def validate_probtype(probtype, pdparam):
N = 100000
# Check to see if mean negative log likelihood == differential entropy
Mval = np.repeat(pdparam[None, :], N, axis=0)
M = probtype.param_placeholder([N])
X = probtype.sample_placeholder([N])
pd = probtype.pdfromflat(M)
calcloglik = U.function([X, M], pd.logp(X))
calcent = U.function([M], pd.entropy())
Xval = tf.get_default_session().run(pd.sample(), feed_dict={M: Mval})
logliks = calcloglik(Xval, Mval)
entval_ll = -logliks.mean() #pylint: disable=E1101
entval_ll_stderr = logliks.std() / np.sqrt(N) #pylint: disable=E1101
entval = calcent(Mval).mean() #pylint: disable=E1101
assert np.abs(entval - entval_ll) < 3 * entval_ll_stderr # within 3 sigmas
# Check to see if kldiv[p,q] = - ent[p] - E_p[log q]
M2 = probtype.param_placeholder([N])
pd2 = probtype.pdfromflat(M2)
q = pdparam + np.random.randn(pdparam.size) * 0.1
Mval2 = np.repeat(q[None, :], N, axis=0)
calckl = U.function([M, M2], pd.kl(pd2))
klval = calckl(Mval, Mval2).mean() #pylint: disable=E1101
logliks = calcloglik(Xval, Mval2)
klval_ll = -entval - logliks.mean() #pylint: disable=E1101
klval_ll_stderr = logliks.std() / np.sqrt(N) #pylint: disable=E1101
assert np.abs(klval - klval_ll) < 3 * klval_ll_stderr # within 3 sigmas
print('ok on', probtype, pdparam)
def test_mpi_adam():
"""
tests the MpiAdam object's functionality
"""
np.random.seed(0)
tf.set_random_seed(0)
a_var = tf.Variable(np.random.randn(3).astype('float32'))
b_var = tf.Variable(np.random.randn(2, 5).astype('float32'))
loss = tf.reduce_sum(tf.square(a_var)) + tf.reduce_sum(tf.sin(b_var))
learning_rate = 1e-2
update_op = tf.train.AdamOptimizer(learning_rate).minimize(loss)
do_update = tf_utils.function([], loss, updates=[update_op])
tf.get_default_session().run(tf.global_variables_initializer())
for step in range(10):
print(step, do_update())
tf.set_random_seed(0)
tf.get_default_session().run(tf.global_variables_initializer())
var_list = [a_var, b_var]
lossandgrad = tf_utils.function(
[], [loss, tf_utils.flatgrad(loss, var_list)], updates=[update_op])
adam = MpiAdam(var_list)
for step in range(10):
loss, grad = lossandgrad()
adam.update(grad, learning_rate)
print(step, loss)
def __init__(self, env, hidden_size, entcoeff=0.001, scope="adversary"):
"""
reward regression from observations and transitions
:param env: (Gym Environment)
:param hidden_size: ([int]) the hidden dimension for the MLP
:param entcoeff: (float) the entropy loss weight
:param scope: (str) tensorflow variable scope
"""
self.scope = scope
self.observation_shape = env.observation_space.shape
self.actions_shape = env.action_space.shape
self.input_shape = tuple([
o + a for o, a in zip(self.observation_shape, self.actions_shape)
])
self.num_actions = env.action_space.shape[0]
self.hidden_size = hidden_size
self.build_ph()
# Build grpah
generator_logits = self.build_graph(self.generator_obs_ph,
self.generator_acs_ph,
reuse=False)
expert_logits = self.build_graph(self.expert_obs_ph,
self.expert_acs_ph,
reuse=True)
# Build accuracy
generator_acc = tf.reduce_mean(
tf.to_float(tf.nn.sigmoid(generator_logits) < 0.5))
expert_acc = tf.reduce_mean(
tf.to_float(tf.nn.sigmoid(expert_logits) > 0.5))
# Build regression loss
# let x = logits, z = targets.
# z * -log(sigmoid(x)) + (1 - z) * -log(1 - sigmoid(x))
generator_loss = tf.nn.sigmoid_cross_entropy_with_logits(
logits=generator_logits, labels=tf.zeros_like(generator_logits))
generator_loss = tf.reduce_mean(generator_loss)
expert_loss = tf.nn.sigmoid_cross_entropy_with_logits(
logits=expert_logits, labels=tf.ones_like(expert_logits))
expert_loss = tf.reduce_mean(expert_loss)
# Build entropy loss
logits = tf.concat([generator_logits, expert_logits], 0)
entropy = tf.reduce_mean(logit_bernoulli_entropy(logits))
entropy_loss = -entcoeff * entropy
# Loss + Accuracy terms
self.losses = [
generator_loss, expert_loss, entropy, entropy_loss, generator_acc,
expert_acc
]
self.loss_name = [
"generator_loss", "expert_loss", "entropy", "entropy_loss",
"generator_acc", "expert_acc"
]
self.total_loss = generator_loss + expert_loss + entropy_loss
# Build Reward for policy
self.reward_op = -tf.log(1 - tf.nn.sigmoid(generator_logits) + 1e-8)
var_list = self.get_trainable_variables()
self.lossandgrad = tf_util.function([
self.generator_obs_ph, self.generator_acs_ph, self.expert_obs_ph,
self.expert_acs_ph
], self.losses + [tf_util.flatgrad(self.total_loss, var_list)])
def build_act(make_obs_ph, q_func, num_actions, scope="deepq", reuse=None):
"""Creates the act function:
:param make_obs_ph: (function (str): TensorFlow Tensor) a function that take a name and creates a placeholder of
input with that name
:param q_func: (function (TensorFlow Tensor, int, str, bool): TensorFlow Tensor)
the model that takes the following inputs:
observation_in: object
the output of observation placeholder
num_actions: int
number of actions
scope: str
reuse: bool
should be passed to outer variable scope
and returns a tensor of shape (batch_size, num_actions) with values of every action.
:param num_actions: (int) number of actions.
:param scope: (str or VariableScope) optional scope for variable_scope.
:param reuse: (bool) whether or not the variables should be reused. To be able to reuse the scope must be given.
:return: (function (TensorFlow Tensor, bool, float): TensorFlow Tensor) act function to select and action given
observation. See the top of the file for details.
"""
with tf.variable_scope(scope, reuse=reuse):
observations_ph = make_obs_ph("observation")
stochastic_ph = tf.placeholder(tf.bool, (), name="stochastic")
update_eps_ph = tf.placeholder(tf.float32, (), name="update_eps")
eps = tf.get_variable("eps", (),
initializer=tf.constant_initializer(0))
q_values = q_func(observations_ph.get(), num_actions, scope="q_func")
deterministic_actions = tf.argmax(q_values, axis=1)
batch_size = tf.shape(observations_ph.get())[0]
random_actions = tf.random_uniform(tf.stack([batch_size]),
minval=0,
maxval=num_actions,
dtype=tf.int64)
chose_random = tf.random_uniform(
tf.stack([batch_size]), minval=0, maxval=1, dtype=tf.float32) < eps
stochastic_actions = tf.where(chose_random, random_actions,
deterministic_actions)
output_actions = tf.cond(stochastic_ph, lambda: stochastic_actions,
lambda: deterministic_actions)
update_eps_expr = eps.assign(
tf.cond(update_eps_ph >= 0, lambda: update_eps_ph, lambda: eps))
_act = tf_utils.function(
inputs=[observations_ph, stochastic_ph, update_eps_ph],
outputs=output_actions,
givens={
update_eps_ph: -1.0,
stochastic_ph: True
},
updates=[update_eps_expr])
def act(obs, stochastic=True, update_eps=-1):
return _act(obs, stochastic, update_eps)
return act
def build_act(q_func,
ob_space,
ac_space,
stochastic_ph,
update_eps_ph,
sess,
obs_phs=None):
"""
Creates the act function:
:param q_func: (DQNPolicy) the policy
:param ob_space: (Gym Space) The observation space of the environment
:param ac_space: (Gym Space) The action space of the environment
:param stochastic_ph: (TensorFlow Tensor) the stochastic placeholder
:param update_eps_ph: (TensorFlow Tensor) the update_eps placeholder
:param sess: (TensorFlow session) The current TensorFlow session
:return: (function (TensorFlow Tensor, bool, float): TensorFlow Tensor, (TensorFlow Tensor, TensorFlow Tensor)
act function to select and action given observation (See the top of the file for details),
A tuple containing the observation placeholder and the processed observation placeholder respectivly.
"""
eps = tf.get_variable("eps", (), initializer=tf.constant_initializer(0))
policy = q_func(sess, ob_space, ac_space, 1, 1, None, obs_phs=obs_phs)
obs_phs = (policy.obs_ph, policy.processed_obs)
#mask = tf.one_hot(mask_idx, depth=541, dtype=tf.float32)
mask = tf.placeholder(dtype=tf.float32)
feasible_q_values = tf.math.multiply(policy.q_values, mask)
deterministic_actions = tf.argmax(feasible_q_values, axis=1)
batch_size = tf.shape(policy.obs_ph)[0]
n_actions = ac_space.nvec if isinstance(ac_space,
MultiDiscrete) else ac_space.n
random_actions = tf.random_uniform(tf.stack([batch_size]),
minval=0,
maxval=n_actions,
dtype=tf.int64)
chose_random = tf.random_uniform(
tf.stack([batch_size]), minval=0, maxval=1, dtype=tf.float32) < eps
stochastic_actions = tf.where(chose_random, random_actions,
deterministic_actions)
output_actions = tf.cond(stochastic_ph, lambda: stochastic_actions,
lambda: deterministic_actions)
update_eps_expr = eps.assign(
tf.cond(update_eps_ph >= 0, lambda: update_eps_ph, lambda: eps))
_act = tf_util.function(
inputs=[policy.obs_ph, stochastic_ph, update_eps_ph, mask],
outputs=output_actions,
givens={
update_eps_ph: -1.0,
stochastic_ph: True
},
updates=[update_eps_expr])
def act(obs, stochastic=False, update_eps=-1, mask=None):
return _act(obs, stochastic, update_eps, mask)
return act, obs_phs
def validate_probtype(probtype, pdparam):
"""
validate probability distribution types
:param probtype: (ProbabilityDistributionType) the type to validate
:param pdparam: ([float]) the flat probabilities to test
"""
number_samples = 100000
# Check to see if mean negative log likelihood == differential entropy
mval = np.repeat(pdparam[None, :], number_samples, axis=0)
mval_ph = probtype.param_placeholder([number_samples])
action_mask_ph = probtype.param_placeholder([number_samples])
action_mask_ph = tf.placeholder_with_default(
tf.zeros_like(action_mask_ph), shape=np.shape(action_mask_ph))
xval_ph = probtype.sample_placeholder([number_samples])
proba_distribution = probtype.proba_distribution_from_flat(mval_ph)
calcloglik = tf_util.function([xval_ph, mval_ph],
proba_distribution.logp(xval_ph))
calcent = tf_util.function([mval_ph], proba_distribution.entropy())
xval = tf.get_default_session().run(proba_distribution.sample(),
feed_dict={mval_ph: mval})
logliks = calcloglik(xval, mval)
entval_ll = -logliks.mean()
entval_ll_stderr = logliks.std() / np.sqrt(number_samples)
entval = calcent(mval).mean()
assert np.abs(entval - entval_ll) < 3 * entval_ll_stderr # within 3 sigmas
# Check to see if kldiv[p,q] = - ent[p] - E_p[log q]
mval2_ph = probtype.param_placeholder([number_samples])
action_mask_ph2 = probtype.param_placeholder([number_samples])
action_mask_ph2 = tf.placeholder_with_default(
tf.zeros_like(action_mask_ph2), shape=np.shape(action_mask_ph2))
pd2 = probtype.proba_distribution_from_flat(mval2_ph)
tmp = pdparam + np.random.randn(pdparam.size) * 0.1
mval2 = np.repeat(tmp[None, :], number_samples, axis=0)
calckl = tf_util.function([mval_ph, mval2_ph], proba_distribution.kl(pd2))
klval = calckl(mval, mval2).mean()
logliks = calcloglik(xval, mval2)
klval_ll = -entval - logliks.mean()
klval_ll_stderr = logliks.std() / np.sqrt(number_samples)
assert np.abs(klval - klval_ll) < 3 * klval_ll_stderr # within 3 sigmas
print('ok on', probtype, pdparam)
def __init__(self, epsilon=1e-2, shape=()):
"""
calulates the running mean and std of a data stream
https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Parallel_algorithm
:param epsilon: (float) helps with arithmetic issues
:param shape: (tuple) the shape of the data stream's output
"""
self._sum = tf.compat.v1.get_variable(
dtype=tf.float64,
shape=shape,
initializer=tf.compat.v1.constant_initializer(0.0),
name="runningsum",
trainable=False)
self._sumsq = tf.compat.v1.get_variable(
dtype=tf.float64,
shape=shape,
initializer=tf.compat.v1.constant_initializer(epsilon),
name="runningsumsq",
trainable=False)
self._count = tf.compat.v1.get_variable(
dtype=tf.float64,
shape=(),
initializer=tf.compat.v1.constant_initializer(epsilon),
name="count",
trainable=False)
self.shape = shape
self.mean = tf.cast(self._sum / self._count, tf.float32)
self.std = tf.sqrt(
tf.maximum(
tf.cast(self._sumsq / self._count, tf.float32) -
tf.square(self.mean), 1e-2))
newsum = tf.compat.v1.placeholder(shape=self.shape,
dtype=tf.float64,
name='sum')
newsumsq = tf.compat.v1.placeholder(shape=self.shape,
dtype=tf.float64,
name='var')
newcount = tf.compat.v1.placeholder(shape=[],
dtype=tf.float64,
name='count')
self.incfiltparams = tf_util.function(
[newsum, newsumsq, newcount], [],
updates=[
tf.compat.v1.assign_add(self._sum, newsum),
tf.compat.v1.assign_add(self._sumsq, newsumsq),
tf.compat.v1.assign_add(self._count, newcount)
])
def test_function():
"""
test the function function in tf_util
"""
with tf.Graph().as_default():
x_ph = tf.placeholder(tf.int32, (), name="x")
y_ph = tf.placeholder(tf.int32, (), name="y")
z_ph = 3 * x_ph + 2 * y_ph
linear_fn = function([x_ph, y_ph], z_ph, givens={y_ph: 0})
with single_threaded_session():
initialize()
assert linear_fn(2) == 6
assert linear_fn(2, 2) == 10
def test_multikwargs():
"""
test the function function in tf_util
"""
with tf.Graph().as_default():
x_ph = tf.placeholder(tf.int32, (), name="x")
with tf.variable_scope("other"):
x2_ph = tf.placeholder(tf.int32, (), name="x")
z_ph = 3 * x_ph + 2 * x2_ph
linear_fn = function([x_ph, x2_ph], z_ph, givens={x2_ph: 0})
with single_threaded_session():
initialize()
assert linear_fn(2) == 6
assert linear_fn(2, 2) == 10
def build_act(q_func, ob_space, ac_space, stochastic_ph, update_eps_ph, sess):
"""
Creates the act function:
:param q_func: (DQNPolicy) the policy
:param ob_space: (Gym Space) The observation space of the environment
:param ac_space: (Gym Space) The action space of the environment
:param stochastic_ph: (TensorFlow Tensor) the stochastic placeholder
:param update_eps_ph: (TensorFlow Tensor) the update_eps placeholder
:param sess: (TensorFlow session) The current TensorFlow session
:return: (function (TensorFlow Tensor, bool, float): TensorFlow Tensor, (TensorFlow Tensor, TensorFlow Tensor)
act function to select and action given observation (See the top of the file for details),
A tuple containing the observation placeholder and the processed observation placeholder respectivly.
"""
eps = tf.get_variable("eps", (), initializer=tf.constant_initializer(0))
policy = q_func(sess, ob_space, ac_space, 1, 1, None)
obs_phs = (policy.obs_ph, policy.processed_obs)
deterministic_actions = tf.argmax(tf.add(policy.q_values, policy.action_mask_ph), axis=1)
batch_size = tf.shape(policy.obs_ph)[0]
random_actions = tf.distributions.Categorical(probs=policy.action_mask_probs_ph, dtype=tf.int64).sample()
chose_random = tf.random_uniform(tf.stack([batch_size]), minval=0, maxval=1, dtype=tf.float32) < eps
stochastic_actions = tf.where(chose_random, random_actions, deterministic_actions)
output_actions = tf.cond(stochastic_ph, lambda: stochastic_actions, lambda: deterministic_actions)
update_eps_expr = eps.assign(tf.cond(update_eps_ph >= 0, lambda: update_eps_ph, lambda: eps))
_act = tf_util.function(inputs=[policy.obs_ph, stochastic_ph, update_eps_ph, policy.action_mask_ph,
policy.action_mask_probs_ph],
outputs=output_actions,
givens={update_eps_ph: -1.0, stochastic_ph: True},
updates=[update_eps_expr])
def act(obs, stochastic=True, update_eps=-1, action_mask=None):
if action_mask is not None:
return _act(obs, stochastic, update_eps, policy.prepare_action_mask(action_mask), action_mask)
else:
return _act(obs, stochastic, update_eps)
return act, obs_phs
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)
def learn(env,
policy,
value_fn,
gamma,
lam,
timesteps_per_batch,
num_timesteps,
animate=False,
callback=None,
desired_kl=0.002):
"""
Trains an ACKTR model.
:param env: (Gym environment) The environment to learn from
:param policy: (Object) The policy model to use (MLP, CNN, LSTM, ...)
:param value_fn: (Object) The value function model to use (MLP, CNN, LSTM, ...)
:param gamma: (float) The discount value
:param lam: (float) the tradeoff between exploration and exploitation
:param timesteps_per_batch: (int) the number of timesteps for each batch
:param num_timesteps: (int) the total number of timesteps to run
:param animate: (bool) if render env
:param callback: (function) called every step, used for logging and saving
:param desired_kl: (float) the Kullback leibler weight for the loss
"""
obfilter = ZFilter(env.observation_space.shape)
max_pathlength = env.spec.timestep_limit
stepsize = tf.Variable(initial_value=np.float32(np.array(0.03)),
name='stepsize')
inputs, loss, loss_sampled = policy.update_info
optim = kfac.KfacOptimizer(learning_rate=stepsize,
cold_lr=stepsize * (1 - 0.9),
momentum=0.9,
kfac_update=2,
epsilon=1e-2,
stats_decay=0.99,
async_eigen_decomp=1,
cold_iter=1,
weight_decay_dict=policy.wd_dict,
max_grad_norm=None)
pi_var_list = []
for var in tf.trainable_variables():
if "pi" in var.name:
pi_var_list.append(var)
update_op, q_runner = optim.minimize(loss,
loss_sampled,
var_list=pi_var_list)
do_update = tf_util.function(inputs, update_op)
tf_util.initialize()
# start queue runners
enqueue_threads = []
coord = tf.train.Coordinator()
for queue_runner in [q_runner, value_fn.q_runner]:
assert queue_runner is not None
enqueue_threads.extend(
queue_runner.create_threads(tf.get_default_session(),
coord=coord,
start=True))
i = 0
timesteps_so_far = 0
while True:
if timesteps_so_far > num_timesteps:
break
logger.log("********** Iteration %i ************" % i)
# Collect paths until we have enough timesteps
timesteps_this_batch = 0
paths = []
while True:
path = rollout(env,
policy,
max_pathlength,
animate=(len(paths) == 0 and (i % 10 == 0)
and animate),
obfilter=obfilter)
paths.append(path)
timesteps_this_batch += path["reward"].shape[0]
timesteps_so_far += path["reward"].shape[0]
if timesteps_this_batch > timesteps_per_batch:
break
# Estimate advantage function
vtargs = []
advs = []
for path in paths:
rew_t = path["reward"]
return_t = common.discount(rew_t, gamma)
vtargs.append(return_t)
vpred_t = value_fn.predict(path)
vpred_t = np.append(vpred_t,
0.0 if path["terminated"] else vpred_t[-1])
delta_t = rew_t + gamma * vpred_t[1:] - vpred_t[:-1]
adv_t = common.discount(delta_t, gamma * lam)
advs.append(adv_t)
# Update value function
value_fn.fit(paths, vtargs)
# Build arrays for policy update
ob_no = np.concatenate([path["observation"] for path in paths])
action_na = np.concatenate([path["action"] for path in paths])
oldac_dist = np.concatenate([path["action_dist"] for path in paths])
adv_n = np.concatenate(advs)
standardized_adv_n = (adv_n - adv_n.mean()) / (adv_n.std() + 1e-8)
# Policy update
do_update(ob_no, action_na, standardized_adv_n)
min_stepsize = np.float32(1e-8)
max_stepsize = np.float32(1e0)
# Adjust stepsize
kl_loss = policy.compute_kl(ob_no, oldac_dist)
if kl_loss > desired_kl * 2:
logger.log("kl too high")
tf.assign(stepsize, tf.maximum(min_stepsize,
stepsize / 1.5)).eval()
elif kl_loss < desired_kl / 2:
logger.log("kl too low")
tf.assign(stepsize, tf.minimum(max_stepsize,
stepsize * 1.5)).eval()
else:
logger.log("kl just right!")
logger.record_tabular(
"EpRewMean", np.mean([path["reward"].sum() for path in paths]))
logger.record_tabular(
"EpRewSEM",
np.std([
path["reward"].sum() / np.sqrt(len(paths)) for path in paths
]))
logger.record_tabular(
"EpLenMean", np.mean([path["reward"].shape[0] for path in paths]))
logger.record_tabular("KL", kl_loss)
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()) == False:
break
logger.dump_tabular()
i += 1
coord.request_stop()
coord.join(enqueue_threads)
def __init__(self, ob_dim, ac_dim):
"""
Create a gaussian MLP policy
:param ob_dim: (int) Observation dimention
:param ac_dim: (int) action dimention
"""
# Here we'll construct a bunch of expressions, which will be used in two places:
# (1) When sampling actions
# (2) When computing loss functions, for the policy update
# Variables specific to (1) have the word "sampled" in them,
# whereas variables specific to (2) have the word "old" in them
ob_no = tf.placeholder(tf.float32, shape=[None, ob_dim * 2],
name="ob") # batch of observations
oldac_na = tf.placeholder(
tf.float32, shape=[None, ac_dim],
name="ac") # batch of actions previous actions
# batch of actions previous action distributions
oldac_dist = tf.placeholder(tf.float32,
shape=[None, ac_dim * 2],
name="oldac_dist")
adv_n = tf.placeholder(tf.float32, shape=[None],
name="adv") # advantage function estimate
wd_dict = {}
layer_1 = tf.nn.tanh(
dense(ob_no,
64,
"h1",
weight_init=tf_util.normc_initializer(1.0),
bias_init=0.0,
weight_loss_dict=wd_dict))
layer_2 = tf.nn.tanh(
dense(layer_1,
64,
"h2",
weight_init=tf_util.normc_initializer(1.0),
bias_init=0.0,
weight_loss_dict=wd_dict))
mean_na = dense(layer_2,
ac_dim,
"mean",
weight_init=tf_util.normc_initializer(0.1),
bias_init=0.0,
weight_loss_dict=wd_dict) # Mean control output
self.wd_dict = wd_dict
# Variance on outputs
self.logstd_1a = logstd_1a = tf.get_variable("logstd", [ac_dim],
tf.float32,
tf.zeros_initializer())
logstd_1a = tf.expand_dims(logstd_1a, 0)
std_1a = tf.exp(logstd_1a)
std_na = tf.tile(std_1a, [tf.shape(mean_na)[0], 1])
ac_dist = tf.concat([
tf.reshape(mean_na, [-1, ac_dim]),
tf.reshape(std_na, [-1, ac_dim])
], 1)
# This is the sampled action we'll perform.
sampled_ac_na = tf.random_normal(tf.shape(
ac_dist[:, ac_dim:])) * ac_dist[:, ac_dim:] + ac_dist[:, :ac_dim]
logprobsampled_n = -tf.reduce_sum(tf.log(
ac_dist[:, ac_dim:]), axis=1) - 0.5 * tf.log(
2.0 * np.pi) * ac_dim - 0.5 * tf.reduce_sum(
tf.square(ac_dist[:, :ac_dim] - sampled_ac_na) /
(tf.square(ac_dist[:, ac_dim:])),
axis=1) # Logprob of sampled action
logprob_n = -tf.reduce_sum(
tf.log(ac_dist[:, ac_dim:]), axis=1
) - 0.5 * tf.log(2.0 * np.pi) * ac_dim - 0.5 * tf.reduce_sum(
tf.square(ac_dist[:, :ac_dim] - oldac_na) /
(tf.square(ac_dist[:, ac_dim:])),
axis=1
) # Logprob of previous actions under CURRENT policy (whereas oldlogprob_n is under OLD policy)
kl_loss = tf.reduce_mean(kl_div(oldac_dist, ac_dist, ac_dim))
# kl = .5 * tf.reduce_mean(tf.square(logprob_n - oldlogprob_n))
# Approximation of KL divergence between old policy used to generate actions,
# and new policy used to compute logprob_n
surr = -tf.reduce_mean(
adv_n * logprob_n
) # Loss function that we'll differentiate to get the policy gradient
surr_sampled = -tf.reduce_mean(logprob_n) # Sampled loss of the policy
# Generate a new action and its logprob
self._act = tf_util.function(
[ob_no], [sampled_ac_na, ac_dist, logprobsampled_n])
# self.compute_kl = U.function([ob_no, oldac_na, oldlogprob_n], kl)
# Compute (approximate) KL divergence between old policy and new policy
self.compute_kl = tf_util.function([ob_no, oldac_dist], kl_loss)
# Input and output variables needed for computing loss
self.update_info = ((ob_no, oldac_na, adv_n), surr, surr_sampled)
tf_util.initialize() # Initialize uninitialized TF variables
def build_train(q_func,
ob_space,
ac_space,
sess,
num_actions,
num_action_streams,
batch_size,
learning_rate=5e-4,
aggregator='reduceLocalMean',
optimizer_name="Adam",
grad_norm_clipping=None,
gamma=0.99,
double_q=True,
scope="bdq",
reuse=None,
losses_version=2,
independent=False,
dueling=True,
target_version="mean",
loss_type="L2",
full_tensorboard_log=False):
"""Creates the act function:
Parameters
----------
make_obs_ph: str -> tf.placeholder or TfInput
a function that takes a name and creates a placeholder of input with that name
q_func: (tf.Variable, int, str, bool) -> tf.Variable
the model that takes the following inputs:
observation_in: object
the output of observation placeholder
num_actions: int
number of actions
scope: str
reuse: bool
should be passed to outer variable scope
and returns a tensor of shape (batch_size, num_actions) with values of every action.
num_actions: int
total number of sub-actions to be represented at the output
num_action_streams: int
specifies the number of action branches in action value (or advantage) function representation
batch_size: int
size of the sampled mini-batch from the replay buffer
reuse: bool
whether or not to reuse the graph variables
optimizer: tf.train.Optimizer
optimizer to use for deep Q-learning
grad_norm_clipping: float or None
clip graident norms to this value. If None no clipping is performed.
gamma: float
discount rate.
double_q: bool
if true will use Double Q-Learning (https://arxiv.org/abs/1509.06461).
In general it is a good idea to keep it enabled. BDQ uses it.
scope: str or VariableScope
optional scope for variable_scope.
reuse: bool or None
whether or not the variables should be reused. To be able to reuse the scope must be given.
losses_version: int
specifies the version number for merging of losses across the branches
version 2 is the best-performing loss used for BDQ.
Returns
-------
act: (tf.Variable, bool, float) -> tf.Variable
function to select an action given an observation.
` See the top of the file for details.
train: (object, np.array, np.array, object, np.array, np.array) -> np.array
optimize the error in Bellman's equation.
` See the top of the file for details.
update_target: () -> ()
copy the parameters from optimized Q function to the target Q function.
` See the top of the file for details.
debug: {str: function}
a bunch of functions to print debug data like q_values.
"""
assert independent and losses_version == 4 or not independent, 'independent needs to be used along with loss v4'
assert independent and target_version == "indep" or not independent, 'independent needs to be used along with independent TD targets'
with tf.variable_scope("input", reuse=reuse):
stochastic_ph = tf.placeholder(tf.bool, (), name="stochastic")
update_eps_ph = tf.placeholder(tf.float32, (), name="update_eps")
with tf.variable_scope(scope, reuse=reuse):
act_f, obs_phs = build_act(q_func, ob_space, ac_space, sess,
num_actions, num_action_streams,
stochastic_ph, update_eps_ph)
# Q-network evaluation
with tf.variable_scope(
"step_model",
reuse=True,
custom_getter=tf_util.outer_scope_getter("step_model")):
step_model = q_func(sess,
ob_space,
ac_space,
1,
1,
None,
num_actions,
reuse=True,
obs_phs=obs_phs) # reuse parameters from act
q_func_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=tf.get_variable_scope().name +
"/model")
# Target Q-network evalution
with tf.variable_scope("target_q_func", reuse=False):
target_policy = q_func(sess,
ob_space,
ac_space,
1,
1,
None,
num_actions,
reuse=False)
target_q_func_vars = tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES,
scope=tf.get_variable_scope().name + "/target_q_func")
# compute estimate of best possible value starting from state at t + 1
double_obs_ph = target_policy.obs_ph
if double_q:
with tf.variable_scope(
"q_func",
reuse=True,
custom_getter=tf_util.outer_scope_getter("q_func")):
selection_q_tp1 = q_func(sess,
ob_space,
ac_space,
1,
1,
None,
num_actions,
reuse=True)
double_obs_ph = selection_q_tp1.obs_ph
else:
selection_q_tp1 = target_policy
num_actions_pad = num_actions // num_action_streams
with tf.variable_scope("loss", reuse=reuse):
act_t_ph = tf.placeholder(tf.int32, [None, num_action_streams],
name="action")
rew_t_ph = tf.placeholder(tf.float32, [None], name="reward")
done_mask_ph = tf.placeholder(tf.float32, [None], name="done")
importance_weights_ph = tf.placeholder(tf.float32, [None],
name="weight")
q_values = []
for dim in range(num_action_streams):
selected_a = tf.squeeze(
tf.slice(act_t_ph, [0, dim], [batch_size, 1])) # TODO better?
q_values.append(
tf.reduce_sum(tf.one_hot(selected_a, num_actions_pad) *
step_model.q_values[dim],
axis=1))
if target_version == "indep":
target_q_values = []
for dim in range(num_action_streams):
selected_a = tf.argmax(selection_q_tp1.q_values[dim], axis=1)
selected_q = tf.reduce_sum(
tf.one_hot(selected_a, num_actions_pad) *
target_policy.q_values[dim],
axis=1)
masked_selected_q = (1.0 - done_mask_ph) * selected_q
target_q = rew_t_ph + gamma * masked_selected_q
target_q_values.append(target_q)
elif target_version == "max":
for dim in range(num_action_streams):
selected_a = tf.argmax(selection_q_tp1.q_values[dim], axis=1)
selected_q = tf.reduce_sum(
tf.one_hot(selected_a, num_actions_pad) *
target_policy.q_values[dim],
axis=1)
masked_selected_q = (1.0 - done_mask_ph) * selected_q
if dim == 0:
max_next_q_values = masked_selected_q
else:
max_next_q_values = tf.maximum(max_next_q_values,
masked_selected_q)
target_q_values = [rew_t_ph + gamma * max_next_q_values
] * num_action_streams # TODO better?
elif target_version == "mean":
for dim in range(num_action_streams):
selected_a = tf.argmax(selection_q_tp1.q_values[dim], axis=1)
selected_q = tf.reduce_sum(
tf.one_hot(selected_a, num_actions_pad) *
target_policy.q_values[dim],
axis=1)
masked_selected_q = (1.0 - done_mask_ph) * selected_q
if dim == 0:
mean_next_q_values = masked_selected_q
else:
mean_next_q_values += masked_selected_q
mean_next_q_values /= num_action_streams
target_q_values = [rew_t_ph + gamma * mean_next_q_values
] * num_action_streams # TODO better?
else:
assert False, 'unsupported target version ' + str(target_version)
if optimizer_name == "Adam":
optimizer = tf.train.AdamOptimizer(learning_rate)
else:
assert False, 'unsupported optimizer ' + str(optimizer_name)
if loss_type == "L2":
loss_function = tf.square
elif loss_type == "Huber":
loss_function = tf_util.huber_loss
else:
assert False, 'unsupported loss type ' + str(loss_type)
if losses_version == 1:
mean_q_value = sum(q_values) / num_action_streams
mean_target_q_value = sum(target_q_values) / num_action_streams
td_error = mean_q_value - tf.stop_gradient(mean_target_q_value)
loss = loss_function(td_error)
weighted_mean_loss = tf.reduce_mean(importance_weights_ph * loss)
optimize_expr = minimize_and_clip(
optimizer,
weighted_mean_loss,
var_list=q_func_vars,
total_n_streams=(num_action_streams + (1 if dueling else 0)),
clip_val=grad_norm_clipping)
optimize_expr = [optimize_expr]
tf.summary.scalar("loss", weighted_mean_loss)
else:
stream_losses = []
for dim in range(num_action_streams):
dim_td_error = q_values[dim] - tf.stop_gradient(
target_q_values[dim])
dim_loss = loss_function(dim_td_error)
# Scaling of learning based on importance sampling weights is optional, either way works
stream_losses.append(
tf.reduce_mean(dim_loss *
importance_weights_ph)) # with scaling
#stream_losses.append(tf.reduce_mean(dim_loss)) # without scaling
if dim == 0:
td_error = tf.abs(dim_td_error)
else:
td_error += tf.abs(dim_td_error)
#td_error /= num_action_streams
mean_loss = sum(stream_losses) / num_action_streams
optimize_expr = minimize_and_clip(
optimizer,
mean_loss,
var_list=q_func_vars,
total_n_streams=(num_action_streams + (1 if dueling else 0)),
clip_val=grad_norm_clipping)
optimize_expr = [optimize_expr]
tf.summary.scalar("loss", mean_loss)
## FIXME tf summary scalars are wrong. They are not matching the original code.
tf.summary.scalar("td_error", tf.reduce_mean(td_error))
if full_tensorboard_log:
tf.summary.histogram("td_error", td_error)
# Target Q-network parameters are periodically updated with the Q-network's
update_target_expr = []
for var, var_target in zip(
sorted(q_func_vars, key=lambda v: v.name),
sorted(target_q_func_vars, key=lambda v: v.name)):
update_target_expr.append(var_target.assign(var))
update_target_expr = tf.group(*update_target_expr)
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('rewards', tf.reduce_mean(rew_t_ph))
tf.summary.scalar('importance_weights',
tf.reduce_mean(importance_weights_ph))
if full_tensorboard_log:
tf.summary.histogram('rewards', rew_t_ph)
tf.summary.histogram('importance_weights', importance_weights_ph)
if tf_util.is_image(obs_phs[0]):
tf.summary.image('observation', obs_phs[0])
elif len(obs_phs[0].shape) == 1:
tf.summary.histogram('observation', obs_phs[0])
# optimize_expr = optimizer.apply_gradients(gradients)
summary = tf.summary.merge_all()
train = tf_util.function(
inputs=[
obs_phs[0],
act_t_ph,
rew_t_ph,
target_policy.obs_ph, #obs_tp1_input,
double_obs_ph,
done_mask_ph,
importance_weights_ph
],
outputs=[summary, td_error],
updates=[optimize_expr])
update_target = tf_util.function([], [], updates=[update_target_expr])
# q_values = tf_util.function([obs_phs[0]], step_model)
return act_f, train, update_target, step_model #{'q_values': q_values}
def __init__(self,
observation_space,
action_space,
hidden_size,
entcoeff=0.001,
scope="adversary",
normalize=True):
"""
Reward regression from observations and transitions
:param observation_space: (gym.spaces)
:param action_space: (gym.spaces)
:param hidden_size: ([int]) the hidden dimension for the MLP
:param entcoeff: (float) the entropy loss weight
:param scope: (str) tensorflow variable scope
:param normalize: (bool) Whether to normalize the reward or not
"""
# TODO: support images properly (using a CNN)
self.scope = scope
self.observation_shape = observation_space.shape
self.actions_shape = action_space.shape
if isinstance(action_space, gym.spaces.Box):
# Continuous action space
self.discrete_actions = False
self.n_actions = action_space.shape[0]
elif isinstance(action_space, gym.spaces.Discrete):
self.n_actions = action_space.n
self.discrete_actions = True
else:
raise ValueError(
'Action space not supported: {}'.format(action_space))
self.hidden_size = hidden_size
self.normalize = normalize
self.obs_rms = None
# Placeholders
self.generator_obs_ph = tf.compat.v1.placeholder(
observation_space.dtype, (None, ) + self.observation_shape,
name="observations_ph")
self.generator_acs_ph = tf.compat.v1.placeholder(
action_space.dtype, (None, ) + self.actions_shape,
name="actions_ph")
self.expert_obs_ph = tf.compat.v1.placeholder(
observation_space.dtype, (None, ) + self.observation_shape,
name="expert_observations_ph")
self.expert_acs_ph = tf.compat.v1.placeholder(
action_space.dtype, (None, ) + self.actions_shape,
name="expert_actions_ph")
# Build graph
generator_logits = self.build_graph(self.generator_obs_ph,
self.generator_acs_ph,
reuse=False)
expert_logits = self.build_graph(self.expert_obs_ph,
self.expert_acs_ph,
reuse=True)
# Build accuracy
generator_acc = tf.reduce_mean(input_tensor=tf.cast(
tf.nn.sigmoid(generator_logits) < 0.5, tf.float32))
expert_acc = tf.reduce_mean(input_tensor=tf.cast(
tf.nn.sigmoid(expert_logits) > 0.5, tf.float32))
# Build regression loss
# let x = logits, z = targets.
# z * -log(sigmoid(x)) + (1 - z) * -log(1 - sigmoid(x))
generator_loss = tf.nn.sigmoid_cross_entropy_with_logits(
logits=generator_logits, labels=tf.zeros_like(generator_logits))
generator_loss = tf.reduce_mean(input_tensor=generator_loss)
expert_loss = tf.nn.sigmoid_cross_entropy_with_logits(
logits=expert_logits, labels=tf.ones_like(expert_logits))
expert_loss = tf.reduce_mean(input_tensor=expert_loss)
# Build entropy loss
logits = tf.concat([generator_logits, expert_logits], 0)
entropy = tf.reduce_mean(input_tensor=logit_bernoulli_entropy(logits))
entropy_loss = -entcoeff * entropy
# Loss + Accuracy terms
self.losses = [
generator_loss, expert_loss, entropy, entropy_loss, generator_acc,
expert_acc
]
self.loss_name = [
"generator_loss", "expert_loss", "entropy", "entropy_loss",
"generator_acc", "expert_acc"
]
self.total_loss = generator_loss + expert_loss + entropy_loss
# Build Reward for policy
self.reward_op = -tf.math.log(1 - tf.nn.sigmoid(generator_logits) +
1e-8)
var_list = self.get_trainable_variables()
self.lossandgrad = tf_util.function([
self.generator_obs_ph, self.generator_acs_ph, self.expert_obs_ph,
self.expert_acs_ph
], self.losses + [tf_util.flatgrad(self.total_loss, var_list)])
def setup_model(self):
# prevent import loops
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()
print("number of workers are", self.nworkers)
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)
# 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 phi
with tf.variable_scope("phi", reuse=False):
self.policy_phi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
# Network for phi old
with tf.variable_scope("oldphi", reuse=False):
self.policy_phi_old = 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)
#kloldnew = self.policy_pi.proba_distribution.kl(old_policy.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))
vf_phi_err = tf.reduce_mean(
tf.square(self.policy_phi.value_flat - ret))
vf_phi_old_err = tf.reduce_mean(
tf.square(self.policy_phi_old.value_flat))
# 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
]
all_var_oldpi_list = tf_util.get_trainable_vars("oldpi")
var_oldpi_list = [
v for v in all_var_oldpi_list
if "/vf" not in v.name and "/q/" not in v.name
]
all_var_phi_list = tf_util.get_trainable_vars("phi")
vf_phi_var_list = [
v for v in all_var_phi_list if "/pi" not in v.name
and "/logstd" not in v.name and "/q" not in v.name
]
all_var_phi_old_list = tf_util.get_trainable_vars("oldphi")
vf_phi_old_var_list = [
v for v in all_var_phi_old_list if "/pi" not in v.name
and "/logstd" not in v.name and "/q" not in v.name
]
#print("vars", vf_var_list)
self.policy_vars = all_var_list
self.oldpolicy_vars = all_var_oldpi_list
print("all var list", all_var_list)
print("phi vars", vf_phi_var_list)
print("phi old vars", vf_phi_old_var_list)
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, 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))
self.compute_vf_phi_lossandgrad = tf_util.function(
[observation, self.policy_phi.obs_ph, ret],
tf_util.flatgrad(vf_phi_err, vf_phi_var_list))
self.compute_vf_loss = tf_util.function(
[observation, old_policy.obs_ph, ret], vferr)
self.compute_vf_phi_loss = tf_util.function(
[observation, self.policy_phi.obs_ph, ret], vf_phi_err)
#self.compute_vf_phi_old_loss = tf_util.function([self.policy_phi_old.obs_ph], vf_phi_old_err)
#self.phi_old_obs = np.array([-0.012815 , -0.02076313, 0.07524705, 0.09407324, 0.0901745 , -0.09339058, 0.03544853, -0.03297224])
#self.phi_old_obs = self.phi_old_obs.reshape((1, 8))
update_phi_old_expr = []
for var, var_target in zip(
sorted(vf_phi_var_list, key=lambda v: v.name),
sorted(vf_phi_old_var_list, key=lambda v: v.name)):
update_phi_old_expr.append(var_target.assign(var))
update_phi_old_expr = tf.group(*update_phi_old_expr)
self.update_phi_old = tf_util.function(
[], [], updates=[update_phi_old_expr])
@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
@contextmanager
def temp_seed(seed):
state = np.random.get_state()
np.random.seed(seed)
try:
yield
finally:
np.random.set_state(state)
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)
self.vf_phi_adam = MpiAdam(vf_phi_var_list, sess=self.sess)
self.vfadam.sync()
self.vf_phi_adam.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))
self.timed = timed
self.allmean = allmean
self.temp_seed = temp_seed
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")
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 build_act(q_func, ob_space, ac_space, stochastic_ph, update_eps_ph, sess, action_filter=None, OnExpo=False, threshold=None): # NKAM
"""
Creates the act function:
:param q_func: (DQNPolicy) the policy
:param ob_space: (Gym Space) The observation space of the environment
:param ac_space: (Gym Space) The action space of the environment
:param stochastic_ph: (TensorFlow Tensor) the stochastic placeholder
:param update_eps_ph: (TensorFlow Tensor) the update_eps placeholder
:param sess: (TensorFlow session) The current TensorFlow session
:param action_filter: (function (TensorFlow Tensor): TensorFlow Tensor) filters actions according to a criterion
:return: (function (TensorFlow Tensor, bool, float): TensorFlow Tensor, (TensorFlow Tensor, TensorFlow Tensor)
act function to select and action given observation (See the top of the file for details),
A tuple containing the observation placeholder and the processed observation placeholder respectivly.
"""
eps = tf.get_variable("eps", (), initializer=tf.constant_initializer(0))
policy = q_func(sess, ob_space, ac_space, 1, 1, None)
obs_phs = (policy.obs_ph, policy.processed_obs)
batch_size = tf.shape(policy.obs_ph)[0]
n_actions = ac_space.nvec if isinstance(ac_space, MultiDiscrete) else ac_space.n
if action_filter != None:
actions_mask = tf.py_func(func=action_filter, inp=[tf.constant(n_actions)], Tout=[tf.bool for _ in range(0, n_actions)])
total_possible_actions = tf.reduce_sum(tf.cast(actions_mask, dtype=tf.int64))
actions_mask = tf.reshape(tf.tile(actions_mask, [batch_size]), [batch_size, tf.shape(actions_mask)[0]])
if threshold != None:
legal_q_values = tf.boolean_mask(policy.q_values, actions_mask)
max_legal_q_value = tf.math.reduce_max(legal_q_values)
max_legal_q_value = tf.math.scalar_mul(threshold, max_legal_q_value)
threshold_tensor = tf.tile(tf.reshape(max_legal_q_value, (1,)), tf.constant(n_actions, dtype=tf.int64, shape=(1,)))
threshold_tensor = tf.reshape(tf.tile(threshold_tensor, [batch_size]), [batch_size, tf.shape(threshold_tensor)[0]])
threshold_mask = tf.math.greater(policy.q_values, threshold_tensor)
actions_mask = tf.math.logical_or(actions_mask, threshold_mask)
# TODO: how to avoid getting stuck???
# if actions_mask == "empty":
# actions_mask = None
else:
actions_mask = tf.constant([True for _ in range(0, n_actions)], dtype=tf.bool)
total_possible_actions = n_actions
actions_mask = tf.reshape(tf.tile(actions_mask, [batch_size]), [batch_size, tf.shape(actions_mask)[0]])
q_values_mask = tf.constant([-float('inf') for _ in range(0, n_actions)])
q_values_mask = tf.reshape(tf.tile(q_values_mask, [batch_size]), [batch_size, tf.shape(q_values_mask)[0]])
masked_q_values = tf.where(actions_mask, policy.q_values, q_values_mask)
deterministic_actions = tf.argmax(masked_q_values, axis=1)
possible_actions = tf.boolean_mask(tf.constant([action for action in range(0, n_actions)], dtype=tf.int64), actions_mask[0])
masked_random_actions = tf.random_uniform(tf.stack([batch_size]), minval=0, maxval=total_possible_actions, dtype=tf.int64)
if not OnExpo:
random_actions = tf.random_uniform(tf.stack([batch_size]), minval=0, maxval=n_actions, dtype=tf.int64)
else:
random_actions = tf.gather(possible_actions, masked_random_actions)
# random_actions = tf.gather(possible_actions, masked_random_actions, batch_dims=0)
chose_random = tf.random_uniform(tf.stack([batch_size]), minval=0, maxval=1, dtype=tf.float32) < eps
stochastic_actions = tf.where(chose_random, random_actions, deterministic_actions)
output_actions = tf.cond(stochastic_ph, lambda: stochastic_actions, lambda: deterministic_actions)
update_eps_expr = eps.assign(tf.cond(update_eps_ph >= 0, lambda: update_eps_ph, lambda: eps))
_act = tf_util.function(inputs=[policy.obs_ph, stochastic_ph, update_eps_ph],
outputs=output_actions,
givens={update_eps_ph: -1.0, stochastic_ph: True},
updates=[update_eps_expr])
def act(obs, stochastic=True, update_eps=-1):
return _act(obs, stochastic, update_eps)
return act, obs_phs
def _init(self,
ob_space,
ac_space,
hid_size,
num_hid_layers,
gaussian_fixed_var=True):
# observation_space 와 action_space
# ob_space = env.observation_space
# ac_space = env.action_space
obs, pdtype = self.get_obs_and_pdtype(ob_space, ac_space)
# return ::
# obs :: placeholder
# pdtype :: action_space의 distribution정보를 담은 객체
with tf.variable_scope("obfilter"):
self.ob_rms = RunningMeanStd(shape=ob_space.shape)
# DeepMind의 observation normalization
obz = tf.clip_by_value((obs - self.ob_rms.mean) / self.ob_rms.std,
-5.0, 5.0)
# ===========================[Value function prediction Model]====================================================
# dense()는 tf.nn.bias_add(tf.matmul(input_tensor, weight), bias) 를 리턴해준다.
last_out = obz
for i in range(num_hid_layers):
last_out = tf.nn.tanh(
dense(last_out,
hid_size,
"vffc%i" % (i + 1),
weight_init=tf_util.normc_initializer(1.0)))
self.vpred = dense(last_out,
1,
"vffinal",
weight_init=tf_util.normc_initializer(1.0))[:, 0]
# dense(last, 1) 의 경우 return shape 가 [None, 1]이므로, [None]을 얻기 위해서 [:, 0]를 해준다.
# ===========================[Policy function MOdel]==============================================================
last_out = obz
for i in range(num_hid_layers):
last_out = tf.nn.tanh(
dense(last_out,
hid_size,
"polfc%i" % (i + 1),
weight_init=tf_util.normc_initializer(1.0)))
# 선택한 gym 환경의 action space가 Box 형태인 경우 다음과 같이
# action distribution의 mean 과 std 를 concate 한 값을
# action output으로 사용한다.
if gaussian_fixed_var and isinstance(ac_space, gym.spaces.Box):
mean = dense(last_out,
pdtype.param_shape()[0] // 2, "polfinal",
tf_util.normc_initializer(0.01))
logstd = tf.get_variable(name="logstd",
shape=[1, pdtype.param_shape()[0] // 2],
initializer=tf.zeros_initializer())
pdparam = tf.concat([mean, mean * 0.0 + logstd], axis=1)
else:
pdparam = dense(last_out,
pdtype.param_shape()[0], "polfinal",
tf_util.normc_initializer(0.01))
# 이 모델에서의 action output은 다음 proba_distribution 에 담기도록 합니다.
self.proba_distribution = pdtype.proba_distribution_from_flat(pdparam)
self.state_in = []
self.state_out = []
# ppo1/mlp_policy class를 상송했으므로
# 해당 클래스에서 정의되어 있는 act 함수를 재정의
# _act를 정의해주면
# act()를 통해서 action / value 를 얻을수 있다.
self.stochastic_ph = tf.placeholder(dtype=tf.bool,
shape=(),
name="stochastic")
action = tf_util.switch(self.stochastic_ph,
self.proba_distribution.sample(),
self.proba_distribution.mode())
self.action = action
self._act = tf_util.function([self.stochastic_ph, obs],
[action, self.vpred])
def _init(self,
ob_space,
ac_space,
hid_size,
num_hid_layers,
gaussian_fixed_var=True,
num_options=2,
dc=0):
assert isinstance(ob_space, gym.spaces.Box)
self.dc = dc
self.num_options = num_options
self.pdtype = pdtype = make_pdtype(ac_space)
sequence_length = None
ob = U.get_placeholder(name="ob",
dtype=tf.float32,
shape=[sequence_length] + list(ob_space.shape))
option = U.get_placeholder(name="option", dtype=tf.int32, shape=[None])
with tf.variable_scope("obfilter"):
self.ob_rms = RunningMeanStd(shape=ob_space.shape)
# normalization
obz = tf.clip_by_value((ob - self.ob_rms.mean) / self.ob_rms.std, -5.0,
5.0)
last_out = obz
for i in range(num_hid_layers):
last_out = tf.nn.tanh(
tf.layers.dense(last_out,
hid_size,
kernel_initializer=U.normc_initializer(1.0),
name="vffc%i" % (i + 1)))
self.vpred = dense3D2(last_out,
1,
"vffinal",
option,
num_options=num_options,
weight_init=U.normc_initializer(1.0))[:, 0]
self.tpred = tf.nn.sigmoid(
dense3D2(tf.stop_gradient(last_out),
1,
"termhead",
option,
num_options=num_options,
weight_init=U.normc_initializer(1.0)))[:, 0]
termination_sample = tf.greater(
self.tpred, tf.random_uniform(shape=tf.shape(self.tpred),
maxval=1.))
# input to policy
last_out = obz
for i in range(num_hid_layers):
last_out = tf.nn.tanh(
tf.layers.dense(last_out,
hid_size,
name="polfc%i" % (i + 1),
kernel_initializer=U.normc_initializer(1.0)))
if gaussian_fixed_var and isinstance(ac_space, gym.spaces.Box):
mean = dense3D2(last_out,
pdtype.param_shape()[0] // 2,
"polfinal",
option,
num_options=num_options,
weight_init=U.normc_initializer(0.01))
logstd = tf.get_variable(
name="logstd",
shape=[num_options, 1,
pdtype.param_shape()[0] // 2],
initializer=tf.zeros_initializer())
pdparam = tf.concat([mean, mean * 0.0 + logstd[option[0]]], axis=1)
else:
pdparam = tf.layers.dense(
last_out,
pdtype.param_shape()[0],
name="polfinal",
kernel_initializer=U.normc_initializer(0.01))
# select option
self.op_pi = tf.nn.softmax(
tf.layers.dense(tf.stop_gradient(last_out),
num_options,
name="OPfc%i" % (i + 1),
kernel_initializer=U.normc_initializer(1.0)))
self.pd = pdtype.pdfromflat(pdparam)
self.state_in = []
self.state_out = []
stochastic = tf.placeholder(dtype=tf.bool, shape=())
ac = U.switch(stochastic, self.pd.sample(), self.pd.mode())
self._act = U.function([stochastic, ob, option],
[ac, self.vpred, last_out, logstd])
self._get_v = U.function([ob, option], [self.vpred])
self.get_term = U.function([ob, option], [termination_sample])
self.get_tpred = U.function([ob, option], [self.tpred])
self.get_vpred = U.function([ob, option], [self.vpred])
self._get_op = U.function([ob], [self.op_pi])
def learn(env,
policy_func,
dataset,
optim_batch_size=128,
max_iters=1e4,
adam_epsilon=1e-5,
optim_stepsize=3e-4,
ckpt_dir=None,
task_name=None,
verbose=False):
"""
Learn a behavior clone policy, and return the save location
:param env: (Gym Environment) the environment
:param policy_func: (function (str, Gym Space, Gym Space): TensorFlow Tensor) creates the policy
:param dataset: (Dset or MujocoDset) the dataset manager
:param optim_batch_size: (int) the batch size
:param max_iters: (int) the maximum number of iterations
:param adam_epsilon: (float) the epsilon value for the adam optimizer
:param optim_stepsize: (float) the optimizer stepsize
:param ckpt_dir: (str) the save directory, can be None for temporary directory
:param task_name: (str) the save name, can be None for saving directly to the directory name
:param verbose: (bool)
:return: (str) the save location for the TensorFlow model
"""
val_per_iter = int(max_iters / 10)
ob_space = env.observation_space
ac_space = env.action_space
policy = policy_func("pi", ob_space,
ac_space) # Construct network for new policy
# placeholder
obs_ph = policy.obs_ph
action_ph = policy.pdtype.sample_placeholder([None])
stochastic_ph = policy.stochastic_ph
loss = tf.reduce_mean(tf.square(action_ph - policy.ac))
var_list = policy.get_trainable_variables()
adam = MpiAdam(var_list, epsilon=adam_epsilon)
lossandgrad = tf_util.function([obs_ph, action_ph, stochastic_ph],
[loss] + [tf_util.flatgrad(loss, var_list)])
tf_util.initialize()
adam.sync()
logger.log("Pretraining with Behavior Cloning...")
for iter_so_far in tqdm(range(int(max_iters))):
ob_expert, ac_expert = dataset.get_next_batch(optim_batch_size,
'train')
train_loss, grad = lossandgrad(ob_expert, ac_expert, True)
adam.update(grad, optim_stepsize)
if verbose and iter_so_far % val_per_iter == 0:
ob_expert, ac_expert = dataset.get_next_batch(-1, 'val')
val_loss, _ = lossandgrad(ob_expert, ac_expert, True)
logger.log("Training loss: {}, Validation loss: {}".format(
train_loss, val_loss))
if ckpt_dir is None:
savedir_fname = tempfile.TemporaryDirectory().name
else:
savedir_fname = os.path.join(ckpt_dir, task_name)
tf_util.save_state(savedir_fname, var_list=policy.get_variables())
return savedir_fname
def setup_model(self):
with SetVerbosity(self.verbose):
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)
# 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.compat.v1.variable_scope("oldpi", reuse=False):
old_pi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
with tf.compat.v1.variable_scope("loss", reuse=False):
# Target advantage function (if applicable)
atarg = tf.compat.v1.placeholder(dtype=tf.float32,
shape=[None])
# Empirical return
ret = tf.compat.v1.placeholder(dtype=tf.float32,
shape=[None])
# learning rate multiplier, updated with schedule
lrmult = tf.compat.v1.placeholder(name='lrmult',
dtype=tf.float32,
shape=[])
# Annealed cliping parameter epislon
clip_param = self.clip_param * lrmult
obs_ph = self.policy_pi.obs_ph
action_ph = self.policy_pi.pdtype.sample_placeholder(
[None])
kloldnew = old_pi.proba_distribution.kl(
self.policy_pi.proba_distribution)
ent = self.policy_pi.proba_distribution.entropy()
meankl = tf.reduce_mean(input_tensor=kloldnew)
meanent = tf.reduce_mean(input_tensor=ent)
pol_entpen = (-self.entcoeff) * meanent
# pnew / pold
ratio = tf.exp(
self.policy_pi.proba_distribution.logp(action_ph) -
old_pi.proba_distribution.logp(action_ph))
# surrogate from conservative policy iteration
surr1 = ratio * atarg
surr2 = tf.clip_by_value(ratio, 1.0 - clip_param,
1.0 + clip_param) * atarg
# PPO's pessimistic surrogate (L^CLIP)
pol_surr = -tf.reduce_mean(
input_tensor=tf.minimum(surr1, surr2))
vf_loss = tf.reduce_mean(
input_tensor=tf.square(self.policy_pi.value_flat -
ret))
total_loss = pol_surr + pol_entpen + vf_loss
losses = [pol_surr, pol_entpen, vf_loss, meankl, meanent]
self.loss_names = [
"pol_surr", "pol_entpen", "vf_loss", "kl", "ent"
]
tf.compat.v1.summary.scalar('entropy_loss', pol_entpen)
tf.compat.v1.summary.scalar('policy_gradient_loss',
pol_surr)
tf.compat.v1.summary.scalar('value_function_loss', vf_loss)
tf.compat.v1.summary.scalar('approximate_kullback-leibler',
meankl)
tf.compat.v1.summary.scalar('clip_factor', clip_param)
tf.compat.v1.summary.scalar('loss', total_loss)
self.params = tf_util.get_trainable_vars("model")
self.assign_old_eq_new = tf_util.function(
[], [],
updates=[
tf.compat.v1.assign(oldv, newv)
for (oldv, newv) in zipsame(
tf_util.get_globals_vars("oldpi"),
tf_util.get_globals_vars("model"))
])
with tf.compat.v1.variable_scope("Adam_mpi", reuse=False):
self.adam = MpiAdam(self.params,
epsilon=self.adam_epsilon,
sess=self.sess)
with tf.compat.v1.variable_scope("input_info", reuse=False):
tf.compat.v1.summary.scalar(
'discounted_rewards', tf.reduce_mean(input_tensor=ret))
tf.compat.v1.summary.scalar(
'learning_rate',
tf.reduce_mean(input_tensor=self.optim_stepsize))
tf.compat.v1.summary.scalar(
'advantage', tf.reduce_mean(input_tensor=atarg))
tf.compat.v1.summary.scalar(
'clip_range',
tf.reduce_mean(input_tensor=self.clip_param))
if self.full_tensorboard_log:
tf.compat.v1.summary.histogram('discounted_rewards',
ret)
tf.compat.v1.summary.histogram('learning_rate',
self.optim_stepsize)
tf.compat.v1.summary.histogram('advantage', atarg)
tf.compat.v1.summary.histogram('clip_range',
self.clip_param)
if tf_util.is_image(self.observation_space):
tf.compat.v1.summary.image('observation', obs_ph)
else:
tf.compat.v1.summary.histogram(
'observation', obs_ph)
self.step = self.policy_pi.step
self.proba_step = self.policy_pi.proba_step
self.initial_state = self.policy_pi.initial_state
tf_util.initialize(sess=self.sess)
self.summary = tf.compat.v1.summary.merge_all()
self.lossandgrad = tf_util.function(
[obs_ph, old_pi.obs_ph, action_ph, atarg, ret, lrmult],
[self.summary,
tf_util.flatgrad(total_loss, self.params)] + losses)
self.compute_losses = tf_util.function(
[obs_ph, old_pi.obs_ph, action_ph, atarg, ret, lrmult],
losses)
def build_act_with_param_noise(q_func, ob_space, ac_space, stochastic_ph, update_eps_ph, sess,
param_noise_filter_func=None):
"""
Creates the act function with support for parameter space noise exploration (https://arxiv.org/abs/1706.01905):
:param q_func: (DQNPolicy) the policy
:param ob_space: (Gym Space) The observation space of the environment
:param ac_space: (Gym Space) The action space of the environment
:param stochastic_ph: (TensorFlow Tensor) the stochastic placeholder
:param update_eps_ph: (TensorFlow Tensor) the update_eps placeholder
:param sess: (TensorFlow session) The current TensorFlow session
:param param_noise_filter_func: (function (TensorFlow Tensor): bool) function that decides whether or not a
variable should be perturbed. Only applicable if param_noise is True. If set to None, default_param_noise_filter
is used by default.
:return: (function (TensorFlow Tensor, bool, float): TensorFlow Tensor, (TensorFlow Tensor, TensorFlow Tensor)
act function to select and action given observation (See the top of the file for details),
A tuple containing the observation placeholder and the processed observation placeholder respectivly.
"""
if param_noise_filter_func is None:
param_noise_filter_func = default_param_noise_filter
update_param_noise_threshold_ph = tf.placeholder(tf.float32, (), name="update_param_noise_threshold")
update_param_noise_scale_ph = tf.placeholder(tf.bool, (), name="update_param_noise_scale")
reset_ph = tf.placeholder(tf.bool, (), name="reset")
eps = tf.get_variable("eps", (), initializer=tf.constant_initializer(0))
param_noise_scale = tf.get_variable("param_noise_scale", (), initializer=tf.constant_initializer(0.01),
trainable=False)
param_noise_threshold = tf.get_variable("param_noise_threshold", (), initializer=tf.constant_initializer(0.05),
trainable=False)
# Unmodified Q.
policy = q_func(sess, ob_space, ac_space, 1, 1, None)
obs_phs = (policy.obs_ph, policy.processed_obs)
# Perturbable Q used for the actual rollout.
with tf.variable_scope("perturbed_model", reuse=False):
perturbable_policy = q_func(sess, ob_space, ac_space, 1, 1, None, obs_phs=obs_phs)
def perturb_vars(original_scope, perturbed_scope):
"""
We have to wrap this code into a function due to the way tf.cond() works.
See https://stackoverflow.com/questions/37063952/confused-by-the-behavior-of-tf-cond for a more detailed
discussion.
:param original_scope: (str or VariableScope) the original scope.
:param perturbed_scope: (str or VariableScope) the perturbed scope.
:return: (TensorFlow Operation)
"""
all_vars = scope_vars(absolute_scope_name(original_scope))
all_perturbed_vars = scope_vars(absolute_scope_name(perturbed_scope))
assert len(all_vars) == len(all_perturbed_vars)
perturb_ops = []
for var, perturbed_var in zip(all_vars, all_perturbed_vars):
if param_noise_filter_func(perturbed_var):
# Perturb this variable.
operation = tf.assign(perturbed_var,
var + tf.random_normal(shape=tf.shape(var), mean=0.,
stddev=param_noise_scale))
else:
# Do not perturb, just assign.
operation = tf.assign(perturbed_var, var)
perturb_ops.append(operation)
assert len(perturb_ops) == len(all_vars)
return tf.group(*perturb_ops)
# Set up functionality to re-compute `param_noise_scale`. This perturbs yet another copy
# of the network and measures the effect of that perturbation in action space. If the perturbation
# is too big, reduce scale of perturbation, otherwise increase.
with tf.variable_scope("adaptive_model", reuse=False):
adaptive_policy = q_func(sess, ob_space, ac_space, 1, 1, None, obs_phs=obs_phs)
perturb_for_adaption = perturb_vars(original_scope="model", perturbed_scope="adaptive_model/model")
kl_loss = tf.reduce_sum(
tf.nn.softmax(policy.q_values) *
(tf.log(tf.nn.softmax(policy.q_values)) - tf.log(tf.nn.softmax(adaptive_policy.q_values))),
axis=-1)
mean_kl = tf.reduce_mean(kl_loss)
def update_scale():
"""
update the scale expression
:return: (TensorFlow Tensor) the updated scale expression
"""
with tf.control_dependencies([perturb_for_adaption]):
update_scale_expr = tf.cond(mean_kl < param_noise_threshold,
lambda: param_noise_scale.assign(param_noise_scale * 1.01),
lambda: param_noise_scale.assign(param_noise_scale / 1.01),
)
return update_scale_expr
# Functionality to update the threshold for parameter space noise.
update_param_noise_thres_expr = param_noise_threshold.assign(
tf.cond(update_param_noise_threshold_ph >= 0, lambda: update_param_noise_threshold_ph,
lambda: param_noise_threshold))
# Put everything together.
perturbed_deterministic_actions = tf.argmax(perturbable_policy.q_values, axis=1)
deterministic_actions = tf.argmax(policy.q_values, axis=1)
batch_size = tf.shape(policy.obs_ph)[0]
n_actions = ac_space.nvec if isinstance(ac_space, MultiDiscrete) else ac_space.n
random_actions = tf.random_uniform(tf.stack([batch_size]), minval=0, maxval=n_actions, dtype=tf.int64)
chose_random = tf.random_uniform(tf.stack([batch_size]), minval=0, maxval=1, dtype=tf.float32) < eps
perturbed_stochastic_actions = tf.where(chose_random, random_actions, perturbed_deterministic_actions)
stochastic_actions = tf.where(chose_random, random_actions, deterministic_actions)
perturbed_output_actions = tf.cond(stochastic_ph, lambda: perturbed_stochastic_actions,
lambda: deterministic_actions)
output_actions = tf.cond(stochastic_ph, lambda: stochastic_actions, lambda: deterministic_actions)
update_eps_expr = eps.assign(tf.cond(update_eps_ph >= 0, lambda: update_eps_ph, lambda: eps))
updates = [
update_eps_expr,
tf.cond(reset_ph, lambda: perturb_vars(original_scope="model", perturbed_scope="perturbed_model/model"),
lambda: tf.group(*[])),
tf.cond(update_param_noise_scale_ph, lambda: update_scale(), lambda: tf.Variable(0., trainable=False)),
update_param_noise_thres_expr,
]
_act = tf_util.function(inputs=[policy.obs_ph, stochastic_ph, update_eps_ph],
outputs=output_actions,
givens={update_eps_ph: -1.0, stochastic_ph: True},
updates=[update_eps_expr])
_perturbed_act = tf_util.function(
inputs=[policy.obs_ph, stochastic_ph, update_eps_ph, reset_ph, update_param_noise_threshold_ph,
update_param_noise_scale_ph],
outputs=perturbed_output_actions,
givens={update_eps_ph: -1.0, stochastic_ph: True, reset_ph: False, update_param_noise_threshold_ph: False,
update_param_noise_scale_ph: False},
updates=updates)
def act(obs, reset=None, update_param_noise_threshold=None, update_param_noise_scale=None, stochastic=True,
update_eps=-1):
"""
get the action from the current observation
:param obs: (Any) Observation that can be feed into the output of make_obs_ph
:param reset: (bool) reset the perturbed policy by sampling a new perturbation
:param update_param_noise_threshold: (float) the desired threshold for the difference between
non-perturbed and perturbed policy
:param update_param_noise_scale: (bool) whether or not to update the scale of the noise for the next time
it is re-perturbed
:param stochastic: (bool) if set to False all the actions are always deterministic (default False)
:param update_eps: (float) update epsilon a new value, if negative not update happens
(default: no update)
:return: (TensorFlow Tensor) tensor of dtype tf.int64 and shape (BATCH_SIZE,) with an action to be
performed for every element of the batch.
"""
if reset is None or update_param_noise_threshold is None or update_param_noise_scale is None:
return _act(obs, stochastic, update_eps)
else:
return _perturbed_act(obs, stochastic, update_eps, reset, update_param_noise_threshold,
update_param_noise_scale)
return act, obs_phs
def build_act_mod(q_func, ob_space, ac_space, stochastic_ph, update_eps_ph,
sess):
"""
Original from build_graph.py from stable_baselines.deepq. This
modified version returns the full sorted action array instead of
the single best action.
Creates the act function:
:param q_func: (DQNPolicy) the policy
:param ob_space: (Gym Space) The observation space of the environment
:param ac_space: (Gym Space) The action space of the environment
:param stochastic_ph: (TensorFlow Tensor) the stochastic placeholder
:param update_eps_ph: (TensorFlow Tensor) the update_eps placeholder
:param sess: (TensorFlow session) The current TensorFlow session
:return: (function (TensorFlow Tensor, bool, float): TensorFlow Tensor, (TensorFlow Tensor, TensorFlow Tensor)
act function to select and action given observation (See the top of the file for details),
A tuple containing the observation placeholder and the processed observation placeholder respectively.
"""
eps = tf.get_variable("eps", (), initializer=tf.constant_initializer(0))
policy = q_func(sess, ob_space, ac_space, 1, 1, None)
obs_phs = (policy.obs_ph, policy.processed_obs)
####################################################################
# MODIFICATION:
# Get all sorted q_values instead of just the best one. Have to
# cast to int64, since it comes back as int32 which is incompatible
# with the random_actions, which is int64.
deterministic_actions = tf.cast(
tf.argsort(policy.q_values, axis=1, direction='DESCENDING'), tf.int64)
####################################################################
# ORIGINAL:
# Note that the original comes out as int64 since that's the default
# from argmax.
# deterministic_actions = tf.argmax(policy.q_values, axis=1)
batch_size = tf.shape(policy.obs_ph)[0]
n_actions = ac_space.nvec if isinstance(ac_space,
MultiDiscrete) else ac_space.n
####################################################################
# MODIFICATION:
# TODO: It feels wrong not to use "batch_size" like was done in the
# original, but this seems to be working.
random_actions = tf.random_uniform((1, n_actions),
minval=0,
maxval=n_actions,
dtype=tf.int64)
####################################################################
# ORIGINAL:
# random_actions = tf.random_uniform(tf.stack([batch_size]), minval=0, maxval=n_actions, dtype=tf.int64)
chose_random = tf.random_uniform(
tf.stack([batch_size]), minval=0, maxval=1, dtype=tf.float32) < eps
stochastic_actions = tf.where(chose_random, random_actions,
deterministic_actions)
output_actions = tf.cond(stochastic_ph, lambda: stochastic_actions,
lambda: deterministic_actions)
update_eps_expr = eps.assign(
tf.cond(update_eps_ph >= 0, lambda: update_eps_ph, lambda: eps))
_act = tf_util.function(
inputs=[policy.obs_ph, stochastic_ph, update_eps_ph],
outputs=output_actions,
givens={
update_eps_ph: -1.0,
stochastic_ph: True
},
updates=[update_eps_expr])
def act(obs, stochastic=True, update_eps=-1):
return _act(obs, stochastic, update_eps)
return act, obs_phs
def build_train(make_obs_ph,
q_func,
num_actions,
optimizer,
grad_norm_clipping=None,
gamma=1.0,
double_q=True,
scope="deepq",
reuse=None,
param_noise=False,
param_noise_filter_func=None):
"""
Creates the train function:
:param make_obs_ph: (function (str): TensorFlow Tensor) a function that takes a name and creates a placeholder of
input with that name
:param q_func: (function (TensorFlow Tensor, int, str, bool): TensorFlow Tensor)
the model that takes the following inputs:
- observation_in: (Any) the output of observation placeholder
- num_actions: int number of actions
- scope: (str)
- reuse: (bool)
should be passed to outer variable scope and returns a tensor of shape (batch_size, num_actions)
with values of every action.
:param num_actions: (int) number of actions
:param reuse: (bool) whether or not to reuse the graph variables
:param optimizer: (tf.train.Optimizer) optimizer to use for the Q-learning objective.
:param grad_norm_clipping: (float) clip gradient norms to this value. If None no clipping is performed.
:param gamma: (float) discount rate.
:param double_q: (bool) if true will use Double Q Learning (https://arxiv.org/abs/1509.06461). In general it is a
good idea to keep it enabled.
:param scope: (str or VariableScope) optional scope for variable_scope.
:param reuse: (bool) whether or not the variables should be reused. To be able to reuse the scope must be given.
:param param_noise: (bool) whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905)
:param param_noise_filter_func: (function (TensorFlow Tensor): bool) function that decides whether or not a
variable should be perturbed. Only applicable if param_noise is True. If set to None, default_param_noise_filter
is used by default.
:return: (tuple)
act: (function (TensorFlow Tensor, bool, float): TensorFlow Tensor) function to select and action given
observation. See the top of the file for details.
train: (function (Any, numpy float, numpy float, Any, numpy bool, numpy float): numpy float)
optimize the error in Bellman's equation. See the top of the file for details.
update_target: (function) copy the parameters from optimized Q function to the target Q function.
See the top of the file for details.
debug: ({str: function}) a bunch of functions to print debug data like q_values.
"""
if param_noise:
act_f = build_act_with_param_noise(
make_obs_ph,
q_func,
num_actions,
scope=scope,
reuse=reuse,
param_noise_filter_func=param_noise_filter_func)
else:
act_f = build_act(make_obs_ph,
q_func,
num_actions,
scope=scope,
reuse=reuse)
with tf.variable_scope(scope, reuse=reuse):
# set up placeholders
obs_t_input = make_obs_ph("obs_t")
act_t_ph = tf.placeholder(tf.int32, [None], name="action")
rew_t_ph = tf.placeholder(tf.float32, [None], name="reward")
obs_tp1_input = make_obs_ph("obs_tp1")
done_mask_ph = tf.placeholder(tf.float32, [None], name="done")
importance_weights_ph = tf.placeholder(tf.float32, [None],
name="weight")
# q network evaluation
q_t = q_func(obs_t_input.get(),
num_actions,
scope="q_func",
reuse=True) # reuse parameters from act
q_func_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=tf.get_variable_scope().name +
"/q_func")
# target q network evalution
q_tp1 = q_func(obs_tp1_input.get(), num_actions, scope="target_q_func")
target_q_func_vars = tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES,
scope=tf.get_variable_scope().name + "/target_q_func")
# q scores for actions which we know were selected in the given state.
q_t_selected = tf.reduce_sum(q_t * tf.one_hot(act_t_ph, num_actions),
1)
# compute estimate of best possible value starting from state at t + 1
if double_q:
q_tp1_using_online_net = q_func(obs_tp1_input.get(),
num_actions,
scope="q_func",
reuse=True)
q_tp1_best_using_online_net = tf.argmax(q_tp1_using_online_net, 1)
q_tp1_best = tf.reduce_sum(
q_tp1 * tf.one_hot(q_tp1_best_using_online_net, num_actions),
1)
else:
q_tp1_best = tf.reduce_max(q_tp1, 1)
q_tp1_best_masked = (1.0 - done_mask_ph) * q_tp1_best
# compute RHS of bellman equation
q_t_selected_target = rew_t_ph + gamma * q_tp1_best_masked
# compute the error (potentially clipped)
td_error = q_t_selected - tf.stop_gradient(q_t_selected_target)
errors = tf_utils.huber_loss(td_error)
weighted_error = tf.reduce_mean(importance_weights_ph * errors)
# compute optimization op (potentially with gradient clipping)
if grad_norm_clipping is not None:
gradients = optimizer.compute_gradients(weighted_error,
var_list=q_func_vars)
for i, (grad, var) in enumerate(gradients):
if grad is not None:
gradients[i] = (tf.clip_by_norm(grad,
grad_norm_clipping), var)
optimize_expr = optimizer.apply_gradients(gradients)
else:
optimize_expr = optimizer.minimize(weighted_error,
var_list=q_func_vars)
# update_target_fn will be called periodically to copy Q network to target Q network
update_target_expr = []
for var, var_target in zip(
sorted(q_func_vars, key=lambda v: v.name),
sorted(target_q_func_vars, key=lambda v: v.name)):
update_target_expr.append(var_target.assign(var))
update_target_expr = tf.group(*update_target_expr)
# Create callable functions
train = tf_utils.function(inputs=[
obs_t_input, act_t_ph, rew_t_ph, obs_tp1_input, done_mask_ph,
importance_weights_ph
],
outputs=td_error,
updates=[optimize_expr])
update_target = tf_utils.function([], [], updates=[update_target_expr])
q_values = tf_utils.function([obs_t_input], q_t)
return act_f, train, update_target, {'q_values': q_values}
def learn(
env,
policy_func,
*,
timesteps_per_batch, # timesteps per actor per update
clip_param,
entcoeff, # clipping parameter epsilon, entropy coeff
optim_epochs,
optim_stepsize,
optim_batchsize, # optimization hypers
gamma,
lam, # advantage estimation
max_timesteps=0,
max_episodes=0,
max_iters=0,
max_seconds=0, # time constraint
callback=None, # you can do anything in the callback, since it takes locals(), globals()
adam_epsilon=1e-5,
schedule='constant', # annealing for stepsize parameters (epsilon and adam)
num_options=2,
app='',
saves=False,
wsaves=False,
epoch=-1,
seed=1,
dc=0):
optim_batchsize_ideal = optim_batchsize
np.random.seed(seed)
tf.set_random_seed(seed)
# env._seed(seed)
gamename = env.spec.id[:-3].lower()
gamename += 'seed' + str(seed)
gamename += app
dirname = '{}_{}opts_saves/'.format(gamename, num_options)
if wsaves:
first = True
if not os.path.exists(dirname):
os.makedirs(dirname)
first = False
# while os.path.exists(dirname) and first:
# dirname += '0'
files = ['pposgd_simple.py', 'mlp_policy.py', 'run_main.py']
for i in range(len(files)):
src = os.path.expanduser('~/baselines/baselines/ppo1/') + files[i]
dest = os.path.expanduser('~/baselines/baselines/ppo1/') + dirname
shutil.copy2(src, dest)
# ----------------------------------------
ob_space = env.observation_space
ac_space = env.action_space
pi = policy_func("pi", ob_space,
ac_space) # Construct network for new policy
oldpi = policy_func("oldpi", ob_space, ac_space) # Network for old policy
atarg = tf.placeholder(
dtype=tf.float32,
shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
# option = tf.placeholder(dtype=tf.int32, shape=[None])
lrmult = tf.placeholder(
name='lrmult', dtype=tf.float32,
shape=[]) # learning rate multiplier, updated with schedule
clip_param = clip_param * lrmult # Annealed cliping parameter epislon
# pdb.set_trace()
ob = U.get_placeholder_cached(name="ob")
option = U.get_placeholder_cached(name="option")
term_adv = U.get_placeholder(name='term_adv',
dtype=tf.float32,
shape=[None])
ac = pi.pdtype.sample_placeholder([None])
kloldnew = oldpi.pd.kl(pi.pd)
ent = pi.pd.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
pol_entpen = (-entcoeff) * meanent
ratio = tf.exp(pi.pd.logp(ac) - oldpi.pd.logp(ac)) # pnew / pold
surr1 = ratio * atarg # surrogate from conservative policy iteration
surr2 = tf.clip_by_value(ratio, 1.0 - clip_param,
1.0 + clip_param) * atarg #
pol_surr = -tf.reduce_mean(tf.minimum(
surr1, surr2)) # PPO's pessimistic surrogate (L^CLIP)
vf_loss = tf.reduce_mean(tf.square(pi.vpred - ret))
total_loss = pol_surr + pol_entpen + vf_loss
losses = [pol_surr, pol_entpen, vf_loss, meankl, meanent]
loss_names = ["pol_surr", "pol_entpen", "vf_loss", "kl", "ent"]
term_loss = pi.tpred * term_adv
log_pi = tf.log(tf.clip_by_value(pi.op_pi, 1e-20, 1.0))
entropy = -tf.reduce_sum(pi.op_pi * log_pi, reduction_indices=1)
op_loss = -tf.reduce_sum(log_pi[0][option[0]] * atarg + entropy * 0.1)
total_loss += op_loss
var_list = pi.get_trainable_variables()
term_list = var_list[6:8]
lossandgrad = U.function([ob, ac, atarg, ret, lrmult, option, term_adv],
losses + [U.flatgrad(total_loss, var_list)])
termloss = U.function([ob, option, term_adv],
[U.flatgrad(term_loss, var_list)
]) # Since we will use a different step size.
adam = MpiAdam(var_list, epsilon=adam_epsilon)
assign_old_eq_new = U.function(
[], [],
updates=[
tf.assign(oldv, newv)
for (oldv,
newv) in zipsame(oldpi.get_variables(), pi.get_variables())
])
compute_losses = U.function([ob, ac, atarg, ret, lrmult, option], losses)
U.initialize()
adam.sync()
saver = tf.train.Saver(max_to_keep=10000)
results = []
if saves:
results = open(
gamename + '_' + str(num_options) + 'opts_' + '_results.csv', 'w')
out = 'epoch,avg_reward'
for opt in range(num_options):
out += ',option {} dur'.format(opt)
for opt in range(num_options):
out += ',option {} std'.format(opt)
for opt in range(num_options):
out += ',option {} term'.format(opt)
for opt in range(num_options):
out += ',option {} adv'.format(opt)
out += '\n'
results.write(out)
# results.write('epoch,avg_reward,option 1 dur, option 2 dur, option 1 term, option 2 term\n')
results.flush()
if epoch >= 0:
dirname = '{}_{}opts_saves/'.format(gamename, num_options)
print("Loading weights from iteration: " + str(epoch))
filename = dirname + '{}_epoch_{}.ckpt'.format(gamename, epoch)
saver.restore(U.get_session(), filename)
episodes_so_far = 0
timesteps_so_far = 0
global iters_so_far
iters_so_far = 0
tstart = time.time()
lenbuffer = deque(maxlen=100) # rolling buffer for episode lengths
rewbuffer = deque(maxlen=100) # rolling buffer for episode rewards
assert sum(
[max_iters > 0, max_timesteps > 0, max_episodes > 0,
max_seconds > 0]) == 1, "Only one time constraint permitted"
seg_gen = traj_segment_generator(pi,
env,
timesteps_per_batch,
stochastic=True,
num_options=num_options,
saves=saves,
results=results,
rewbuffer=rewbuffer,
dc=dc)
datas = [0 for _ in range(num_options)]
while True:
if callback: callback(locals(), globals())
if max_timesteps and timesteps_so_far >= max_timesteps:
break
elif max_episodes and episodes_so_far >= max_episodes:
break
elif max_iters and iters_so_far >= max_iters:
break
elif max_seconds and time.time() - tstart >= max_seconds:
break
if schedule == 'constant':
cur_lrmult = 1.0
elif schedule == 'linear':
cur_lrmult = max(1.0 - float(timesteps_so_far) / max_timesteps, 0)
else:
raise NotImplementedError
logger.log("********** Iteration %i ************" % iters_so_far)
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, gamma, lam)
opt_d = []
for i in range(num_options):
dur = np.mean(
seg['opt_dur'][i]) if len(seg['opt_dur'][i]) > 0 else 0.
opt_d.append(dur)
std = []
for i in range(num_options):
logstd = np.mean(
seg['logstds'][i]) if len(seg['logstds'][i]) > 0 else 0.
std.append(np.exp(logstd))
print("mean opt dur:", opt_d)
print("mean op pol:", np.mean(np.array(seg['optpol_p']), axis=0))
print("mean term p:", np.mean(np.array(seg['term_p']), axis=0))
print("mean value val:", np.mean(np.array(seg['value_val']), axis=0))
ob, ac, opts, atarg, tdlamret = seg["ob"], seg["ac"], seg["opts"], seg[
"adv"], seg["tdlamret"]
vpredbefore = seg["vpred"] # predicted value function before udpate
atarg = (atarg - atarg.mean()
) / atarg.std() # standardized advantage function estimate
if hasattr(pi, "ob_rms"):
pi.ob_rms.update(ob) # update running mean/std for policy
assign_old_eq_new() # set old parameter values to new parameter values
if iters_so_far % 5 == 0 and wsaves:
print("weights are saved...")
filename = dirname + '{}_epoch_{}.ckpt'.format(
gamename, iters_so_far)
save_path = saver.save(U.get_session(), filename)
min_batch = 160
t_advs = [[] for _ in range(num_options)]
for opt in range(num_options):
indices = np.where(opts == opt)[0]
print("batch size:", indices.size)
opt_d[opt] = indices.size
if not indices.size:
t_advs[opt].append(0.)
continue
# This part is only necessasry when we use options.
# We proceed to these verifications in order not to discard any collected trajectories.
if datas[opt] != 0:
if (indices.size < min_batch and datas[opt].n > min_batch):
datas[opt] = Dataset(dict(ob=ob[indices],
ac=ac[indices],
atarg=atarg[indices],
vtarg=tdlamret[indices]),
shuffle=not pi.recurrent)
t_advs[opt].append(0.)
continue
elif indices.size + datas[opt].n < min_batch:
# pdb.set_trace()
oldmap = datas[opt].data_map
cat_ob = np.concatenate((oldmap['ob'], ob[indices]))
cat_ac = np.concatenate((oldmap['ac'], ac[indices]))
cat_atarg = np.concatenate(
(oldmap['atarg'], atarg[indices]))
cat_vtarg = np.concatenate(
(oldmap['vtarg'], tdlamret[indices]))
datas[opt] = Dataset(dict(ob=cat_ob,
ac=cat_ac,
atarg=cat_atarg,
vtarg=cat_vtarg),
shuffle=not pi.recurrent)
t_advs[opt].append(0.)
continue
elif (indices.size + datas[opt].n > min_batch and datas[opt].n
< min_batch) or (indices.size > min_batch
and datas[opt].n < min_batch):
oldmap = datas[opt].data_map
cat_ob = np.concatenate((oldmap['ob'], ob[indices]))
cat_ac = np.concatenate((oldmap['ac'], ac[indices]))
cat_atarg = np.concatenate(
(oldmap['atarg'], atarg[indices]))
cat_vtarg = np.concatenate(
(oldmap['vtarg'], tdlamret[indices]))
datas[opt] = d = Dataset(dict(ob=cat_ob,
ac=cat_ac,
atarg=cat_atarg,
vtarg=cat_vtarg),
shuffle=not pi.recurrent)
if (indices.size > min_batch and datas[opt].n > min_batch):
datas[opt] = d = Dataset(dict(ob=ob[indices],
ac=ac[indices],
atarg=atarg[indices],
vtarg=tdlamret[indices]),
shuffle=not pi.recurrent)
elif datas[opt] == 0:
datas[opt] = d = Dataset(dict(ob=ob[indices],
ac=ac[indices],
atarg=atarg[indices],
vtarg=tdlamret[indices]),
shuffle=not pi.recurrent)
optim_batchsize = optim_batchsize or ob.shape[0]
optim_epochs = np.clip(
np.int(10 * (indices.size /
(timesteps_per_batch / num_options))), 10,
10) if num_options > 1 else optim_epochs
print("optim epochs:", optim_epochs)
logger.log("Optimizing...")
# Here we do a bunch of optimization epochs over the data
for _ in range(optim_epochs):
losses = [
] # list of tuples, each of which gives the loss for a minibatch
for batch in d.iterate_once(optim_batchsize):
tadv, nodc_adv = pi.get_term_adv(batch["ob"], [opt])
tadv = tadv if num_options > 1 else np.zeros_like(tadv)
t_advs[opt].append(nodc_adv)
*newlosses, grads = lossandgrad(batch["ob"], batch["ac"],
batch["atarg"],
batch["vtarg"], cur_lrmult,
[opt], tadv)
termg = termloss(batch["ob"], [opt], tadv)
adam.update(termg[0], 5e-7 * cur_lrmult)
adam.update(grads, optim_stepsize * cur_lrmult)
losses.append(newlosses)
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))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
timesteps_so_far += sum(lens)
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", timesteps_so_far)
logger.record_tabular("TimeElapsed", time.time() - tstart)
if MPI.COMM_WORLD.Get_rank() == 0:
logger.dump_tabular()
if saves:
out = "{},{}"
for _ in range(num_options):
out += ",{},{},{},{}"
out += "\n"
info = [iters_so_far, np.mean(rewbuffer)]
for i in range(num_options):
info.append(opt_d[i])
for i in range(num_options):
info.append(std[i])
for i in range(num_options):
info.append(np.mean(np.array(seg['term_p']), axis=0)[i])
for i in range(num_options):
info.append(np.mean(t_advs[i]))
results.write(out.format(*info))
results.flush()
def _init(self,
ob_space,
ac_space,
hid_size,
num_hid_layers,
gaussian_fixed_var=True):
obs, pdtype = self.get_obs_and_pdtype(ob_space, ac_space)
with tf.variable_scope("obfilter"):
self.ob_rms = RunningMeanStd(shape=ob_space.shape)
obz = tf.clip_by_value((obs - self.ob_rms.mean) / self.ob_rms.std,
-5.0, 5.0)
last_out = obz
for i in range(num_hid_layers):
last_out = tf.nn.tanh(
dense(last_out,
hid_size,
"vffc%i" % (i + 1),
weight_init=tf_util.normc_initializer(1.0)))
self.vpred = dense(last_out,
1,
"vffinal",
weight_init=tf_util.normc_initializer(1.0))[:, 0]
last_out = obz
for i in range(num_hid_layers):
last_out = tf.nn.tanh(
dense(last_out,
hid_size,
"polfc%i" % (i + 1),
weight_init=tf_util.normc_initializer(1.0)))
if gaussian_fixed_var and isinstance(ac_space, gym.spaces.Box):
mean = dense(last_out,
pdtype.param_shape()[0] // 2, "polfinal",
tf_util.normc_initializer(0.01))
logstd = tf.get_variable(name="logstd",
shape=[1, pdtype.param_shape()[0] // 2],
initializer=tf.zeros_initializer())
pdparam = tf.concat([mean, mean * 0.0 + logstd], axis=1)
else:
pdparam = dense(last_out,
pdtype.param_shape()[0], "polfinal",
tf_util.normc_initializer(0.01))
self.proba_distribution = pdtype.proba_distribution_from_flat(pdparam)
self.state_in = []
self.state_out = []
# change for BC
self.stochastic_ph = tf.placeholder(dtype=tf.bool,
shape=(),
name="stochastic")
action = tf_util.switch(self.stochastic_ph,
self.proba_distribution.sample(),
self.proba_distribution.mode())
self.action = action
self._act = tf_util.function([self.stochastic_ph, obs],
[action, self.vpred])
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 __init__(self, ob_dim, ac_dim, verbose=1):
"""
Create an MLP policy for a value function
:param ob_dim: (int) Observation dimention
:param ac_dim: (int) action dimention
:param verbose: (int) verbosity level
"""
obs_ph = tf.placeholder(tf.float32,
shape=[None, ob_dim * 2 + ac_dim * 2 + 2
]) # batch of observations
vtarg_n = tf.placeholder(tf.float32, shape=[None], name='vtarg')
wd_dict = {}
layer_1 = tf.nn.elu(
dense(obs_ph,
64,
"h1",
weight_init=tf_util.normc_initializer(1.0),
bias_init=0,
weight_loss_dict=wd_dict))
layer_2 = tf.nn.elu(
dense(layer_1,
64,
"h2",
weight_init=tf_util.normc_initializer(1.0),
bias_init=0,
weight_loss_dict=wd_dict))
vpred_n = dense(layer_2,
1,
"hfinal",
weight_init=tf_util.normc_initializer(1.0),
bias_init=0,
weight_loss_dict=wd_dict)[:, 0]
sample_vpred_n = vpred_n + tf.random_normal(tf.shape(vpred_n))
wd_loss = tf.get_collection("vf_losses", None)
loss = tf.reduce_mean(tf.square(vpred_n - vtarg_n)) + tf.add_n(wd_loss)
loss_sampled = tf.reduce_mean(
tf.square(vpred_n - tf.stop_gradient(sample_vpred_n)))
self._predict = tf_util.function([obs_ph], vpred_n)
optim = kfac.KfacOptimizer(learning_rate=0.001,
cold_lr=0.001 * (1 - 0.9),
momentum=0.9,
clip_kl=0.3,
epsilon=0.1,
stats_decay=0.95,
async=1,
kfac_update=2,
cold_iter=50,
weight_decay_dict=wd_dict,
max_grad_norm=None,
verbose=verbose)
vf_var_list = []
for var in tf.trainable_variables():
if "vf" in var.name:
vf_var_list.append(var)
update_op, self.q_runner = optim.minimize(loss,
loss_sampled,
var_list=vf_var_list)
self.do_update = tf_util.function([obs_ph, vtarg_n], update_op) # pylint: disable=E1101
tf_util.initialize() # Initialize uninitialized TF variables
def build_train(q_func, ob_space, ac_space, optimizer, sess, grad_norm_clipping=None, gamma=1.0, double_q=True,
scope="deepq", reuse=None, param_noise=False, param_noise_filter_func=None):
"""
Creates the train function:
:param q_func: (DQNPolicy) the policy
:param ob_space: (Gym Space) The observation space of the environment
:param ac_space: (Gym Space) The action space of the environment
:param reuse: (bool) whether or not to reuse the graph variables
:param optimizer: (tf.train.Optimizer) optimizer to use for the Q-learning objective.
:param sess: (TensorFlow session) The current TensorFlow session
:param grad_norm_clipping: (float) clip gradient norms to this value. If None no clipping is performed.
:param gamma: (float) discount rate.
:param double_q: (bool) if true will use Double Q Learning (https://arxiv.org/abs/1509.06461). In general it is a
good idea to keep it enabled.
:param scope: (str or VariableScope) optional scope for variable_scope.
:param reuse: (bool) whether or not the variables should be reused. To be able to reuse the scope must be given.
:param param_noise: (bool) whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905)
:param param_noise_filter_func: (function (TensorFlow Tensor): bool) function that decides whether or not a
variable should be perturbed. Only applicable if param_noise is True. If set to None, default_param_noise_filter
is used by default.
:return: (tuple)
act: (function (TensorFlow Tensor, bool, float): TensorFlow Tensor) function to select and action given
observation. See the top of the file for details.
train: (function (Any, numpy float, numpy float, Any, numpy bool, numpy float): numpy float)
optimize the error in Bellman's equation. See the top of the file for details.
update_target: (function) copy the parameters from optimized Q function to the target Q function.
See the top of the file for details.
step_model: (DQNPolicy) Policy for evaluation
"""
n_actions = ac_space.nvec if isinstance(ac_space, MultiDiscrete) else ac_space.n
with tf.variable_scope("input", reuse=reuse):
stochastic_ph = tf.placeholder(tf.bool, (), name="stochastic")
update_eps_ph = tf.placeholder(tf.float32, (), name="update_eps")
with tf.variable_scope(scope, reuse=reuse):
if param_noise:
act_f, obs_phs = build_act_with_param_noise(q_func, ob_space, ac_space, stochastic_ph, update_eps_ph, sess,
param_noise_filter_func=param_noise_filter_func)
else:
act_f, obs_phs = build_act(q_func, ob_space, ac_space, stochastic_ph, update_eps_ph, sess)
# q network evaluation
with tf.variable_scope("step_model", reuse=True, custom_getter=tf_util.outer_scope_getter("step_model")):
step_model = q_func(sess, ob_space, ac_space, 1, 1, None, reuse=True, obs_phs=obs_phs)
q_func_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=tf.get_variable_scope().name + "/model")
# target q network evaluation
with tf.variable_scope("target_q_func", reuse=False):
target_policy = q_func(sess, ob_space, ac_space, 1, 1, None, reuse=False)
target_q_func_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=tf.get_variable_scope().name + "/target_q_func")
# compute estimate of best possible value starting from state at t + 1
double_q_values = None
double_obs_ph = target_policy.obs_ph
if double_q:
with tf.variable_scope("double_q", reuse=True, custom_getter=tf_util.outer_scope_getter("double_q")):
double_policy = q_func(sess, ob_space, ac_space, 1, 1, None, reuse=True)
double_q_values = double_policy.q_values
double_obs_ph = double_policy.obs_ph
with tf.variable_scope("loss", reuse=reuse):
# set up placeholders
act_t_ph = tf.placeholder(tf.int32, [None], name="action")
rew_t_ph = tf.placeholder(tf.float32, [None], name="reward")
done_mask_ph = tf.placeholder(tf.float32, [None], name="done")
importance_weights_ph = tf.placeholder(tf.float32, [None], name="weight")
# q scores for actions which we know were selected in the given state.
q_t_selected = tf.reduce_sum(step_model.q_values * tf.one_hot(act_t_ph, n_actions), axis=1)
# compute estimate of best possible value starting from state at t + 1
if double_q:
q_tp1_best_using_online_net = tf.argmax(double_q_values, axis=1)
q_tp1_best = tf.reduce_sum(target_policy.q_values * tf.one_hot(q_tp1_best_using_online_net, n_actions), axis=1)
else:
q_tp1_best = tf.reduce_max(target_policy.q_values, axis=1)
q_tp1_best_masked = (1.0 - done_mask_ph) * q_tp1_best
# compute RHS of bellman equation
q_t_selected_target = rew_t_ph + gamma * q_tp1_best_masked
# compute the error (potentially clipped)
td_error = q_t_selected - tf.stop_gradient(q_t_selected_target)
errors = tf_util.huber_loss(td_error)
weighted_error = tf.reduce_mean(importance_weights_ph * errors)
tf.summary.scalar("td_error", tf.reduce_mean(td_error))
tf.summary.histogram("td_error", td_error)
tf.summary.scalar("loss", weighted_error)
# update_target_fn will be called periodically to copy Q network to target Q network
update_target_expr = []
for var, var_target in zip(sorted(q_func_vars, key=lambda v: v.name),
sorted(target_q_func_vars, key=lambda v: v.name)):
update_target_expr.append(var_target.assign(var))
update_target_expr = tf.group(*update_target_expr)
# compute optimization op (potentially with gradient clipping)
gradients = optimizer.compute_gradients(weighted_error, var_list=q_func_vars)
if grad_norm_clipping is not None:
for i, (grad, var) in enumerate(gradients):
if grad is not None:
gradients[i] = (tf.clip_by_norm(grad, grad_norm_clipping), var)
with tf.variable_scope("input_info", reuse=False):
# Edit: Add action
tf.summary.scalar('actions', tf.reduce_mean(act_t_ph))
tf.summary.histogram('actions', act_t_ph)
tf.summary.scalar('rewards', tf.reduce_mean(rew_t_ph))
tf.summary.histogram('rewards', rew_t_ph)
tf.summary.scalar('importance_weights', tf.reduce_mean(importance_weights_ph))
tf.summary.histogram('importance_weights', importance_weights_ph)
# Valid image: RGB, RGBD, GrayScale
is_image = len(obs_phs[0].shape) == 3 and obs_phs[0].shape[-1] in [1, 3, 4]
if is_image:
tf.summary.image('observation', obs_phs[0])
elif len(obs_phs[0].shape) == 1:
tf.summary.histogram('observation', obs_phs[0])
optimize_expr = optimizer.apply_gradients(gradients)
summary = tf.summary.merge_all()
# Create callable functions
train = tf_util.function(
inputs=[
obs_phs[0],
act_t_ph,
rew_t_ph,
target_policy.obs_ph,
double_obs_ph,
done_mask_ph,
importance_weights_ph
],
outputs=[summary, td_error],
updates=[optimize_expr]
)
update_target = tf_util.function([], [], updates=[update_target_expr])
return act_f, train, update_target, step_model
def build_act(q_func, ob_space, ac_space, sess, num_actions,
num_action_streams, stochastic_ph, update_eps_ph):
"""Creates the act function:
Parameters
----------
make_obs_ph: str -> tf.placeholder or TfInput
a function that takes a name and creates a placeholder of input with that name
q_func: (tf.Variable, int, str, bool) -> tf.Variable
the model that takes the following inputs:
observation_in: object
the output of observation placeholder
num_actions: int
number of actions
scope: str
reuse: bool
should be passed to outer variable scope
and returns a tensor of shape (batch_size, num_actions) with values of every action.
num_actions: int
total number of sub-actions to be represented at the output
num_action_streams: int
specifies the number of action branches in action value (or advantage) function representation
scope: str or VariableScope
optional scope for variable_scope.
reuse: bool or None
whether or not the variables should be reused. To be able to reuse the scope must be given.
Returns
-------
act: (tf.Variable, bool, float) -> tf.Variable
function to select an action given observation.
` See the top of the file for details.
"""
eps = tf.get_variable("eps", (), initializer=tf.constant_initializer(0))
policy = q_func(sess, ob_space, ac_space, 1, 1, None, num_actions)
obs_phs = (policy.obs_ph, policy.processed_obs)
assert (num_action_streams >=
1), "number of action branches is not acceptable, has to be >=1"
# TODO better: enable non-uniform number of sub-actions per joint
num_actions_pad = num_actions // num_action_streams # number of sub-actions per action dimension
output_actions = []
for dim in range(num_action_streams):
q_values_batch = policy.q_values[dim][
0] # TODO better: does not allow evaluating actions over a whole batch
deterministic_action = tf.argmax(q_values_batch)
random_action = tf.random_uniform([],
minval=0,
maxval=num_actions //
num_action_streams,
dtype=tf.int64)
chose_random = tf.random_uniform(
[], minval=0, maxval=1, dtype=tf.float32) < eps
stochastic_action = tf.cond(chose_random, lambda: random_action,
lambda: deterministic_action)
output_action = tf.cond(stochastic_ph, lambda: stochastic_action,
lambda: deterministic_action)
output_actions.append(output_action)
update_eps_expr = eps.assign(
tf.cond(update_eps_ph >= 0, lambda: update_eps_ph, lambda: eps))
_act = tf_util.function(
inputs=[policy.obs_ph, stochastic_ph, update_eps_ph],
outputs=output_actions,
givens={
update_eps_ph: -1.0,
stochastic_ph: True
},
updates=[update_eps_expr])
def act(obs, stochastic=True, update_eps=-1):
return _act(obs, stochastic, update_eps)
return act, obs_phs