def main(args):
"""
start training the model
:param args: (ArgumentParser) the training argument
"""
with tf_util.make_session(num_cpu=1):
set_global_seeds(args.seed)
env = gym.make(args.env_id)
def policy_fn(name, ob_space, ac_space, reuse=False, placeholders=None, sess=None):
return mlp_policy.MlpPolicy(name=name, ob_space=ob_space, ac_space=ac_space, reuse=reuse, sess=sess,
hid_size=args.policy_hidden_size, num_hid_layers=2, placeholders=placeholders)
#========================================================================================================
env = bench.Monitor(env, logger.get_dir() and
os.path.join(logger.get_dir(), "monitor.json"))
env.seed(args.seed)
gym.logger.setLevel(logging.WARN)
# 길고 긴 task name을 받아옵니다. ===========================================================================
task_name = get_task_name(args)
args.checkpoint_dir = os.path.join(args.checkpoint_dir, task_name)
args.log_dir = os.path.join(args.log_dir, task_name)
# =======================================================================================================
if args.task == 'train':
dataset = MujocoDset(expert_path=args.expert_path, traj_limitation=args.traj_limitation)
#discriminator 네트워크 생성
reward_giver = TransitionClassifier(env, args.adversary_hidden_size, entcoeff=args.adversary_entcoeff)
#policy 네트워크 학습
#policy network 가 policy_fn 으로 선언되어있다는 것에 주의
train(env, args.seed, policy_fn, reward_giver,
dataset, args.algo, args.g_step, args.d_step,
args.policy_entcoeff, args.num_timesteps, args.save_per_iter, args.checkpoint_dir, args.pretrained, args.bc_max_iter, task_name)
# ======================================= 이것은 나중에 이해하보는 거시여=============================================
# 학습된 모델을 evaluate 할 때 사용합니다.
elif args.task == 'evaluate':
runner(env,
policy_fn,
args.load_model_path,
timesteps_per_batch=1024,
number_trajs=10,
stochastic_policy=args.stochastic_policy,
save=args.save_sample
)
else:
raise NotImplementedError
env.close()
class TRPO(ActorCriticRLModel):
"""
Trust Region Policy Optimization (https://arxiv.org/abs/1502.05477)
:param policy: (ActorCriticPolicy or str) The policy model to use (MlpPolicy, CnnPolicy, CnnLstmPolicy, ...)
:param env: (Gym environment or str) The environment to learn from (if registered in Gym, can be str)
:param gamma: (float) the discount value
:param timesteps_per_batch: (int) the number of timesteps to run per batch (horizon)
:param max_kl: (float) the kullback leiber loss threshold
:param cg_iters: (int) the number of iterations for the conjugate gradient calculation
:param lam: (float) GAE factor
:param entcoeff: (float) the weight for the entropy loss
:param cg_damping: (float) the compute gradient dampening factor
:param vf_stepsize: (float) the value function stepsize
:param vf_iters: (int) the value function's number iterations for learning
:param verbose: (int) the verbosity level: 0 none, 1 training information, 2 tensorflow debug
:param tensorboard_log: (str) the log location for tensorboard (if None, no logging)
:param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance
:param policy_kwargs: (dict) additional arguments to be passed to the policy on creation
:param full_tensorboard_log: (bool) enable additional logging when using tensorboard
WARNING: this logging can take a lot of space quickly
"""
def __init__(self,
policy,
env,
gamma=0.99,
timesteps_per_batch=1024,
max_kl=0.01,
cg_iters=10,
lam=0.98,
entcoeff=0.0,
cg_damping=1e-2,
vf_stepsize=3e-4,
vf_iters=3,
verbose=0,
tensorboard_log=None,
_init_setup_model=True,
policy_kwargs=None,
full_tensorboard_log=False):
super(TRPO, self).__init__(policy=policy,
env=env,
verbose=verbose,
requires_vec_env=False,
_init_setup_model=_init_setup_model,
policy_kwargs=policy_kwargs)
self.using_gail = False
self.timesteps_per_batch = timesteps_per_batch
self.cg_iters = cg_iters
self.cg_damping = cg_damping
self.gamma = gamma
self.lam = lam
self.max_kl = max_kl
self.vf_iters = vf_iters
self.vf_stepsize = vf_stepsize
self.entcoeff = entcoeff
self.tensorboard_log = tensorboard_log
self.full_tensorboard_log = full_tensorboard_log
# GAIL Params
self.hidden_size_adversary = 100
self.adversary_entcoeff = 1e-3
self.expert_dataset = None
self.g_step = 1
self.d_step = 1
self.d_stepsize = 3e-4
self.graph = None
self.sess = None
self.policy_pi = None
self.loss_names = None
self.assign_old_eq_new = None
self.compute_losses = None
self.compute_lossandgrad = None
self.compute_fvp = None
self.compute_vflossandgrad = None
self.d_adam = None
self.vfadam = None
self.get_flat = None
self.set_from_flat = None
self.timed = None
self.allmean = None
self.nworkers = None
self.rank = None
self.reward_giver = None
self.step = None
self.proba_step = None
self.initial_state = None
self.params = None
self.summary = None
self.episode_reward = None
if _init_setup_model:
self.setup_model()
def _get_pretrain_placeholders(self):
policy = self.policy_pi
action_ph = policy.pdtype.sample_placeholder([None])
if isinstance(self.action_space, gym.spaces.Discrete):
return policy.obs_ph, action_ph, policy.policy
return policy.obs_ph, action_ph, policy.deterministic_action
def setup_model(self):
# prevent import loops
from stable_baselines.gail.adversary import TransitionClassifier
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the TRPO model must be " \
"an instance of common.policies.ActorCriticPolicy."
self.nworkers = MPI.COMM_WORLD.Get_size()
self.rank = MPI.COMM_WORLD.Get_rank()
np.set_printoptions(precision=3)
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf_util.single_threaded_session(graph=self.graph)
if self.using_gail:
self.reward_giver = TransitionClassifier(
self.observation_space,
self.action_space,
self.hidden_size_adversary,
entcoeff=self.adversary_entcoeff)
# Construct network for new policy
self.policy_pi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
# Network for old policy
with tf.variable_scope("oldpi", reuse=False):
old_policy = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
with tf.variable_scope("loss", reuse=False):
atarg = tf.placeholder(dtype=tf.float32, shape=[
None
]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32,
shape=[None]) # Empirical return
observation = self.policy_pi.obs_ph
action = self.policy_pi.pdtype.sample_placeholder([None])
kloldnew = old_policy.proba_distribution.kl(
self.policy_pi.proba_distribution)
ent = self.policy_pi.proba_distribution.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
entbonus = self.entcoeff * meanent
vferr = tf.reduce_mean(
tf.square(self.policy_pi.value_fn[:, 0] - ret))
# advantage * pnew / pold
ratio = tf.exp(
self.policy_pi.proba_distribution.logp(action) -
old_policy.proba_distribution.logp(action))
surrgain = tf.reduce_mean(ratio * atarg)
optimgain = surrgain + entbonus
losses = [optimgain, meankl, entbonus, surrgain, meanent]
self.loss_names = [
"optimgain", "meankl", "entloss", "surrgain", "entropy"
]
dist = meankl
all_var_list = tf_util.get_trainable_vars("model")
var_list = [
v for v in all_var_list
if "/vf" not in v.name and "/q/" not in v.name
]
vf_var_list = [
v for v in all_var_list
if "/pi" not in v.name and "/logstd" not in v.name
]
self.get_flat = tf_util.GetFlat(var_list, sess=self.sess)
self.set_from_flat = tf_util.SetFromFlat(var_list,
sess=self.sess)
klgrads = tf.gradients(dist, var_list)
flat_tangent = tf.placeholder(dtype=tf.float32,
shape=[None],
name="flat_tan")
shapes = [var.get_shape().as_list() for var in var_list]
start = 0
tangents = []
for shape in shapes:
var_size = tf_util.intprod(shape)
tangents.append(
tf.reshape(flat_tangent[start:start + var_size],
shape))
start += var_size
gvp = tf.add_n([
tf.reduce_sum(grad * tangent)
for (grad, tangent) in zipsame(klgrads, tangents)
]) # pylint: disable=E1111
fvp = tf_util.flatgrad(gvp, var_list)
tf.summary.scalar('entropy_loss', meanent)
tf.summary.scalar('policy_gradient_loss', optimgain)
tf.summary.scalar('value_function_loss', surrgain)
tf.summary.scalar('approximate_kullback-leiber', meankl)
tf.summary.scalar(
'loss',
optimgain + meankl + entbonus + surrgain + meanent)
self.assign_old_eq_new = \
tf_util.function([], [], updates=[tf.assign(oldv, newv) for (oldv, newv) in
zipsame(tf_util.get_globals_vars("oldpi"),
tf_util.get_globals_vars("model"))])
self.compute_losses = tf_util.function(
[observation, old_policy.obs_ph, action, atarg],
losses)
self.compute_fvp = tf_util.function([
flat_tangent, observation, old_policy.obs_ph, action,
atarg
], fvp)
self.compute_vflossandgrad = tf_util.function(
[observation, old_policy.obs_ph, ret],
tf_util.flatgrad(vferr, vf_var_list))
@contextmanager
def timed(msg):
if self.rank == 0 and self.verbose >= 1:
print(colorize(msg, color='magenta'))
start_time = time.time()
yield
print(
colorize("done in {:.3f} seconds".format(
(time.time() - start_time)),
color='magenta'))
else:
yield
def allmean(arr):
assert isinstance(arr, np.ndarray)
out = np.empty_like(arr)
MPI.COMM_WORLD.Allreduce(arr, out, op=MPI.SUM)
out /= self.nworkers
return out
tf_util.initialize(sess=self.sess)
th_init = self.get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
self.set_from_flat(th_init)
with tf.variable_scope("Adam_mpi", reuse=False):
self.vfadam = MpiAdam(vf_var_list, sess=self.sess)
if self.using_gail:
self.d_adam = MpiAdam(
self.reward_giver.get_trainable_variables(),
sess=self.sess)
self.d_adam.sync()
self.vfadam.sync()
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards',
tf.reduce_mean(ret))
tf.summary.scalar('learning_rate',
tf.reduce_mean(self.vf_stepsize))
tf.summary.scalar('advantage', tf.reduce_mean(atarg))
tf.summary.scalar('kl_clip_range',
tf.reduce_mean(self.max_kl))
if self.full_tensorboard_log:
tf.summary.histogram('discounted_rewards', ret)
tf.summary.histogram('learning_rate', self.vf_stepsize)
tf.summary.histogram('advantage', atarg)
tf.summary.histogram('kl_clip_range', self.max_kl)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', observation)
else:
tf.summary.histogram('observation', observation)
self.timed = timed
self.allmean = allmean
self.step = self.policy_pi.step
self.proba_step = self.policy_pi.proba_step
self.initial_state = self.policy_pi.initial_state
self.params = find_trainable_variables("model")
if self.using_gail:
self.params.extend(
self.reward_giver.get_trainable_variables())
self.summary = tf.summary.merge_all()
self.compute_lossandgrad = \
tf_util.function([observation, old_policy.obs_ph, action, atarg, ret],
[self.summary, tf_util.flatgrad(optimgain, var_list)] + losses)
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100,
tb_log_name="TRPO",
reset_num_timesteps=True):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn(seed)
with self.sess.as_default():
seg_gen = traj_segment_generator(
self.policy_pi,
self.env,
self.timesteps_per_batch,
reward_giver=self.reward_giver,
gail=self.using_gail)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
t_start = time.time()
len_buffer = deque(
maxlen=40) # rolling buffer for episode lengths
reward_buffer = deque(
maxlen=40) # rolling buffer for episode rewards
self.episode_reward = np.zeros((self.n_envs, ))
true_reward_buffer = None
if self.using_gail:
true_reward_buffer = deque(maxlen=40)
# Initialize dataloader
batchsize = self.timesteps_per_batch // self.d_step
self.expert_dataset.init_dataloader(batchsize)
# Stats not used for now
# TODO: replace with normal tb logging
# g_loss_stats = Stats(loss_names)
# d_loss_stats = Stats(reward_giver.loss_name)
# ep_stats = Stats(["True_rewards", "Rewards", "Episode_length"])
while True:
if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
if callback(locals(), globals()) is False:
break
if total_timesteps and timesteps_so_far >= total_timesteps:
break
logger.log("********** Iteration %i ************" %
iters_so_far)
def fisher_vector_product(vec):
return self.allmean(
self.compute_fvp(
vec, *fvpargs,
sess=self.sess)) + self.cg_damping * vec
# ------------------ Update G ------------------
logger.log("Optimizing Policy...")
# g_step = 1 when not using GAIL
mean_losses = None
vpredbefore = None
tdlamret = None
observation = None
action = None
seg = None
for k in range(self.g_step):
with self.timed("sampling"):
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, self.gamma, self.lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
observation, action, atarg, tdlamret = seg["ob"], seg[
"ac"], seg["adv"], seg["tdlamret"]
vpredbefore = seg[
"vpred"] # predicted value function before update
atarg = (atarg - atarg.mean()) / atarg.std(
) # standardized advantage function estimate
# true_rew is the reward without discount
if writer is not None:
self.episode_reward = total_episode_reward_logger(
self.episode_reward, seg["true_rew"].reshape(
(self.n_envs, -1)), seg["dones"].reshape(
(self.n_envs, -1)), writer,
self.num_timesteps)
args = seg["ob"], seg["ob"], seg["ac"], atarg
fvpargs = [arr[::5] for arr in args]
self.assign_old_eq_new(sess=self.sess)
with self.timed("computegrad"):
steps = self.num_timesteps + (k + 1) * (
seg["total_timestep"] / self.g_step)
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata(
) if self.full_tensorboard_log else None
# run loss backprop with summary, and save the metadata (memory, compute time, ...)
if writer is not None:
summary, grad, *lossbefore = self.compute_lossandgrad(
*args,
tdlamret,
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
if self.full_tensorboard_log:
writer.add_run_metadata(
run_metadata, 'step%d' % steps)
writer.add_summary(summary, steps)
else:
_, grad, *lossbefore = self.compute_lossandgrad(
*args,
tdlamret,
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
lossbefore = self.allmean(np.array(lossbefore))
grad = self.allmean(grad)
if np.allclose(grad, 0):
logger.log("Got zero gradient. not updating")
else:
with self.timed("conjugate_gradient"):
stepdir = conjugate_gradient(
fisher_vector_product,
grad,
cg_iters=self.cg_iters,
verbose=self.rank == 0
and self.verbose >= 1)
assert np.isfinite(stepdir).all()
shs = .5 * stepdir.dot(
fisher_vector_product(stepdir))
# abs(shs) to avoid taking square root of negative values
lagrange_multiplier = np.sqrt(
abs(shs) / self.max_kl)
# logger.log("lagrange multiplier:", lm, "gnorm:", np.linalg.norm(g))
fullstep = stepdir / lagrange_multiplier
expectedimprove = grad.dot(fullstep)
surrbefore = lossbefore[0]
stepsize = 1.0
thbefore = self.get_flat()
thnew = None
for _ in range(10):
thnew = thbefore + fullstep * stepsize
self.set_from_flat(thnew)
mean_losses = surr, kl_loss, *_ = self.allmean(
np.array(
self.compute_losses(*args,
sess=self.sess)))
improve = surr - surrbefore
logger.log("Expected: %.3f Actual: %.3f" %
(expectedimprove, improve))
if not np.isfinite(mean_losses).all():
logger.log(
"Got non-finite value of losses -- bad!"
)
elif kl_loss > self.max_kl * 1.5:
logger.log(
"violated KL constraint. shrinking step."
)
elif improve < 0:
logger.log(
"surrogate didn't improve. shrinking step."
)
else:
logger.log("Stepsize OK!")
break
stepsize *= .5
else:
logger.log("couldn't compute a good step")
self.set_from_flat(thbefore)
if self.nworkers > 1 and iters_so_far % 20 == 0:
# list of tuples
paramsums = MPI.COMM_WORLD.allgather(
(thnew.sum(), self.vfadam.getflat().sum()))
assert all(
np.allclose(ps, paramsums[0])
for ps in paramsums[1:])
with self.timed("vf"):
for _ in range(self.vf_iters):
# NOTE: for recurrent policies, use shuffle=False?
for (mbob, mbret) in dataset.iterbatches(
(seg["ob"], seg["tdlamret"]),
include_final_partial_batch=False,
batch_size=128,
shuffle=True):
grad = self.allmean(
self.compute_vflossandgrad(
mbob, mbob, mbret, sess=self.sess))
self.vfadam.update(grad, self.vf_stepsize)
for (loss_name, loss_val) in zip(self.loss_names,
mean_losses):
logger.record_tabular(loss_name, loss_val)
logger.record_tabular(
"explained_variance_tdlam_before",
explained_variance(vpredbefore, tdlamret))
if self.using_gail:
# ------------------ Update D ------------------
logger.log("Optimizing Discriminator...")
logger.log(fmt_row(13, self.reward_giver.loss_name))
assert len(observation) == self.timesteps_per_batch
batch_size = self.timesteps_per_batch // self.d_step
# NOTE: uses only the last g step for observation
d_losses = [
] # list of tuples, each of which gives the loss for a minibatch
# NOTE: for recurrent policies, use shuffle=False?
for ob_batch, ac_batch in dataset.iterbatches(
(observation, action),
include_final_partial_batch=False,
batch_size=batch_size,
shuffle=True):
ob_expert, ac_expert = self.expert_dataset.get_next_batch(
)
# update running mean/std for reward_giver
if self.reward_giver.normalize:
self.reward_giver.obs_rms.update(
np.concatenate((ob_batch, ob_expert), 0))
# Reshape actions if needed when using discrete actions
if isinstance(self.action_space,
gym.spaces.Discrete):
if len(ac_batch.shape) == 2:
ac_batch = ac_batch[:, 0]
if len(ac_expert.shape) == 2:
ac_expert = ac_expert[:, 0]
*newlosses, grad = self.reward_giver.lossandgrad(
ob_batch, ac_batch, ob_expert, ac_expert)
self.d_adam.update(self.allmean(grad),
self.d_stepsize)
d_losses.append(newlosses)
logger.log(fmt_row(13, np.mean(d_losses, axis=0)))
# lr: lengths and rewards
lr_local = (seg["ep_lens"], seg["ep_rets"],
seg["ep_true_rets"]) # local values
list_lr_pairs = MPI.COMM_WORLD.allgather(
lr_local) # list of tuples
lens, rews, true_rets = map(flatten_lists,
zip(*list_lr_pairs))
true_reward_buffer.extend(true_rets)
else:
# lr: lengths and rewards
lr_local = (seg["ep_lens"], seg["ep_rets"]
) # local values
list_lr_pairs = MPI.COMM_WORLD.allgather(
lr_local) # list of tuples
lens, rews = map(flatten_lists, zip(*list_lr_pairs))
len_buffer.extend(lens)
reward_buffer.extend(rews)
if len(len_buffer) > 0:
logger.record_tabular("EpLenMean", np.mean(len_buffer))
logger.record_tabular("EpRewMean",
np.mean(reward_buffer))
if self.using_gail:
logger.record_tabular("EpTrueRewMean",
np.mean(true_reward_buffer))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
current_it_timesteps = MPI.COMM_WORLD.allreduce(
seg["total_timestep"])
timesteps_so_far += current_it_timesteps
self.num_timesteps += current_it_timesteps
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", self.num_timesteps)
logger.record_tabular("TimeElapsed", time.time() - t_start)
if self.verbose >= 1 and self.rank == 0:
logger.dump_tabular()
return self
def save(self, save_path):
if self.using_gail and self.expert_dataset is not None:
# Exit processes to pickle the dataset
self.expert_dataset.prepare_pickling()
data = {
"gamma": self.gamma,
"timesteps_per_batch": self.timesteps_per_batch,
"max_kl": self.max_kl,
"cg_iters": self.cg_iters,
"lam": self.lam,
"entcoeff": self.entcoeff,
"cg_damping": self.cg_damping,
"vf_stepsize": self.vf_stepsize,
"vf_iters": self.vf_iters,
"hidden_size_adversary": self.hidden_size_adversary,
"adversary_entcoeff": self.adversary_entcoeff,
"expert_dataset": self.expert_dataset,
"g_step": self.g_step,
"d_step": self.d_step,
"d_stepsize": self.d_stepsize,
"using_gail": self.using_gail,
"verbose": self.verbose,
"policy": self.policy,
"observation_space": self.observation_space,
"action_space": self.action_space,
"n_envs": self.n_envs,
"_vectorize_action": self._vectorize_action,
"policy_kwargs": self.policy_kwargs
}
params = self.sess.run(self.params)
self._save_to_file(save_path, data=data, params=params)
def setup_model(self):
# prevent import loops
from stable_baselines.gail.adversary import TransitionClassifier
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the TRPO model must be " \
"an instance of common.policies.ActorCriticPolicy."
self.nworkers = MPI.COMM_WORLD.Get_size()
self.rank = MPI.COMM_WORLD.Get_rank()
np.set_printoptions(precision=3)
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf_util.single_threaded_session(graph=self.graph)
if self.using_gail:
self.reward_giver = TransitionClassifier(
self.observation_space,
self.action_space,
self.hidden_size_adversary,
entcoeff=self.adversary_entcoeff)
# Construct network for new policy
self.policy_pi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
# Network for old policy
with tf.variable_scope("oldpi", reuse=False):
old_policy = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
with tf.variable_scope("loss", reuse=False):
atarg = tf.placeholder(dtype=tf.float32, shape=[
None
]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32,
shape=[None]) # Empirical return
observation = self.policy_pi.obs_ph
action = self.policy_pi.pdtype.sample_placeholder([None])
kloldnew = old_policy.proba_distribution.kl(
self.policy_pi.proba_distribution)
ent = self.policy_pi.proba_distribution.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
entbonus = self.entcoeff * meanent
vferr = tf.reduce_mean(
tf.square(self.policy_pi.value_fn[:, 0] - ret))
# advantage * pnew / pold
ratio = tf.exp(
self.policy_pi.proba_distribution.logp(action) -
old_policy.proba_distribution.logp(action))
surrgain = tf.reduce_mean(ratio * atarg)
optimgain = surrgain + entbonus
losses = [optimgain, meankl, entbonus, surrgain, meanent]
self.loss_names = [
"optimgain", "meankl", "entloss", "surrgain", "entropy"
]
dist = meankl
all_var_list = tf_util.get_trainable_vars("model")
var_list = [
v for v in all_var_list
if "/vf" not in v.name and "/q/" not in v.name
]
vf_var_list = [
v for v in all_var_list
if "/pi" not in v.name and "/logstd" not in v.name
]
self.get_flat = tf_util.GetFlat(var_list, sess=self.sess)
self.set_from_flat = tf_util.SetFromFlat(var_list,
sess=self.sess)
klgrads = tf.gradients(dist, var_list)
flat_tangent = tf.placeholder(dtype=tf.float32,
shape=[None],
name="flat_tan")
shapes = [var.get_shape().as_list() for var in var_list]
start = 0
tangents = []
for shape in shapes:
var_size = tf_util.intprod(shape)
tangents.append(
tf.reshape(flat_tangent[start:start + var_size],
shape))
start += var_size
gvp = tf.add_n([
tf.reduce_sum(grad * tangent)
for (grad, tangent) in zipsame(klgrads, tangents)
]) # pylint: disable=E1111
fvp = tf_util.flatgrad(gvp, var_list)
tf.summary.scalar('entropy_loss', meanent)
tf.summary.scalar('policy_gradient_loss', optimgain)
tf.summary.scalar('value_function_loss', surrgain)
tf.summary.scalar('approximate_kullback-leiber', meankl)
tf.summary.scalar(
'loss',
optimgain + meankl + entbonus + surrgain + meanent)
self.assign_old_eq_new = \
tf_util.function([], [], updates=[tf.assign(oldv, newv) for (oldv, newv) in
zipsame(tf_util.get_globals_vars("oldpi"),
tf_util.get_globals_vars("model"))])
self.compute_losses = tf_util.function(
[observation, old_policy.obs_ph, action, atarg],
losses)
self.compute_fvp = tf_util.function([
flat_tangent, observation, old_policy.obs_ph, action,
atarg
], fvp)
self.compute_vflossandgrad = tf_util.function(
[observation, old_policy.obs_ph, ret],
tf_util.flatgrad(vferr, vf_var_list))
@contextmanager
def timed(msg):
if self.rank == 0 and self.verbose >= 1:
print(colorize(msg, color='magenta'))
start_time = time.time()
yield
print(
colorize("done in {:.3f} seconds".format(
(time.time() - start_time)),
color='magenta'))
else:
yield
def allmean(arr):
assert isinstance(arr, np.ndarray)
out = np.empty_like(arr)
MPI.COMM_WORLD.Allreduce(arr, out, op=MPI.SUM)
out /= self.nworkers
return out
tf_util.initialize(sess=self.sess)
th_init = self.get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
self.set_from_flat(th_init)
with tf.variable_scope("Adam_mpi", reuse=False):
self.vfadam = MpiAdam(vf_var_list, sess=self.sess)
if self.using_gail:
self.d_adam = MpiAdam(
self.reward_giver.get_trainable_variables(),
sess=self.sess)
self.d_adam.sync()
self.vfadam.sync()
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards',
tf.reduce_mean(ret))
tf.summary.scalar('learning_rate',
tf.reduce_mean(self.vf_stepsize))
tf.summary.scalar('advantage', tf.reduce_mean(atarg))
tf.summary.scalar('kl_clip_range',
tf.reduce_mean(self.max_kl))
if self.full_tensorboard_log:
tf.summary.histogram('discounted_rewards', ret)
tf.summary.histogram('learning_rate', self.vf_stepsize)
tf.summary.histogram('advantage', atarg)
tf.summary.histogram('kl_clip_range', self.max_kl)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', observation)
else:
tf.summary.histogram('observation', observation)
self.timed = timed
self.allmean = allmean
self.step = self.policy_pi.step
self.proba_step = self.policy_pi.proba_step
self.initial_state = self.policy_pi.initial_state
self.params = find_trainable_variables("model")
if self.using_gail:
self.params.extend(
self.reward_giver.get_trainable_variables())
self.summary = tf.summary.merge_all()
self.compute_lossandgrad = \
tf_util.function([observation, old_policy.obs_ph, action, atarg, ret],
[self.summary, tf_util.flatgrad(optimgain, var_list)] + losses)
def setup_model(self):
# prevent import loops
from stable_baselines.gail.adversary import TransitionClassifier
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the TRPO model must be " \
"an instance of common.policies.ActorCriticPolicy."
self.nworkers = MPI.COMM_WORLD.Get_size()
self.rank = MPI.COMM_WORLD.Get_rank()
np.set_printoptions(precision=3)
self.graph = tf.Graph()
with self.graph.as_default():
self.set_random_seed(self.seed)
self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess, graph=self.graph)
if self.using_gail:
self.reward_giver = TransitionClassifier(self.observation_space, self.action_space,
self.hidden_size_adversary,
entcoeff=self.adversary_entcoeff)
# Penalty related variable
with tf.variable_scope('penalty'):
cur_cost_ph = tf.placeholder(dtype=tf.float32, shape=[None]) # episodic cost
param_init = np.log(max(np.exp(self.penalty_init) - 1, 1e-8))
penalty_param = tf.get_variable('penalty_param',
initializer=float(param_init),
trainable=True,
dtype=tf.float32)
penalty = tf.nn.softplus(penalty_param)
penalty_loss = tf.reduce_mean(-penalty_param * (cur_cost_ph - self.cost_lim))
# Construct network for new policy
self.policy_pi = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs, 1,
None, reuse=False, **self.policy_kwargs)
# Network for old policy
with tf.variable_scope("oldpi", reuse=False):
old_policy = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs, 1,
None, reuse=False, **self.policy_kwargs)
# # Network for safety value function
# with tf.variable_Scope("vc",reuse=False):
# self.cost_value = MLPValue(self.sess, self.observation_spacem, self.n_envs, 1, None)
with tf.variable_scope("loss", reuse=False):
atarg = tf.placeholder(dtype=tf.float32, shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
catarg = tf.placeholder(dtype=tf.float32, shape=[None]) # Target cost advantage function
cret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical cost
observation = self.policy_pi.obs_ph
action = self.policy_pi.pdtype.sample_placeholder([None])
kloldnew = old_policy.proba_distribution.kl(self.policy_pi.proba_distribution)
ent = self.policy_pi.proba_distribution.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
entbonus = self.entcoeff * meanent
vferr = tf.reduce_mean(tf.square(self.policy_pi.value_flat - ret))
vcerr = tf.reduce_mean(tf.square(self.policy_pi.vcf_flat - cret))
# advantage * pnew / pold
ratio = tf.exp(self.policy_pi.proba_distribution.logp(action) -
old_policy.proba_distribution.logp(action))
surrgain = tf.reduce_mean(ratio * atarg)
# Surrogate for cost function
surrcost = tf.reduce_mean(ratio * catarg)
optimgain = surrgain + entbonus
# Include surr_cost in pi_objective
optimgain -= penalty * surrcost
optimgain /= (1 + penalty)
# # Loss function for pi is negative of pi_objective
# optimgain = -optimgain # Should we??
losses = [optimgain, meankl, entbonus, surrgain, meanent, surrcost]
self.loss_names = ["optimgain", "meankl", "entloss", "surrgain", "entropy", "surrcost"]
dist = meankl
all_var_list = tf_util.get_trainable_vars("model")
var_list = [v for v in all_var_list if "/vf" not in v.name and "/q/" not in v.name and "/vcf" not in v.name] # policy parameters
vf_var_list = [v for v in all_var_list if "/pi" not in v.name and "/logstd" not in v.name and "/vcf" not in v.name] # value parameters
vcf_var_list = [v for v in all_var_list if "/pi" not in v.name and "/logstd" not in v.name and "/vf" not in v.name] # cost value parameters
self.get_flat = tf_util.GetFlat(var_list, sess=self.sess)
self.set_from_flat = tf_util.SetFromFlat(var_list, sess=self.sess)
klgrads = tf.gradients(dist, var_list)
flat_tangent = tf.placeholder(dtype=tf.float32, shape=[None], name="flat_tan")
shapes = [var.get_shape().as_list() for var in var_list]
start = 0
tangents = []
for shape in shapes:
var_size = tf_util.intprod(shape)
tangents.append(tf.reshape(flat_tangent[start: start + var_size], shape))
start += var_size
gvp = tf.add_n([tf.reduce_sum(grad * tangent)
for (grad, tangent) in zipsame(klgrads, tangents)]) # pylint: disable=E1111
# Fisher vector products
fvp = tf_util.flatgrad(gvp, var_list)
tf.summary.scalar('penalty_loss', penalty_loss)
tf.summary.scalar('entropy_loss', meanent)
tf.summary.scalar('policy_gradient_loss', optimgain)
tf.summary.scalar('value_function_loss', surrgain)
tf.summary.scalar('constraint_cost_function_loss', surrcost)
tf.summary.scalar('approximate_kullback-leibler', meankl)
tf.summary.scalar('loss', optimgain + meankl + entbonus + surrgain + meanent + surrcost + penalty_loss)
self.assign_old_eq_new = \
tf_util.function([], [], updates=[tf.assign(oldv, newv) for (oldv, newv) in
zipsame(tf_util.get_globals_vars("oldpi"),
tf_util.get_globals_vars("model"))])
self.compute_losses = tf_util.function([observation, old_policy.obs_ph, action, atarg, catarg], losses)
self.compute_fvp = tf_util.function([flat_tangent, observation, old_policy.obs_ph, action, atarg, catarg],
fvp) # Why need all inputs? Might for implementation easiness
# self.compute_vflossandgrad = tf_util.function([observation, old_policy.obs_ph, ret],
# tf_util.flatgrad(vferr, vf_var_list)) # Why need old_policy.obs_ph? Doesn't seem to be used
# self.compute_vcflossandgrad = tf_util.function([observation, old_policy.obs_ph, cret],
# tf_util.flatgrad(vcerr, vcf_var_list))
self.compute_vflossandgrad = tf_util.function([observation, old_policy.obs_ph, ret, cret],
[tf_util.flatgrad(vferr, vf_var_list), tf_util.flatgrad(vcerr, vcf_var_list)])
self.compute_lagrangiangrad = tf_util.function([cur_cost_ph],
tf_util.flatgrad(penalty_loss, [penalty_param]))
@contextmanager
def timed(msg):
if self.rank == 0 and self.verbose >= 1:
print(colorize(msg, color='magenta'))
start_time = time.time()
yield
print(colorize("done in {:.3f} seconds".format((time.time() - start_time)),
color='magenta'))
else:
yield
def allmean(arr):
assert isinstance(arr, np.ndarray)
out = np.empty_like(arr)
MPI.COMM_WORLD.Allreduce(arr, out, op=MPI.SUM)
out /= self.nworkers
return out
tf_util.initialize(sess=self.sess)
th_init = self.get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
self.set_from_flat(th_init)
with tf.variable_scope("Adam_mpi", reuse=False):
self.vfadam = MpiAdam(vf_var_list, sess=self.sess)
if self.using_gail:
self.d_adam = MpiAdam(self.reward_giver.get_trainable_variables(), sess=self.sess)
self.d_adam.sync()
self.vfadam.sync()
# optimizer for constraint costs value function
self.vcadam = MpiAdam(vcf_var_list, sess=self.sess)
self.vcadam.sync()
# optimizer for lagragian value of safe RL
self.penaltyadam = MpiAdam([penalty_param], sess=self.sess)
self.penaltyadam.sync()
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards', tf.reduce_mean(ret))
tf.summary.scalar('discounted_costs', tf.reduce_mean(cret))
tf.summary.scalar('learning_rate', tf.reduce_mean(self.vf_stepsize))
tf.summary.scalar('advantage', tf.reduce_mean(atarg))
tf.summary.scalar('cost_advantage', tf.reduce_mean(catarg))
tf.summary.scalar('kl_clip_range', tf.reduce_mean(self.max_kl))
if self.full_tensorboard_log:
tf.summary.histogram('discounted_rewards', ret)
tf.summary.histogram('discounted_rewards', cret)
tf.summary.histogram('learning_rate', self.vf_stepsize)
tf.summary.histogram('penalty_learning_rate', self.penalty_lr)
tf.summary.histogram('advantage', atarg)
tf.summary.histogram('cost_advantage', catarg)
tf.summary.histogram('kl_clip_range', self.max_kl)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', observation)
else:
tf.summary.histogram('observation', observation)
self.timed = timed
self.allmean = allmean
self.step = self.policy_pi.step
self.proba_step = self.policy_pi.proba_step
self.initial_state = self.policy_pi.initial_state
self.params = tf_util.get_trainable_vars("model") + tf_util.get_trainable_vars("oldpi")
if self.using_gail:
self.params.extend(self.reward_giver.get_trainable_variables())
self.summary = tf.summary.merge_all()
self.compute_lossandgrad = \
tf_util.function([observation, old_policy.obs_ph, action, atarg, catarg, ret, cret, cur_cost_ph],
[self.summary, tf_util.flatgrad(optimgain, var_list)] + losses)
class PPO2(ActorCriticRLModel):
"""
Proximal Policy Optimization algorithm (GPU version).
Paper: https://arxiv.org/abs/1707.06347
:param policy: (ActorCriticPolicy or str) The policy model to use (MlpPolicy, CnnPolicy, CnnLstmPolicy, ...)
:param env: (Gym environment or str) The environment to learn from (if registered in Gym, can be str)
:param gamma: (float) Discount factor
:param n_steps: (int) The number of steps to run for each environment per update
(i.e. batch size is n_steps * n_env where n_env is number of environment copies running in parallel)
:param ent_coef: (float) Entropy coefficient for the loss calculation
:param learning_rate: (float or callable) The learning rate, it can be a function
:param vf_coef: (float) Value function coefficient for the loss calculation
:param max_grad_norm: (float) The maximum value for the gradient clipping
:param lam: (float) Factor for trade-off of bias vs variance for Generalized Advantage Estimator
:param nminibatches: (int) Number of training minibatches per update. For recurrent policies,
the number of environments run in parallel should be a multiple of nminibatches.
:param noptepochs: (int) Number of epoch when optimizing the surrogate
:param cliprange: (float or callable) Clipping parameter, it can be a function
:param cliprange_vf: (float or callable) Clipping parameter for the value function, it can be a function.
This is a parameter specific to the OpenAI implementation. If None is passed (default),
then `cliprange` (that is used for the policy) will be used.
IMPORTANT: this clipping depends on the reward scaling.
To deactivate value function clipping (and recover the original PPO implementation),
you have to pass a negative value (e.g. -1).
:param verbose: (int) the verbosity level: 0 none, 1 training information, 2 tensorflow debug
:param tensorboard_log: (str) the log location for tensorboard (if None, no logging)
:param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance
:param policy_kwargs: (dict) additional arguments to be passed to the policy on creation
:param full_tensorboard_log: (bool) enable additional logging when using tensorboard
WARNING: this logging can take a lot of space quickly
:param seed: (int) Seed for the pseudo-random generators (python, numpy, tensorflow).
If None (default), use random seed. Note that if you want completely deterministic
results, you must set `n_cpu_tf_sess` to 1.
:param n_cpu_tf_sess: (int) The number of threads for TensorFlow operations
If None, the number of cpu of the current machine will be used.
"""
def __init__(self, policy, env, gamma=0.99, n_steps=128, ent_coef=0.01, learning_rate=2.5e-4, vf_coef=0.5,
max_grad_norm=0.5, lam=0.95, nminibatches=4, noptepochs=4, cliprange=0.2, cliprange_vf=None,
verbose=0, tensorboard_log=None, _init_setup_model=True, policy_kwargs=None,
full_tensorboard_log=False, seed=None, n_cpu_tf_sess=None):
super(PPO2, self).__init__(policy=policy, env=env, verbose=verbose, requires_vec_env=True,
_init_setup_model=_init_setup_model, policy_kwargs=policy_kwargs,
seed=seed, n_cpu_tf_sess=n_cpu_tf_sess)
self.learning_rate = learning_rate
self.cliprange = cliprange
self.cliprange_vf = cliprange_vf
self.n_steps = n_steps
self.ent_coef = ent_coef
self.vf_coef = vf_coef
self.max_grad_norm = max_grad_norm
self.gamma = gamma
self.lam = lam
self.nminibatches = nminibatches
self.noptepochs = noptepochs
self.tensorboard_log = tensorboard_log
self.full_tensorboard_log = full_tensorboard_log
# GAIL Params
self.hidden_size_adversary = 100
self.adversary_entcoeff = 1e-3
self.expert_dataset = None
self.g_step = 1
self.d_step = 1
self.d_stepsize = 3e-4
self.graph = None
self.sess = None
self.action_ph = None
self.advs_ph = None
self.rewards_ph = None
self.old_neglog_pac_ph = None
self.old_vpred_ph = None
self.learning_rate_ph = None
self.clip_range_ph = None
self.entropy = None
self.vf_loss = None
self.pg_loss = None
self.approxkl = None
self.clipfrac = None
self.params = None
self._train = None
self.loss_names = None
self.train_model = None
self.act_model = None
self.step = None
self.proba_step = None
self.value = None
self.initial_state = None
self.n_batch = None
self.summary = None
self.episode_reward = None
self.reward_giver = None
self.using_gail = False
if _init_setup_model:
self.setup_model()
def _get_pretrain_placeholders(self):
policy = self.act_model
if isinstance(self.action_space, gym.spaces.Discrete):
return policy.obs_ph, self.action_ph, policy.policy
return policy.obs_ph, self.action_ph, policy.deterministic_action
def setup_model(self):
from stable_baselines.gail.adversary import TransitionClassifier
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the PPO2 model must be " \
"an instance of common.policies.ActorCriticPolicy."
self.n_batch = self.n_envs * self.n_steps
self.graph = tf.Graph()
with self.graph.as_default():
self.set_random_seed(self.seed)
self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess, graph=self.graph)
if self.using_gail:
self.reward_giver = TransitionClassifier(self.observation_space, self.action_space,
self.hidden_size_adversary,
normalize=False,
entcoeff=self.adversary_entcoeff)
n_batch_step = None
n_batch_train = None
if issubclass(self.policy, RecurrentActorCriticPolicy):
assert self.n_envs % self.nminibatches == 0, "For recurrent policies, "\
"the number of environments run in parallel should be a multiple of nminibatches."
n_batch_step = self.n_envs
n_batch_train = self.n_batch // self.nminibatches
act_model = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs, 1,
n_batch_step, reuse=False, **self.policy_kwargs)
with tf.variable_scope("train_model", reuse=True,
custom_getter=tf_util.outer_scope_getter("train_model")):
train_model = self.policy(self.sess, self.observation_space, self.action_space,
self.n_envs // self.nminibatches, self.n_steps, n_batch_train,
reuse=True, **self.policy_kwargs)
with tf.variable_scope("loss", reuse=False):
self.action_ph = train_model.pdtype.sample_placeholder([None], name="action_ph")
self.advs_ph = tf.placeholder(tf.float32, [None], name="advs_ph")
self.rewards_ph = tf.placeholder(tf.float32, [None], name="rewards_ph")
self.old_neglog_pac_ph = tf.placeholder(tf.float32, [None], name="old_neglog_pac_ph")
self.old_vpred_ph = tf.placeholder(tf.float32, [None], name="old_vpred_ph")
self.learning_rate_ph = tf.placeholder(tf.float32, [], name="learning_rate_ph")
self.clip_range_ph = tf.placeholder(tf.float32, [], name="clip_range_ph")
neglogpac = train_model.proba_distribution.neglogp(self.action_ph)
self.entropy = tf.reduce_mean(train_model.proba_distribution.entropy())
vpred = train_model.value_flat
# Value function clipping: not present in the original PPO
if self.cliprange_vf is None:
# Default behavior (legacy from OpenAI baselines):
# use the same clipping as for the policy
self.clip_range_vf_ph = self.clip_range_ph
self.cliprange_vf = self.cliprange
elif isinstance(self.cliprange_vf, (float, int)) and self.cliprange_vf < 0:
# Original PPO implementation: no value function clipping
self.clip_range_vf_ph = None
else:
# Last possible behavior: clipping range
# specific to the value function
self.clip_range_vf_ph = tf.placeholder(tf.float32, [], name="clip_range_vf_ph")
if self.clip_range_vf_ph is None:
# No clipping
vpred_clipped = train_model.value_flat
else:
# Clip the different between old and new value
# NOTE: this depends on the reward scaling
vpred_clipped = self.old_vpred_ph + \
tf.clip_by_value(train_model.value_flat - self.old_vpred_ph,
- self.clip_range_vf_ph, self.clip_range_vf_ph)
vf_losses1 = tf.square(vpred - self.rewards_ph)
vf_losses2 = tf.square(vpred_clipped - self.rewards_ph)
self.vf_loss = .5 * tf.reduce_mean(tf.maximum(vf_losses1, vf_losses2))
ratio = tf.exp(self.old_neglog_pac_ph - neglogpac)
pg_losses = -self.advs_ph * ratio
pg_losses2 = -self.advs_ph * tf.clip_by_value(ratio, 1.0 - self.clip_range_ph, 1.0 +
self.clip_range_ph)
self.pg_loss = tf.reduce_mean(tf.maximum(pg_losses, pg_losses2))
self.approxkl = .5 * tf.reduce_mean(tf.square(neglogpac - self.old_neglog_pac_ph))
self.clipfrac = tf.reduce_mean(tf.cast(tf.greater(tf.abs(ratio - 1.0),
self.clip_range_ph), tf.float32))
loss = self.pg_loss - self.entropy * self.ent_coef + self.vf_loss * self.vf_coef
tf.summary.scalar('entropy_loss', self.entropy)
tf.summary.scalar('policy_gradient_loss', self.pg_loss)
tf.summary.scalar('value_function_loss', self.vf_loss)
tf.summary.scalar('approximate_kullback-leibler', self.approxkl)
tf.summary.scalar('clip_factor', self.clipfrac)
tf.summary.scalar('loss', loss)
with tf.variable_scope('model'):
self.params = tf.trainable_variables()
# Update params with GAIL params too
if self.using_gail:
self.params.extend(self.reward_giver.get_trainable_variables())
if self.full_tensorboard_log:
for var in self.params:
tf.summary.histogram(var.name, var)
grads = tf.gradients(loss, self.params)
if self.max_grad_norm is not None:
grads, _grad_norm = tf.clip_by_global_norm(grads, self.max_grad_norm)
grads = list(zip(grads, self.params))
trainer = tf.train.AdamOptimizer(learning_rate=self.learning_rate_ph, epsilon=1e-5)
self._train = trainer.apply_gradients(grads)
self.loss_names = ['policy_loss', 'value_loss', 'policy_entropy', 'approxkl', 'clipfrac']
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards', tf.reduce_mean(self.rewards_ph))
tf.summary.scalar('learning_rate', tf.reduce_mean(self.learning_rate_ph))
tf.summary.scalar('advantage', tf.reduce_mean(self.advs_ph))
tf.summary.scalar('clip_range', tf.reduce_mean(self.clip_range_ph))
if self.clip_range_vf_ph is not None:
tf.summary.scalar('clip_range_vf', tf.reduce_mean(self.clip_range_vf_ph))
tf.summary.scalar('old_neglog_action_probabilty', tf.reduce_mean(self.old_neglog_pac_ph))
tf.summary.scalar('old_value_pred', tf.reduce_mean(self.old_vpred_ph))
if self.full_tensorboard_log:
tf.summary.histogram('discounted_rewards', self.rewards_ph)
tf.summary.histogram('learning_rate', self.learning_rate_ph)
tf.summary.histogram('advantage', self.advs_ph)
tf.summary.histogram('clip_range', self.clip_range_ph)
tf.summary.histogram('old_neglog_action_probabilty', self.old_neglog_pac_ph)
tf.summary.histogram('old_value_pred', self.old_vpred_ph)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', train_model.obs_ph)
else:
tf.summary.histogram('observation', train_model.obs_ph)
self.train_model = train_model
self.act_model = act_model
self.step = act_model.step
self.proba_step = act_model.proba_step
self.value = act_model.value
self.initial_state = act_model.initial_state
tf.global_variables_initializer().run(session=self.sess) # pylint: disable=E1101
self.summary = tf.summary.merge_all()
def _train_step(self, learning_rate, cliprange, obs, returns, masks, actions, values, neglogpacs, update,
writer, states=None, cliprange_vf=None):
"""
Training of PPO2 Algorithm
:param learning_rate: (float) learning rate
:param cliprange: (float) Clipping factor
:param obs: (np.ndarray) The current observation of the environment
:param returns: (np.ndarray) the rewards
:param masks: (np.ndarray) The last masks for done episodes (used in recurent policies)
:param actions: (np.ndarray) the actions
:param values: (np.ndarray) the values
:param neglogpacs: (np.ndarray) Negative Log-likelihood probability of Actions
:param update: (int) the current step iteration
:param writer: (TensorFlow Summary.writer) the writer for tensorboard
:param states: (np.ndarray) For recurrent policies, the internal state of the recurrent model
:return: policy gradient loss, value function loss, policy entropy,
approximation of kl divergence, updated clipping range, training update operation
:param cliprange_vf: (float) Clipping factor for the value function
"""
advs = returns - values
advs = (advs - advs.mean()) / (advs.std() + 1e-8)
td_map = {self.train_model.obs_ph: obs, self.action_ph: actions,
self.advs_ph: advs, self.rewards_ph: returns,
self.learning_rate_ph: learning_rate, self.clip_range_ph: cliprange,
self.old_neglog_pac_ph: neglogpacs, self.old_vpred_ph: values}
if states is not None:
td_map[self.train_model.states_ph] = states
td_map[self.train_model.dones_ph] = masks
if cliprange_vf is not None and cliprange_vf >= 0:
td_map[self.clip_range_vf_ph] = cliprange_vf
if states is None:
update_fac = self.n_batch // self.nminibatches // self.noptepochs + 1
else:
update_fac = self.n_batch // self.nminibatches // self.noptepochs // self.n_steps + 1
if writer is not None:
# run loss backprop with summary, but once every 10 runs save the metadata (memory, compute time, ...)
if self.full_tensorboard_log and (1 + update) % 10 == 0:
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, policy_loss, value_loss, policy_entropy, approxkl, clipfrac, _ = self.sess.run(
[self.summary, self.pg_loss, self.vf_loss, self.entropy, self.approxkl, self.clipfrac, self._train],
td_map, options=run_options, run_metadata=run_metadata)
writer.add_run_metadata(run_metadata, 'step%d' % (update * update_fac))
else:
summary, policy_loss, value_loss, policy_entropy, approxkl, clipfrac, _ = self.sess.run(
[self.summary, self.pg_loss, self.vf_loss, self.entropy, self.approxkl, self.clipfrac, self._train],
td_map)
writer.add_summary(summary, (update * update_fac))
else:
policy_loss, value_loss, policy_entropy, approxkl, clipfrac, _ = self.sess.run(
[self.pg_loss, self.vf_loss, self.entropy, self.approxkl, self.clipfrac, self._train], td_map)
return policy_loss, value_loss, policy_entropy, approxkl, clipfrac
def learn(self, total_timesteps, callback=None, log_interval=1, tb_log_name="PPO2",
reset_num_timesteps=True):
# Transform to callable if needed
self.learning_rate = get_schedule_fn(self.learning_rate)
self.cliprange = get_schedule_fn(self.cliprange)
cliprange_vf = get_schedule_fn(self.cliprange_vf)
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn()
runner = Runner(env=self.env, model=self, n_steps=self.n_steps, gamma=self.gamma, lam=self.lam)
self.episode_reward = np.zeros((self.n_envs,))
ep_info_buf = deque(maxlen=100)
t_first_start = time.time()
n_updates = total_timesteps // self.n_batch
for update in range(1, n_updates + 1):
assert self.n_batch % self.nminibatches == 0
batch_size = self.n_batch // self.nminibatches
t_start = time.time()
frac = 1.0 - (update - 1.0) / n_updates
lr_now = self.learning_rate(frac)
cliprange_now = self.cliprange(frac)
cliprange_vf_now = cliprange_vf(frac)
# true_reward is the reward without discount
obs, returns, masks, actions, values, neglogpacs, states, ep_infos, true_reward = runner.run()
self.num_timesteps += self.n_batch
ep_info_buf.extend(ep_infos)
mb_loss_vals = []
if states is None: # nonrecurrent version
update_fac = self.n_batch // self.nminibatches // self.noptepochs + 1
inds = np.arange(self.n_batch)
for epoch_num in range(self.noptepochs):
np.random.shuffle(inds)
for start in range(0, self.n_batch, batch_size):
timestep = self.num_timesteps // update_fac + ((self.noptepochs * self.n_batch + epoch_num *
self.n_batch + start) // batch_size)
end = start + batch_size
mbinds = inds[start:end]
slices = (arr[mbinds] for arr in (obs, returns, masks, actions, values, neglogpacs))
mb_loss_vals.append(self._train_step(lr_now, cliprange_now, *slices, writer=writer,
update=timestep, cliprange_vf=cliprange_vf_now))
else: # recurrent version
update_fac = self.n_batch // self.nminibatches // self.noptepochs // self.n_steps + 1
assert self.n_envs % self.nminibatches == 0
env_indices = np.arange(self.n_envs)
flat_indices = np.arange(self.n_envs * self.n_steps).reshape(self.n_envs, self.n_steps)
envs_per_batch = batch_size // self.n_steps
for epoch_num in range(self.noptepochs):
np.random.shuffle(env_indices)
for start in range(0, self.n_envs, envs_per_batch):
timestep = self.num_timesteps // update_fac + ((self.noptepochs * self.n_envs + epoch_num *
self.n_envs + start) // envs_per_batch)
end = start + envs_per_batch
mb_env_inds = env_indices[start:end]
mb_flat_inds = flat_indices[mb_env_inds].ravel()
slices = (arr[mb_flat_inds] for arr in (obs, returns, masks, actions, values, neglogpacs))
mb_states = states[mb_env_inds]
mb_loss_vals.append(self._train_step(lr_now, cliprange_now, *slices, update=timestep,
writer=writer, states=mb_states,
cliprange_vf=cliprange_vf_now))
loss_vals = np.mean(mb_loss_vals, axis=0)
t_now = time.time()
fps = int(self.n_batch / (t_now - t_start))
if writer is not None:
self.episode_reward = total_episode_reward_logger(self.episode_reward,
true_reward.reshape((self.n_envs, self.n_steps)),
masks.reshape((self.n_envs, self.n_steps)),
writer, self.num_timesteps)
if self.verbose >= 1 and (update % log_interval == 0 or update == 1):
explained_var = explained_variance(values, returns)
logger.logkv("serial_timesteps", update * self.n_steps)
logger.logkv("n_updates", update)
logger.logkv("total_timesteps", self.num_timesteps)
logger.logkv("fps", fps)
logger.logkv("explained_variance", float(explained_var))
if len(ep_info_buf) > 0 and len(ep_info_buf[0]) > 0:
logger.logkv('ep_reward_mean', safe_mean([ep_info['r'] for ep_info in ep_info_buf]))
logger.logkv('ep_len_mean', safe_mean([ep_info['l'] for ep_info in ep_info_buf]))
logger.logkv('time_elapsed', t_start - t_first_start)
for (loss_val, loss_name) in zip(loss_vals, self.loss_names):
logger.logkv(loss_name, loss_val)
logger.dumpkvs()
if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
if callback(locals(), globals()) is False:
break
return self
def save(self, save_path, cloudpickle=False):
data = {
"gamma": self.gamma,
"n_steps": self.n_steps,
"vf_coef": self.vf_coef,
"ent_coef": self.ent_coef,
"max_grad_norm": self.max_grad_norm,
"learning_rate": self.learning_rate,
"lam": self.lam,
"nminibatches": self.nminibatches,
"noptepochs": self.noptepochs,
"cliprange": self.cliprange,
"cliprange_vf": self.cliprange_vf,
"verbose": self.verbose,
"policy": self.policy,
"observation_space": self.observation_space,
"action_space": self.action_space,
"n_envs": self.n_envs,
"n_cpu_tf_sess": self.n_cpu_tf_sess,
"seed": self.seed,
"_vectorize_action": self._vectorize_action,
"policy_kwargs": self.policy_kwargs
}
params_to_save = self.get_parameters()
self._save_to_file(save_path, data=data, params=params_to_save, cloudpickle=cloudpickle)
def setup_model(self):
from stable_baselines.gail.adversary import TransitionClassifier
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the PPO2 model must be " \
"an instance of common.policies.ActorCriticPolicy."
self.n_batch = self.n_envs * self.n_steps
self.graph = tf.Graph()
with self.graph.as_default():
self.set_random_seed(self.seed)
self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess, graph=self.graph)
if self.using_gail:
self.reward_giver = TransitionClassifier(self.observation_space, self.action_space,
self.hidden_size_adversary,
normalize=False,
entcoeff=self.adversary_entcoeff)
n_batch_step = None
n_batch_train = None
if issubclass(self.policy, RecurrentActorCriticPolicy):
assert self.n_envs % self.nminibatches == 0, "For recurrent policies, "\
"the number of environments run in parallel should be a multiple of nminibatches."
n_batch_step = self.n_envs
n_batch_train = self.n_batch // self.nminibatches
act_model = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs, 1,
n_batch_step, reuse=False, **self.policy_kwargs)
with tf.variable_scope("train_model", reuse=True,
custom_getter=tf_util.outer_scope_getter("train_model")):
train_model = self.policy(self.sess, self.observation_space, self.action_space,
self.n_envs // self.nminibatches, self.n_steps, n_batch_train,
reuse=True, **self.policy_kwargs)
with tf.variable_scope("loss", reuse=False):
self.action_ph = train_model.pdtype.sample_placeholder([None], name="action_ph")
self.advs_ph = tf.placeholder(tf.float32, [None], name="advs_ph")
self.rewards_ph = tf.placeholder(tf.float32, [None], name="rewards_ph")
self.old_neglog_pac_ph = tf.placeholder(tf.float32, [None], name="old_neglog_pac_ph")
self.old_vpred_ph = tf.placeholder(tf.float32, [None], name="old_vpred_ph")
self.learning_rate_ph = tf.placeholder(tf.float32, [], name="learning_rate_ph")
self.clip_range_ph = tf.placeholder(tf.float32, [], name="clip_range_ph")
neglogpac = train_model.proba_distribution.neglogp(self.action_ph)
self.entropy = tf.reduce_mean(train_model.proba_distribution.entropy())
vpred = train_model.value_flat
# Value function clipping: not present in the original PPO
if self.cliprange_vf is None:
# Default behavior (legacy from OpenAI baselines):
# use the same clipping as for the policy
self.clip_range_vf_ph = self.clip_range_ph
self.cliprange_vf = self.cliprange
elif isinstance(self.cliprange_vf, (float, int)) and self.cliprange_vf < 0:
# Original PPO implementation: no value function clipping
self.clip_range_vf_ph = None
else:
# Last possible behavior: clipping range
# specific to the value function
self.clip_range_vf_ph = tf.placeholder(tf.float32, [], name="clip_range_vf_ph")
if self.clip_range_vf_ph is None:
# No clipping
vpred_clipped = train_model.value_flat
else:
# Clip the different between old and new value
# NOTE: this depends on the reward scaling
vpred_clipped = self.old_vpred_ph + \
tf.clip_by_value(train_model.value_flat - self.old_vpred_ph,
- self.clip_range_vf_ph, self.clip_range_vf_ph)
vf_losses1 = tf.square(vpred - self.rewards_ph)
vf_losses2 = tf.square(vpred_clipped - self.rewards_ph)
self.vf_loss = .5 * tf.reduce_mean(tf.maximum(vf_losses1, vf_losses2))
ratio = tf.exp(self.old_neglog_pac_ph - neglogpac)
pg_losses = -self.advs_ph * ratio
pg_losses2 = -self.advs_ph * tf.clip_by_value(ratio, 1.0 - self.clip_range_ph, 1.0 +
self.clip_range_ph)
self.pg_loss = tf.reduce_mean(tf.maximum(pg_losses, pg_losses2))
self.approxkl = .5 * tf.reduce_mean(tf.square(neglogpac - self.old_neglog_pac_ph))
self.clipfrac = tf.reduce_mean(tf.cast(tf.greater(tf.abs(ratio - 1.0),
self.clip_range_ph), tf.float32))
loss = self.pg_loss - self.entropy * self.ent_coef + self.vf_loss * self.vf_coef
tf.summary.scalar('entropy_loss', self.entropy)
tf.summary.scalar('policy_gradient_loss', self.pg_loss)
tf.summary.scalar('value_function_loss', self.vf_loss)
tf.summary.scalar('approximate_kullback-leibler', self.approxkl)
tf.summary.scalar('clip_factor', self.clipfrac)
tf.summary.scalar('loss', loss)
with tf.variable_scope('model'):
self.params = tf.trainable_variables()
# Update params with GAIL params too
if self.using_gail:
self.params.extend(self.reward_giver.get_trainable_variables())
if self.full_tensorboard_log:
for var in self.params:
tf.summary.histogram(var.name, var)
grads = tf.gradients(loss, self.params)
if self.max_grad_norm is not None:
grads, _grad_norm = tf.clip_by_global_norm(grads, self.max_grad_norm)
grads = list(zip(grads, self.params))
trainer = tf.train.AdamOptimizer(learning_rate=self.learning_rate_ph, epsilon=1e-5)
self._train = trainer.apply_gradients(grads)
self.loss_names = ['policy_loss', 'value_loss', 'policy_entropy', 'approxkl', 'clipfrac']
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards', tf.reduce_mean(self.rewards_ph))
tf.summary.scalar('learning_rate', tf.reduce_mean(self.learning_rate_ph))
tf.summary.scalar('advantage', tf.reduce_mean(self.advs_ph))
tf.summary.scalar('clip_range', tf.reduce_mean(self.clip_range_ph))
if self.clip_range_vf_ph is not None:
tf.summary.scalar('clip_range_vf', tf.reduce_mean(self.clip_range_vf_ph))
tf.summary.scalar('old_neglog_action_probabilty', tf.reduce_mean(self.old_neglog_pac_ph))
tf.summary.scalar('old_value_pred', tf.reduce_mean(self.old_vpred_ph))
if self.full_tensorboard_log:
tf.summary.histogram('discounted_rewards', self.rewards_ph)
tf.summary.histogram('learning_rate', self.learning_rate_ph)
tf.summary.histogram('advantage', self.advs_ph)
tf.summary.histogram('clip_range', self.clip_range_ph)
tf.summary.histogram('old_neglog_action_probabilty', self.old_neglog_pac_ph)
tf.summary.histogram('old_value_pred', self.old_vpred_ph)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', train_model.obs_ph)
else:
tf.summary.histogram('observation', train_model.obs_ph)
self.train_model = train_model
self.act_model = act_model
self.step = act_model.step
self.proba_step = act_model.proba_step
self.value = act_model.value
self.initial_state = act_model.initial_state
tf.global_variables_initializer().run(session=self.sess) # pylint: disable=E1101
self.summary = tf.summary.merge_all()
class TRPO(BaseRLModel):
def __init__(self,
policy,
env,
gamma=0.99,
timesteps_per_batch=1024,
max_kl=0.01,
cg_iters=10,
lam=0.98,
entcoeff=0.0,
cg_damping=1e-2,
vf_stepsize=3e-4,
vf_iters=3,
verbose=0,
_init_setup_model=True):
"""
learns a TRPO policy using the given environment
:param policy: (function (str, Gym Space, Gym Space, bool): MLPPolicy) policy generator
:param env: (Gym environment or str) The environment to learn from (if registered in Gym, can be str)
:param gamma: (float) the discount value
:param timesteps_per_batch: (int) the number of timesteps to run per batch (horizon)
:param max_kl: (float) the kullback leiber loss threshold
:param cg_iters: (int) the number of iterations for the conjugate gradient calculation
:param lam: (float) GAE factor
:param entcoeff: (float) the weight for the entropy loss
:param cg_damping: (float) the compute gradient dampening factor
:param vf_stepsize: (float) the value function stepsize
:param vf_iters: (int) the value function's number iterations for learning
:param verbose: (int) the verbosity level: 0 none, 1 training information, 2 tensorflow debug
:param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance
"""
super(TRPO, self).__init__(policy=policy,
env=env,
requires_vec_env=False,
verbose=verbose)
self.using_gail = False
self.timesteps_per_batch = timesteps_per_batch
self.cg_iters = cg_iters
self.cg_damping = cg_damping
self.gamma = gamma
self.lam = lam
self.max_kl = max_kl
self.vf_iters = vf_iters
self.vf_stepsize = vf_stepsize
self.entcoeff = entcoeff
# GAIL Params
self.pretrained_weight = None
self.hidden_size_adversary = 100
self.adversary_entcoeff = 1e-3
self.expert_dataset = None
self.save_per_iter = 1
self.checkpoint_dir = "/tmp/gail/ckpt/"
self.g_step = 1
self.d_step = 1
self.task_name = "task_name"
self.d_stepsize = 3e-4
self.graph = None
self.sess = None
self.policy_pi = None
self.loss_names = None
self.assign_old_eq_new = None
self.compute_losses = None
self.compute_lossandgrad = None
self.compute_fvp = None
self.compute_vflossandgrad = None
self.d_adam = None
self.vfadam = None
self.get_flat = None
self.set_from_flat = None
self.timed = None
self.allmean = None
self.nworkers = None
self.rank = None
self.reward_giver = None
self.step = None
self.proba_step = None
self.initial_state = None
self.params = None
if _init_setup_model:
self.setup_model()
def setup_model(self):
# prevent import loops
from stable_baselines.gail.adversary import TransitionClassifier
with SetVerbosity(self.verbose):
self.nworkers = MPI.COMM_WORLD.Get_size()
self.rank = MPI.COMM_WORLD.Get_rank()
np.set_printoptions(precision=3)
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf_util.single_threaded_session(graph=self.graph)
if self.using_gail:
self.reward_giver = TransitionClassifier(
self.env,
self.hidden_size_adversary,
entcoeff=self.adversary_entcoeff)
# Construct network for new policy
with tf.variable_scope("pi", reuse=False):
self.policy_pi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False)
# Network for old policy
with tf.variable_scope("oldpi", reuse=False):
old_policy = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False)
atarg = tf.placeholder(
dtype=tf.float32,
shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32,
shape=[None]) # Empirical return
observation = self.policy_pi.obs_ph
action = self.policy_pi.pdtype.sample_placeholder([None])
kloldnew = old_policy.proba_distribution.kl(
self.policy_pi.proba_distribution)
ent = self.policy_pi.proba_distribution.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
entbonus = self.entcoeff * meanent
vferr = tf.reduce_mean(
tf.square(self.policy_pi.value_fn[:, 0] - ret))
# advantage * pnew / pold
ratio = tf.exp(
self.policy_pi.proba_distribution.logp(action) -
old_policy.proba_distribution.logp(action))
surrgain = tf.reduce_mean(ratio * atarg)
optimgain = surrgain + entbonus
losses = [optimgain, meankl, entbonus, surrgain, meanent]
self.loss_names = [
"optimgain", "meankl", "entloss", "surrgain", "entropy"
]
dist = meankl
all_var_list = tf_util.get_trainable_vars("pi")
var_list = [
v for v in all_var_list
if "/vf" not in v.name and "/q/" not in v.name
]
vf_var_list = [
v for v in all_var_list
if "/pi" not in v.name and "/logstd" not in v.name
]
self.vfadam = MpiAdam(vf_var_list, sess=self.sess)
self.get_flat = tf_util.GetFlat(var_list, sess=self.sess)
self.set_from_flat = tf_util.SetFromFlat(var_list,
sess=self.sess)
if self.using_gail:
self.d_adam = MpiAdam(
self.reward_giver.get_trainable_variables())
klgrads = tf.gradients(dist, var_list)
flat_tangent = tf.placeholder(dtype=tf.float32,
shape=[None],
name="flat_tan")
shapes = [var.get_shape().as_list() for var in var_list]
start = 0
tangents = []
for shape in shapes:
var_size = tf_util.intprod(shape)
tangents.append(
tf.reshape(flat_tangent[start:start + var_size],
shape))
start += var_size
gvp = tf.add_n([
tf.reduce_sum(grad * tangent)
for (grad, tangent) in zipsame(klgrads, tangents)
]) # pylint: disable=E1111
fvp = tf_util.flatgrad(gvp, var_list)
self.assign_old_eq_new = tf_util.function(
[], [],
updates=[
tf.assign(oldv, newv) for (
oldv,
newv) in zipsame(tf_util.get_globals_vars("oldpi"),
tf_util.get_globals_vars("pi"))
])
self.compute_losses = tf_util.function(
[observation, old_policy.obs_ph, action, atarg], losses)
self.compute_lossandgrad = tf_util.function(
[observation, old_policy.obs_ph, action, atarg],
losses + [tf_util.flatgrad(optimgain, var_list)])
self.compute_fvp = tf_util.function([
flat_tangent, observation, old_policy.obs_ph, action, atarg
], fvp)
self.compute_vflossandgrad = tf_util.function(
[observation, old_policy.obs_ph, ret],
tf_util.flatgrad(vferr, vf_var_list))
@contextmanager
def timed(msg):
if self.rank == 0 and self.verbose >= 1:
print(colorize(msg, color='magenta'))
start_time = time.time()
yield
print(
colorize("done in %.3f seconds" %
(time.time() - start_time),
color='magenta'))
else:
yield
def allmean(arr):
assert isinstance(arr, np.ndarray)
out = np.empty_like(arr)
MPI.COMM_WORLD.Allreduce(arr, out, op=MPI.SUM)
out /= self.nworkers
return out
tf_util.initialize(sess=self.sess)
th_init = self.get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
self.set_from_flat(th_init)
if self.using_gail:
self.d_adam.sync()
self.vfadam.sync()
self.timed = timed
self.allmean = allmean
self.step = self.policy_pi.step
self.proba_step = self.policy_pi.proba_step
self.initial_state = self.policy_pi.initial_state
self.params = find_trainable_variables("pi")
if self.using_gail:
self.params.extend(
self.reward_giver.get_trainable_variables())
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100):
with SetVerbosity(self.verbose):
self._setup_learn(seed)
with self.sess.as_default():
seg_gen = traj_segment_generator(
self.policy_pi,
self.env,
self.timesteps_per_batch,
reward_giver=self.reward_giver,
gail=self.using_gail)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
t_start = time.time()
lenbuffer = deque(
maxlen=40) # rolling buffer for episode lengths
rewbuffer = deque(
maxlen=40) # rolling buffer for episode rewards
true_rewbuffer = None
if self.using_gail:
true_rewbuffer = deque(maxlen=40)
# Stats not used for now
# g_loss_stats = Stats(loss_names)
# d_loss_stats = Stats(reward_giver.loss_name)
# ep_stats = Stats(["True_rewards", "Rewards", "Episode_length"])
# if provide pretrained weight
if self.pretrained_weight is not None:
tf_util.load_state(
self.pretrained_weight,
var_list=tf_util.get_globals_vars("pi"),
sess=self.sess)
while True:
if callback:
callback(locals(), globals())
if total_timesteps and timesteps_so_far >= total_timesteps:
break
logger.log("********** Iteration %i ************" %
iters_so_far)
def fisher_vector_product(vec):
return self.allmean(
self.compute_fvp(
vec, *fvpargs,
sess=self.sess)) + self.cg_damping * vec
# ------------------ Update G ------------------
logger.log("Optimizing Policy...")
# g_step = 1 when not using GAIL
mean_losses = None
vpredbefore = None
tdlamret = None
observation = None
action = None
seg = None
for _ in range(self.g_step):
with self.timed("sampling"):
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, self.gamma, self.lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
observation, action, atarg, tdlamret = seg["ob"], seg[
"ac"], seg["adv"], seg["tdlamret"]
vpredbefore = seg[
"vpred"] # predicted value function before udpate
atarg = (atarg - atarg.mean()) / atarg.std(
) # standardized advantage function estimate
args = seg["ob"], seg["ob"], seg["ac"], atarg
fvpargs = [arr[::5] for arr in args]
self.assign_old_eq_new(sess=self.sess)
with self.timed("computegrad"):
*lossbefore, grad = self.compute_lossandgrad(
*args, sess=self.sess)
lossbefore = self.allmean(np.array(lossbefore))
grad = self.allmean(grad)
if np.allclose(grad, 0):
logger.log("Got zero gradient. not updating")
else:
with self.timed("cg"):
stepdir = conjugate_gradient(
fisher_vector_product,
grad,
cg_iters=self.cg_iters,
verbose=self.rank == 0
and self.verbose >= 1)
assert np.isfinite(stepdir).all()
shs = .5 * stepdir.dot(
fisher_vector_product(stepdir))
# abs(shs) to avoid taking square root of negative values
lagrange_multiplier = np.sqrt(
abs(shs) / self.max_kl)
# logger.log("lagrange multiplier:", lm, "gnorm:", np.linalg.norm(g))
fullstep = stepdir / lagrange_multiplier
expectedimprove = grad.dot(fullstep)
surrbefore = lossbefore[0]
stepsize = 1.0
thbefore = self.get_flat()
thnew = None
for _ in range(10):
thnew = thbefore + fullstep * stepsize
self.set_from_flat(thnew)
mean_losses = surr, kl_loss, *_ = self.allmean(
np.array(
self.compute_losses(*args,
sess=self.sess)))
improve = surr - surrbefore
logger.log("Expected: %.3f Actual: %.3f" %
(expectedimprove, improve))
if not np.isfinite(mean_losses).all():
logger.log(
"Got non-finite value of losses -- bad!"
)
elif kl_loss > self.max_kl * 1.5:
logger.log(
"violated KL constraint. shrinking step."
)
elif improve < 0:
logger.log(
"surrogate didn't improve. shrinking step."
)
else:
logger.log("Stepsize OK!")
break
stepsize *= .5
else:
logger.log("couldn't compute a good step")
self.set_from_flat(thbefore)
if self.nworkers > 1 and iters_so_far % 20 == 0:
# list of tuples
paramsums = MPI.COMM_WORLD.allgather(
(thnew.sum(), self.vfadam.getflat().sum()))
assert all(
np.allclose(ps, paramsums[0])
for ps in paramsums[1:])
with self.timed("vf"):
for _ in range(self.vf_iters):
for (mbob, mbret) in dataset.iterbatches(
(seg["ob"], seg["tdlamret"]),
include_final_partial_batch=False,
batch_size=128):
grad = self.allmean(
self.compute_vflossandgrad(
mbob, mbob, mbret, sess=self.sess))
self.vfadam.update(grad, self.vf_stepsize)
for (loss_name, loss_val) in zip(self.loss_names,
mean_losses):
logger.record_tabular(loss_name, loss_val)
logger.record_tabular(
"ev_tdlam_before",
explained_variance(vpredbefore, tdlamret))
if self.using_gail:
# ------------------ Update D ------------------
logger.log("Optimizing Discriminator...")
logger.log(fmt_row(13, self.reward_giver.loss_name))
ob_expert, ac_expert = self.expert_dataset.get_next_batch(
len(observation))
batch_size = len(observation) // self.d_step
d_losses = [
] # list of tuples, each of which gives the loss for a minibatch
for ob_batch, ac_batch in dataset.iterbatches(
(observation, action),
include_final_partial_batch=False,
batch_size=batch_size):
ob_expert, ac_expert = self.expert_dataset.get_next_batch(
len(ob_batch))
# update running mean/std for reward_giver
if hasattr(self.reward_giver, "obs_rms"):
self.reward_giver.obs_rms.update(
np.concatenate((ob_batch, ob_expert), 0))
*newlosses, grad = self.reward_giver.lossandgrad(
ob_batch, ac_batch, ob_expert, ac_expert)
self.d_adam.update(self.allmean(grad),
self.d_stepsize)
d_losses.append(newlosses)
logger.log(fmt_row(13, np.mean(d_losses, axis=0)))
lrlocal = (seg["ep_lens"], seg["ep_rets"],
seg["ep_true_rets"]) # local values
listoflrpairs = MPI.COMM_WORLD.allgather(
lrlocal) # list of tuples
lens, rews, true_rets = map(flatten_lists,
zip(*listoflrpairs))
true_rewbuffer.extend(true_rets)
else:
lrlocal = (seg["ep_lens"], seg["ep_rets"]
) # local values
listoflrpairs = MPI.COMM_WORLD.allgather(
lrlocal) # list of tuples
lens, rews = map(flatten_lists, zip(*listoflrpairs))
lenbuffer.extend(lens)
rewbuffer.extend(rews)
logger.record_tabular("EpLenMean", np.mean(lenbuffer))
logger.record_tabular("EpRewMean", np.mean(rewbuffer))
if self.using_gail:
logger.record_tabular("EpTrueRewMean",
np.mean(true_rewbuffer))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
timesteps_so_far += seg["total_timestep"]
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", timesteps_so_far)
logger.record_tabular("TimeElapsed", time.time() - t_start)
if self.verbose >= 1 and self.rank == 0:
logger.dump_tabular()
return self
def predict(self, observation, state=None, mask=None):
if state is None:
state = self.initial_state
if mask is None:
mask = [False for _ in range(self.n_envs)]
observation = np.array(observation).reshape(
(-1, ) + self.observation_space.shape)
actions, _, states, _ = self.step(observation, state, mask)
return actions, states
def action_probability(self, observation, state=None, mask=None):
if state is None:
state = self.initial_state
if mask is None:
mask = [False for _ in range(self.n_envs)]
observation = np.array(observation).reshape(
(-1, ) + self.observation_space.shape)
return self.proba_step(observation, state, mask)
def save(self, save_path):
data = {
"gamma": self.gamma,
"timesteps_per_batch": self.timesteps_per_batch,
"max_kl": self.max_kl,
"cg_iters": self.cg_iters,
"lam": self.lam,
"entcoeff": self.entcoeff,
"cg_damping": self.cg_damping,
"vf_stepsize": self.vf_stepsize,
"vf_iters": self.vf_iters,
"pretrained_weight": self.pretrained_weight,
"reward_giver": self.reward_giver,
"expert_dataset": self.expert_dataset,
"save_per_iter": self.save_per_iter,
"checkpoint_dir": self.checkpoint_dir,
"g_step": self.g_step,
"d_step": self.d_step,
"task_name": self.task_name,
"d_stepsize": self.d_stepsize,
"using_gail": self.using_gail,
"verbose": self.verbose,
"policy": self.policy,
"observation_space": self.observation_space,
"action_space": self.action_space,
"n_envs": self.n_envs,
"_vectorize_action": self._vectorize_action
}
params = self.sess.run(self.params)
self._save_to_file(save_path, data=data, params=params)
@classmethod
def load(cls, load_path, env=None, **kwargs):
data, params = cls._load_from_file(load_path)
model = cls(policy=data["policy"], env=None, _init_setup_model=False)
model.__dict__.update(data)
model.__dict__.update(kwargs)
model.set_env(env)
model.setup_model()
restores = []
for param, loaded_p in zip(model.params, params):
restores.append(param.assign(loaded_p))
model.sess.run(restores)
return model
def setup_model(self):
# prevent import loops
from stable_baselines.gail.adversary import TransitionClassifier
from stable_baselines.mdal.adversary import TabularAdversaryTF, NeuralAdversaryTRPO
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the MDPO model must be " \
"an instance of common.policies.ActorCriticPolicy."
self.nworkers = MPI.COMM_WORLD.Get_size()
self.rank = MPI.COMM_WORLD.Get_rank()
np.set_printoptions(precision=3)
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf_util.single_threaded_session(graph=self.graph)
# self._setup_learn(self.seed)
self._setup_learn()
if self.using_gail:
self.reward_giver = TransitionClassifier(self.observation_space, self.action_space,
self.hidden_size_adversary,
entcoeff=self.adversary_entcoeff)
elif self.using_mdal:
if self.neural:
self.reward_giver = NeuralAdversaryTRPO(self.sess, self.observation_space, self.action_space,
self.hidden_size_adversary,
entcoeff=self.adversary_entcoeff)
else:
self.reward_giver = TabularAdversaryTF(self.sess, self.observation_space, self.action_space,
self.hidden_size_adversary,
entcoeff=self.adversary_entcoeff,
expert_features=self.expert_dataset.successor_features,
exploration_bonus=self.exploration_bonus,
bonus_coef=self.bonus_coef,
t_c=self.t_c,
is_action_features=self.is_action_features)
# Construct network for new policy
self.policy_pi = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs, 1,
None, reuse=False, **self.policy_kwargs)
# Network for old policy
with tf.variable_scope("oldpi", reuse=False):
self.old_policy = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs, 1,
None, reuse=False, **self.policy_kwargs)
# Network for fitting closed form
with tf.variable_scope("closedpi", reuse=False):
self.closed_policy = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs, 1,
None, reuse=False, **self.policy_kwargs)
with tf.variable_scope("loss", reuse=False):
self.atarg = tf.placeholder(dtype=tf.float32, shape=[None]) # Target advantage function (if applicable)
self.vtarg = tf.placeholder(dtype=tf.float32, shape=[None])
self.ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
self.learning_rate_ph = tf.placeholder(dtype=tf.float32, shape=[], name="learning_rate_ph")
self.outer_learning_rate_ph = tf.placeholder(dtype=tf.float32, shape=[], name="outer_learning_rate_ph")
self.old_vpred_ph = tf.placeholder(dtype=tf.float32, shape=[None], name="old_vpred_ph")
self.clip_range_vf_ph = tf.placeholder(dtype=tf.float32, shape=[], name="clip_range_ph")
observation = self.policy_pi.obs_ph
self.action = self.policy_pi.pdtype.sample_placeholder([None])
if self.tsallis_q == 1.0:
kloldnew = self.policy_pi.proba_distribution.kl(self.old_policy.proba_distribution)
ent = self.policy_pi.proba_distribution.entropy()
meankl = tf.reduce_mean(kloldnew)
else:
logp_pi = self.policy_pi.proba_distribution.logp(self.action)
logp_pi_old = self.old_policy.proba_distribution.logp(self.action)
ent = self.policy_pi.proba_distribution.entropy()
#kloldnew = self.policy_pi.proba_distribution.kl_tsallis(self.old_policy.proba_distribution, self.tsallis_q)
tsallis_q = 2.0 - self.tsallis_q
meankl = tf.reduce_mean(tf_log_q(tf.exp(logp_pi), tsallis_q) - tf_log_q(tf.exp(logp_pi_old), tsallis_q)) #tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
entbonus = self.entcoeff * meanent
if self.cliprange_vf is None:
vpred_clipped = self.policy_pi.value_flat
else:
vpred_clipped = self.old_vpred_ph + \
tf.clip_by_value(self.policy_pi.value_flat - self.old_vpred_ph,
- self.clip_range_vf_ph, self.clip_range_vf_ph)
vf_losses1 = tf.square(self.policy_pi.value_flat - self.ret)
vf_losses2 = tf.square(vpred_clipped - self.ret)
vferr = tf.reduce_mean(tf.maximum(vf_losses1, vf_losses2))
# advantage * pnew / pold
ratio = tf.exp(self.policy_pi.proba_distribution.logp(self.action) -
self.old_policy.proba_distribution.logp(self.action))
if self.method == "multistep-SGD":
surrgain = tf.reduce_mean(ratio * self.atarg) - meankl / self.learning_rate_ph
elif self.method == "closedreverse-KL":
surrgain = tf.reduce_mean(tf.exp(self.atarg) * self.policy_pi.proba_distribution.logp(self.action))
else:
policygain = tf.reduce_mean(tf.exp(self.atarg) * tf.log(self.closed_policy.proba_distribution.mean))
surrgain = tf.reduce_mean(ratio * self.atarg) - tf.reduce_mean(self.learning_rate_ph * ratio * self.policy_pi.proba_distribution.logp(self.action))
optimgain = surrgain #+ entbonus - self.learning_rate_ph * meankl
losses = [optimgain, meankl, entbonus, surrgain, meanent]
self.loss_names = ["optimgain", "meankl", "entloss", "surrgain", "entropy"]
dist = meankl
all_var_list = tf_util.get_trainable_vars("model")
var_list = [v for v in all_var_list if "/vf" not in v.name and "/q/" not in v.name]
vf_var_list = [v for v in all_var_list if "/pi" not in v.name and "/logstd" not in v.name]
print("policy vars", var_list)
all_closed_var_list = tf_util.get_trainable_vars("closedpi")
closed_var_list = [v for v in all_closed_var_list if "/vf" not in v.name and "/q" not in v.name]
self.get_flat = tf_util.GetFlat(var_list, sess=self.sess)
self.set_from_flat = tf_util.SetFromFlat(var_list, sess=self.sess)
klgrads = tf.gradients(dist, var_list)
flat_tangent = tf.placeholder(dtype=tf.float32, shape=[None], name="flat_tan")
shapes = [var.get_shape().as_list() for var in var_list]
start = 0
tangents = []
for shape in shapes:
var_size = tf_util.intprod(shape)
tangents.append(tf.reshape(flat_tangent[start: start + var_size], shape))
start += var_size
gvp = tf.add_n([tf.reduce_sum(grad * tangent)
for (grad, tangent) in zipsame(klgrads, tangents)]) # pylint: disable=E1111
fvp = tf_util.flatgrad(gvp, var_list)
# tf.summary.scalar('entropy_loss', meanent)
# tf.summary.scalar('policy_gradient_loss', optimgain)
# tf.summary.scalar('value_function_loss', surrgain)
# tf.summary.scalar('approximate_kullback-leibler', meankl)
# tf.summary.scalar('loss', optimgain + meankl + entbonus + surrgain + meanent)
self.assign_old_eq_new = \
tf_util.function([], [], updates=[tf.assign(oldv, newv) for (oldv, newv) in
zipsame(tf_util.get_globals_vars("oldpi"),
tf_util.get_globals_vars("model"))])
self.compute_losses = tf_util.function([observation, self.old_policy.obs_ph, self.action, self.atarg, self.learning_rate_ph, self.vtarg], losses)
self.compute_fvp = tf_util.function([flat_tangent, observation, self.old_policy.obs_ph, self.action, self.atarg],
fvp)
self.compute_vflossandgrad = tf_util.function([observation, self.old_policy.obs_ph, self.ret, self.old_vpred_ph, self.clip_range_vf_ph],
tf_util.flatgrad(vferr, vf_var_list))
grads = tf.gradients(-optimgain, var_list)
grads, _grad_norm = tf.clip_by_global_norm(grads, 0.5)
trainer = tf.train.AdamOptimizer(learning_rate=self.outer_learning_rate_ph, epsilon=1e-5)
# trainer = tf.train.AdamOptimizer(learning_rate=3e-4, epsilon=1e-5)
grads = list(zip(grads, var_list))
self._train = trainer.apply_gradients(grads)
@contextmanager
def timed(msg):
if self.rank == 0 and self.verbose >= 1:
# print(colorize(msg, color='magenta'))
# start_time = time.time()
yield
# print(colorize("done in {:.3f} seconds".format((time.time() - start_time)),
# color='magenta'))
else:
yield
def allmean(arr):
assert isinstance(arr, np.ndarray)
out = np.empty_like(arr)
MPI.COMM_WORLD.Allreduce(arr, out, op=MPI.SUM)
out /= self.nworkers
return out
tf_util.initialize(sess=self.sess)
th_init = self.get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
self.set_from_flat(th_init)
with tf.variable_scope("Adam_mpi", reuse=False):
self.vfadam = MpiAdam(vf_var_list, sess=self.sess)
if self.using_gail or self.using_mdal:
self.d_adam = MpiAdam(self.reward_giver.get_trainable_variables(), sess=self.sess)
self.d_adam.sync()
self.vfadam.sync()
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards', tf.reduce_mean(self.ret))
tf.summary.scalar('learning_rate', tf.reduce_mean(self.vf_stepsize))
tf.summary.scalar('advantage', tf.reduce_mean(self.atarg))
tf.summary.scalar('kl_clip_range', tf.reduce_mean(self.max_kl))
if self.full_tensorboard_log:
tf.summary.histogram('discounted_rewards', self.ret)
tf.summary.histogram('learning_rate', self.vf_stepsize)
tf.summary.histogram('advantage', self.atarg)
tf.summary.histogram('kl_clip_range', self.max_kl)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', observation)
else:
tf.summary.histogram('observation', observation)
self.timed = timed
self.allmean = allmean
self.step = self.policy_pi.step
self.proba_step = self.policy_pi.proba_step
self.initial_state = self.policy_pi.initial_state
self.params = tf_util.get_trainable_vars("model") + tf_util.get_trainable_vars("oldpi")
if self.using_gail:
self.params.extend(self.reward_giver.get_trainable_variables())
self.summary = tf.summary.merge_all()
self.compute_lossandgrad = \
tf_util.function([observation, self.old_policy.obs_ph, self.action, self.atarg, self.ret, self.learning_rate_ph, self.vtarg, self.closed_policy.obs_ph],
[self.summary, tf_util.flatgrad(optimgain, var_list)] + losses)