def policy_fn(nbatch=None,
nsteps=None,
sess=None,
observ_placeholder=None):
ob_space = env.observation_space
X = observ_placeholder if observ_placeholder is not None else observation_placeholder(
ob_space, batch_size=nbatch)
extra_tensors = {}
if normalize_observations and X.dtype == tf.float32:
encoded_x, rms = _normalize_clip_observation(X)
extra_tensors['rms'] = rms
else:
encoded_x = X
encoded_x = encode_observation(ob_space, encoded_x)
with tf.variable_scope('pi', reuse=tf.AUTO_REUSE):
policy_latent = policy_network(encoded_x)
if isinstance(policy_latent, tuple):
policy_latent, recurrent_tensors = policy_latent
if recurrent_tensors is not None:
# recurrent architecture, need a few more steps
nenv = nbatch // nsteps
assert nenv > 0, 'Bad input for recurrent policy: batch size {} smaller than nsteps {}'.format(
nbatch, nsteps)
policy_latent, recurrent_tensors = policy_network(
encoded_x, nenv)
extra_tensors.update(recurrent_tensors)
_v_net = value_network
if _v_net is None or _v_net == 'shared':
vf_latent = policy_latent
else:
if _v_net == 'copy':
_v_net = policy_network
else:
assert callable(_v_net)
with tf.variable_scope('vf', reuse=tf.AUTO_REUSE):
# TODO recurrent architectures are not supported with value_network=copy yet
vf_latent = _v_net(encoded_x)
policy = PolicyWithValue(env=env,
observations=X,
latent=policy_latent,
vf_latent=vf_latent,
sess=sess,
estimate_q=estimate_q,
**extra_tensors)
return policy
def __init__(self, sess, ob_space, ac_space, nenv, nsteps, nstack, reuse=False):
nbatch = nenv * nsteps
nh, nw, nc = ob_space.shape
ob_shape = (nbatch, nh, nw, nc * nstack)
nact = ac_space.n
X = tf.placeholder(tf.uint8, ob_shape) # obs
with tf.variable_scope("model", reuse=reuse):
h = nature_cnn(X)
pi_logits = fc(h, 'pi', nact, init_scale=0.01)
pi = tf.nn.softmax(pi_logits)
q = fc(h, 'q', nact)
a = sample(tf.nn.softmax(pi_logits)) # could change this to use self.pi instead
self.initial_state = [] # not stateful
self.X = X
self.pi = pi # actual policy params now
self.pi_logits = pi_logits
self.q = q
self.vf = q
def step(ob, *args, **kwargs):
# returns actions, mus, states
a0, pi0 = sess.run([a, pi], {X: ob})
return a0, pi0, [] # dummy state
def out(ob, *args, **kwargs):
pi0, q0 = sess.run([pi, q], {X: ob})
return pi0, q0
def act(ob, *args, **kwargs):
return sess.run(a, {X: ob})
self.step = step
self.out = out
self.act = act
def __init__(self, sess, ob_space, ac_space, nenv, nsteps, nstack, reuse=False, nlstm=256):
nbatch = nenv * nsteps
nh, nw, nc = ob_space.shape
ob_shape = (nbatch, nh, nw, nc * nstack)
nact = ac_space.n
X = tf.placeholder(tf.uint8, ob_shape) # obs
M = tf.placeholder(tf.float32, [nbatch]) #mask (done t-1)
S = tf.placeholder(tf.float32, [nenv, nlstm*2]) #states
with tf.variable_scope("model", reuse=reuse):
h = nature_cnn(X)
# lstm
xs = batch_to_seq(h, nenv, nsteps)
ms = batch_to_seq(M, nenv, nsteps)
h5, snew = lstm(xs, ms, S, 'lstm1', nh=nlstm)
h5 = seq_to_batch(h5)
pi_logits = fc(h5, 'pi', nact, init_scale=0.01)
pi = tf.nn.softmax(pi_logits)
q = fc(h5, 'q', nact)
a = sample(pi_logits) # could change this to use self.pi instead
self.initial_state = np.zeros((nenv, nlstm*2), dtype=np.float32)
self.X = X
self.M = M
self.S = S
self.pi = pi # actual policy params now
self.q = q
def step(ob, state, mask, *args, **kwargs):
# returns actions, mus, states
a0, pi0, s = sess.run([a, pi, snew], {X: ob, S: state, M: mask})
return a0, pi0, s
self.step = step
def __init__(self, scalar_keys=[], histogram_keys=[]):
self.scalar_keys = scalar_keys
self.histogram_keys = histogram_keys
self.scalar_summaries = []
self.scalar_summaries_ph = []
self.histogram_summaries_ph = []
self.histogram_summaries = []
with tf.variable_scope('summary'):
for k in scalar_keys:
ph = tf.placeholder('float32',
None,
name=k + '.scalar.summary')
sm = tf.summary.scalar(k + '.scalar.summary', ph)
self.scalar_summaries_ph.append(ph)
self.scalar_summaries.append(sm)
for k in histogram_keys:
ph = tf.placeholder('float32',
None,
name=k + '.histogram.summary')
sm = tf.summary.scalar(k + '.histogram.summary', ph)
self.histogram_summaries_ph.append(ph)
self.histogram_summaries.append(sm)
self.summaries = tf.summary.merge(self.scalar_summaries +
self.histogram_summaries)
def __init__(self, epsilon=1e-4, shape=(), scope=''):
sess = get_session()
self._new_mean = tf.placeholder(shape=shape, dtype=tf.float64)
self._new_var = tf.placeholder(shape=shape, dtype=tf.float64)
self._new_count = tf.placeholder(shape=(), dtype=tf.float64)
with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
self._mean = tf.get_variable('mean',
initializer=np.zeros(
shape, 'float64'),
dtype=tf.float64)
self._var = tf.get_variable('std',
initializer=np.ones(shape, 'float64'),
dtype=tf.float64)
self._count = tf.get_variable('count',
initializer=np.full((), epsilon,
'float64'),
dtype=tf.float64)
self.update_ops = tf.group([
self._var.assign(self._new_var),
self._mean.assign(self._new_mean),
self._count.assign(self._new_count)
])
sess.run(tf.variables_initializer([self._mean, self._var,
self._count]))
self.sess = sess
self._set_mean_var_count()
def model(inpt, num_actions, scope, reuse=False):
"""This model takes as input an observation and returns values of all actions."""
with tf.variable_scope(scope, reuse=reuse):
out = inpt
out = layers.fully_connected(out,
num_outputs=64,
activation_fn=tf.nn.tanh)
out = layers.fully_connected(out,
num_outputs=num_actions,
activation_fn=None)
return out
def build_graph(self, obs_ph, acs_ph, reuse=False):
with tf.variable_scope(self.scope):
if reuse:
tf.get_variable_scope().reuse_variables()
with tf.variable_scope("obfilter"):
self.obs_rms = RunningMeanStd(shape=self.observation_shape)
obs = (obs_ph - self.obs_rms.mean / self.obs_rms.std)
_input = tf.concat(
[obs, acs_ph],
axis=1) # concatenate the two input -> form a transition
p_h1 = tf.contrib.layers.fully_connected(_input,
self.hidden_size,
activation_fn=tf.nn.tanh)
p_h2 = tf.contrib.layers.fully_connected(p_h1,
self.hidden_size,
activation_fn=tf.nn.tanh)
logits = tf.contrib.layers.fully_connected(
p_h2, 1, activation_fn=tf.identity)
return logits
def test_multikwargs():
with tf.Graph().as_default():
x = tf.placeholder(tf.int32, (), name="x")
with tf.variable_scope("other"):
x2 = tf.placeholder(tf.int32, (), name="x")
z = 3 * x + 2 * x2
lin = function([x, x2], z, givens={x2: 0})
with single_threaded_session():
initialize()
assert lin(2) == 6
assert lin(2, 2) == 10
def __init__(self, inputs_tf, dimo, dimg, dimu, max_u, o_stats, g_stats,
hidden, layers, **kwargs):
"""The actor-critic network and related training pytorch.
Args:
inputs_tf (dict of tensors): all necessary inputs for the network: the
observation (o), the goal (g), and the action (u)
dimo (int): the dimension of the observations
dimg (int): the dimension of the goals
dimu (int): the dimension of the actions
max_u (float): the maximum magnitude of actions; action outputs will be scaled
accordingly
o_stats (baselines.her.Normalizer): normalizer for observations
g_stats (baselines.her.Normalizer): normalizer for goals
hidden (int): number of hidden units that should be used in hidden layers
layers (int): number of hidden layers
"""
self.o_tf = inputs_tf['o']
self.g_tf = inputs_tf['g']
self.u_tf = inputs_tf['u']
# Prepare inputs for actor and critic.
o = self.o_stats.normalize(self.o_tf)
g = self.g_stats.normalize(self.g_tf)
input_pi = tf.concat(axis=1, values=[o, g]) # for actor
# Networks.
with tf.variable_scope('pi'):
self.pi_tf = self.max_u * tf.tanh(
nn(input_pi, [self.hidden] * self.layers + [self.dimu]))
with tf.variable_scope('Q'):
# for policy training
input_Q = tf.concat(axis=1, values=[o, g, self.pi_tf / self.max_u])
self.Q_pi_tf = nn(input_Q, [self.hidden] * self.layers + [1])
# for critic training
input_Q = tf.concat(axis=1, values=[o, g, self.u_tf / self.max_u])
self._input_Q = input_Q # exposed for tests
self.Q_tf = nn(input_Q, [self.hidden] * self.layers + [1],
reuse=True)
def dense(x, size, name, weight_init=None, bias_init=0, weight_loss_dict=None, reuse=None):
with tf.variable_scope(name, reuse=reuse):
assert (len(tf.get_variable_scope().name.split('/')) == 2)
w = tf.get_variable("w", [x.get_shape()[1], size], initializer=weight_init)
b = tf.get_variable("b", [size], initializer=tf.constant_initializer(bias_init))
weight_decay_fc = 3e-4
if weight_loss_dict is not None:
weight_decay = tf.multiply(tf.nn.l2_loss(w), weight_decay_fc, name='weight_decay_loss')
if weight_loss_dict is not None:
weight_loss_dict[w] = weight_decay_fc
weight_loss_dict[b] = 0.0
tf.add_to_collection(tf.get_variable_scope().name.split('/')[0] + '_' + 'losses', weight_decay)
return tf.nn.bias_add(tf.matmul(x, w), b)
def _init(self,
ob_space,
ac_space,
hid_size,
num_hid_layers,
gaussian_fixed_var=True):
assert isinstance(ob_space, gym.spaces.Box)
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))
with tf.variable_scope("obfilter"):
self.ob_rms = RunningMeanStd(shape=ob_space.shape)
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(
dense(last_out,
hid_size,
"vffc%i" % (i + 1),
weight_init=U.normc_initializer(1.0)))
self.vpred = dense(last_out,
1,
"vffinal",
weight_init=U.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=U.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",
U.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",
U.normc_initializer(0.01))
self.pd = pdtype.pdfromflat(pdparam)
self.state_in = []
self.state_out = []
# change for BC
stochastic = U.get_placeholder(name="stochastic",
dtype=tf.bool,
shape=())
ac = U.switch(stochastic, self.pd.sample(), self.pd.mode())
self.ac = ac
self._act = U.function([stochastic, ob], [ac, self.vpred])
def __init__(self, name, reuse=False, *args, **kwargs):
with tf.variable_scope(name):
if reuse:
tf.get_variable_scope().reuse_variables()
self._init(*args, **kwargs)
self.scope = tf.get_variable_scope().name
def __init__(self, policy, env, nsteps,
ent_coef=0.01, vf_coef=0.5, max_grad_norm=0.5, lr=7e-4,
alpha=0.99, epsilon=1e-5, total_timesteps=int(80e6), lrschedule='linear'):
sess = tf_util.get_session()
nenvs = env.num_envs
nbatch = nenvs*nsteps
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# step_model is used for sampling
step_model = policy(nenvs, 1, sess)
# train_model is used to train our network
train_model = policy(nbatch, nsteps, sess)
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
# Calculate the loss
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
# Policy loss
neglogpac = train_model.pd.neglogp(A)
# L = A(s,a) * -logpi(a|s)
pg_loss = tf.reduce_mean(ADV * neglogpac)
# Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
entropy = tf.reduce_mean(train_model.pd.entropy())
# Value loss
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
loss = pg_loss - entropy*ent_coef + vf_loss * vf_coef
# Update parameters using loss
# 1. Get the model parameters
params = find_trainable_variables("a2c_model")
# 2. Calculate the gradients
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
# Clip the gradients (normalize)
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
# zip aggregate each gradient with parameters associated
# For instance zip(ABCD, xyza) => Ax, By, Cz, Da
# 3. Make op for one policy and value update step of A2C
trainer = tf.train.RMSPropOptimizer(learning_rate=LR, decay=alpha, epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
# Here we calculate advantage A(s,a) = R + yV(s') - V(s)
# rewards = R + yV(s')
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {train_model.X:obs, A:actions, ADV:advs, R:rewards, LR:cur_lr}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train],
td_map
)
return policy_loss, value_loss, policy_entropy
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
def __call__(self, obs, action, reuse=False):
with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
x = tf.concat([obs, action], axis=-1) # this assumes observation and action can be concatenated
x = self.network_builder(x)
x = tf.layers.dense(x, 1, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
return x
def __call__(self, obs, reuse=False):
with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
x = self.network_builder(obs)
x = tf.layers.dense(x, self.nb_actions, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
x = tf.nn.tanh(x)
return x
def __init__(self,
policy,
ob_space,
ac_space,
nenvs,
total_timesteps,
nprocs=32,
nsteps=20,
ent_coef=0.01,
vf_coef=0.5,
vf_fisher_coef=1.0,
lr=0.25,
max_grad_norm=0.5,
kfac_clip=0.001,
lrschedule='linear',
is_async=True):
self.sess = sess = get_session()
nbatch = nenvs * nsteps
A = tf.placeholder(ac_space.dtype, [
nbatch,
] + list(ac_space.shape))
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
PG_LR = tf.placeholder(tf.float32, [])
VF_LR = tf.placeholder(tf.float32, [])
with tf.variable_scope('acktr_model', reuse=tf.AUTO_REUSE):
self.model = step_model = policy(nenvs, 1, sess=sess)
self.model2 = train_model = policy(nenvs * nsteps,
nsteps,
sess=sess)
neglogpac = train_model.pd.neglogp(A)
self.logits = train_model.pi
##training loss
pg_loss = tf.reduce_mean(ADV * neglogpac)
entropy = tf.reduce_mean(train_model.pd.entropy())
pg_loss = pg_loss - ent_coef * entropy
vf_loss = tf.losses.mean_squared_error(tf.squeeze(train_model.vf), R)
train_loss = pg_loss + vf_coef * vf_loss
##Fisher loss construction
self.pg_fisher = pg_fisher_loss = -tf.reduce_mean(neglogpac)
sample_net = train_model.vf + tf.random_normal(tf.shape(
train_model.vf))
self.vf_fisher = vf_fisher_loss = -vf_fisher_coef * tf.reduce_mean(
tf.pow(train_model.vf - tf.stop_gradient(sample_net), 2))
self.joint_fisher = joint_fisher_loss = pg_fisher_loss + vf_fisher_loss
self.params = params = find_trainable_variables("acktr_model")
self.grads_check = grads = tf.gradients(train_loss, params)
with tf.device('/gpu:0'):
self.optim = optim = kfac.KfacOptimizer(learning_rate=PG_LR, clip_kl=kfac_clip,\
momentum=0.9, kfac_update=1, epsilon=0.01,\
stats_decay=0.99, is_async=is_async, cold_iter=10, max_grad_norm=max_grad_norm)
# update_stats_op = optim.compute_and_apply_stats(joint_fisher_loss, var_list=params)
optim.compute_and_apply_stats(joint_fisher_loss, var_list=params)
train_op, q_runner = optim.apply_gradients(list(zip(grads,
params)))
self.q_runner = q_runner
self.lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(obs)):
cur_lr = self.lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
PG_LR: cur_lr,
VF_LR: cur_lr
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, train_op], td_map)
return policy_loss, value_loss, policy_entropy
self.train = train
self.save = functools.partial(save_variables, sess=sess)
self.load = functools.partial(load_variables, sess=sess)
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=sess)
def learn(*,
network,
env,
total_timesteps,
timesteps_per_batch=1024, # what to train on
max_kl=0.001,
cg_iters=10,
gamma=0.99,
lam=1.0, # advantage estimation
seed=None,
ent_coef=0.0,
cg_damping=1e-2,
vf_stepsize=3e-4,
vf_iters=3,
max_episodes=0, max_iters=0, # time constraint
callback=None,
load_path=None,
**network_kwargs
):
'''
learn a policy function with TRPO algorithm
Parameters:
----------
network neural network to learn. Can be either string ('mlp', 'cnn', 'lstm', 'lnlstm' for basic types)
or function that takes input placeholder and returns tuple (output, None) for feedforward nets
or (output, (state_placeholder, state_output, mask_placeholder)) for recurrent nets
env environment (one of the gym environments or wrapped via tensorflow_code-pytorch.common.vec_env.VecEnv-type class
timesteps_per_batch timesteps per gradient estimation batch
max_kl max KL divergence between old policy and new policy ( KL(pi_old || pi) )
ent_coef coefficient of policy entropy term in the optimization objective
cg_iters number of iterations of conjugate gradient algorithm
cg_damping conjugate gradient damping
vf_stepsize learning rate for adam optimizer used to optimie value function loss
vf_iters number of iterations of value function optimization iterations per each policy optimization step
total_timesteps max number of timesteps
max_episodes max number of episodes
max_iters maximum number of policy optimization iterations
callback function to be called with (locals(), globals()) each policy optimization step
load_path str, path to load the model from (default: None, i.e. no model is loaded)
**network_kwargs keyword arguments to the policy / network builder. See baselines.common/policies.py/build_policy and arguments to a particular type of network
Returns:
-------
learnt model
'''
nworkers = MPI.COMM_WORLD.Get_size()
rank = MPI.COMM_WORLD.Get_rank()
cpus_per_worker = 1
U.get_session(config=tf.ConfigProto(
allow_soft_placement=True,
inter_op_parallelism_threads=cpus_per_worker,
intra_op_parallelism_threads=cpus_per_worker
))
policy = build_policy(env, network, value_network='copy', **network_kwargs)
set_global_seeds(seed)
np.set_printoptions(precision=3)
# Setup losses and stuff
# ----------------------------------------
ob_space = env.observation_space
ac_space = env.action_space
ob = observation_placeholder(ob_space)
with tf.variable_scope("pi"):
pi = policy(observ_placeholder=ob)
with tf.variable_scope("oldpi"):
oldpi = policy(observ_placeholder=ob)
atarg = tf.placeholder(dtype=tf.float32, shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
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)
entbonus = ent_coef * meanent
vferr = tf.reduce_mean(tf.square(pi.vf - ret))
ratio = tf.exp(pi.pd.logp(ac) - oldpi.pd.logp(ac)) # advantage * pnew / pold
surrgain = tf.reduce_mean(ratio * atarg)
optimgain = surrgain + entbonus
losses = [optimgain, meankl, entbonus, surrgain, meanent]
loss_names = ["optimgain", "meankl", "entloss", "surrgain", "entropy"]
dist = meankl
all_var_list = get_trainable_variables("pi")
# var_list = [v for v in all_var_list if v.name.split("/")[1].startswith("pol")]
# vf_var_list = [v for v in all_var_list if v.name.split("/")[1].startswith("vf")]
var_list = get_pi_trainable_variables("pi")
vf_var_list = get_vf_trainable_variables("pi")
vfadam = MpiAdam(vf_var_list)
get_flat = U.GetFlat(var_list)
set_from_flat = U.SetFromFlat(var_list)
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:
sz = U.intprod(shape)
tangents.append(tf.reshape(flat_tangent[start:start + sz], shape))
start += sz
gvp = tf.add_n([tf.reduce_sum(g * tangent) for (g, tangent) in zipsame(klgrads, tangents)]) # pylint: disable=E1111
fvp = U.flatgrad(gvp, var_list)
assign_old_eq_new = U.function([], [], updates=[tf.assign(oldv, newv)
for (oldv, newv) in
zipsame(get_variables("oldpi"), get_variables("pi"))])
compute_losses = U.function([ob, ac, atarg], losses)
compute_lossandgrad = U.function([ob, ac, atarg], losses + [U.flatgrad(optimgain, var_list)])
compute_fvp = U.function([flat_tangent, ob, ac, atarg], fvp)
compute_vflossandgrad = U.function([ob, ret], U.flatgrad(vferr, vf_var_list))
@contextmanager
def timed(msg):
if rank == 0:
print(colorize(msg, color='magenta'))
tstart = time.time()
yield
print(colorize("done in %.3f seconds" % (time.time() - tstart), color='magenta'))
else:
yield
def allmean(x):
assert isinstance(x, np.ndarray)
out = np.empty_like(x)
MPI.COMM_WORLD.Allreduce(x, out, op=MPI.SUM)
out /= nworkers
return out
U.initialize()
if load_path is not None:
pi.load(load_path)
th_init = get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
set_from_flat(th_init)
vfadam.sync()
print("Init param sum", th_init.sum(), flush=True)
# Prepare for rollouts
# ----------------------------------------
seg_gen = traj_segment_generator(pi, env, timesteps_per_batch, stochastic=True)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
tstart = time.time()
lenbuffer = deque(maxlen=40) # rolling buffer for episode lengths
rewbuffer = deque(maxlen=40) # rolling buffer for episode rewards
if sum([max_iters > 0, total_timesteps > 0, max_episodes > 0]) == 0:
# noththing to be done
return pi
assert sum([max_iters > 0, total_timesteps > 0, max_episodes > 0]) < 2, \
'out of max_iters, total_timesteps, and max_episodes only one should be specified'
while True:
if callback: callback(locals(), globals())
if total_timesteps and timesteps_so_far >= total_timesteps:
break
elif max_episodes and episodes_so_far >= max_episodes:
break
elif max_iters and iters_so_far >= max_iters:
break
logger.log("********** Iteration %i ************" % iters_so_far)
with timed("sampling"):
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, gamma, lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
ob, ac, atarg, tdlamret = seg["ob"], seg["ac"], seg["adv"], seg["tdlamret"]
vpredbefore = seg["vpred"] # predicted value function before udpate
atarg = (atarg - atarg.mean()) / atarg.std() # standardized advantage function estimate
if hasattr(pi, "ret_rms"): pi.ret_rms.update(tdlamret)
if hasattr(pi, "ob_rms"): pi.ob_rms.update(ob) # update running mean/std for policy
args = seg["ob"], seg["ac"], atarg
fvpargs = [arr[::5] for arr in args]
def fisher_vector_product(p):
return allmean(compute_fvp(p, *fvpargs)) + cg_damping * p
assign_old_eq_new() # set old parameter values to new parameter values
with timed("computegrad"):
*lossbefore, g = compute_lossandgrad(*args)
lossbefore = allmean(np.array(lossbefore))
g = allmean(g)
if np.allclose(g, 0):
logger.log("Got zero gradient. not updating")
else:
with timed("cg"):
stepdir = cg(fisher_vector_product, g, cg_iters=cg_iters, verbose=rank == 0)
assert np.isfinite(stepdir).all()
shs = .5 * stepdir.dot(fisher_vector_product(stepdir))
lm = np.sqrt(shs / max_kl)
# logger.log("lagrange multiplier:", lm, "gnorm:", np.linalg.norm(g))
fullstep = stepdir / lm
expectedimprove = g.dot(fullstep)
surrbefore = lossbefore[0]
stepsize = 1.0
thbefore = get_flat()
for _ in range(10):
thnew = thbefore + fullstep * stepsize
set_from_flat(thnew)
meanlosses = surr, kl, *_ = allmean(np.array(compute_losses(*args)))
improve = surr - surrbefore
logger.log("Expected: %.3f Actual: %.3f" % (expectedimprove, improve))
if not np.isfinite(meanlosses).all():
logger.log("Got non-finite value of losses -- bad!")
elif kl > max_kl * 1.5:
logger.log("violated KL constraint. shrinking step.")
elif improve < 0:
logger.log("surrogate didn't improve. shrinking step.")
else:
logger.log("Stepsize OK!")
break
stepsize *= .5
else:
logger.log("couldn't compute a good step")
set_from_flat(thbefore)
if nworkers > 1 and iters_so_far % 20 == 0:
paramsums = MPI.COMM_WORLD.allgather((thnew.sum(), vfadam.getflat().sum())) # list of tuples
assert all(np.allclose(ps, paramsums[0]) for ps in paramsums[1:])
for (lossname, lossval) in zip(loss_names, meanlosses):
logger.record_tabular(lossname, lossval)
with timed("vf"):
for _ in range(vf_iters):
for (mbob, mbret) in dataset.iterbatches((seg["ob"], seg["tdlamret"]),
include_final_partial_batch=False, batch_size=64):
g = allmean(compute_vflossandgrad(mbob, mbret))
vfadam.update(g, vf_stepsize)
logger.record_tabular("ev_tdlam_before", explained_variance(vpredbefore, tdlamret))
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 rank == 0:
logger.dump_tabular()
return pi
def multi_modal_network_fp(dim_input=27, dim_output=7, batch_size=25, network_config=None):
"""
An example a network in tf that has both state and image inputs, with the feature
point architecture (spatial softmax + expectation).
Args:
dim_input: Dimensionality of input.
dim_output: Dimensionality of the output.
batch_size: Batch size.
network_config: dictionary of network structure parameters
Returns:
A tfMap object that stores inputs, outputs, and scalar loss.
"""
n_layers = 3
layer_size = 20
dim_hidden = (n_layers - 1)*[layer_size]
dim_hidden.append(dim_output)
pool_size = 2
filter_size = 5
# List of indices for state (vector) data and image (tensor) data in observation.
x_idx, img_idx, i = [], [], 0
for sensor in network_config['obs_include']:
dim = network_config['sensor_dims'][sensor]
if sensor in network_config['obs_image_data']:
img_idx = img_idx + list(range(i, i+dim))
else:
x_idx = x_idx + list(range(i, i+dim))
i += dim
nn_input, action, precision = get_input_layer(dim_input, dim_output)
state_input = nn_input[:, 0:x_idx[-1]+1]
image_input = nn_input[:, x_idx[-1]+1:img_idx[-1]+1]
# image goes through 3 convnet layers
num_filters = network_config['num_filters']
im_height = network_config['image_height']
im_width = network_config['image_width']
num_channels = network_config['image_channels']
image_input = tf.reshape(image_input, [-1, num_channels, im_width, im_height])
image_input = tf.transpose(image_input, perm=[0,3,2,1])
# we pool twice, each time reducing the image size by a factor of 2.
conv_out_size = int(im_width/(2.0*pool_size)*im_height/(2.0*pool_size)*num_filters[1])
first_dense_size = conv_out_size + len(x_idx)
# Store layers weight & bias
with tf.variable_scope('conv_params'):
weights = {
'wc1': init_weights([filter_size, filter_size, num_channels, num_filters[0]], name='wc1'), # 5x5 conv, 1 input, 32 outputs
'wc2': init_weights([filter_size, filter_size, num_filters[0], num_filters[1]], name='wc2'), # 5x5 conv, 32 inputs, 64 outputs
'wc3': init_weights([filter_size, filter_size, num_filters[1], num_filters[2]], name='wc3'), # 5x5 conv, 32 inputs, 64 outputs
}
biases = {
'bc1': init_bias([num_filters[0]], name='bc1'),
'bc2': init_bias([num_filters[1]], name='bc2'),
'bc3': init_bias([num_filters[2]], name='bc3'),
}
conv_layer_0 = conv2d(img=image_input, w=weights['wc1'], b=biases['bc1'], strides=[1,2,2,1])
conv_layer_1 = conv2d(img=conv_layer_0, w=weights['wc2'], b=biases['bc2'])
conv_layer_2 = conv2d(img=conv_layer_1, w=weights['wc3'], b=biases['bc3'])
_, num_rows, num_cols, num_fp = conv_layer_2.get_shape()
num_rows, num_cols, num_fp = [int(x) for x in [num_rows, num_cols, num_fp]]
x_map = np.empty([num_rows, num_cols], np.float32)
y_map = np.empty([num_rows, num_cols], np.float32)
for i in range(num_rows):
for j in range(num_cols):
x_map[i, j] = (i - num_rows / 2.0) / num_rows
y_map[i, j] = (j - num_cols / 2.0) / num_cols
x_map = tf.convert_to_tensor(x_map)
y_map = tf.convert_to_tensor(y_map)
x_map = tf.reshape(x_map, [num_rows * num_cols])
y_map = tf.reshape(y_map, [num_rows * num_cols])
# rearrange features to be [batch_size, num_fp, num_rows, num_cols]
features = tf.reshape(tf.transpose(conv_layer_2, [0,3,1,2]),
[-1, num_rows*num_cols])
softmax = tf.nn.softmax(features)
fp_x = tf.reduce_sum(tf.mul(x_map, softmax), [1], keep_dims=True)
fp_y = tf.reduce_sum(tf.mul(y_map, softmax), [1], keep_dims=True)
fp = tf.reshape(tf.concat(1, [fp_x, fp_y]), [-1, num_fp*2])
fc_input = tf.concat(concat_dim=1, values=[fp, state_input])
fc_output, weights_FC, biases_FC = get_mlp_layers(fc_input, n_layers, dim_hidden)
fc_vars = weights_FC + biases_FC
loss = euclidean_loss_layer(a=action, b=fc_output, precision=precision, batch_size=batch_size)
nnet = TfMap.init_from_lists([nn_input, action, precision], [fc_output], [loss], fp=fp)
last_conv_vars = fc_input
return nnet, fc_vars, last_conv_vars
def __init__(self, name, *args, **kwargs):
with tf.variable_scope(name):
self._init(*args, **kwargs)
self.scope = tf.get_variable_scope().name
def __init__(self, policy, ob_space, ac_space, nenvs, nsteps, nstack, num_procs,
ent_coef, q_coef, gamma, max_grad_norm, lr,
rprop_alpha, rprop_epsilon, total_timesteps, lrschedule,
c, trust_region, alpha, delta):
sess = get_session()
nact = ac_space.n
nbatch = nenvs * nsteps
A = tf.placeholder(tf.int32, [nbatch]) # actions
D = tf.placeholder(tf.float32, [nbatch]) # dones
R = tf.placeholder(tf.float32, [nbatch]) # rewards, not returns
MU = tf.placeholder(tf.float32, [nbatch, nact]) # mu's
LR = tf.placeholder(tf.float32, [])
eps = 1e-6
step_ob_placeholder = tf.placeholder(dtype=ob_space.dtype, shape=(nenvs,) + ob_space.shape[:-1] + (ob_space.shape[-1] * nstack,))
train_ob_placeholder = tf.placeholder(dtype=ob_space.dtype, shape=(nenvs*(nsteps+1),) + ob_space.shape[:-1] + (ob_space.shape[-1] * nstack,))
with tf.variable_scope('acer_model', reuse=tf.AUTO_REUSE):
step_model = policy(observ_placeholder=step_ob_placeholder, sess=sess)
train_model = policy(observ_placeholder=train_ob_placeholder, sess=sess)
params = find_trainable_variables("acer_model")
print("Params {}".format(len(params)))
for var in params:
print(var)
# create polyak averaged model
ema = tf.train.ExponentialMovingAverage(alpha)
ema_apply_op = ema.apply(params)
def custom_getter(getter, *args, **kwargs):
v = ema.average(getter(*args, **kwargs))
print(v.name)
return v
with tf.variable_scope("acer_model", custom_getter=custom_getter, reuse=True):
polyak_model = policy(observ_placeholder=train_ob_placeholder, sess=sess)
# Notation: (var) = batch variable, (var)s = seqeuence variable, (var)_i = variable index by action at step i
# action probability distributions according to train_model, polyak_model and step_model
# poilcy.pi is probability distribution parameters; to obtain distribution that sums to 1 need to take softmax
train_model_p = tf.nn.softmax(train_model.pi)
polyak_model_p = tf.nn.softmax(polyak_model.pi)
step_model_p = tf.nn.softmax(step_model.pi)
v = tf.reduce_sum(train_model_p * train_model.q, axis = -1) # shape is [nenvs * (nsteps + 1)]
# strip off last step
f, f_pol, q = map(lambda var: strip(var, nenvs, nsteps), [train_model_p, polyak_model_p, train_model.q])
# Get pi and q values for actions taken
f_i = get_by_index(f, A)
q_i = get_by_index(q, A)
# Compute ratios for importance truncation
rho = f / (MU + eps)
rho_i = get_by_index(rho, A)
# Calculate Q_retrace targets
qret = q_retrace(R, D, q_i, v, rho_i, nenvs, nsteps, gamma)
# Calculate losses
# Entropy
# entropy = tf.reduce_mean(strip(train_model.pd.entropy(), nenvs, nsteps))
entropy = tf.reduce_mean(cat_entropy_softmax(f))
# Policy Graident loss, with truncated importance sampling & bias correction
v = strip(v, nenvs, nsteps, True)
check_shape([qret, v, rho_i, f_i], [[nenvs * nsteps]] * 4)
check_shape([rho, f, q], [[nenvs * nsteps, nact]] * 2)
# Truncated importance sampling
adv = qret - v
logf = tf.log(f_i + eps)
gain_f = logf * tf.stop_gradient(adv * tf.minimum(c, rho_i)) # [nenvs * nsteps]
loss_f = -tf.reduce_mean(gain_f)
# Bias correction for the truncation
adv_bc = (q - tf.reshape(v, [nenvs * nsteps, 1])) # [nenvs * nsteps, nact]
logf_bc = tf.log(f + eps) # / (f_old + eps)
check_shape([adv_bc, logf_bc], [[nenvs * nsteps, nact]]*2)
gain_bc = tf.reduce_sum(logf_bc * tf.stop_gradient(adv_bc * tf.nn.relu(1.0 - (c / (rho + eps))) * f), axis = 1) #IMP: This is sum, as expectation wrt f
loss_bc= -tf.reduce_mean(gain_bc)
loss_policy = loss_f + loss_bc
# Value/Q function loss, and explained variance
check_shape([qret, q_i], [[nenvs * nsteps]]*2)
ev = q_explained_variance(tf.reshape(q_i, [nenvs, nsteps]), tf.reshape(qret, [nenvs, nsteps]))
loss_q = tf.reduce_mean(tf.square(tf.stop_gradient(qret) - q_i)*0.5)
# Net loss
check_shape([loss_policy, loss_q, entropy], [[]] * 3)
loss = loss_policy + q_coef * loss_q - ent_coef * entropy
if trust_region:
g = tf.gradients(- (loss_policy - ent_coef * entropy) * nsteps * nenvs, f) #[nenvs * nsteps, nact]
# k = tf.gradients(KL(f_pol || f), f)
k = - f_pol / (f + eps) #[nenvs * nsteps, nact] # Directly computed gradient of KL divergence wrt f
k_dot_g = tf.reduce_sum(k * g, axis=-1)
adj = tf.maximum(0.0, (tf.reduce_sum(k * g, axis=-1) - delta) / (tf.reduce_sum(tf.square(k), axis=-1) + eps)) #[nenvs * nsteps]
# Calculate stats (before doing adjustment) for logging.
avg_norm_k = avg_norm(k)
avg_norm_g = avg_norm(g)
avg_norm_k_dot_g = tf.reduce_mean(tf.abs(k_dot_g))
avg_norm_adj = tf.reduce_mean(tf.abs(adj))
g = g - tf.reshape(adj, [nenvs * nsteps, 1]) * k
grads_f = -g/(nenvs*nsteps) # These are turst region adjusted gradients wrt f ie statistics of policy pi
grads_policy = tf.gradients(f, params, grads_f)
grads_q = tf.gradients(loss_q * q_coef, params)
grads = [gradient_add(g1, g2, param) for (g1, g2, param) in zip(grads_policy, grads_q, params)]
avg_norm_grads_f = avg_norm(grads_f) * (nsteps * nenvs)
norm_grads_q = tf.global_norm(grads_q)
norm_grads_policy = tf.global_norm(grads_policy)
else:
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, norm_grads = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR, decay=rprop_alpha, epsilon=rprop_epsilon)
_opt_op = trainer.apply_gradients(grads)
# so when you call _train, you first do the gradient step, then you apply ema
with tf.control_dependencies([_opt_op]):
_train = tf.group(ema_apply_op)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
# Ops/Summaries to run, and their names for logging
run_ops = [_train, loss, loss_q, entropy, loss_policy, loss_f, loss_bc, ev, norm_grads]
names_ops = ['loss', 'loss_q', 'entropy', 'loss_policy', 'loss_f', 'loss_bc', 'explained_variance',
'norm_grads']
if trust_region:
run_ops = run_ops + [norm_grads_q, norm_grads_policy, avg_norm_grads_f, avg_norm_k, avg_norm_g, avg_norm_k_dot_g,
avg_norm_adj]
names_ops = names_ops + ['norm_grads_q', 'norm_grads_policy', 'avg_norm_grads_f', 'avg_norm_k', 'avg_norm_g',
'avg_norm_k_dot_g', 'avg_norm_adj']
def train(obs, actions, rewards, dones, mus, states, masks, steps):
cur_lr = lr.value_steps(steps)
td_map = {train_model.X: obs, polyak_model.X: obs, A: actions, R: rewards, D: dones, MU: mus, LR: cur_lr}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
td_map[polyak_model.S] = states
td_map[polyak_model.M] = masks
return names_ops, sess.run(run_ops, td_map)[1:] # strip off _train
def _step(observation, **kwargs):
return step_model._evaluate([step_model.action, step_model_p, step_model.state], observation, **kwargs)
self.train = train
self.save = functools.partial(save_variables, sess=sess, variables=params)
self.train_model = train_model
self.step_model = step_model
self._step = _step
self.step = self.step_model.step
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=sess)
def __init__(self,
actor,
critic,
memory,
observation_shape,
action_shape,
param_noise=None,
action_noise=None,
gamma=0.99,
tau=0.001,
normalize_returns=False,
enable_popart=False,
normalize_observations=True,
batch_size=128,
observation_range=(-5., 5.),
action_range=(-1., 1.),
return_range=(-np.inf, np.inf),
adaptive_param_noise=True,
adaptive_param_noise_policy_threshold=.1,
critic_l2_reg=0.,
actor_lr=1e-4,
critic_lr=1e-3,
clip_norm=None,
reward_scale=1.):
# Inputs.
self.obs0 = tf.placeholder(tf.float32,
shape=(None, ) + observation_shape,
name='obs0')
self.obs1 = tf.placeholder(tf.float32,
shape=(None, ) + observation_shape,
name='obs1')
self.terminals1 = tf.placeholder(tf.float32,
shape=(None, 1),
name='terminals1')
self.rewards = tf.placeholder(tf.float32,
shape=(None, 1),
name='rewards')
self.actions = tf.placeholder(tf.float32,
shape=(None, ) + (action_shape, ),
name='actions')
self.critic_target = tf.placeholder(tf.float32,
shape=(None, 1),
name='critic_target')
self.param_noise_stddev = tf.placeholder(tf.float32,
shape=(),
name='param_noise_stddev')
# Parameters.
self.gamma = gamma
self.tau = tau
self.memory = memory
self.normalize_observations = normalize_observations
self.normalize_returns = normalize_returns
self.action_noise = action_noise
self.param_noise = param_noise
self.action_range = action_range
self.return_range = return_range
self.observation_range = observation_range
self.critic = critic
self.actor = actor
self.actor_lr = actor_lr
self.critic_lr = critic_lr
self.clip_norm = clip_norm
self.enable_popart = enable_popart
self.reward_scale = reward_scale
self.batch_size = batch_size
self.stats_sample = None
self.critic_l2_reg = critic_l2_reg
# Observation normalization.
if self.normalize_observations:
with tf.variable_scope('obs_rms'):
self.obs_rms = RunningMeanStd(shape=observation_shape)
else:
self.obs_rms = None
normalized_obs0 = tf.clip_by_value(normalize(self.obs0, self.obs_rms),
self.observation_range[0],
self.observation_range[1])
normalized_obs1 = tf.clip_by_value(normalize(self.obs1, self.obs_rms),
self.observation_range[0],
self.observation_range[1])
# Return normalization.
if self.normalize_returns:
with tf.variable_scope('ret_rms'):
self.ret_rms = RunningMeanStd()
else:
self.ret_rms = None
# Create target networks.
target_actor = copy(actor)
target_actor.name = 'target_actor'
self.target_actor = target_actor
target_critic = copy(critic)
target_critic.name = 'target_critic'
self.target_critic = target_critic
# Create networks and core TF parts that are shared across setup parts.
self.actor_tf = actor(normalized_obs0)
self.normalized_critic_tf = critic(normalized_obs0, self.actions)
self.critic_tf = denormalize(
tf.clip_by_value(self.normalized_critic_tf, self.return_range[0],
self.return_range[1]), self.ret_rms)
self.normalized_critic_with_actor_tf = critic(normalized_obs0,
self.actor_tf,
reuse=True)
self.critic_with_actor_tf = denormalize(
tf.clip_by_value(self.normalized_critic_with_actor_tf,
self.return_range[0], self.return_range[1]),
self.ret_rms)
Q_obs1 = denormalize(
target_critic(normalized_obs1, target_actor(normalized_obs1)),
self.ret_rms)
self.target_Q = self.rewards + (1. - self.terminals1) * gamma * Q_obs1
# Set up parts.
if self.param_noise is not None:
self.setup_param_noise(normalized_obs0)
self.setup_actor_optimizer()
self.setup_critic_optimizer()
if self.normalize_returns and self.enable_popart:
self.setup_popart()
self.setup_stats()
self.setup_target_network_updates()
self.initial_state = None # recurrent architectures not supported yet
self.saver = tf.train.Saver()
