)
self.train = train
self.train_model = train_model
self.act_model = act_model
self.step = act_model.step
self.value = act_model.value
self.initial_state = act_model.initial_state
self.save = save
self.load = load
initialize()
def save_model(self, num_steps):
logger.info("Saving model at step: {}".format(num_steps))
tf_util.save_state(self.model_path)
def learn(*, network, sess, env, nsteps, total_timesteps, ent_coef, lr,
vf_coef=0.5, max_grad_norm=0.5, gamma=0.99, lam=0.95,
log_interval=10, nminibatches=4, noptepochs=4, cliprange=0.2,
save_interval=0, save_path=None, load_path=None, **network_kwargs):
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
mpi_size = comm.Get_size()
#sess = tf.get_default_session()
# tb_writer = TB_Writer(sess)
if isinstance(lr, float): lr = constfn(lr)
else: assert callable(lr)
if isinstance(cliprange, float): cliprange = constfn(cliprange)
else: assert callable(cliprange)
total_timesteps = int(total_timesteps)
nenvs = env.num_envs
ob_space = env.observation_space
ac_space = env.action_space
nbatch = nenvs * nsteps
nbatch_train = nbatch // nminibatches
policy = RandCropCnnPolicy
model = Model(policy=policy, sess=sess, ob_space=ob_space, ac_space=ac_space,
nbatch_act=nenvs, nbatch_train=nbatch_train,
nsteps=nsteps, ent_coef=ent_coef, vf_coef=vf_coef,
max_grad_norm=max_grad_norm)
# utils.load_all_params(sess)
if load_path is not None:
model.load(load_path)
logger.info("Model pramas loaded from save")
runner = Runner(env=env, model=model, nsteps=nsteps, gamma=gamma, lam=lam)
logger.info("Initilizing runner")
epinfobuf10 = deque(maxlen=10)
epinfobuf100 = deque(maxlen=100)
tfirststart = time.time()
active_ep_buf = epinfobuf100
nupdates = total_timesteps//nbatch
logger.info("Running {} updates, each needs {} batches".format(nupdates, nbatch))
mean_rewards = []
datapoints = []
run_t_total = 0
train_t_total = 0
can_save = True
checkpoints = list(range(0,2049,10))
saved_key_checkpoints = [False] * len(checkpoints)
#init_rand = tf.variables_initializer([v for v in tf.global_variables() if 'randcnn' in v.name])
# if Config.SYNC_FROM_ROOT and rank != 0:
# can_save = False
# def save_model(base_name=None):
# base_dict = {'datapoints': datapoints}
# utils.save_params_in_scopes(
# sess, ['model'], Config.get_save_file(base_name=base_name), base_dict)
for update in range(1, nupdates+1):
assert nbatch % nminibatches == 0
nbatch_train = nbatch // nminibatches
tstart = time.time()
frac = 1.0 - (update - 1.0) / nupdates
lrnow = lr(frac)
cliprangenow = cliprange(frac)
logger.info('collecting rollouts...')
run_tstart = time.time()
obs, returns, masks, actions, values, neglogpacs, states, epinfos = runner.run()
epinfobuf10.extend(epinfos)
epinfobuf100.extend(epinfos)
run_elapsed = time.time() - run_tstart
run_t_total += run_elapsed
logger.info('rollouts complete')
mblossvals = []
logger.info('updating parameters...')
train_tstart = time.time()
if states is None: # nonrecurrent version
inds = np.arange(nbatch)
for _ in range(noptepochs):
np.random.shuffle(inds)
for start in range(0, nbatch, nbatch_train):
end = start + nbatch_train
mbinds = inds[start:end]
slices = (arr[mbinds] for arr in (obs, returns, masks, actions, values, neglogpacs))
mblossvals.append(model.train(lrnow, cliprangenow, *slices))
else: # recurrent version
assert nenvs % nminibatches == 0
envinds = np.arange(nenvs)
flatinds = np.arange(nenvs * nsteps).reshape(nenvs, nsteps)
envsperbatch = nbatch_train // nsteps
for _ in range(noptepochs):
np.random.shuffle(envinds)
for start in range(0, nenvs, envsperbatch):
end = start + envsperbatch
mbenvinds = envinds[start:end]
mbflatinds = flatinds[mbenvinds].ravel()
slices = (arr[mbflatinds] for arr in (obs, returns, masks, actions, values, neglogpacs))
mbstates = states[mbenvinds]
mblossvals.append(model.train(lrnow, cliprangenow, *slices, mbstates))
# update the dropout mask
sess.run([model.train_model.dropout_assign_ops])
train_elapsed = time.time() - train_tstart
train_t_total += train_elapsed
logger.info('update complete')
lossvals = np.mean(mblossvals, axis=0)
tnow = time.time()
fps = int(nbatch / (tnow - tstart))
if update % log_interval == 0 or update == 1:
step = update*nbatch
#rew_mean_10 = utils.process_ep_buf(active_ep_buf, tb_writer=tb_writer, suffix='', step=step)
rew_mean_10 = safemean([epinfo['r'] for epinfo in epinfobuf10])
rew_mean_100 = safemean([epinfo['r'] for epinfo in epinfobuf100])
ep_len_mean_10 = np.nanmean([epinfo['l'] for epinfo in epinfobuf10])
ep_len_mean_100 = np.nanmean([epinfo['l'] for epinfo in epinfobuf100])
logger.info('\n----', update)
mean_rewards.append(rew_mean_10)
datapoints.append([step, rew_mean_10])
mean_rewards.append(rew_mean_10)
logger.logkv('eprew10', rew_mean_10)
logger.logkv('eprew100', rew_mean_100)
logger.logkv('eplenmean10', ep_len_mean_10)
logger.logkv('eplenmean100', ep_len_mean_100)
#logger.info('time_elapsed', tnow - tfirststart, run_t_total, train_t_total)
logger.logkv('misc/total_time_elapsed', tnow - tfirststart)
logger.logkv('misc/run_t_total', run_t_total)
logger.logkv('misc/train_t_total', train_t_total)
logger.logkv('nupdate', update)
#logger.info('timesteps', update*nsteps, total_timesteps)
logger.logkv("misc/total_timesteps", update*nbatch)
logger.logkv("misc/serial_timesteps", update*nsteps)
#logger.info('fps', fps)
logger.logkv("fps", fps)
if len(mblossvals):
for (lossval, lossname) in zip(lossvals, model.loss_names):
logger.info(lossname, lossval)
#tb_writer.log_scalar(lossval, lossname)
logger.logkv('loss/' + lossname, lossval)
logger.info('----\n')
logger.dumpkvs()
#if can_save:
if 0: ## not doing checkpoint saving yet
if save_interval and (update % save_interval == 0):
save_model()
for j, checkpoint in enumerate(checkpoints):
if (not saved_key_checkpoints[j]) and (step >= (checkpoint * 1e6)):
saved_key_checkpoints[j] = True
save_model(str(checkpoint) + 'M')
# save_model()
if save_path:
model.save(save_path)
env.close()
return model
def update(self):
#Some logic gathering best ret, rooms etc using MPI.
temp = sum(MPI.COMM_WORLD.allgather(self.local_rooms), [])
temp = sorted(list(set(temp)))
self.rooms = temp
temp = sum(MPI.COMM_WORLD.allgather(self.scores), [])
temp = sorted(list(set(temp)))
self.scores = temp
temp = sum(MPI.COMM_WORLD.allgather([self.local_best_ret]), [])
self.best_ret = max(temp)
eprews = MPI.COMM_WORLD.allgather(
np.mean(list(self.I.statlists["eprew"])))
local_best_rets = MPI.COMM_WORLD.allgather(self.local_best_ret)
n_rooms = sum(MPI.COMM_WORLD.allgather([len(self.local_rooms)]), [])
if MPI.COMM_WORLD.Get_rank() == 0:
logger.info(f"Rooms visited {self.rooms}")
logger.info(f"Best return {self.best_ret}")
logger.info(f"Best local return {sorted(local_best_rets)}")
logger.info(f"eprews {sorted(eprews)}")
logger.info(f"n_rooms {sorted(n_rooms)}")
logger.info(f"Extrinsic coefficient {self.ext_coeff}")
logger.info(f"Gamma {self.gamma}")
logger.info(f"Gamma ext {self.gamma_ext}")
logger.info(f"All scores {sorted(self.scores)}")
#Normalize intrinsic rewards.
rffs_int = np.array(
[self.I.rff_int.update(rew) for rew in self.I.buf_rews_int.T])
self.I.rff_rms_int.update(rffs_int.ravel())
rews_int = self.I.buf_rews_int / np.sqrt(self.I.rff_rms_int.var)
self.mean_int_rew = np.mean(rews_int)
self.max_int_rew = np.max(rews_int)
#Don't normalize extrinsic rewards.
rews_ext = self.I.buf_rews_ext
rewmean, rewstd, rewmax = self.I.buf_rews_int.mean(
), self.I.buf_rews_int.std(), np.max(self.I.buf_rews_int)
#Calculate intrinsic returns and advantages.
lastgaelam = 0
for t in range(self.nsteps - 1, -1, -1): # nsteps-2 ... 0
if self.use_news:
nextnew = self.I.buf_news[:, t +
1] if t + 1 < self.nsteps else self.I.buf_new_last
else:
nextnew = 0.0 #No dones for intrinsic reward.
nextvals = self.I.buf_vpreds_int[:, t +
1] if t + 1 < self.nsteps else self.I.buf_vpred_int_last
nextnotnew = 1 - nextnew
delta = rews_int[:,
t] + self.gamma * nextvals * nextnotnew - self.I.buf_vpreds_int[:,
t]
self.I.buf_advs_int[:,
t] = lastgaelam = delta + self.gamma * self.lam * nextnotnew * lastgaelam
rets_int = self.I.buf_advs_int + self.I.buf_vpreds_int
#Calculate extrinsic returns and advantages.
lastgaelam = 0
for t in range(self.nsteps - 1, -1, -1): # nsteps-2 ... 0
nextnew = self.I.buf_news[:, t +
1] if t + 1 < self.nsteps else self.I.buf_new_last
#Use dones for extrinsic reward.
nextvals = self.I.buf_vpreds_ext[:, t +
1] if t + 1 < self.nsteps else self.I.buf_vpred_ext_last
nextnotnew = 1 - nextnew
delta = rews_ext[:,
t] + self.gamma_ext * nextvals * nextnotnew - self.I.buf_vpreds_ext[:,
t]
self.I.buf_advs_ext[:,
t] = lastgaelam = delta + self.gamma_ext * self.lam * nextnotnew * lastgaelam
rets_ext = self.I.buf_advs_ext + self.I.buf_vpreds_ext
#Combine the extrinsic and intrinsic advantages.
self.I.buf_advs = self.int_coeff * self.I.buf_advs_int + self.ext_coeff * self.I.buf_advs_ext
#Collects info for reporting.
info = dict(
advmean=self.I.buf_advs.mean(),
advstd=self.I.buf_advs.std(),
retintmean=rets_int.mean(), # previously retmean
retintstd=rets_int.std(), # previously retstd
retextmean=rets_ext.mean(), # previously not there
retextstd=rets_ext.std(), # previously not there
rewintmean_unnorm=rewmean, # previously rewmean
rewintmax_unnorm=rewmax, # previously not there
rewintmean_norm=self.mean_int_rew, # previously rewintmean
rewintmax_norm=self.max_int_rew, # previously rewintmax
rewintstd_unnorm=rewstd, # previously rewstd
vpredintmean=self.I.buf_vpreds_int.mean(), # previously vpredmean
vpredintstd=self.I.buf_vpreds_int.std(), # previously vrpedstd
vpredextmean=self.I.buf_vpreds_ext.mean(), # previously not there
vpredextstd=self.I.buf_vpreds_ext.std(), # previously not there
ev_int=np.clip(
explained_variance(self.I.buf_vpreds_int.ravel(),
rets_int.ravel()), -1, None),
ev_ext=np.clip(
explained_variance(self.I.buf_vpreds_ext.ravel(),
rets_ext.ravel()), -1, None),
rooms=SemicolonList(self.rooms),
n_rooms=len(self.rooms),
best_ret=self.best_ret,
reset_counter=self.I.reset_counter)
info[f'mem_available'] = psutil.virtual_memory().available
to_record = {
'acs': self.I.buf_acs,
'rews_int': self.I.buf_rews_int,
'rews_int_norm': rews_int,
'rews_ext': self.I.buf_rews_ext,
'vpred_int': self.I.buf_vpreds_int,
'vpred_ext': self.I.buf_vpreds_ext,
'adv_int': self.I.buf_advs_int,
'adv_ext': self.I.buf_advs_ext,
'ent': self.I.buf_ent,
'ret_int': rets_int,
'ret_ext': rets_ext,
}
if self.I.venvs[0].record_obs:
to_record['obs'] = self.I.buf_obs[None]
self.recorder.record(bufs=to_record, infos=self.I.buf_epinfos)
#Create feeddict for optimization.
envsperbatch = self.I.nenvs // self.nminibatches
ph_buf = [
(self.stochpol.ph_ac, self.I.buf_acs),
(self.ph_ret_int, rets_int),
(self.ph_ret_ext, rets_ext),
(self.ph_oldnlp, self.I.buf_nlps),
(self.ph_adv, self.I.buf_advs),
]
if self.I.mem_state is not NO_STATES:
ph_buf.extend([
(self.stochpol.ph_istate, self.I.seg_init_mem_state),
(self.stochpol.ph_new, self.I.buf_news),
])
verbose = True
if verbose and self.is_log_leader:
samples = np.prod(self.I.buf_advs.shape)
logger.info(
"buffer shape %s, samples_per_mpi=%i, mini_per_mpi=%i, samples=%i, mini=%i "
% (str(self.I.buf_advs.shape), samples, samples //
self.nminibatches, samples * self.comm_train_size,
samples * self.comm_train_size // self.nminibatches))
logger.info(" " * 6 + fmt_row(13, self.loss_names))
epoch = 0
start = 0
#Optimizes on current data for several epochs.
while epoch < self.nepochs:
end = start + envsperbatch
mbenvinds = slice(start, end, None)
fd = {ph: buf[mbenvinds] for (ph, buf) in ph_buf}
fd.update({self.ph_lr: self.lr, self.ph_cliprange: self.cliprange})
fd[self.stochpol.ph_ob[None]] = np.concatenate([
self.I.buf_obs[None][mbenvinds],
self.I.buf_ob_last[None][mbenvinds, None]
], 1)
assert list(fd[self.stochpol.ph_ob[None]].shape) == [self.I.nenvs//self.nminibatches, self.nsteps + 1] + list(self.ob_space.shape), \
[fd[self.stochpol.ph_ob[None]].shape, [self.I.nenvs//self.nminibatches, self.nsteps + 1] + list(self.ob_space.shape)]
fd.update({
self.stochpol.ph_mean: self.stochpol.ob_rms.mean,
self.stochpol.ph_std: self.stochpol.ob_rms.var**0.5
})
ret = tf.get_default_session().run(self._losses + [self._train],
feed_dict=fd)[:-1]
if not self.testing:
lossdict = dict(zip([n for n in self.loss_names], ret), axis=0)
else:
lossdict = {}
#Synchronize the lossdict across mpi processes, otherwise weights may be rolled back on one process but not another.
_maxkl = lossdict.pop('maxkl')
lossdict = dict_gather(self.comm_train, lossdict, op='mean')
maxmaxkl = dict_gather(self.comm_train, {"maxkl": _maxkl},
op='max')
lossdict["maxkl"] = maxmaxkl["maxkl"]
if verbose and self.is_log_leader:
logger.info(
"%i:%03i %s" %
(epoch, start,
fmt_row(13, [lossdict[n] for n in self.loss_names])))
start += envsperbatch
if start == self.I.nenvs:
epoch += 1
start = 0
if self.is_train_leader:
self.I.stats["n_updates"] += 1
info.update([('opt_' + n, lossdict[n]) for n in self.loss_names])
tnow = time.time()
info['tps'] = self.nsteps * self.I.nenvs / (tnow -
self.I.t_last_update)
info['time_elapsed'] = time.time() - self.t0
self.I.t_last_update = tnow
self.stochpol.update_normalization( # Necessary for continuous control tasks with odd obs ranges, only implemented in mlp policy,
ob=self.I.buf_obs # NOTE: not shared via MPI
)
return info
def __init__(self,
sess,
prefix,
env,
nb_epochs,
nb_epoch_cycles,
render_eval,
reward_scale,
render,
param_noise,
actor,
critic,
normalize_returns,
normalize_observations,
critic_l2_reg,
actor_lr,
actor_dis_lr,
critic_lr,
exp_scale,
action_noise,
popart,
gamma,
clip_norm,
nb_train_steps,
nb_dis_train_steps,
nb_rollout_steps,
nb_eval_steps,
batch_size,
dis_batch_size,
memory,
tau=0.01,
eval_env=None,
param_noise_adaption_interval=50):
self.sess = sess
self.prefix = prefix
self.env = env
self.nb_epochs = nb_epochs
self.nb_epoch_cycles = nb_epoch_cycles
self.render_eval = render_eval
self.reward_scale = reward_scale
self.render = render
self.param_noise = param_noise
self.actor = actor
self.critic = critic
self.normalize_returns = normalize_returns
self.normalize_observations = normalize_observations
self.critic_l2_reg = critic_l2_reg
self.actor_lr = actor_lr
self.actor_dis_lr = actor_dis_lr
self.critic_lr = critic_lr
self.exp_scale = exp_scale
self.action_noise = action_noise
self.popart = popart
self.gamma = gamma
self.clip_norm = clip_norm
self.nb_train_steps = nb_train_steps
self.nb_dis_train_steps = nb_dis_train_steps
self.nb_rollout_steps = nb_rollout_steps
self.nb_eval_steps = nb_eval_steps
self.batch_size = batch_size
self.dis_batch_size = dis_batch_size
self.memory = memory
self.tau = tau
self.eval_env = eval_env
self.param_noise_adaption_interval = param_noise_adaption_interval
self.rank = MPI.COMM_WORLD.Get_rank()
assert (np.abs(self.env.action_space.low) == self.env.action_space.high
).all() # we assume symmetric actions.
self.max_action = self.env.action_space.high
logger.info('scaling actions by {} before executing in env'.format(
self.max_action))
self.agent = DDPG(self.prefix,
self.actor,
self.critic,
self.memory,
self.env.observation_space.shape,
self.env.action_space.shape,
gamma=self.gamma,
tau=self.tau,
normalize_returns=self.normalize_returns,
normalize_observations=self.normalize_observations,
batch_size=self.batch_size,
dis_batch_size=self.dis_batch_size,
action_noise=self.action_noise,
param_noise=self.param_noise,
critic_l2_reg=self.critic_l2_reg,
actor_lr=self.actor_lr,
actor_dis_lr=self.actor_dis_lr,
critic_lr=self.critic_lr,
exp_scale=self.exp_scale,
enable_popart=self.popart,
clip_norm=self.clip_norm,
reward_scale=self.reward_scale)
logger.info('Using agent with the following configuration:')
logger.info(str(self.agent.__dict__.items()))
# Set up logging stuff only for a single worker.
if self.rank == 0:
self.saver = tf.train.Saver()
else:
self.saver = None
self.step = 0
self.episode = 0
self.eval_episode_rewards_history = deque(maxlen=100)
self.episode_rewards_history = deque(maxlen=100)
#with U.single_threaded_session() as sess:
#self.sess = U.single_threaded_session()
# Prepare everything.
self.agent.initialize(self.sess)
#self.sess.graph.finalize()
self.agent.reset()
self.obs = self.env.reset()
if self.eval_env is not None:
self.eval_obs = eval_env.reset()
self.done = False
self.episode_reward = 0.
self.episode_step = 0
self.episodes = 0
self.t = 0
self.epoch = 0
self.start_time = time.time()
self.epoch_episode_rewards = []
self.epoch_episode_steps = []
self.epoch_episode_eval_rewards = []
self.epoch_episode_eval_steps = []
self.epoch_start_time = time.time()
self.epoch_actions = []
self.epoch_qs = []
self.epoch_episodes = 0
self.epoch_actor_losses = []
self.epoch_critic_losses = []
self.epoch_adaptive_distances = []
self.eval_episode_rewards = []
self.eval_qs = []
parser.add_argument('--nb-train-steps', type=int, default=50) # per epoch cycle and MPI worker
parser.add_argument('--nb-eval-steps', type=int, default=100) # per epoch cycle and MPI worker
parser.add_argument('--nb-rollout-steps', type=int, default=100) # per epoch cycle and MPI worker
parser.add_argument('--noise-type', type=str, default='adaptive-param_0.2') # choices are adaptive-param_xx, ou_xx, normal_xx, none
parser.add_argument('--logdir', type=str, default=None)
boolean_flag(parser, 'gym-monitor', default=False)
boolean_flag(parser, 'evaluation', default=True)
boolean_flag(parser, 'bind-to-core', default=False)
return vars(parser.parse_args())
if __name__ == '__main__':
args = parse_args()
# Figure out what logdir to use.
if args['logdir'] is None:
args['logdir'] = os.getenv('OPENAI_LOGDIR')
# Print and save arguments.
logger.info('Arguments:')
for key in sorted(args.keys()):
logger.info('{}: {}'.format(key, args[key]))
logger.info('')
if args['logdir']:
with open(os.path.join(args['logdir'], 'args.json'), 'w') as f:
json.dump(args, f)
# Run actual script.
run(**args)
def _create_network(self, reuse=False):
logger.info("Creating a DDPG agent with action space %d x %s..." %
(self.dimu, self.max_u))
self.sess = tf.get_default_session()
if self.sess is None:
self.sess = tf.InteractiveSession()
# running averages
with tf.variable_scope('o_stats') as vs:
if reuse:
vs.reuse_variables()
self.o_stats = Normalizer(self.dimo,
self.norm_eps,
self.norm_clip,
sess=self.sess)
with tf.variable_scope('g_stats') as vs:
if reuse:
vs.reuse_variables()
self.g_stats = Normalizer(self.dimg,
self.norm_eps,
self.norm_clip,
sess=self.sess)
with tf.variable_scope('uo_stats') as vs:
if reuse:
vs.reuse_variables()
self.uo_stats = Normalizer(self.dimo + self.dimg,
self.norm_eps,
self.norm_clip,
sess=self.sess)
with tf.variable_scope('lo_stats') as vs:
if reuse:
vs.reuse_variables()
self.lo_stats = Normalizer(2 * self.dimo,
self.norm_eps,
self.norm_clip,
sess=self.sess)
'''
# mini-batch sampling.
batch = self.staging_tf.get()
batch_tf = OrderedDict([(key, batch[i])
for i, key in enumerate(self.stage_shapes.keys())])
batch_tf['r'] = tf.reshape(batch_tf['r'], [-1, 1])
#choose only the demo buffer samples
mask = np.concatenate((np.zeros(self.batch_size - self.demo_batch_size), np.ones(self.demo_batch_size)), axis = 0)
# networks
with tf.variable_scope('main') as vs:
if reuse:
vs.reuse_variables()
self.main = self.create_actor_critic(batch_tf, net_type='main', **self.__dict__)
vs.reuse_variables()
with tf.variable_scope('target') as vs:
if reuse:
vs.reuse_variables()
target_batch_tf = batch_tf.copy()
target_batch_tf['o'] = batch_tf['o_2']
target_batch_tf['g'] = batch_tf['g_2']
self.target = self.create_actor_critic(
target_batch_tf, net_type='target', **self.__dict__)
vs.reuse_variables()
assert len(self._vars("main")) == len(self._vars("target"))
# loss functions
target_Q_pi_tf = self.target.Q_pi_tf
clip_range = (-self.clip_return, 0. if self.clip_pos_returns else np.inf)
target_tf = tf.clip_by_value(batch_tf['r'] + self.gamma * target_Q_pi_tf, *clip_range)
self.Q_loss_tf = tf.reduce_mean(tf.square(tf.stop_gradient(target_tf) - self.main.Q_tf))
if self.bc_loss ==1 and self.q_filter == 1 : # train with demonstrations and use bc_loss and q_filter both
maskMain = tf.reshape(tf.boolean_mask(self.main.Q_tf > self.main.Q_pi_tf, mask), [-1]) #where is the demonstrator action better than actor action according to the critic? choose those samples only
#define the cloning loss on the actor's actions only on the samples which adhere to the above masks
self.cloning_loss_tf = tf.reduce_sum(tf.square(tf.boolean_mask(tf.boolean_mask((self.main.pi_tf), mask), maskMain, axis=0) - tf.boolean_mask(tf.boolean_mask((batch_tf['u']), mask), maskMain, axis=0)))
self.pi_loss_tf = -self.prm_loss_weight * tf.reduce_mean(self.main.Q_pi_tf) #primary loss scaled by it's respective weight prm_loss_weight
self.pi_loss_tf += self.prm_loss_weight * self.action_l2 * tf.reduce_mean(tf.square(self.main.pi_tf / self.max_u)) #L2 loss on action values scaled by the same weight prm_loss_weight
self.pi_loss_tf += self.aux_loss_weight * self.cloning_loss_tf #adding the cloning loss to the actor loss as an auxilliary loss scaled by its weight aux_loss_weight
elif self.bc_loss == 1 and self.q_filter == 0: # train with demonstrations without q_filter
self.cloning_loss_tf = tf.reduce_sum(tf.square(tf.boolean_mask((self.main.pi_tf), mask) - tf.boolean_mask((batch_tf['u']), mask)))
self.pi_loss_tf = -self.prm_loss_weight * tf.reduce_mean(self.main.Q_pi_tf)
self.pi_loss_tf += self.prm_loss_weight * self.action_l2 * tf.reduce_mean(tf.square(self.main.pi_tf / self.max_u))
self.pi_loss_tf += self.aux_loss_weight * self.cloning_loss_tf
else: #If not training with demonstrations
self.pi_loss_tf = -tf.reduce_mean(self.main.Q_pi_tf)
self.pi_loss_tf += self.action_l2 * tf.reduce_mean(tf.square(self.main.pi_tf / self.max_u))
Q_grads_tf = tf.gradients(self.Q_loss_tf, self._vars('main/Q'))
pi_grads_tf = tf.gradients(self.pi_loss_tf, self._vars('main/pi'))
assert len(self._vars('main/Q')) == len(Q_grads_tf)
assert len(self._vars('main/pi')) == len(pi_grads_tf)
self.Q_grads_vars_tf = zip(Q_grads_tf, self._vars('main/Q'))
self.pi_grads_vars_tf = zip(pi_grads_tf, self._vars('main/pi'))
self.Q_grad_tf = flatten_grads(grads=Q_grads_tf, var_list=self._vars('main/Q'))
self.pi_grad_tf = flatten_grads(grads=pi_grads_tf, var_list=self._vars('main/pi'))
# optimizers
self.Q_adam = MpiAdam(self._vars('main/Q'), scale_grad_by_procs=False)
self.pi_adam = MpiAdam(self._vars('main/pi'), scale_grad_by_procs=False)
'''
# polyak averaging
#self.main_vars = self._vars('main/Q') + self._vars('main/pi')
#self.target_vars = self._vars('target/Q') + self._vars('target/pi')
self.stats_vars = self._global_vars('o_stats') + self._global_vars(
'g_stats')
#self.init_target_net_op = list(
# map(lambda v: v[0].assign(v[1]), zip(self.target_vars, self.main_vars)))
#self.update_target_net_op = list(
# map(lambda v: v[0].assign(self.polyak * v[0] + (1. - self.polyak) * v[1]), zip(self.target_vars, self.main_vars)))
# initialize all variables
tf.variables_initializer(self._global_vars('')).run()
def learn(*,
network,
env,
total_timesteps,
eval_env=None,
seed=None,
nsteps=2048,
ent_coef=0.0,
lr=3e-4,
vf_coef=0.5,
max_grad_norm=0.5,
gamma=0.99,
lam=0.95,
log_interval=10,
nminibatches=4,
noptepochs=4,
cliprange=0.2,
save_interval=0,
load_path=None,
model_fn=None,
update_fn=None,
use_demo=False,
demos=None,
render_env=False,
init_fn=None,
mpi_rank_weight=1,
comm=None,
base_path=None,
**network_kwargs):
'''
Learn policy using PPO algorithm (https://arxiv.org/abs/1707.06347)
Parameters:
----------
network: policy network architecture. Either string (mlp, lstm, lnlstm, cnn_lstm, cnn, cnn_small, conv_only - see baselines.common/models.py for full list)
specifying the standard network architecture, or a function that takes tensorflow tensor as input and returns
tuple (output_tensor, extra_feed) where output tensor is the last network layer output, extra_feed is None for feed-forward
neural nets, and extra_feed is a dictionary describing how to feed state into the network for recurrent neural nets.
See common/models.py/lstm for more details on using recurrent nets in policies
env: baselines.common.vec_env.VecEnv environment. Needs to be vectorized for parallel environment simulation.
The environments produced by gym.make can be wrapped using baselines.common.vec_env.DummyVecEnv class.
nsteps: int number of steps of the vectorized environment per update (i.e. batch size is nsteps * nenv where
nenv is number of environment copies simulated in parallel)
total_timesteps: int number of timesteps (i.e. number of actions taken in the environment)
ent_coef: float policy entropy coefficient in the optimization objective
lr: float or function learning rate, constant or a schedule function [0,1] -> R+ where 1 is beginning of the
training and 0 is the end of the training.
vf_coef: float value function loss coefficient in the optimization objective
max_grad_norm: float or None gradient norm clipping coefficient
gamma: float discounting factor
lam: float advantage estimation discounting factor (lambda in the paper)
log_interval: int number of timesteps between logging events
nminibatches: int number of training minibatches per update. For recurrent policies,
should be smaller or equal than number of environments run in parallel.
noptepochs: int number of training epochs per update
cliprange: float or function clipping range, constant or schedule function [0,1] -> R+ where 1 is beginning of the training
and 0 is the end of the training
save_interval: int number of timesteps between saving events
load_path: str path to load the model from
**network_kwargs: keyword arguments to the policy / network builder. See baselines.common/policies.py/build_policy and arguments to a particular type of network
For instance, 'mlp' network architecture has arguments num_hidden and num_layers.
'''
set_global_seeds(seed)
if isinstance(lr, float): lr = constfn(lr)
else: assert callable(lr)
if isinstance(cliprange, float): cliprange = constfn(cliprange)
else: assert callable(cliprange)
total_timesteps = int(total_timesteps)
policy = build_policy(env, network, **network_kwargs)
# Get the nb of env
nenvs = env.num_envs
# Get state_space and action_space
ob_space = env.observation_space
ac_space = env.action_space
# Calculate the batch_size
nbatch = nenvs * nsteps
nbatch_train = nbatch // nminibatches
is_mpi_root = (MPI is None or MPI.COMM_WORLD.Get_rank() == 0)
# Instantiate the model object (that creates act_model and train_model)
if model_fn is None:
from baselines.ppo2.model import Model
model_fn = Model
model = model_fn(policy=policy,
ob_space=ob_space,
ac_space=ac_space,
nbatch_act=nenvs,
nbatch_train=nbatch_train,
nsteps=nsteps,
ent_coef=ent_coef,
vf_coef=vf_coef,
max_grad_norm=max_grad_norm,
comm=comm,
mpi_rank_weight=mpi_rank_weight)
if save_interval and logger.get_dir():
if base_path is None:
base_path = os.path.dirname(os.path.abspath(__file__))
if not os.path.isdir(osp.join(base_path, "models")):
os.mkdir(osp.join(base_path, "models"))
# with open(osp.join(base_path, "models", 'model.pkl'), 'wb') as fh:
# fh.write(cloudpickle.dumps(model))
if load_path is not None:
model.load(load_path)
# Instantiate the runner object
runner = Runner(env=env,
model=model,
nsteps=nsteps,
gamma=gamma,
lam=lam,
use_demo=use_demo,
demos=demos,
render_env=render_env)
if eval_env is not None:
eval_runner = Runner(env=eval_env,
model=model,
nsteps=nsteps,
gamma=gamma,
lam=lam)
epinfobuf = deque(maxlen=100)
if eval_env is not None:
eval_epinfobuf = deque(maxlen=100)
if init_fn is not None:
init_fn()
# Start total timer
tfirststart = time.perf_counter()
nupdates = total_timesteps // nbatch
# initialize dictionary for data saving
performance = {"reward": []}
for name in model.loss_names:
performance[name] = []
for update in range(1, nupdates + 1):
assert nbatch % nminibatches == 0
# Start timer
tstart = time.perf_counter()
frac = 1.0 - (update - 1.0) / nupdates
# Calculate the learning rate
lrnow = lr(frac)
# Calculate the cliprange
cliprangenow = cliprange(frac)
if update % log_interval == 0 and is_mpi_root:
logger.info('Stepping environment...')
# Get minibatch
obs, returns, masks, actions, values, neglogpacs, states, epinfos = runner.run(
) #pylint: disable=E0632
rewards = runner.mb_rewards
# save reward data
performance["reward"].extend(rewards)
if eval_env is not None:
eval_obs, eval_returns, eval_masks, eval_actions, eval_values, eval_neglogpacs, eval_states, eval_epinfos = eval_runner.run(
) #pylint: disable=E0632
if update % log_interval == 0 and is_mpi_root: logger.info('Done.')
epinfobuf.extend(epinfos)
if eval_env is not None:
eval_epinfobuf.extend(eval_epinfos)
# Here what we're going to do is for each minibatch calculate the loss and append it.
mblossvals = []
if states is None: # nonrecurrent version
# Index of each element of batch_size
# Create the indices array
inds = np.arange(nbatch)
for _ in range(noptepochs):
# Randomize the indexes
np.random.shuffle(inds)
# 0 to batch_size with batch_train_size step
for start in range(0, nbatch, nbatch_train):
end = start + nbatch_train
mbinds = inds[start:end]
slices = (arr[mbinds]
for arr in (obs, returns, masks, actions, values,
neglogpacs))
mblossvals.append(model.train(lrnow, cliprangenow,
*slices))
else: # recurrent version
assert nenvs % nminibatches == 0
envsperbatch = nenvs // nminibatches
envinds = np.arange(nenvs)
flatinds = np.arange(nenvs * nsteps).reshape(nenvs, nsteps)
for _ in range(noptepochs):
np.random.shuffle(envinds)
for start in range(0, nenvs, envsperbatch):
end = start + envsperbatch
mbenvinds = envinds[start:end]
mbflatinds = flatinds[mbenvinds].ravel()
slices = (arr[mbflatinds]
for arr in (obs, returns, masks, actions, values,
neglogpacs))
mbstates = states[mbenvinds]
mblossvals.append(
model.train(lrnow, cliprangenow, *slices, mbstates))
# Feedforward --> get losses --> update
lossvals = np.mean(mblossvals, axis=0)
# End timer
tnow = time.perf_counter()
# Calculate the fps (frame per second)
fps = int(nbatch / (tnow - tstart))
if update_fn is not None:
update_fn(update)
if update % log_interval == 0 or update == 1:
# Calculates if value function is a good predicator of the returns (ev > 1)
# or if it's just worse than predicting nothing (ev =< 0)
ev = explained_variance(values, returns)
logger.logkv("misc/serial_timesteps", update * nsteps)
logger.logkv("misc/nupdates", update)
logger.logkv("misc/total_timesteps", update * nbatch)
logger.logkv("fps", fps)
logger.logkv("misc/explained_variance", float(ev))
logger.logkv('eprewmean',
safemean([epinfo['r'] for epinfo in epinfobuf]))
logger.logkv('eplenmean',
safemean([epinfo['l'] for epinfo in epinfobuf]))
if eval_env is not None:
logger.logkv(
'eval_eprewmean',
safemean([epinfo['r'] for epinfo in eval_epinfobuf]))
logger.logkv(
'eval_eplenmean',
safemean([epinfo['l'] for epinfo in eval_epinfobuf]))
logger.logkv('misc/time_elapsed', tnow - tfirststart)
for (lossval, lossname) in zip(lossvals, model.loss_names):
logger.logkv('loss/' + lossname, lossval)
logger.dumpkvs()
# save loss data
for i in range(len(model.loss_names)):
loss_name = model.loss_names[i]
performance[loss_name].extend([lossvals[i]])
if save_interval and (update % save_interval == 0 or update
== 1) and logger.get_dir() and is_mpi_root:
model_dir = osp.join(base_path, "models")
os.makedirs(model_dir, exist_ok=True)
savepath = osp.join(model_dir, '%.5i' % update)
print('Saving to', savepath)
model.save(savepath)
print("Saved model successfully.")
performance_fname = os.path.join(base_path, "performance.p")
with open(performance_fname, "wb") as f:
pickle.dump(performance, f)
model_dir = osp.join(base_path, "models")
os.makedirs(model_dir, exist_ok=True)
savepath = osp.join(model_dir, '%.5i' % update)
print('Saving to', savepath)
model.save(savepath)
print("Saved model successfully.")
performance_fname = os.path.join(base_path, "performance.p")
with open(performance_fname, "wb") as f:
pickle.dump(performance, f)
return performance
def learn(
network,
env,
seed=None,
total_timesteps=None,
nb_epochs=None, # with default settings, perform 1M steps total
nb_epoch_cycles=20,
nb_rollout_steps=100,
reward_scale=1.0,
render=False,
render_eval=False,
noise_type='adaptive-param_0.2',
normalize_returns=False,
normalize_observations=True,
critic_l2_reg=1e-2,
actor_lr=1e-4,
critic_lr=1e-3,
popart=False,
gamma=0.99,
clip_norm=None,
nb_train_steps=50, # per epoch cycle and MPI worker,
nb_eval_steps=100,
batch_size=64, # per MPI worker
tau=0.01,
eval_env=None,
param_noise_adaption_interval=50,
**network_kwargs):
set_global_seeds(seed)
if total_timesteps is not None:
assert nb_epochs is None
nb_epochs = int(total_timesteps) // (nb_epoch_cycles *
nb_rollout_steps)
else:
nb_epochs = 500
if MPI is not None:
rank = MPI.COMM_WORLD.Get_rank()
else:
rank = 0
nb_actions = env.action_space.shape[-1]
assert (np.abs(env.action_space.low) == env.action_space.high
).all() # we assume symmetric actions.
memory = Memory(limit=int(1e6),
action_shape=env.action_space.shape,
observation_shape=env.observation_space.shape)
critic = Critic(network=network, **network_kwargs)
actor = Actor(nb_actions, network=network, **network_kwargs)
action_noise = None
param_noise = None
if noise_type is not None:
for current_noise_type in noise_type.split(','):
current_noise_type = current_noise_type.strip()
if current_noise_type == 'none':
pass
elif 'adaptive-param' in current_noise_type:
_, stddev = current_noise_type.split('_')
param_noise = AdaptiveParamNoiseSpec(
initial_stddev=float(stddev),
desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = NormalActionNoise(mu=np.zeros(nb_actions),
sigma=float(stddev) *
np.ones(nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(
mu=np.zeros(nb_actions),
sigma=float(stddev) * np.ones(nb_actions))
else:
raise RuntimeError(
'unknown noise type "{}"'.format(current_noise_type))
max_action = env.action_space.high
logger.info(
'scaling actions by {} before executing in env'.format(max_action))
agent = DDPG(actor,
critic,
memory,
env.observation_space.shape,
env.action_space.shape,
gamma=gamma,
tau=tau,
normalize_returns=normalize_returns,
normalize_observations=normalize_observations,
batch_size=batch_size,
action_noise=action_noise,
param_noise=param_noise,
critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr,
critic_lr=critic_lr,
enable_popart=popart,
clip_norm=clip_norm,
reward_scale=reward_scale)
logger.info('Using agent with the following configuration:')
logger.info(str(agent.__dict__.items()))
eval_episode_rewards_history = deque(maxlen=100)
episode_rewards_history = deque(maxlen=100)
sess = U.get_session()
# Prepare everything.
agent.initialize(sess)
sess.graph.finalize()
agent.reset()
obs = env.reset()
if eval_env is not None:
eval_obs = eval_env.reset()
nenvs = obs.shape[0]
episode_reward = np.zeros(nenvs, dtype=np.float32) #vector
episode_step = np.zeros(nenvs, dtype=int) # vector
episodes = 0 #scalar
t = 0 # scalar
epoch = 0
start_time = time.time()
epoch_episode_rewards = []
epoch_episode_steps = []
epoch_actions = []
epoch_qs = []
epoch_episodes = 0
for epoch in range(nb_epochs):
for cycle in range(nb_epoch_cycles):
# Perform rollouts.
if nenvs > 1:
# if simulating multiple envs in parallel, impossible to reset agent at the end of the episode in each
# of the environments, so resetting here instead
agent.reset()
for t_rollout in range(nb_rollout_steps):
# Predict next action.
action, q, _, _ = agent.step(obs,
apply_noise=True,
compute_Q=True)
# Execute next action.
if rank == 0 and render:
env.render()
# max_action is of dimension A, whereas action is dimension (nenvs, A) - the multiplication gets broadcasted to the batch
new_obs, r, done, info = env.step(
max_action * action
) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
# note these outputs are batched from vecenv
t += 1
if rank == 0 and render:
env.render()
episode_reward += r
episode_step += 1
# Book-keeping.
epoch_actions.append(action)
epoch_qs.append(q)
agent.store_transition(
obs, action, r, new_obs, done
) #the batched data will be unrolled in memory.py's append.
obs = new_obs
for d in range(len(done)):
if done[d]:
# Episode done.
epoch_episode_rewards.append(episode_reward[d])
episode_rewards_history.append(episode_reward[d])
epoch_episode_steps.append(episode_step[d])
episode_reward[d] = 0.
episode_step[d] = 0
epoch_episodes += 1
episodes += 1
if nenvs == 1:
agent.reset()
# Train.
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
for t_train in range(nb_train_steps):
# Adapt param noise, if necessary.
if memory.nb_entries >= batch_size and t_train % param_noise_adaption_interval == 0:
distance = agent.adapt_param_noise()
epoch_adaptive_distances.append(distance)
cl, al = agent.train()
epoch_critic_losses.append(cl)
epoch_actor_losses.append(al)
agent.update_target_net()
# Evaluate.
eval_episode_rewards = []
eval_qs = []
if eval_env is not None:
nenvs_eval = eval_obs.shape[0]
eval_episode_reward = np.zeros(nenvs_eval, dtype=np.float32)
for t_rollout in range(nb_eval_steps):
eval_action, eval_q, _, _ = agent.step(eval_obs,
apply_noise=False,
compute_Q=True)
eval_obs, eval_r, eval_done, eval_info = eval_env.step(
max_action * eval_action
) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
if render_eval:
eval_env.render()
eval_episode_reward += eval_r
eval_qs.append(eval_q)
for d in range(len(eval_done)):
if eval_done[d]:
eval_episode_rewards.append(eval_episode_reward[d])
eval_episode_rewards_history.append(
eval_episode_reward[d])
eval_episode_reward[d] = 0.0
if MPI is not None:
mpi_size = MPI.COMM_WORLD.Get_size()
else:
mpi_size = 1
# Log stats.
# XXX shouldn't call np.mean on variable length lists
duration = time.time() - start_time
stats = agent.get_stats()
combined_stats = stats.copy()
combined_stats['rollout/return'] = np.mean(epoch_episode_rewards)
combined_stats['rollout/return_std'] = np.std(epoch_episode_rewards)
combined_stats['rollout/return_history'] = np.mean(
episode_rewards_history)
combined_stats['rollout/return_history_std'] = np.std(
episode_rewards_history)
combined_stats['rollout/episode_steps'] = np.mean(epoch_episode_steps)
combined_stats['rollout/actions_mean'] = np.mean(epoch_actions)
combined_stats['rollout/Q_mean'] = np.mean(epoch_qs)
combined_stats['train/loss_actor'] = np.mean(epoch_actor_losses)
combined_stats['train/loss_critic'] = np.mean(epoch_critic_losses)
combined_stats['train/param_noise_distance'] = np.mean(
epoch_adaptive_distances)
combined_stats['total/duration'] = duration
combined_stats['total/steps_per_second'] = float(t) / float(duration)
combined_stats['total/episodes'] = episodes
combined_stats['rollout/episodes'] = epoch_episodes
combined_stats['rollout/actions_std'] = np.std(epoch_actions)
# Evaluation statistics.
if eval_env is not None:
combined_stats['eval/return'] = eval_episode_rewards
combined_stats['eval/return_history'] = np.mean(
eval_episode_rewards_history)
combined_stats['eval/Q'] = eval_qs
combined_stats['eval/episodes'] = len(eval_episode_rewards)
def as_scalar(x):
if isinstance(x, np.ndarray):
assert x.size == 1
return x[0]
elif np.isscalar(x):
return x
else:
raise ValueError('expected scalar, got %s' % x)
combined_stats_sums = np.array(
[np.array(x).flatten()[0] for x in combined_stats.values()])
if MPI is not None:
combined_stats_sums = MPI.COMM_WORLD.allreduce(combined_stats_sums)
combined_stats = {
k: v / mpi_size
for (k, v) in zip(combined_stats.keys(), combined_stats_sums)
}
# Total statistics.
combined_stats['total/epochs'] = epoch + 1
combined_stats['total/steps'] = t
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
if rank == 0:
logger.dump_tabular()
logger.info('')
logdir = logger.get_dir()
if rank == 0 and logdir:
if hasattr(env, 'get_state'):
with open(os.path.join(logdir, 'env_state.pkl'), 'wb') as f:
pickle.dump(env.get_state(), f)
if eval_env and hasattr(eval_env, 'get_state'):
with open(os.path.join(logdir, 'eval_env_state.pkl'),
'wb') as f:
pickle.dump(eval_env.get_state(), f)
return agent
def view(env,
seed=None,
total_timesteps=None,
reward_scale=1.0,
render=False,
render_eval=False,
noise_type='adaptive-param_0.2',
normalize_returns=False,
normalize_observations=True,
critic_l2_reg=1e-2,
actor_lr=1e-4,
critic_lr=1e-3,
popart=False,
gamma=0.99,
clip_norm=None,
nb_train_steps=50, # per epoch cycle and MPI worker,
nb_eval_steps=100,
nb_save_epochs=None,
batch_size=64, # per MPI worker
tau=0.01,
action_range=(-250.0, 250.0),
observation_range=(-5.0, 5.0),
eval_env=None,
load_path=None,
save_dir=None,
param_noise_adaption_interval=50,
**network_kwargs):
set_global_seeds(seed)
if MPI is not None:
rank = MPI.COMM_WORLD.Get_rank()
else:
rank = 0
memory = Memory(limit=int(1e6))
network_spec = [
{
'layer_type': 'dense',
'units': int (256),
'activation': 'relu',
'nodes_in': ['main'],
'nodes_out': ['main']
},
{
'layer_type': 'dense',
'units': int (128),
'activation': 'relu',
'nodes_in': ['main'],
'nodes_out': ['main']
},
{
'layer_type': 'dense',
'units': int (1),
'activation': 'tanh',
'nodes_in': ['main'],
'nodes_out': ['main']
}
]
vnetwork_spec = [
{
'layer_type': 'concat',
'nodes_in': ['action_movement', 'observation_self'],
'nodes_out': ['main']
},
{
'layer_type': 'dense',
'units': int (256),
'activation': 'relu',
'nodes_in': ['main'],
'nodes_out': ['main']
},
{
'layer_type': 'dense',
'units': int (128),
'activation': 'relu',
'nodes_in': ['main'],
'nodes_out': ['main']
},
{
'layer_type': 'dense',
'units': int (1),
'activation': '',
'nodes_in': ['main'],
'nodes_out': ['main']
}
]
network = DdpgPolicy(scope="ddpg", ob_space=env.observation_space, ac_space=env.action_space, network_spec=network_spec, v_network_spec=vnetwork_spec,
stochastic=False, reuse=False, build_act=True,
trainable_vars=None, not_trainable_vars=None,
gaussian_fixed_var=False, weight_decay=0.0, ema_beta=0.99999,
normalize_observations=normalize_observations, normalize_returns=normalize_returns,
observation_range=observation_range)
target_network = DdpgPolicy(scope="target", ob_space=env.observation_space, ac_space=env.action_space, network_spec=network_spec, v_network_spec=vnetwork_spec,
stochastic=False, reuse=False, build_act=True,
trainable_vars=None, not_trainable_vars=None,
gaussian_fixed_var=False, weight_decay=0.0, ema_beta=0.99999,
normalize_observations=normalize_observations, normalize_returns=normalize_returns,
observation_range=observation_range)
max_action = action_range[1]
logger.info('scaling actions by {} before executing in env'.format(max_action))
agent = DDPG(network, target_network, memory, env.observation_space, env.action_space,
gamma=gamma, tau=tau, normalize_returns=normalize_returns, normalize_observations=normalize_observations,
batch_size=batch_size, action_noise=None, param_noise=None, critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr, critic_lr=critic_lr, enable_popart=popart, clip_norm=clip_norm,
reward_scale=reward_scale)
logger.info('Using agent with the following configuration:')
logger.info(str(agent.__dict__.items()))
sess = U.get_session()
loader = functools.partial(load_variables, sess=sess)
if load_path != None:
loader(load_path)
# Prepare everything.
agent.initialize(sess)
sess.graph.finalize()
nenvs = env.num_envs
obs = env.reset()
n_agents = obs['observation_self'].shape[0]
for epoch in range(total_timesteps):
agent.reset()
obs = env.reset()
while True:
action, q, _, _ = agent.step(obs, apply_noise=False, compute_Q=False)
# max_action is of dimension A, whereas action is dimension (nenvs, A) - the multiplication gets broadcasted to the batch
for k, v in action.items():
action[k] *= max_action
nenvs_actions = []
for i in range(nenvs):
nenv_action = {'action_movement' : action['action_movement'][i*n_agents:(i + 1)*n_agents]}
nenvs_actions.append(nenv_action)
print(nenvs_actions)
obs, r, done, info = env.step(nenvs_actions)
env.render()
if True in done:
break
def run(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
# Create envs.
env = gym.make(env_id)
env = bench.Monitor(
env,
logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
if evaluation and rank == 0:
eval_env = gym.make(env_id)
eval_env = bench.Monitor(eval_env,
os.path.join(logger.get_dir(), 'gym_eval'))
env = bench.Monitor(env, None)
else:
eval_env = None
# Parse noise_type
action_noise = None
param_noise = None
nb_actions = env.action_space.shape[-1]
for current_noise_type in noise_type.split(','):
current_noise_type = current_noise_type.strip()
if current_noise_type == 'none':
pass
elif 'adaptive-param' in current_noise_type:
_, stddev = current_noise_type.split('_')
param_noise = AdaptiveParamNoiseSpec(
initial_stddev=float(stddev),
desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = NormalActionNoise(mu=np.zeros(nb_actions),
sigma=float(stddev) *
np.ones(nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(
mu=np.zeros(nb_actions),
sigma=float(stddev) * np.ones(nb_actions))
else:
raise RuntimeError(
'unknown noise type "{}"'.format(current_noise_type))
# Configure components.
memory = Memory(limit=int(1e6),
action_shape=env.action_space.shape,
observation_shape=env.observation_space.shape)
critic = Critic(layer_norm=layer_norm)
actor = Actor(nb_actions, layer_norm=layer_norm)
# Seed everything to make things reproducible.
seed = seed + 1000000 * rank
logger.info('rank {}: seed={}, logdir={}'.format(rank, seed,
logger.get_dir()))
tf.reset_default_graph()
set_global_seeds(seed)
env.seed(seed)
if eval_env is not None:
eval_env.seed(seed)
# Disable logging for rank != 0 to avoid noise.
if rank == 0:
start_time = time.time()
training.train(env=env,
eval_env=eval_env,
param_noise=param_noise,
action_noise=action_noise,
actor=actor,
critic=critic,
memory=memory,
**kwargs)
env.close()
if eval_env is not None:
eval_env.close()
if rank == 0:
logger.info('total runtime: {}s'.format(time.time() - start_time))
def main():
num_envs = 64
learning_rate = 5e-4
ent_coef = .01
gamma = .999
lam = .95
nsteps = 256
nminibatches = 8
ppo_epochs = 3
clip_range = .2
total_timesteps = 1_000_000 ## now this counts steps in testing runs
use_vf_clipping = True
## From random_ppo.py
max_grad_norm = 0.5
vf_coef = 0.5
L2_WEIGHT = 10e-4
FM_COEFF = 0.002
REAL_THRES = 0.1
parser = argparse.ArgumentParser(
description='Process procgen testing arguments.')
parser.add_argument('--env_name', type=str, default='fruitbot')
parser.add_argument(
'--distribution_mode',
type=str,
default='easy',
choices=["easy", "hard", "exploration", "memory", "extreme"])
parser.add_argument('--num_levels', type=int, default=1000)
## default starting_level set to 50 to test on unseen levels!
parser.add_argument('--start_level', type=int, default=1000)
parser.add_argument('--run_id', '-id', type=int, default=0)
parser.add_argument('--load_id', type=int, default=0)
parser.add_argument('--nrollouts', '-nroll', type=int, default=0)
args = parser.parse_args()
args.total_timesteps = total_timesteps
if args.nrollouts:
total_timesteps = int(args.nrollouts * num_envs * nsteps)
run_ID = 'run_' + str(args.run_id).zfill(2)
run_ID += '_load{}'.format(args.load_id)
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
mpi_rank_weight = 0
num_levels = args.num_levels
log_comm = comm.Split(0, 0)
format_strs = ['csv', 'stdout', 'log'] if log_comm.Get_rank() == 0 else []
logpath = join(LOG_DIR, run_ID)
if not os.path.exists(logpath):
os.system("mkdir -p %s" % logpath)
fpath = join(logpath, 'args_{}.json'.format(run_ID))
with open(fpath, 'w') as fh:
json.dump(vars(args), fh, indent=4, sort_keys=True)
print("\nSaved args at:\n\t{}\n".format(fpath))
logger.configure(dir=logpath, format_strs=format_strs)
logger.info("creating environment")
venv = ProcgenEnv(num_envs=num_envs,
env_name=args.env_name,
num_levels=num_levels,
start_level=args.start_level,
distribution_mode=args.distribution_mode)
venv = VecExtractDictObs(venv, "rgb")
venv = VecMonitor(
venv=venv,
filename=None,
keep_buf=100,
)
venv = VecNormalize(venv=venv, ob=False)
logger.info("creating tf session")
setup_mpi_gpus()
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True #pylint: disable=E1101
sess = tf.compat.v1.Session(config=config)
sess.__enter__()
logger.info("Testing")
## Modified based on random_ppo.learn
env = venv
nenvs = env.num_envs
ob_space = env.observation_space
ac_space = env.action_space
nbatch = nenvs * nsteps
nbatch_train = nbatch // nminibatches
nrollouts = total_timesteps // nbatch
network = lambda x: build_impala_cnn(x, depths=[16, 32, 32], emb_size=256)
policy = build_policy(env, network)
model = Model(policy=policy,
ob_space=ob_space,
ac_space=ac_space,
nbatch_act=nenvs,
nbatch_train=nbatch_train,
nsteps=nsteps,
ent_coef=ent_coef,
vf_coef=vf_coef,
max_grad_norm=max_grad_norm)
LOAD_PATH = "log/vanilla/saved_vanilla_v{}.tar".format(args.load_id)
model.load(LOAD_PATH)
logger.info("Model pramas loaded from save")
runner = Runner(env=env, model=model, nsteps=nsteps, gamma=gamma, lam=lam)
epinfobuf10 = deque(maxlen=10)
epinfobuf100 = deque(maxlen=100)
# tfirststart = time.time() ## Not doing timing yet
# active_ep_buf = epinfobuf100
mean_rewards = []
datapoints = []
for rollout in range(1, nrollouts + 1):
logger.info('collecting rollouts {}...'.format(rollout))
obs, returns, masks, actions, values, neglogpacs, states, epinfos = runner.run(
) ## differnent from random_ppo!
epinfobuf10.extend(epinfos)
epinfobuf100.extend(epinfos)
rew_mean_10 = safemean([epinfo['r'] for epinfo in epinfobuf10])
rew_mean_100 = safemean([epinfo['r'] for epinfo in epinfobuf100])
ep_len_mean_10 = np.nanmean([epinfo['l'] for epinfo in epinfobuf10])
ep_len_mean_100 = np.nanmean([epinfo['l'] for epinfo in epinfobuf100])
logger.info('\n----', rollout)
mean_rewards.append(rew_mean_10)
logger.logkv('eprew10', rew_mean_10)
logger.logkv('eprew100', rew_mean_100)
logger.logkv('eplenmean10', ep_len_mean_10)
logger.logkv('eplenmean100', ep_len_mean_100)
logger.logkv("misc/total_timesteps", rollout * nbatch)
logger.info('----\n')
logger.dumpkvs()
env.close()
print("Rewards history: ", mean_rewards)
return mean_rewards
def train(env,
nb_epochs,
nb_epoch_cycles,
render_eval,
reward_scale,
render,
param_noise,
actor,
critics,
normalize_returns,
normalize_observations,
critic_l2_reg,
actor_lr,
critic_lr,
action_noise,
popart,
gamma,
clip_norm,
nb_train_steps,
nb_rollout_steps,
nb_eval_steps,
batch_size,
memory,
tau=0.005,
eval_env=None,
param_noise_adaption_interval=50,
initial_random_steps=1e4,
policy_and_target_update_period=1,
use_mpi_adam=False,
stop_actor_steps=None,
stop_critic_steps=None,
**kwargs):
if kwargs is not None:
logger.info("Warning: redundant hyper-parameters: " + str(kwargs))
rank = MPI.COMM_WORLD.Get_rank()
assert (np.abs(env.action_space.low) == env.action_space.high
).all() # we assume symmetric actions.
max_action = env.action_space.high
logger.info(
'scaling actions by {} before executing in env'.format(max_action))
agent = EnsembleDDPG(actor,
critics,
memory,
env.observation_space.shape,
env.action_space.shape,
gamma=gamma,
tau=tau,
normalize_returns=normalize_returns,
normalize_observations=normalize_observations,
batch_size=batch_size,
action_noise=action_noise,
param_noise=param_noise,
critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr,
critic_lr=critic_lr,
enable_popart=popart,
clip_norm=clip_norm,
reward_scale=reward_scale,
use_mpi_adam=use_mpi_adam)
logger.info('Using agent with the following configuration:')
logger.info(str(agent.__dict__.items()))
# Set up logging stuff only for a single worker.
if rank == 0:
saver = tf.train.Saver()
else:
saver = None
step = 0
episode = 0
eval_episode_rewards_history = deque(maxlen=100)
episode_rewards_history = deque(maxlen=100)
with U.single_threaded_session() as sess:
# Prepare everything.
agent.initialize(sess)
sess.graph.finalize()
agent.reset()
obs = env.reset()
if eval_env is not None:
eval_obs = eval_env.reset()
done = False
episode_reward = 0.
episode_step = 0
episodes = 0
t = 0
epoch = 0
start_time = time.time()
epoch_episode_rewards = []
epoch_episode_steps = []
epoch_episode_eval_rewards = []
epoch_episode_eval_steps = []
epoch_start_time = time.time()
epoch_actions = []
epoch_qs = []
epoch_episodes = 0
for epoch in range(nb_epochs):
for cycle in range(nb_epoch_cycles):
# Perform rollouts.
for t_rollout in range(nb_rollout_steps):
# Predict next action.
if t < initial_random_steps:
action = np.random.uniform(low=-1,
high=1,
size=env.action_space.shape)
else:
action, _ = agent.pi(obs,
apply_noise=True,
compute_Q=False)
assert action.shape == env.action_space.shape
# Execute next action.
if rank == 0 and render:
env.render()
assert max_action.shape == action.shape
new_obs, r, done, info = env.step(
max_action * action
) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
t += 1
if rank == 0 and render:
env.render()
episode_reward += r
episode_step += 1
# Book-keeping.
epoch_actions.append(action)
agent.store_transition(obs, action, r, new_obs, done)
obs = new_obs
if done:
# Episode done.
epoch_episode_rewards.append(episode_reward)
episode_rewards_history.append(episode_reward)
epoch_episode_steps.append(episode_step)
episode_reward = 0.
episode_step = 0
epoch_episodes += 1
episodes += 1
agent.reset()
obs = env.reset()
# Train.
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
for t_train in range(nb_train_steps):
# Adapt param noise, if necessary.
if memory.nb_entries >= batch_size and t_train % param_noise_adaption_interval == 0:
distance = agent.adapt_param_noise()
epoch_adaptive_distances.append(distance)
# use take_update flag to control actor training
take_update = (t_train %
policy_and_target_update_period == 0)
stop_actor_training = False
stop_critic_training = False
if stop_actor_steps is not None and t > stop_actor_steps:
stop_actor_training = True
if stop_critic_steps is not None and t > stop_critic_steps:
stop_critic_training = True
cl, al = agent.train(take_update, stop_critic_training,
stop_actor_training)
epoch_critic_losses.append(cl)
# use take_update flag to control target updating
if take_update:
epoch_actor_losses.append(al)
agent.update_target_net()
# Evaluate.
eval_episode_rewards = []
eval_qs = []
if eval_env is not None:
eval_episode_reward = 0.
for t_rollout in range(nb_eval_steps):
eval_action, eval_q = agent.pi(eval_obs,
apply_noise=False,
compute_Q=True)
eval_obs, eval_r, eval_done, eval_info = eval_env.step(
max_action * eval_action
) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
if render_eval:
eval_env.render()
eval_episode_reward += eval_r
eval_qs.append(eval_q)
if eval_done:
eval_obs = eval_env.reset()
eval_episode_rewards.append(eval_episode_reward)
eval_episode_rewards_history.append(
eval_episode_reward)
eval_episode_reward = 0.
mpi_size = MPI.COMM_WORLD.Get_size()
# Log stats.
# XXX shouldn't call np.mean on variable length lists
duration = time.time() - start_time
stats = agent.get_stats()
combined_stats = stats.copy()
combined_stats['rollout/return'] = np.mean(epoch_episode_rewards)
combined_stats['rollout/return_history'] = np.mean(
episode_rewards_history)
combined_stats['rollout/return_latest'] = np.mean(
epoch_episode_rewards[-1])
combined_stats['rollout/episode_steps'] = np.mean(
epoch_episode_steps)
combined_stats['rollout/actions_mean'] = np.mean(epoch_actions)
combined_stats['train/loss_actor'] = np.mean(epoch_actor_losses)
combined_stats['train/loss_critic'] = np.mean(epoch_critic_losses)
combined_stats['train/param_noise_distance'] = np.mean(
epoch_adaptive_distances)
combined_stats['total/duration'] = duration
combined_stats['total/steps_per_second'] = float(t) / float(
duration)
combined_stats['total/episodes'] = episodes
combined_stats['rollout/episodes'] = epoch_episodes
combined_stats['rollout/actions_std'] = np.std(epoch_actions)
# Evaluation statistics.
if eval_env is not None:
combined_stats['eval/return'] = eval_episode_rewards
combined_stats['eval/return_history'] = np.mean(
eval_episode_rewards_history)
combined_stats['eval/episodes'] = len(eval_episode_rewards)
def as_scalar(x):
if isinstance(x, np.ndarray):
assert x.size == 1
return x[0]
elif np.isscalar(x):
return x
else:
raise ValueError('expected scalar, got %s' % x)
combined_stats_sums = MPI.COMM_WORLD.allreduce(
np.array([as_scalar(x) for x in combined_stats.values()]))
combined_stats = {
k: v / mpi_size
for (k, v) in zip(combined_stats.keys(), combined_stats_sums)
}
# Total statistics.
combined_stats['total/epochs'] = epoch + 1
combined_stats['total/steps'] = t
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
logger.dump_tabular()
logger.info('')
logdir = logger.get_dir()
if rank == 0 and logdir:
if hasattr(env, 'get_state'):
with open(os.path.join(logdir, 'env_state.pkl'),
'wb') as f:
pickle.dump(env.get_state(), f)
if eval_env and hasattr(eval_env, 'get_state'):
with open(os.path.join(logdir, 'eval_env_state.pkl'),
'wb') as f:
pickle.dump(eval_env.get_state(), f)
def log_params(params, logger=logger):
for key in sorted(params.keys()):
logger.info('{}: {}'.format(key, params[key]))
def run(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
# Create envs.
env = gym.make(env_id)
env = bench.Monitor(env, logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
if evaluation and rank==0:
eval_env = gym.make(env_id)
eval_env = bench.Monitor(eval_env, os.path.join(logger.get_dir(), 'gym_eval'))
env = bench.Monitor(env, None)
else:
eval_env = None
# Parse noise_type
action_noise = None
param_noise = None
nb_actions = env.action_space.shape[-1]
for current_noise_type in noise_type.split(','):
current_noise_type = current_noise_type.strip()
if current_noise_type == 'none':
pass
elif 'adaptive-param' in current_noise_type:
_, stddev = current_noise_type.split('_')
param_noise = AdaptiveParamNoiseSpec(initial_stddev=float(stddev), desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = NormalActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
else:
raise RuntimeError('unknown noise type "{}"'.format(current_noise_type))
# Configure components.
memory = Memory(limit=int(1e6), action_shape=env.action_space.shape, observation_shape=env.observation_space.shape)
critic = Critic(layer_norm=layer_norm)
actor = Actor(nb_actions, layer_norm=layer_norm)
# Seed everything to make things reproducible.
seed = seed + 1000000 * rank
logger.info('rank {}: seed={}, logdir={}'.format(rank, seed, logger.get_dir()))
tf.reset_default_graph()
set_global_seeds(seed)
env.seed(seed)
if eval_env is not None:
eval_env.seed(seed)
# Disable logging for rank != 0 to avoid noise.
if rank == 0:
start_time = time.time()
training.train(env=env, eval_env=eval_env, param_noise=param_noise,
action_noise=action_noise, actor=actor, critic=critic, memory=memory, **kwargs)
env.close()
if eval_env is not None:
eval_env.close()
if rank == 0:
logger.info('total runtime: {}s'.format(time.time() - start_time))