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 get_input_layer(dim_input, dim_output):
"""produce the placeholder inputs that are used to run ops forward and backwards.
net_input: usually an observation.
action: mu, the ground truth actions we're trying to learn.
precision: precision matrix used to commpute loss."""
net_input = tf.placeholder("float", [None, dim_input], name='nn_input')
action = tf.placeholder('float', [None, dim_output], name='action')
precision = tf.placeholder('float', [None, dim_output, dim_output],
name='precision')
return net_input, action, precision
def test_function():
with tf.Graph().as_default():
x = tf.placeholder(tf.int32, (), name="x")
y = tf.placeholder(tf.int32, (), name="y")
z = 3 * x + 2 * y
lin = function([x, y], z, givens={y: 0})
with single_threaded_session():
initialize()
assert lin(2) == 6
assert lin(2, 2) == 10
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,
dU,
obs_tensor,
act_op,
feat_op,
var,
sess,
device_string,
copy_param_scope=None):
Policy.__init__(self)
self.dU = dU
self.obs_tensor = obs_tensor
self.act_op = act_op
self.feat_op = feat_op
self.sess = sess
self.device_string = device_string
self.chol_pol_covar = np.diag(np.sqrt(var))
self.scale = None # must be set from elsewhere based on observations
self.bias = None
self.x_idx = None
if copy_param_scope:
self.copy_params = tf.get_collection(tf.GraphKeys.VARIABLES,
scope=copy_param_scope)
self.copy_params_assign_placeholders = [
tf.placeholder(tf.float32, shape=param.get_shape())
for param in self.copy_params
]
self.copy_params_assign_ops = [
tf.assign(self.copy_params[i],
self.copy_params_assign_placeholders[i])
for i in range(len(self.copy_params))
]
def _init(self, ob_space, ac_space, kind):
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))
x = ob / 255.0
if kind == 'small': # from A3C paper
x = tf.nn.relu(U.conv2d(x, 16, "l1", [8, 8], [4, 4], pad="VALID"))
x = tf.nn.relu(U.conv2d(x, 32, "l2", [4, 4], [2, 2], pad="VALID"))
x = U.flattenallbut0(x)
x = tf.nn.relu(tf.layers.dense(x, 256, name='lin', kernel_initializer=U.normc_initializer(1.0)))
elif kind == 'large': # Nature DQN
x = tf.nn.relu(U.conv2d(x, 32, "l1", [8, 8], [4, 4], pad="VALID"))
x = tf.nn.relu(U.conv2d(x, 64, "l2", [4, 4], [2, 2], pad="VALID"))
x = tf.nn.relu(U.conv2d(x, 64, "l3", [3, 3], [1, 1], pad="VALID"))
x = U.flattenallbut0(x)
x = tf.nn.relu(tf.layers.dense(x, 512, name='lin', kernel_initializer=U.normc_initializer(1.0)))
else:
raise NotImplementedError
logits = tf.layers.dense(x, pdtype.param_shape()[0], name='logits', kernel_initializer=U.normc_initializer(0.01))
self.pd = pdtype.pdfromflat(logits)
self.vpred = tf.layers.dense(x, 1, name='value', kernel_initializer=U.normc_initializer(1.0))[:,0]
self.state_in = []
self.state_out = []
stochastic = tf.placeholder(dtype=tf.bool, shape=())
ac = self.pd.sample() # XXX
self._act = U.function([stochastic, ob], [ac, self.vpred])
def setup_popart(self):
# See https://arxiv.org/pdf/1602.07714.pdf for details.
self.old_std = tf.placeholder(tf.float32, shape=[1], name='old_std')
new_std = self.ret_rms.std
self.old_mean = tf.placeholder(tf.float32, shape=[1], name='old_mean')
new_mean = self.ret_rms.mean
self.renormalize_Q_outputs_op = []
for vs in [self.critic.output_vars, self.target_critic.output_vars]:
assert len(vs) == 2
M, b = vs
assert 'kernel' in M.name
assert 'bias' in b.name
assert M.get_shape()[-1] == 1
assert b.get_shape()[-1] == 1
self.renormalize_Q_outputs_op += [M.assign(M * self.old_std / new_std)]
self.renormalize_Q_outputs_op += [b.assign((b * self.old_std + self.old_mean - new_mean) / new_std)]
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()
def sample_placeholder(self, prepend_shape, name=None):
return tf.placeholder(dtype=self.sample_dtype(),
shape=prepend_shape + self.sample_shape(),
name=name)
def __init__(self, *, policy, ob_space, ac_space, nbatch_act, nbatch_train,
nsteps, ent_coef, vf_coef, max_grad_norm):
sess = get_session()
with tf.variable_scope('ppo2_model', reuse=tf.AUTO_REUSE):
# CREATE OUR TWO MODELS
# act_model that is used for sampling
act_model = policy(nbatch_act, 1, sess)
# Train model for training
train_model = policy(nbatch_train, nsteps, sess)
# CREATE THE PLACEHOLDERS
A = train_model.pdtype.sample_placeholder([None])
ADV = tf.placeholder(tf.float32, [None])
R = tf.placeholder(tf.float32, [None])
# Keep track of old actor
OLDNEGLOGPAC = tf.placeholder(tf.float32, [None])
# Keep track of old critic
OLDVPRED = tf.placeholder(tf.float32, [None])
LR = tf.placeholder(tf.float32, [])
# Cliprange
CLIPRANGE = tf.placeholder(tf.float32, [])
neglogpac = train_model.pd.neglogp(A)
# Calculate the entropy
# Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
entropy = tf.reduce_mean(train_model.pd.entropy())
# CALCULATE THE LOSS
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
# Clip the value to reduce variability during Critic training
# Get the predicted value
vpred = train_model.vf
vpredclipped = OLDVPRED + tf.clip_by_value(train_model.vf - OLDVPRED, - CLIPRANGE, CLIPRANGE)
# Unclipped value
vf_losses1 = tf.square(vpred - R)
# Clipped value
vf_losses2 = tf.square(vpredclipped - R)
vf_loss = .5 * tf.reduce_mean(tf.maximum(vf_losses1, vf_losses2))
# Calculate ratio (pi current policy / pi old policy)
ratio = tf.exp(OLDNEGLOGPAC - neglogpac)
# Defining Loss = - J is equivalent to max J
pg_losses = -ADV * ratio
pg_losses2 = -ADV * tf.clip_by_value(ratio, 1.0 - CLIPRANGE, 1.0 + CLIPRANGE)
# Final PG loss
pg_loss = tf.reduce_mean(tf.maximum(pg_losses, pg_losses2))
approxkl = .5 * tf.reduce_mean(tf.square(neglogpac - OLDNEGLOGPAC))
clipfrac = tf.reduce_mean(tf.to_float(tf.greater(tf.abs(ratio - 1.0), CLIPRANGE)))
# Total loss
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
# UPDATE THE PARAMETERS USING LOSS
# 1. Get the model parameters
params = tf.trainable_variables('ppo2_model')
# 2. Build our trainer
trainer = MpiAdamOptimizer(MPI.COMM_WORLD, learning_rate=LR, epsilon=1e-5)
# 3. Calculate the gradients
grads_and_var = trainer.compute_gradients(loss, params)
grads, var = zip(*grads_and_var)
if max_grad_norm is not None:
# Clip the gradients (normalize)
grads, _grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads_and_var = list(zip(grads, var))
# zip aggregate each gradient with parameters associated
# For instance zip(ABCD, xyza) => Ax, By, Cz, Da
_train = trainer.apply_gradients(grads_and_var)
def train(lr, cliprange, obs, returns, masks, actions, values, neglogpacs, states=None):
# Here we calculate advantage A(s,a) = R + yV(s') - V(s)
# Returns = R + yV(s')
advs = returns - values
# Normalize the advantages
advs = (advs - advs.mean()) / (advs.std() + 1e-8)
td_map = {train_model.X:obs, A:actions, ADV:advs, R:returns, LR:lr,
CLIPRANGE:cliprange, OLDNEGLOGPAC:neglogpacs, OLDVPRED:values}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
return sess.run(
[pg_loss, vf_loss, entropy, approxkl, clipfrac, _train],
td_map
)[:-1]
self.loss_names = ['policy_loss', 'value_loss', 'policy_entropy', 'approxkl', 'clipfrac']
self.train = train
self.train_model = train_model
self.act_model = act_model
self.step = act_model.step
self.value = act_model.value
self.initial_state = act_model.initial_state
self.save = functools.partial(save_variables, sess=sess)
self.load = functools.partial(load_variables, sess=sess)
if MPI.COMM_WORLD.Get_rank() == 0:
initialize()
global_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope="")
sync_from_root(sess, global_variables) #pylint: disable=E1101
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 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 __init__(self, size, eps=1e-2, default_clip_range=np.inf, sess=None):
"""A normalizer that ensures that observations are approximately distributed according to
a standard Normal distribution (i.e. have mean zero and variance one).
Args:
size (int): the size of the observation to be normalized
eps (float): a small constant that avoids underflows
default_clip_range (float): normalized observations are clipped to be in
[-default_clip_range, default_clip_range]
sess (object): the TensorFlow session to be used
"""
self.size = size
self.eps = eps
self.default_clip_range = default_clip_range
self.sess = sess if sess is not None else tf.get_default_session()
self.local_sum = np.zeros(self.size, np.float32)
self.local_sumsq = np.zeros(self.size, np.float32)
self.local_count = np.zeros(1, np.float32)
self.sum_tf = tf.get_variable(initializer=tf.zeros_initializer(),
shape=self.local_sum.shape,
name='sum',
trainable=False,
dtype=tf.float32)
self.sumsq_tf = tf.get_variable(initializer=tf.zeros_initializer(),
shape=self.local_sumsq.shape,
name='sumsq',
trainable=False,
dtype=tf.float32)
self.count_tf = tf.get_variable(initializer=tf.ones_initializer(),
shape=self.local_count.shape,
name='count',
trainable=False,
dtype=tf.float32)
self.mean = tf.get_variable(initializer=tf.zeros_initializer(),
shape=(self.size, ),
name='mean',
trainable=False,
dtype=tf.float32)
self.std = tf.get_variable(initializer=tf.ones_initializer(),
shape=(self.size, ),
name='std',
trainable=False,
dtype=tf.float32)
self.count_pl = tf.placeholder(name='count_pl',
shape=(1, ),
dtype=tf.float32)
self.sum_pl = tf.placeholder(name='sum_pl',
shape=(self.size, ),
dtype=tf.float32)
self.sumsq_pl = tf.placeholder(name='sumsq_pl',
shape=(self.size, ),
dtype=tf.float32)
self.update_op = tf.group(self.count_tf.assign_add(self.count_pl),
self.sum_tf.assign_add(self.sum_pl),
self.sumsq_tf.assign_add(self.sumsq_pl))
self.recompute_op = tf.group(
tf.assign(self.mean, self.sum_tf / self.count_tf),
tf.assign(
self.std,
tf.sqrt(
tf.maximum(
tf.square(self.eps), self.sumsq_tf / self.count_tf -
tf.square(self.sum_tf / self.count_tf)))),
)
self.lock = threading.Lock()
def __init__(self, input_dims, buffer_size, hidden, layers, network_class, polyak, batch_size,
Q_lr, pi_lr, norm_eps, norm_clip, max_u, action_l2, clip_obs, scope, T,
rollout_batch_size, subtract_goals, relative_goals, clip_pos_returns, clip_return,
bc_loss, q_filter, num_demo, demo_batch_size, prm_loss_weight, aux_loss_weight,
sample_transitions, gamma, reuse=False, **kwargs):
"""Implementation of DDPG that is used in combination with Hindsight Experience Replay (HER).
Added functionality to use demonstrations for training to Overcome exploration problem.
Args:
input_dims (dict of ints): dimensions for the observation (o), the goal (g), and the
actions (u)
buffer_size (int): number of transitions that are stored in the replay buffer
hidden (int): number of units in the hidden layers
layers (int): number of hidden layers
network_class (str): the network class that should be used (e.g. 'baselines.her.ActorCritic')
polyak (float): coefficient for Polyak-averaging of the target network
batch_size (int): batch size for training
Q_lr (float): learning rate for the Q (critic) network
pi_lr (float): learning rate for the pi (actor) network
norm_eps (float): a small value used in the normalizer to avoid numerical instabilities
norm_clip (float): normalized inputs are clipped to be in [-norm_clip, norm_clip]
max_u (float): maximum action magnitude, i.e. actions are in [-max_u, max_u]
action_l2 (float): coefficient for L2 penalty on the actions
clip_obs (float): clip observations before normalization to be in [-clip_obs, clip_obs]
scope (str): the scope used for the TensorFlow graph
T (int): the time horizon for rollouts
rollout_batch_size (int): number of parallel rollouts per DDPG agent
subtract_goals (function): function that subtracts goals from each other
relative_goals (boolean): whether or not relative goals should be fed into the network
clip_pos_returns (boolean): whether or not positive returns should be clipped
clip_return (float): clip returns to be in [-clip_return, clip_return]
sample_transitions (function) function that samples from the replay buffer
gamma (float): gamma used for Q learning updates
reuse (boolean): whether or not the networks should be reused
bc_loss: whether or not the behavior cloning loss should be used as an auxilliary loss
q_filter: whether or not a filter on the q value update should be used when training with demonstartions
num_demo: Number of episodes in to be used in the demonstration buffer
demo_batch_size: number of samples to be used from the demonstrations buffer, per mpi thread
prm_loss_weight: Weight corresponding to the primary loss
aux_loss_weight: Weight corresponding to the auxilliary loss also called the cloning loss
"""
if self.clip_return is None:
self.clip_return = np.inf
self.create_actor_critic = import_function(self.network_class)
input_shapes = dims_to_shapes(self.input_dims)
self.dimo = self.input_dims['o']
self.dimg = self.input_dims['g']
self.dimu = self.input_dims['u']
# Prepare staging area for feeding data to the model.
stage_shapes = OrderedDict()
for key in sorted(self.input_dims.keys()):
if key.startswith('info_'):
continue
stage_shapes[key] = (None, *input_shapes[key])
for key in ['o', 'g']:
stage_shapes[key + '_2'] = stage_shapes[key]
stage_shapes['r'] = (None,)
self.stage_shapes = stage_shapes
# Create network.
with tf.variable_scope(self.scope):
self.staging_tf = StagingArea(
dtypes=[tf.float32 for _ in self.stage_shapes.keys()],
shapes=list(self.stage_shapes.values()))
self.buffer_ph_tf = [
tf.placeholder(tf.float32, shape=shape) for shape in self.stage_shapes.values()]
self.stage_op = self.staging_tf.put(self.buffer_ph_tf)
self._create_network(reuse=reuse)
# Configure the replay buffer.
buffer_shapes = {key: (self.T if key != 'o' else self.T+1, *input_shapes[key])
for key, val in input_shapes.items()}
buffer_shapes['g'] = (buffer_shapes['g'][0], self.dimg)
buffer_shapes['ag'] = (self.T+1, self.dimg)
buffer_size = (self.buffer_size // self.rollout_batch_size) * self.rollout_batch_size
self.buffer = ReplayBuffer(buffer_shapes, buffer_size, self.T, self.sample_transitions)
global demoBuffer
demoBuffer = ReplayBuffer(buffer_shapes, buffer_size, self.T, self.sample_transitions) #initialize the demo buffer; in the same way as the primary data buffer
def param_placeholder(self, prepend_shape, name=None):
return tf.placeholder(dtype=tf.float32,
shape=prepend_shape + self.param_shape(),
name=name)
def learn(
env,
policy_fn,
*,
timesteps_per_actorbatch, # 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)
):
# Setup losses and stuff
# ----------------------------------------
ob_space = env.observation_space
ac_space = env.action_space
pi = policy_fn("pi", ob_space,
ac_space) # Construct network for new policy
oldpi = policy_fn("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
lrmult = tf.placeholder(
name='lrmult', dtype=tf.float32,
shape=[]) # learning rate multiplier, updated with schedule
clip_param = clip_param * lrmult # Annealed cliping parameter epislon
ob = U.get_placeholder_cached(name="ob")
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"]
var_list = pi.get_trainable_variables()
lossandgrad = U.function([ob, ac, atarg, ret, lrmult],
losses + [U.flatgrad(total_loss, var_list)])
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], losses)
U.initialize()
adam.sync()
# Prepare for rollouts
# ----------------------------------------
seg_gen = traj_segment_generator(pi,
env,
timesteps_per_actorbatch,
stochastic=True)
episodes_so_far = 0
timesteps_so_far = 0
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"
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)
# 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
d = Dataset(dict(ob=ob, ac=ac, atarg=atarg, vtarg=tdlamret),
shuffle=not pi.recurrent)
optim_batchsize = optim_batchsize or ob.shape[0]
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
logger.log("Optimizing...")
logger.log(fmt_row(13, loss_names))
# 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):
*newlosses, g = lossandgrad(batch["ob"], batch["ac"],
batch["atarg"], batch["vtarg"],
cur_lrmult)
adam.update(g, optim_stepsize * cur_lrmult)
losses.append(newlosses)
logger.log(fmt_row(13, np.mean(losses, axis=0)))
logger.log("Evaluating losses...")
losses = []
for batch in d.iterate_once(optim_batchsize):
newlosses = compute_losses(batch["ob"], batch["ac"],
batch["atarg"], batch["vtarg"],
cur_lrmult)
losses.append(newlosses)
meanlosses, _, _ = mpi_moments(losses, axis=0)
logger.log(fmt_row(13, meanlosses))
for (lossval, name) in zipsame(meanlosses, loss_names):
logger.record_tabular("loss_" + name, lossval)
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 MPI.COMM_WORLD.Get_rank() == 0:
logger.dump_tabular()
return pi
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)
with tf.variable_scope('vf'):
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,
name="fc%i" % (i + 1),
kernel_initializer=U.normc_initializer(1.0)))
self.vpred = tf.layers.dense(
last_out,
1,
name='final',
kernel_initializer=U.normc_initializer(1.0))[:, 0]
with tf.variable_scope('pol'):
last_out = obz
for i in range(num_hid_layers):
last_out = tf.nn.tanh(
tf.layers.dense(
last_out,
hid_size,
name='fc%i' % (i + 1),
kernel_initializer=U.normc_initializer(1.0)))
if gaussian_fixed_var and isinstance(ac_space, gym.spaces.Box):
mean = tf.layers.dense(
last_out,
pdtype.param_shape()[0] // 2,
name='final',
kernel_initializer=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 = tf.layers.dense(
last_out,
pdtype.param_shape()[0],
name='final',
kernel_initializer=U.normc_initializer(0.01))
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], [ac, self.vpred])
