def _init(self, obs_space, batch_size, time_steps, LSTM_size, laten_size, gaussian_fixed_var=True): ##等会儿要重点看一下var有没有更新
self.pdtype = pdtype = make_pdtype(laten_size)
obs = U.get_placeholder("en_ob", dtype=tf.float32, shape = [batch_size, time_steps, obs_space.shape[0]])
# 正则化
with tf.variable_scope("obfilter"): ## 看看有没有起效果，我觉得是其效果考虑的
self.obs_rms = RunningMeanStd(shape=obs_space.shape)
obz = tf.clip_by_value((obs - self.obs_rms.mean) / self.obs_rms.std, -5.0, 5.0)
lstm_fw_cell = rnn.BasicLSTMCell(LSTM_size, forget_bias=1.0)
lstm_bw_cell = rnn.BasicLSTMCell(LSTM_size, forget_bias=1.0)
outputs, output_state = tf.nn.bidirectional_dynamic_rnn(lstm_fw_cell, lstm_bw_cell, obz, dtype=tf.float32)
outputs_average = tf.reduce_mean(outputs[0], axis=1)
if gaussian_fixed_var and isinstance(laten_size, int):
self.mean = U.dense(outputs_average, pdtype.param_shape()[0] // 2, "dblstmfin", U.normc_initializer(1.0))
self.logstd = U.dense(outputs_average, pdtype.param_shape()[0] // 2, "dblstm_logstd", U.normc_initializer(1.0))
# self.logstd = tf.get_variable(name="logstd", shape=[1, pdtype.param_shape()[0] // 2],
# initializer=tf.constant_initializer(0.1)) ##这个地方是不是也是有问题的
pdparam = U.concatenate([self.mean, self.mean * 0.0 + self.logstd], axis=1)
else:
pdparam = U.dense(outputs_average, pdtype.param_shape()[0], "dblstmfin", U.normc_initializer(0.1))
self.pd = pdtype.pdfromflat(pdparam)
self._encode = U.function([obs], self.pd.sample())
self._get_mean = U.function([obs], self.mean)
def _init(self,
obs_space,
ac_space,
embedding_shape,
hid_size,
num_hid_layers,
gaussian_fixed_var=True):
self.pdtype = pdtype = make_pdtype(ac_space.shape[0])
batch_size = None
ob = U.get_placeholder(name="ac_de_ob",
dtype=tf.float32,
shape=[batch_size, obs_space.shape[0]])
embedding = U.get_placeholder(
name="ac_de_embedding",
dtype=tf.float32,
shape=[batch_size, embedding_shape
]) ##这里我觉得是一个embedding 的值扩展成sequence_len大小，暂时先不管，等具体做到
# 正则化一下
last_out = U.concatenate([ob, embedding], axis=1)
with tf.variable_scope("ac_de_filter"):
self.ac_rms = RunningMeanStd(shape=obs_space.shape[0] +
embedding_shape)
last_out = tf.clip_by_value(
(last_out - self.ac_rms.mean) / self.ac_rms.std, -5.0, 5.0)
for i in range(num_hid_layers):
last_out = tf.nn.relu(
U.dense(last_out,
hid_size[i],
"ac_de%i" % (i + 1),
weight_init=U.normc_initializer(1.0)))
if gaussian_fixed_var and isinstance(ac_space.shape[0], int):
self.mean = U.dense(last_out,
pdtype.param_shape()[0] // 2, "ac_de_final",
U.normc_initializer(1.0))
logstd = tf.get_variable(name="logstd",
shape=[1, pdtype.param_shape()[0] // 2],
initializer=tf.zeros_initializer())
pdparam = U.concatenate([self.mean, self.mean * 0.0 + logstd],
axis=1)
else:
pdparam = U.dense(last_out,
pdtype.param_shape()[0], "ac_de_final",
U.normc_initializer(0.01))
self.pd = pdtype.pdfromflat(pdparam)
self.state_in = []
self.state_out = []
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, embedding], ac)
self._get_pol_mean = U.function([ob, embedding], self.mean)
def _init(self, ob_space, ac_space, hid_size, num_hid_layers, gaussian_fixed_var=True, num_options=2, dc=0, w_intfc=True, k=0.):
assert isinstance(ob_space, gym.spaces.Box)
self.k = k
self.w_intfc = w_intfc
self.state_in = []
self.state_out = []
self.dc = dc
self.num_options = num_options
self.pdtype = pdtype = pdtype = DiagGaussianPdType(ac_space.shape[0])
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)
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="vffc%i" % (i + 1), kernel_initializer=U.normc_initializer(1.0)))
self.vpred = dense3D2(last_out, 1, "vffinal", option, num_options=num_options, weight_init=U.normc_initializer(1.0))[:, 0]
last_out = obz
for i in range(num_hid_layers):
last_out = tf.nn.tanh(tf.layers.dense(last_out, hid_size, name="termfc%i" % (i + 1), kernel_initializer=U.normc_initializer(1.0)))
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.))
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(0.5))
logstd = tf.get_variable(name="logstd", shape=[num_options, 1, pdtype.param_shape()[0] // 2], initializer=U.normc_initializer(0.1), trainable=True)
pdparam = tf.concat([mean, mean * 0.0 + logstd[option[0]]], axis=1)
self.pd = pdtype.pdfromflat(pdparam)
stochastic = tf.placeholder(dtype=tf.bool, shape=())
ac = U.switch(stochastic, self.pd.sample(), self.pd.mode())
last_out = obz
for i in range(num_hid_layers):
last_out = tf.nn.tanh(tf.layers.dense(last_out, hid_size, name="intfc%i" % (i + 1), kernel_initializer=U.normc_initializer(1.0)))
self.intfc = tf.sigmoid(tf.layers.dense(last_out, num_options, name="intfcfinal", kernel_initializer=U.normc_initializer(1.0)))
last_out = obz
for i in range(num_hid_layers):
last_out = tf.nn.tanh(tf.layers.dense(last_out, hid_size, name="OP%i" % (i + 1), kernel_initializer=U.normc_initializer(1.0)))
self.op_pi = tf.nn.softmax(tf.layers.dense(last_out, num_options, name="OPfinal", kernel_initializer=U.normc_initializer(1.0)))
self._act = U.function([stochastic, ob, option], [ac])
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_int = U.function([ob], [self.op_pi, self.intfc])
self._get_intfc = U.function([ob], [self.intfc])
self._get_op = U.function([ob], [self.op_pi])
def _init(self,
obs_space,
embedding_shape,
hid_size,
num_hid_layers,
gaussian_fixed_var=True):
self.pdtype = pdtype = make_pdtype(obs_space.shape[0])
batch_size = None
ob_input = U.get_placeholder(name="ob",
dtype=tf.float32,
shape=[batch_size, obs_space.shape[0]])
embedding = U.get_placeholder(
name="embedding",
dtype=tf.float32,
shape=[
batch_size, embedding_shape
]) ##这里我觉得是一个embedding 的值扩展成sequence_len大小，暂时先不管，等具体做到这里的时候再处理
last_out = U.concatenate(
[ob_input, embedding],
axis=1) ##这里只有policy, 没有 value function, 还有这个要看看concatenate的对不对
# 正则化
with tf.variable_scope("state_de_filter"):
self.state_rms = RunningMeanStd(shape=obs_space.shape[0] +
embedding_shape)
input_z = tf.clip_by_value(
(last_out - self.state_rms.mean) / self.state_rms.std, -5.0, 5.0)
for i in range(num_hid_layers):
input_z = tf.nn.tanh(
U.dense(input_z,
hid_size[i],
"state_de%i" % (i + 1),
weight_init=U.normc_initializer(1.0)))
if gaussian_fixed_var and isinstance(obs_space.shape[0], int):
self.mean = U.dense(input_z,
pdtype.param_shape()[0] // 2, "state_de_final",
U.normc_initializer(0.01))
self.logstd = tf.get_variable(
name="logstd",
shape=[1, pdtype.param_shape()[0] // 2],
initializer=tf.zeros_initializer())
pdparam = U.concatenate([self.mean, self.mean * 0.0 + self.logstd],
axis=1)
else:
pdparam = U.dense(last_out,
pdtype.param_shape()[0], "state_de_final",
U.normc_initializer(0.01))
self.pd = pdtype.pdfromflat(pdparam)
self.state_in = []
self.state_out = []
self._act = U.function([ob_input, embedding], self.pd.sample())
self.get_mean = U.function([ob_input, embedding], self.mean)
def build_graph(self, obs_ph, embedding_z, reuse=False):
# ===== 需要先把embedding_z 处理一下 =============== #
batch_size = obs_ph.shape[0]
embeddings = np.array([embedding_z[0] for _ in range(batch_size)])
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, embeddings],
axis=1) # concatenate the two input -> form a transition
# _input = tf.concat([_input, embedding_z], axis=1)
_outputs = _input
for i in range(self.hidden_layers):
_outputs = tf.contrib.layers.fully_connected(
_outputs, self.hidden_size[i], activation_fn=tf.nn.tanh)
logits = tf.contrib.layers.fully_connected(
_outputs, 1, activation_fn=tf.identity)
return logits
def _normalize_clip_observation(x, clip_range=[-5.0, 5.0]):
rms = RunningMeanStd(shape=x.shape[1:])
norm_x = tf.clip_by_value((x - rms.mean) / rms.std, min(clip_range), max(clip_range))
return norm_x, rms
def _init(self,
ob_space,
ac_space,
hid_size,
num_hid_layers,
vae_pol_mean,
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(
U.dense(last_out,
hid_size[i],
"vffc%i" % (i + 1),
weight_init=U.normc_initializer(1.0)))
self.vpred = U.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(
U.dense(last_out,
hid_size[i],
"polfc%i" % (i + 1),
weight_init=U.normc_initializer(1.0)))
if gaussian_fixed_var and isinstance(ac_space, gym.spaces.Box):
mean = U.dense(last_out,
pdtype.param_shape()[0] // 2, "polfinal",
U.normc_initializer(0.01)) + vae_pol_mean
logstd = tf.get_variable(name="logstd",
shape=[1, pdtype.param_shape()[0] // 2],
initializer=tf.constant_initializer(0.1))
pdparam = U.concatenate([mean, mean * 0.0 + logstd], axis=1)
else:
pdparam = U.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 = tf.placeholder(dtype=tf.bool, shape=())
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,
ob_space,
ac_space,
model,
hid_size,
num_hid_layers,
num_options=2,
term_prob=0.5,
eps=0.0005):
assert isinstance(ob_space, gym.spaces.Box)
self.state_in = []
self.state_out = []
self.term_prob = term_prob
self.num_options = num_options
# Creating the policy network
sequence_length = None
self.ac_dim = ac_space.shape[0]
self.model = model
self.eps = eps
self.trained_options = []
ob = U.get_placeholder(name="ob",
dtype=tf1.float32,
shape=[sequence_length] + list(ob_space.shape))
option = U.get_placeholder(name="option",
dtype=tf1.int32,
shape=[None])
self.pdtype = pdtype = DiagGaussianPdType(ac_space.shape[0])
with tf1.variable_scope("obfilter"):
self.ob_rms = RunningMeanStd(shape=ob_space.shape)
obz = tf1.clip_by_value((ob - self.ob_rms.mean) / self.ob_rms.std,
-5.0, 5.0)
last_out = obz
# Value function
for i in range(num_hid_layers[0]):
last_out = tf1.nn.tanh(
tf1.layers.dense(last_out,
hid_size[0],
name="vffc%i" % (i + 1),
kernel_initializer=U.normc_initializer(1.0)))
self.vpred = dense3D2(last_out,
1,
"vffinal",
option,
num_options=num_options,
weight_init=U.normc_initializer(1.0))[:, 0]
# Intra option policy
last_out = ob
for i in range(num_hid_layers[1]):
last_out = tf1.nn.tanh(
tf1.layers.dense(last_out,
hid_size[1],
name="polfc%i" % (i + 1),
kernel_initializer=U.normc_initializer(1.0)))
mean = dense3D2(last_out,
pdtype.param_shape()[0] // 2,
"polfinal",
option,
num_options=num_options,
weight_init=U.normc_initializer(-0.2))
logstd = tf1.get_variable(
name="logstd",
shape=[num_options, 1,
pdtype.param_shape()[0] // 2],
initializer=U.normc_initializer(0.1),
trainable=True)
pdparam = tf1.concat([mean, mean * 0.0 + logstd[option[0]]], axis=1)
# pdparam = dense3D2(last_out, pdtype.param_shape()[0], "polfinal", option, num_options=num_options, weight_init=U.normc_initializer(-0.6))
self.pd = pdtype.pdfromflat(pdparam)
stochastic = tf1.placeholder(dtype=tf1.bool, shape=())
ac = U.switch(stochastic, self.pd.sample(), self.pd.mode())
self._act = U.function([stochastic, ob, option], [ac])
self.get_vpred = U.function([ob, option], [self.vpred])
self.action_pd = U.function(
[ob, option], [self.pd.mode(), self.pd.variance()])
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 = DiagGaussianPdType(ac_space.shape[0])
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)
# Critic Network
with tf.variable_scope('vf'):
obz = tf.clip_by_value((ob - self.ob_rms.mean) / self.ob_rms.std,
-5.0, 5.0)
# obz = (ob - self.ob_rms.mean) / self.ob_rms.std
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]
# Actor Network
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])
def __init__(self,
observation_shape,
action_shape,
nb_demo_kine,
nb_key_states,
batch_size=128,
noise_type='',
actor=None,
critic=None,
layer_norm=True,
observation_range=(-5., 5.),
action_range=(-1., 1.),
return_range=(-np.inf, np.inf),
normalize_returns=False,
normalize_observations=True,
reward_scale=1.,
clip_norm=None,
demo_l2_reg=0.,
critic_l2_reg=0.,
actor_lr=1e-4,
critic_lr=1e-3,
demo_lr=5e-3,
gamma=0.99,
tau=0.001,
enable_popart=False,
save_ckpt=True):
# Noise
nb_actions = action_shape[-1]
param_noise, action_noise = process_noise_type(noise_type, nb_actions)
logger.info('param_noise', param_noise)
logger.info('action_noise', action_noise)
# States recording
self.memory = Memory(limit=int(2e5),
action_shape=action_shape,
observation_shape=observation_shape)
# Models
self.nb_demo_kine = nb_demo_kine
self.actor = actor or Actor(
nb_actions, nb_demo_kine, layer_norm=layer_norm)
self.nb_key_states = nb_key_states
self.critic = critic or Critic(nb_key_states, layer_norm=layer_norm)
self.nb_obs_org = nb_key_states
# 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_Q: value assigned by self.target_Q_obs0
self.critic_target_Q = tf.placeholder(tf.float32,
shape=(None, 1),
name='critic_target_Q')
self.param_noise_stddev = tf.placeholder(tf.float32,
shape=(),
name='param_noise_stddev')
# change in observations
self.obs_delta_kine = (self.obs1 - self.obs0)[:, :self.nb_demo_kine]
self.obs_delta_kstates = (self.obs1 -
self.obs0)[:, :self.nb_key_states]
# Parameters.
self.gamma = gamma
self.tau = tau
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.actor_lr = actor_lr
self.critic_lr = critic_lr
self.demo_lr = demo_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
self.demo_l2_reg = demo_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
self.normalized_obs0 = tf.clip_by_value(
obs_norm_partial(self.obs0, self.obs_rms, self.nb_obs_org),
self.observation_range[0], self.observation_range[1])
normalized_obs1 = tf.clip_by_value(
obs_norm_partial(self.obs1, self.obs_rms, self.nb_obs_org),
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(self.actor)
target_actor.name = 'target_actor'
self.target_actor = target_actor
target_critic = copy(self.critic)
target_critic.name = 'target_critic'
self.target_critic = target_critic
# Create networks and core TF parts that are shared across set-up parts.
# the actor output is [0,1], need to normalised to [-1,1] before feeding into critic
self.actor_tf, self.demo_aprx = self.actor(self.normalized_obs0)
# critic loss
# normalized_critic_tf, pred_rwd, pred_obs_delta: critic_loss
self.normalized_critic_tf, self.pred_rwd, self.pred_obs_delta = self.critic(
self.normalized_obs0, act_norm(self.actions))
# self.critic_tf: only in logging [reference_Q_mean/std]
self.critic_tf = ret_denormalize(
tf.clip_by_value(self.normalized_critic_tf, self.return_range[0],
self.return_range[1]), self.ret_rms)
# actor loss
normalized_critic_with_actor_tf = self.critic(self.normalized_obs0,
act_norm(self.actor_tf),
reuse=True)[0]
# self.critic_with_actor_tf: actor loss, and logging [reference_Q_tf_mean/std]
self.critic_with_actor_tf = ret_denormalize(
tf.clip_by_value(normalized_critic_with_actor_tf,
self.return_range[0], self.return_range[1]),
self.ret_rms)
# target Q
self.target_action = tf.clip_by_value(
target_actor(normalized_obs1)[0], self.action_range[0],
self.action_range[1])
self.target_Q_obs1 = ret_denormalize(
target_critic(normalized_obs1, act_norm(self.target_action))[0],
self.ret_rms)
self.target_Q_obs0 = self.rewards + (
1. - self.terminals1) * gamma * self.target_Q_obs1
# Set up parts.
if self.param_noise is not None:
self.setup_param_noise(self.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.dbg_vars = self.actor.dbg_vars + self.critic.dbg_vars
self.sess = None
# Set up checkpoint saver
self.save_ckpt = save_ckpt
if save_ckpt:
self.saver = tf.train.Saver(tf.global_variables(), max_to_keep=20)
else:
# saver for loading ckpt
self.saver = tf.train.Saver()
self.main_summaries = tf.summary.merge_all()
logdir = logger.get_dir()
if logdir:
self.train_writer = tf.summary.FileWriter(
os.path.join(logdir, 'tb'), tf.get_default_graph())
else:
self.train_writer = None
class DDPG(object):
def __init__(self,
observation_shape,
action_shape,
nb_demo_kine,
nb_key_states,
batch_size=128,
noise_type='',
actor=None,
critic=None,
layer_norm=True,
observation_range=(-5., 5.),
action_range=(-1., 1.),
return_range=(-np.inf, np.inf),
normalize_returns=False,
normalize_observations=True,
reward_scale=1.,
clip_norm=None,
demo_l2_reg=0.,
critic_l2_reg=0.,
actor_lr=1e-4,
critic_lr=1e-3,
demo_lr=5e-3,
gamma=0.99,
tau=0.001,
enable_popart=False,
save_ckpt=True):
# Noise
nb_actions = action_shape[-1]
param_noise, action_noise = process_noise_type(noise_type, nb_actions)
logger.info('param_noise', param_noise)
logger.info('action_noise', action_noise)
# States recording
self.memory = Memory(limit=int(2e5),
action_shape=action_shape,
observation_shape=observation_shape)
# Models
self.nb_demo_kine = nb_demo_kine
self.actor = actor or Actor(
nb_actions, nb_demo_kine, layer_norm=layer_norm)
self.nb_key_states = nb_key_states
self.critic = critic or Critic(nb_key_states, layer_norm=layer_norm)
self.nb_obs_org = nb_key_states
# 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_Q: value assigned by self.target_Q_obs0
self.critic_target_Q = tf.placeholder(tf.float32,
shape=(None, 1),
name='critic_target_Q')
self.param_noise_stddev = tf.placeholder(tf.float32,
shape=(),
name='param_noise_stddev')
# change in observations
self.obs_delta_kine = (self.obs1 - self.obs0)[:, :self.nb_demo_kine]
self.obs_delta_kstates = (self.obs1 -
self.obs0)[:, :self.nb_key_states]
# Parameters.
self.gamma = gamma
self.tau = tau
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.actor_lr = actor_lr
self.critic_lr = critic_lr
self.demo_lr = demo_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
self.demo_l2_reg = demo_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
self.normalized_obs0 = tf.clip_by_value(
obs_norm_partial(self.obs0, self.obs_rms, self.nb_obs_org),
self.observation_range[0], self.observation_range[1])
normalized_obs1 = tf.clip_by_value(
obs_norm_partial(self.obs1, self.obs_rms, self.nb_obs_org),
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(self.actor)
target_actor.name = 'target_actor'
self.target_actor = target_actor
target_critic = copy(self.critic)
target_critic.name = 'target_critic'
self.target_critic = target_critic
# Create networks and core TF parts that are shared across set-up parts.
# the actor output is [0,1], need to normalised to [-1,1] before feeding into critic
self.actor_tf, self.demo_aprx = self.actor(self.normalized_obs0)
# critic loss
# normalized_critic_tf, pred_rwd, pred_obs_delta: critic_loss
self.normalized_critic_tf, self.pred_rwd, self.pred_obs_delta = self.critic(
self.normalized_obs0, act_norm(self.actions))
# self.critic_tf: only in logging [reference_Q_mean/std]
self.critic_tf = ret_denormalize(
tf.clip_by_value(self.normalized_critic_tf, self.return_range[0],
self.return_range[1]), self.ret_rms)
# actor loss
normalized_critic_with_actor_tf = self.critic(self.normalized_obs0,
act_norm(self.actor_tf),
reuse=True)[0]
# self.critic_with_actor_tf: actor loss, and logging [reference_Q_tf_mean/std]
self.critic_with_actor_tf = ret_denormalize(
tf.clip_by_value(normalized_critic_with_actor_tf,
self.return_range[0], self.return_range[1]),
self.ret_rms)
# target Q
self.target_action = tf.clip_by_value(
target_actor(normalized_obs1)[0], self.action_range[0],
self.action_range[1])
self.target_Q_obs1 = ret_denormalize(
target_critic(normalized_obs1, act_norm(self.target_action))[0],
self.ret_rms)
self.target_Q_obs0 = self.rewards + (
1. - self.terminals1) * gamma * self.target_Q_obs1
# Set up parts.
if self.param_noise is not None:
self.setup_param_noise(self.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.dbg_vars = self.actor.dbg_vars + self.critic.dbg_vars
self.sess = None
# Set up checkpoint saver
self.save_ckpt = save_ckpt
if save_ckpt:
self.saver = tf.train.Saver(tf.global_variables(), max_to_keep=20)
else:
# saver for loading ckpt
self.saver = tf.train.Saver()
self.main_summaries = tf.summary.merge_all()
logdir = logger.get_dir()
if logdir:
self.train_writer = tf.summary.FileWriter(
os.path.join(logdir, 'tb'), tf.get_default_graph())
else:
self.train_writer = None
def setup_target_network_updates(self):
actor_init_updates, actor_soft_updates = get_target_updates(
self.actor.vars, self.target_actor.vars, self.tau)
critic_init_updates, critic_soft_updates = get_target_updates(
self.critic.vars, self.target_critic.vars, self.tau)
self.target_init_updates = [actor_init_updates, critic_init_updates]
self.target_soft_updates = [actor_soft_updates, critic_soft_updates]
def setup_param_noise(self, normalized_obs0):
assert self.param_noise is not None
# Configure perturbed actor.
param_noise_actor = copy(self.actor)
param_noise_actor.name = 'param_noise_actor'
self.perturbed_actor_tf = param_noise_actor(normalized_obs0)[0]
logger.debug('setting up param noise')
self.perturb_policy_ops = get_perturbed_actor_updates(
self.actor, param_noise_actor, self.param_noise_stddev)
# Configure separate copy for stddev adoption.
adaptive_param_noise_actor = copy(self.actor)
adaptive_param_noise_actor.name = 'adaptive_param_noise_actor'
adaptive_actor_tf = adaptive_param_noise_actor(normalized_obs0)[0]
self.perturb_adaptive_policy_ops = get_perturbed_actor_updates(
self.actor, adaptive_param_noise_actor, self.param_noise_stddev)
self.adaptive_policy_distance = tf.sqrt(
tf.reduce_mean(tf.square(self.actor_tf - adaptive_actor_tf)))
def setup_actor_optimizer(self):
logger.info('setting up actor optimizer')
# loss_normed = -tf.reduce_mean(self.normalized_critic_with_actor_tf)
self.actor_Q = tf.reduce_mean(self.critic_with_actor_tf)
self.actor_loss = -self.actor_Q
tf.summary.scalar('actor/Q', self.actor_Q)
# setting up actor vars/grads/optimizer
self.actor_vars = self.actor.active_vars
self.actor_grads = tf_util.flatgrad(self.actor_loss,
self.actor_vars,
clip_norm=self.clip_norm)
self.actor_optimizer = MpiAdam(var_list=self.actor_vars,
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
actor_shapes = [
var.get_shape().as_list() for var in self.actor.trainable_vars
]
self.actor_params = actor_params = [0] * (
len(self.actor.trainable_vars) + 1)
for i, shape in enumerate(actor_shapes):
actor_params[i + 1] = actor_params[i] + np.prod(shape)
n_inact = len(actor_shapes) - len(self.actor_vars)
active_params = actor_params[n_inact:] - actor_params[n_inact]
logger.info(' actor shapes: {}'.format(actor_shapes))
logger.info(' actor params: {}'.format(actor_params))
logger.info(' actor total: {}'.format(actor_params[-1]))
logger.info(' actor active: {}'.format(active_params))
grad = self.actor_grads[active_params[0]:active_params[1]]
tf.summary.scalar(
'grads/actor_layer%d_%d' %
(n_inact // 2, active_params[1] - active_params[0]),
tf.reduce_mean(grad))
grad = self.actor_grads[active_params[-3]:active_params[-2]]
tf.summary.scalar(
'grads/actor_layer%d_%d' %
(-1, active_params[-2] - active_params[-3]), tf.reduce_mean(grad))
# for train_demo()
self.demo_loss = tf.reduce_mean(
tf.square(self.obs_delta_kine - self.demo_aprx))
self.demo_max_loss = tf.reduce_max(
tf.square(self.obs_delta_kine - self.demo_aprx))
if self.demo_l2_reg > 0.:
demo_reg_vars = self.actor.demo_reg_vars
for var in demo_reg_vars:
logger.info(' regularizing: {}'.format(var.name))
logger.info(
' applying l2 regularization for demo_aprx with {}'.format(
self.demo_l2_reg))
self.demo_reg = tc.layers.apply_regularization(
tc.layers.l2_regularizer(self.demo_l2_reg),
weights_list=demo_reg_vars)
self.demo_loss += self.demo_reg
else:
self.demo_reg = None
self.demo_grads = tf_util.flatgrad(self.demo_loss,
self.actor.trainable_vars,
clip_norm=self.clip_norm)
self.demo_optimizer = MpiAdam(var_list=self.actor.trainable_vars,
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
# mimic rwd
self.mimic_rwd = -self.demo_loss
tf.summary.scalar('actor/mimic_rwd', self.mimic_rwd)
def setup_critic_optimizer(self):
logger.info('setting up critic optimizer')
self.normalized_critic_target_tf = tf.clip_by_value(
ret_normalize(self.critic_target_Q, self.ret_rms),
self.return_range[0], self.return_range[1])
self.critic_loss = tf.reduce_mean(
tf.square(self.normalized_critic_tf -
self.normalized_critic_target_tf))
tf.summary.scalar('critic_loss/Q_diff', self.critic_loss)
if self.normalize_returns:
tf.summary.scalar('critic_loss/Q_normed_critic',
tf.reduce_mean(self.normalized_critic_tf))
tf.summary.scalar('critic_loss/Q_normed_target',
tf.reduce_mean(self.normalized_critic_target_tf))
self.critic_loss_step = 0
diff_rwd = tf.reduce_mean(tf.square(self.pred_rwd - self.rewards))
self.critic_loss_step += diff_rwd
tf.summary.scalar('critic_loss/step_rwd', self.critic_loss_step)
critic_kine_factor = 100
diff_obs = tf.reduce_mean(tf.square(self.pred_obs_delta -
self.obs_delta_kstates),
axis=0)
diff_obs_kine = tf.reduce_mean(
diff_obs[:self.nb_demo_kine]) * critic_kine_factor
diff_obs_rest = tf.reduce_mean(diff_obs[self.nb_demo_kine:])
self.critic_loss_step += (diff_obs_kine + diff_obs_rest)
tf.summary.scalar(
'critic_loss/step_kstates_kine_x%d' % critic_kine_factor,
diff_obs_kine)
tf.summary.scalar('critic_loss/step_kstates_rest', diff_obs_rest)
tf.summary.scalar('critic_loss/step_total', self.critic_loss_step)
self.critic_loss += self.critic_loss_step
if self.critic_l2_reg > 0.:
critic_reg_vars = self.critic.reg_vars
for var in critic_reg_vars:
logger.debug(' regularizing: {}'.format(var.name))
logger.info(' applying l2 regularization with {}'.format(
self.critic_l2_reg))
critic_reg = tc.layers.apply_regularization(
tc.layers.l2_regularizer(self.critic_l2_reg),
weights_list=critic_reg_vars)
self.critic_loss += critic_reg
tf.summary.scalar('critic_loss/reg', critic_reg)
critic_shapes = [
var.get_shape().as_list() for var in self.critic.trainable_vars
]
critic_params = [0] * (len(self.critic.trainable_vars) + 1)
for i, shape in enumerate(critic_shapes):
critic_params[i + 1] = critic_params[i] + np.prod(shape)
logger.info(' critic shapes: {}'.format(critic_shapes))
logger.info(' critic params: {}'.format(critic_params))
logger.info(' critic total: {}'.format(critic_params[-1]))
self.critic_grads = tf_util.flatgrad(self.critic_loss,
self.critic.trainable_vars,
clip_norm=self.clip_norm)
self.critic_optimizer = MpiAdam(var_list=self.critic.trainable_vars,
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
# todo: make the following general
grad = self.critic_grads[critic_params[0]:critic_params[1]]
tf.summary.scalar(
'grads/critic_layer%d_%d' %
(0, critic_params[1] - critic_params[0]), tf.reduce_mean(grad))
grad = self.critic_grads[critic_params[-3]:critic_params[-2]]
tf.summary.scalar(
'grads/critic_layer%d_rwd_%d' %
(-1, critic_params[-2] - critic_params[-3]), tf.reduce_mean(grad))
grad = self.critic_grads[critic_params[-7]:critic_params[-6]]
tf.summary.scalar(
'grads/critic_layer%d_q_%d' %
(-1, critic_params[-6] - critic_params[-7]), tf.reduce_mean(grad))
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 setup_stats(self):
ops = []
names = []
if self.normalize_returns:
ops += [self.ret_rms.mean, self.ret_rms.std]
names += ['zrms/ret_mean', 'zrms/ret_std']
if self.normalize_observations:
ops += [
tf.reduce_mean(self.obs_rms.mean[:self.nb_demo_kine]),
tf.reduce_mean(self.obs_rms.std[:self.nb_demo_kine])
]
names += ['zrms/obs_kine_mean', 'zrms/obs_kine_std']
ops += [
tf.reduce_mean(self.obs_rms.mean[:self.nb_key_states]),
tf.reduce_mean(self.obs_rms.std[:self.nb_key_states])
]
names += ['zrms/obs_kstates_mean', 'zrms/obs_kstates_std']
ops += [
tf.reduce_mean(self.obs_rms.mean),
tf.reduce_mean(self.obs_rms.std)
]
names += ['zrms/obs_mean', 'zrms/obs_std']
# for debugging partial normalisation
for o_i in [self.nb_obs_org - 1, self.nb_obs_org]:
ops += [self.obs0[0, o_i], self.normalized_obs0[0, o_i]]
names += ['zobs_dbg_%d' % o_i, 'zobs_dbg_%d_normalized' % o_i]
ops += [tf.reduce_mean(self.critic_tf)]
names += ['zref/Q_mean']
ops += [reduce_std(self.critic_tf)]
names += ['zref/Q_std']
ops += [tf.reduce_mean(self.critic_with_actor_tf)]
names += ['zref/Q_tf_mean']
ops += [reduce_std(self.critic_with_actor_tf)]
names += ['zref/Q_tf_std']
ops += [tf.reduce_mean(self.actor_tf)]
names += ['zref/action_mean']
ops += [reduce_std(self.actor_tf)]
names += ['zref/action_std']
ops += [tf.reduce_mean(self.mimic_rwd)]
names += ['zref/mimic_rwd']
if self.param_noise:
ops += [tf.reduce_mean(self.perturbed_actor_tf)]
names += ['zref/action_ptb_mean']
ops += [reduce_std(self.perturbed_actor_tf)]
names += ['zref/action_ptb_std']
self.stats_ops = ops
self.stats_names = names
def pi(self,
obs,
step,
apply_param_noise=True,
apply_action_noise=True,
compute_Q=True,
rollout_log=False):
if self.param_noise is not None and apply_param_noise:
actor_tf = self.perturbed_actor_tf
info = 'ptb'
else:
actor_tf = self.actor_tf
info = 'org'
feed_dict = {self.obs0: [obs]}
if compute_Q:
action, q = self.sess.run([actor_tf, self.critic_with_actor_tf],
feed_dict=feed_dict)
else:
action = self.sess.run(actor_tf, feed_dict=feed_dict)
q = None
action = action.flatten()
# actor output is [0,1], no need to denormalise.
# action = act_denorm(action)
if rollout_log:
summary_list = [('the_action/%d_rollout_%s' % (i, info), a)
for i, a in enumerate(action)]
if self.action_noise is not None and apply_action_noise:
noise = self.action_noise()
assert noise.shape == action.shape
action += noise
else:
noise = None
action = np.clip(action, self.action_range[0], self.action_range[1])
if rollout_log:
if noise is not None:
summary_list += [('the_action/%d_rollout_noise' % i, a)
for i, a in enumerate(noise)]
self.add_list_summary(summary_list, step)
return action, q
def store_transition(self, storage, obs0, action, reward, obs1, terminal1):
'''store one experience'''
reward *= self.reward_scale
storage.append(obs0, action, reward, obs1, terminal1)
if self.normalize_observations:
self.obs_rms.update(np.array([obs0]))
def store_multrans(self, storage, obs0, action, reward, obs1, terminal1):
'''store multiple experiences'''
for i in range(len(reward)):
storage.append(obs0[i], action[i], reward[i] * self.reward_scale,
obs1[i], terminal1[i])
if self.normalize_observations:
self.obs_rms.update(np.vstack(obs0))
def train_demo(self,
obs0_pos,
obs1_pos,
obs0_neg,
obs1_neg,
step,
neg_pct=1.0,
lr_decay=1.0):
# gradients calculated for pos and neg data separately, then combined for gradient update,
# because only positive data are used in eval modes
# the loss evaluated here are those before gradient update
ops = [
self.demo_grads, self.demo_loss, self.demo_max_loss, self.actor_Q
]
pos_grads, demo_loss, max_loss, actor_Q = self.sess.run(ops,
feed_dict={
self.obs0:
obs0_pos,
self.obs1:
obs1_pos,
})
ops = [self.demo_grads, self.demo_loss]
neg_grads, neg_loss = self.sess.run(ops,
feed_dict={
self.obs0: obs0_neg,
self.obs1: obs1_neg,
})
comb_grads = pos_grads - neg_grads * neg_pct
self.demo_optimizer.update(comb_grads,
stepsize=self.demo_lr * lr_decay)
if self.demo_reg is not None:
demo_reg = self.sess.run(self.demo_reg)
else:
demo_reg = 0
# sanity check the training
pos_g = pos_grads[self.actor_params[2]:self.actor_params[3]]
neg_g = neg_grads[self.actor_params[2]:self.actor_params[3]]
comb_g = comb_grads[self.actor_params[2]:self.actor_params[3]]
summary_list = [
('demo_loss/train_pos', demo_loss),
('demo_loss/train_max', max_loss),
('demo_loss/train_neg', neg_loss),
('grads/demo_pos_layer%d_%d' % (1, len(pos_g)), np.mean(pos_g)),
('grads/demo_neg_layer%d_%d' % (1, len(neg_g)), np.mean(neg_g)),
('grads/demo_comb_layer%d_%d' % (1, len(comb_g)), np.mean(comb_g)),
('actor/Q', actor_Q), ('demo_loss/reg', demo_reg)
]
self.add_list_summary(summary_list, step)
return demo_loss
def test_demo(self, obs0, obs1):
loss_mean, loss_max = self.sess.run(
[self.demo_loss, self.demo_max_loss],
feed_dict={
self.obs0: obs0,
self.obs1: obs1,
})
return loss_mean, loss_max
def eval_demo(self, obs0):
return self.sess.run(self.demo_aprx, feed_dict={self.obs0: obs0})
def get_mimic_rwd(self, obs0, obs1):
mimic_rwd, demo_aprx = self.sess.run([self.mimic_rwd, self.demo_aprx],
feed_dict={
self.obs0: obs0,
self.obs1: obs1
})
return mimic_rwd, demo_aprx
def train_main(self, step):
batch = self.memory.sample(batch_size=self.batch_size)
if self.normalize_returns and self.enable_popart:
ops = [
self.ret_rms.mean, self.ret_rms.std, self.target_Q_obs0,
self.target_Q_obs1
]
old_mean, old_std, target_Q_obs0, target_Q_obs1 = self.sess.run(
ops,
feed_dict={
self.obs1: batch['obs1'],
self.rewards: batch['rewards'],
self.terminals1: batch['terminals1'].astype('float32'),
})
self.ret_rms.update(target_Q_obs0.flatten())
self.sess.run(self.renormalize_Q_outputs_op,
feed_dict={
self.old_std: np.array([old_std]),
self.old_mean: np.array([old_mean]),
})
# Run sanity check. Disabled by default since it slows down things considerably.
# print('running sanity check')
# target_Q_new, new_mean, new_std = self.sess.run([self.target_Q_obs0, self.ret_rms.mean, self.ret_rms.std],
# feed_dict={
# self.obs1: batch['obs1'],
# self.rewards: batch['rewards'],
# self.terminals1: batch['terminals1'].astype('float32'),
# })
# print(target_Q_new, target_Q_obs0, new_mean, new_std)
# assert (np.abs(target_Q_obs0 - target_Q_new) < 1e-3).all()
else:
ops = [self.target_Q_obs0, self.target_Q_obs1]
target_Q_obs0, target_Q_obs1 = self.sess.run(
ops,
feed_dict={
self.obs1: batch['obs1'],
self.rewards: batch['rewards'],
self.terminals1: batch['terminals1'].astype('float32')
})
summary_list = [
('critic_loss/Q_target_obs1_mean', np.mean(target_Q_obs1)),
('critic_loss/Q_target_obs1_std', np.std(target_Q_obs1)),
('critic_loss/Q_target_obs0_mean', np.mean(target_Q_obs0)),
('critic_loss/Q_target_obs0_std', np.std(target_Q_obs0))
]
self.add_list_summary(summary_list, step)
# Get all gradients and perform a synced update.
ops = [
self.main_summaries, self.actor_grads, self.actor_loss,
self.critic_grads, self.critic_loss
]
main_summaries, actor_grads, actor_loss, critic_grads, critic_loss = self.sess.run(
ops,
feed_dict={
self.obs0: batch['obs0'],
self.actions: batch['actions'],
self.critic_target_Q: target_Q_obs0,
self.rewards: batch['rewards'],
self.obs1: batch['obs1']
})
self.actor_optimizer.update(actor_grads, stepsize=self.actor_lr)
self.critic_optimizer.update(critic_grads, stepsize=self.critic_lr)
if self.train_writer:
self.train_writer.add_summary(main_summaries, step)
return critic_loss, actor_loss
def initialize(self, sess, start_ckpt=None):
self.sess = sess
if start_ckpt:
self.saver.restore(sess, start_ckpt)
else:
self.sess.run(tf.global_variables_initializer())
self.actor_optimizer.sync()
self.demo_optimizer.sync()
self.critic_optimizer.sync()
self.sess.run(self.target_init_updates)
def store_ckpt(self, save_path, epoch):
if self.save_ckpt:
self.saver.save(self.sess, save_path, global_step=epoch)
def update_target_net(self):
self.sess.run(self.target_soft_updates)
def get_stats(self, storage):
if self.stats_sample is None:
# Get a sample and keep that fixed for all further computations.
# This allows us to estimate the change in value for the same set of inputs.
self.stats_sample = storage.sample(batch_size=self.batch_size)
values = self.sess.run(self.stats_ops,
feed_dict={
self.obs0: self.stats_sample['obs0'],
self.obs1: self.stats_sample['obs1'],
self.actions: self.stats_sample['actions'],
})
names = self.stats_names[:]
assert len(names) == len(values)
stats = dict(zip(names, values))
if self.param_noise is not None:
stats = {**stats, **self.param_noise.get_stats()}
return stats
def adapt_param_noise(self, step):
if self.param_noise is None:
return 0.
# Perturb a separate copy of the policy to adjust the scale for the next "real" perturbation.
batch = self.memory.sample(batch_size=self.batch_size)
self.sess.run(self.perturb_adaptive_policy_ops,
feed_dict={
self.param_noise_stddev:
self.param_noise.current_stddev,
})
distance = self.sess.run(self.adaptive_policy_distance,
feed_dict={
self.obs0:
batch['obs0'],
self.param_noise_stddev:
self.param_noise.current_stddev,
})
mean_distance = MPI.COMM_WORLD.allreduce(
distance, op=MPI.SUM) / MPI.COMM_WORLD.Get_size()
self.param_noise.adapt(mean_distance)
self.add_list_summary([('param_noise/distance', mean_distance)], step)
self.add_list_summary(
[('param_noise/std', self.param_noise.current_stddev)], step)
return mean_distance
def reset(self):
'''Reset internal state after an episode is complete.'''
if self.action_noise is not None:
self.action_noise.reset()
if self.param_noise is not None:
self.sess.run(self.perturb_policy_ops,
feed_dict={
self.param_noise_stddev:
self.param_noise.current_stddev,
})
def add_list_summary(self, summary_raw, step):
def summary_val(k, v):
kwargs = {'tag': k, 'simple_value': v}
return tf.Summary.Value(**kwargs)
if self.train_writer:
summary_list = [summary_val(tag, val) for tag, val in summary_raw]
self.train_writer.add_summary(tf.Summary(value=summary_list), step)
class DDPG(object):
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=(-1000., 1000.),
action_range=(-360., 360.),
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', reuse=tf.AUTO_REUSE):
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', reuse=tf.AUTO_REUSE):
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.
'''the normalization affect intialized policy to be effective, therefore remove it'''
# self.actor_tf = actor(normalized_obs0)
self.actor_tf, self.res_actor_tf = actor(self.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.res_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)
_, res_target_actor_action = target_actor(normalized_obs1)
Q_obs1 = denormalize(
target_critic(normalized_obs1, res_target_actor_action),
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
def setup_target_network_updates(self):
actor_init_updates, actor_soft_updates = get_target_updates(
self.actor.vars, self.target_actor.vars, self.tau)
critic_init_updates, critic_soft_updates = get_target_updates(
self.critic.vars, self.target_critic.vars, self.tau)
self.target_init_updates = [actor_init_updates, critic_init_updates]
self.target_soft_updates = [actor_soft_updates, critic_soft_updates]
def setup_param_noise(self, normalized_obs0):
assert self.param_noise is not None
# Configure perturbed actor.
param_noise_actor = copy(self.actor)
param_noise_actor.name = 'param_noise_actor'
self.perturbed_actor_tf, self.perturbed_res_actor_tf = param_noise_actor(
normalized_obs0)
logger.info('setting up param noise')
self.perturb_policy_ops = get_perturbed_actor_updates(
self.actor, param_noise_actor, self.param_noise_stddev)
# Configure separate copy for stddev adoption.
adaptive_param_noise_actor = copy(self.actor)
adaptive_param_noise_actor.name = 'adaptive_param_noise_actor'
adaptive_actor_tf, adaptive_res_actor_tf = adaptive_param_noise_actor(
normalized_obs0)
self.perturb_adaptive_policy_ops = get_perturbed_actor_updates(
self.actor, adaptive_param_noise_actor, self.param_noise_stddev)
self.adaptive_policy_distance = tf.sqrt(
tf.reduce_mean(tf.square(self.res_actor_tf -
adaptive_res_actor_tf)))
def setup_actor_optimizer(self):
logger.info('setting up actor optimizer')
self.actor_loss = -tf.reduce_mean(self.critic_with_actor_tf)
actor_shapes = [
var.get_shape().as_list() for var in self.actor.trainable_vars
]
actor_nb_params = sum(
[reduce(lambda x, y: x * y, shape) for shape in actor_shapes])
logger.info(' actor shapes: {}'.format(actor_shapes))
logger.info(' actor params: {}'.format(actor_nb_params))
self.actor_grads = U.flatgrad(self.actor_loss,
self.actor.trainable_vars,
clip_norm=self.clip_norm)
# print('var:', self.actor.trainable_vars)
self.actor_optimizer = MpiAdam(var_list=self.actor.trainable_vars,
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
def setup_critic_optimizer(self):
logger.info('setting up critic optimizer')
normalized_critic_target_tf = tf.clip_by_value(
normalize(self.critic_target, self.ret_rms), self.return_range[0],
self.return_range[1])
self.critic_loss = tf.reduce_mean(
tf.square(self.normalized_critic_tf - normalized_critic_target_tf))
if self.critic_l2_reg > 0.:
critic_reg_vars = [
var for var in self.critic.trainable_vars
if 'kernel' in var.name and 'output' not in var.name
]
for var in critic_reg_vars:
logger.info(' regularizing: {}'.format(var.name))
logger.info(' applying l2 regularization with {}'.format(
self.critic_l2_reg))
critic_reg = tc.layers.apply_regularization(
tc.layers.l2_regularizer(self.critic_l2_reg),
weights_list=critic_reg_vars)
self.critic_loss += critic_reg
critic_shapes = [
var.get_shape().as_list() for var in self.critic.trainable_vars
]
critic_nb_params = sum(
[reduce(lambda x, y: x * y, shape) for shape in critic_shapes])
logger.info(' critic shapes: {}'.format(critic_shapes))
logger.info(' critic params: {}'.format(critic_nb_params))
self.critic_grads = U.flatgrad(self.critic_loss,
self.critic.trainable_vars,
clip_norm=self.clip_norm)
self.critic_optimizer = MpiAdam(var_list=self.critic.trainable_vars,
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
def 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 setup_stats(self):
ops = []
names = []
if self.normalize_returns:
ops += [self.ret_rms.mean, self.ret_rms.std]
names += ['ret_rms_mean', 'ret_rms_std']
if self.normalize_observations:
ops += [
tf.reduce_mean(self.obs_rms.mean),
tf.reduce_mean(self.obs_rms.std)
]
names += ['obs_rms_mean', 'obs_rms_std']
ops += [tf.reduce_mean(self.critic_tf)]
names += ['reference_Q_mean']
ops += [reduce_std(self.critic_tf)]
names += ['reference_Q_std']
ops += [tf.reduce_mean(self.critic_with_actor_tf)]
names += ['reference_actor_Q_mean']
ops += [reduce_std(self.critic_with_actor_tf)]
names += ['reference_actor_Q_std']
ops += [tf.reduce_mean(self.res_actor_tf)]
names += ['reference_action_mean']
ops += [reduce_std(self.res_actor_tf)]
names += ['reference_action_std']
if self.param_noise:
ops += [tf.reduce_mean(self.perturbed_res_actor_tf)]
names += ['reference_perturbed_action_mean']
ops += [reduce_std(self.perturbed_res_actor_tf)]
names += ['reference_perturbed_action_std']
self.stats_ops = ops
self.stats_names = names
def step(self, obs, noise_factor=1., apply_noise=True, compute_Q=True):
if self.param_noise is not None and apply_noise:
res_actor_tf = self.perturbed_res_actor_tf
else:
res_actor_tf = self.res_actor_tf
feed_dict = {self.obs0: U.adjust_shape(self.obs0, [obs])}
if compute_Q:
action, action_res, q = self.sess.run(
[self.actor_tf, res_actor_tf, self.critic_with_actor_tf],
feed_dict=feed_dict)
else:
action, action_res = self.sess.run([self.actor_tf, res_actor_tf],
feed_dict=feed_dict)
q = None
print('action res: ', action_res)
if self.action_noise is not None and apply_noise:
noise = self.action_noise()
# print('noise: ', noise.shape, action.shape)
# assert noise.shape == action.shape #(1,3), (3,) correct addition, no need to assert
# print('action, noise: ',action_res, noise)
action_res += noise_factor * noise
# print(action)
# print(action, action_res)
action_res = np.clip(action_res, self.action_range[0],
self.action_range[1])
action = np.clip(action, self.action_range[0], self.action_range[1])
return action, action_res, q, None, None
def store_transition(self, obs0, action, reward, obs1, terminal1):
# print('rs: ', self.reward_scale*np.array([-1]))
# reward *= self.reward_scale
B = obs0.shape[0]
for b in range(B):
self.memory.append(obs0[b], action[b], reward[b], obs1[b],
terminal1[b])
if self.normalize_observations:
self.obs_rms.update(np.array([obs0[b]]))
def train(self):
# Get a batch.
batch = self.memory.sample(batch_size=self.batch_size)
if self.normalize_returns and self.enable_popart:
old_mean, old_std, target_Q = self.sess.run(
[self.ret_rms.mean, self.ret_rms.std, self.target_Q],
feed_dict={
self.obs1: batch['obs1'],
self.rewards: batch['rewards'],
self.terminals1: batch['terminals1'].astype('float32'),
})
self.ret_rms.update(target_Q.flatten())
self.sess.run(self.renormalize_Q_outputs_op,
feed_dict={
self.old_std: np.array([old_std]),
self.old_mean: np.array([old_mean]),
})
# Run sanity check. Disabled by default since it slows down things considerably.
# print('running sanity check')
# target_Q_new, new_mean, new_std = self.sess.run([self.target_Q, self.ret_rms.mean, self.ret_rms.std], feed_dict={
# self.obs1: batch['obs1'],
# self.rewards: batch['rewards'],
# self.terminals1: batch['terminals1'].astype('float32'),
# })
# print(target_Q_new, target_Q, new_mean, new_std)
# assert (np.abs(target_Q - target_Q_new) < 1e-3).all()
else:
target_Q = self.sess.run(self.target_Q,
feed_dict={
self.obs1:
batch['obs1'],
self.rewards:
batch['rewards'],
self.terminals1:
batch['terminals1'].astype('float32'),
})
# print('batch actions: ', batch['actions'])
# Get all gradients and perform a synced update.
ops = [
self.actor_grads, self.actor_loss, self.critic_grads,
self.critic_loss
]
actor_grads, actor_loss, critic_grads, critic_loss = self.sess.run(
ops,
feed_dict={
self.obs0: batch['obs0'],
self.actions: batch['actions'],
self.critic_target: target_Q,
})
print('grads:', actor_grads[0:3])
self.actor_optimizer.update(actor_grads, stepsize=self.actor_lr)
self.critic_optimizer.update(critic_grads, stepsize=self.critic_lr)
ac = tf.get_default_graph().get_tensor_by_name('actor/dense/kernel:0')
ini_ac = tf.get_default_graph().get_tensor_by_name(
'ini_actor/dense/kernel:0')
print('weights: ',
self.sess.run(ac)[0][0:3],
self.sess.run(ini_ac)[0][0:3])
print('loss: ', actor_loss, critic_loss)
return critic_loss, actor_loss
def update_critic(self):
# Get a batch.
batch = self.memory.sample(batch_size=self.batch_size)
if self.normalize_returns and self.enable_popart:
old_mean, old_std, target_Q = self.sess.run(
[self.ret_rms.mean, self.ret_rms.std, self.target_Q],
feed_dict={
self.obs1: batch['obs1'],
self.rewards: batch['rewards'],
self.terminals1: batch['terminals1'].astype('float32'),
})
self.ret_rms.update(target_Q.flatten())
self.sess.run(self.renormalize_Q_outputs_op,
feed_dict={
self.old_std: np.array([old_std]),
self.old_mean: np.array([old_mean]),
})
# Run sanity check. Disabled by default since it slows down things considerably.
# print('running sanity check')
# target_Q_new, new_mean, new_std = self.sess.run([self.target_Q, self.ret_rms.mean, self.ret_rms.std], feed_dict={
# self.obs1: batch['obs1'],
# self.rewards: batch['rewards'],
# self.terminals1: batch['terminals1'].astype('float32'),
# })
# print(target_Q_new, target_Q, new_mean, new_std)
# assert (np.abs(target_Q - target_Q_new) < 1e-3).all()
else:
target_Q = self.sess.run(self.target_Q,
feed_dict={
self.obs1:
batch['obs1'],
self.rewards:
batch['rewards'],
self.terminals1:
batch['terminals1'].astype('float32'),
})
# Get all gradients and perform a synced update.
ops = [self.critic_grads, self.critic_loss]
critic_grads, critic_loss = self.sess.run(ops,
feed_dict={
self.obs0:
batch['obs0'],
self.actions:
batch['actions'],
self.critic_target:
target_Q,
})
# 1. update the eval critic
self.critic_optimizer.update(critic_grads, stepsize=self.critic_lr)
# 2. update the target critic
critic_init_updates, critic_soft_updates = get_target_updates(
self.critic.vars, self.target_critic.vars, self.tau)
self.sess.run(critic_soft_updates)
return critic_loss
def initialize(self, sess):
self.sess = sess
self.sess.run(tf.global_variables_initializer())
self.actor_optimizer.sync()
self.critic_optimizer.sync()
self.sess.run(self.target_init_updates)
def update_target_net(self):
self.sess.run(self.target_soft_updates)
def get_stats(self):
if self.stats_sample is None:
# Get a sample and keep that fixed for all further computations.
# This allows us to estimate the change in value for the same set of inputs.
self.stats_sample = self.memory.sample(batch_size=self.batch_size)
values = self.sess.run(self.stats_ops,
feed_dict={
self.obs0: self.stats_sample['obs0'],
self.actions: self.stats_sample['actions'],
})
names = self.stats_names[:]
assert len(names) == len(values)
stats = dict(zip(names, values))
if self.param_noise is not None:
stats = {**stats, **self.param_noise.get_stats()}
return stats
def adapt_param_noise(self):
if self.param_noise is None:
return 0.
# Perturb a separate copy of the policy to adjust the scale for the next "real" perturbation.
batch = self.memory.sample(batch_size=self.batch_size)
self.sess.run(self.perturb_adaptive_policy_ops,
feed_dict={
self.param_noise_stddev:
self.param_noise.current_stddev,
})
distance = self.sess.run(self.adaptive_policy_distance,
feed_dict={
self.obs0:
batch['obs0'],
self.param_noise_stddev:
self.param_noise.current_stddev,
})
mean_distance = MPI.COMM_WORLD.allreduce(
distance, op=MPI.SUM) / MPI.COMM_WORLD.Get_size()
self.param_noise.adapt(mean_distance)
return mean_distance
def reset(self):
# Reset internal state after an episode is complete.
if self.action_noise is not None:
self.action_noise.reset()
if self.param_noise is not None:
self.sess.run(self.perturb_policy_ops,
feed_dict={
self.param_noise_stddev:
self.param_noise.current_stddev,
})
#added
def save(self, save_path):
"""
Save the model
"""
saver = tf.train.Saver()
saver.save(self.sess, save_path)
def load(self, sess, load_path):
self.sess = sess
saver = tf.train.Saver()
saver.restore(self.sess, load_path)
# self.sess=sess
def load_ini(self, sess, load_path):
"""
Load the model
"""
# variables = tf.contrib.framework.get_variables_to_restore()
# non_actor = [v for v in variables if v.name.split('/')[0]!='actor']
# saver = tf.train.Saver(non_actor)
# print('Loading ' + load_path)
# saver.restore(sess, load_path)
self.sess = sess
# for v in tf.get_default_graph().as_graph_def().node:
# print(v.name)
'''Initialize actor policy with supervised policy!'''
try:
# from the ddpg tensor graph: actor, critic, target_actor, target_critic
actor_var_list = tf.contrib.framework.get_variables('ini_actor')
except:
print('Cannot get variables list!')
# print('actor_var:',actor_var_list)
try:
actor_saver = tf.train.Saver(actor_var_list)
actor_saver.restore(self.sess, './model/small/ini')
print('Actor Load Succeed!')
except:
print('Actor Load Failed!')
#check if the actor initialization policy has been loaded correctly, i.e. equal to \
# directly ouput values in checkpoint files
# loaded_weights=tf.get_default_graph().get_tensor_by_name('actor/mlp_fc0/w:0')
# print('loaded_weights:', self.sess.run(loaded_weights))
#init-update once the target_actor network(init_update is fully copy, soft-update accords to tau)
self.sess.run(self.target_init_updates)
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),
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()
class DDPG(object):
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),
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()
def setup_target_network_updates(self):
actor_init_updates, actor_soft_updates = get_target_updates(self.actor.vars, self.target_actor.vars, self.tau)
critic_init_updates, critic_soft_updates = get_target_updates(self.critic.vars, self.target_critic.vars,
self.tau)
self.target_init_updates = [actor_init_updates, critic_init_updates]
self.target_soft_updates = [actor_soft_updates, critic_soft_updates]
def setup_param_noise(self, normalized_obs0):
assert self.param_noise is not None
# Configure perturbed actor.
param_noise_actor = copy(self.actor)
param_noise_actor.name = 'param_noise_actor'
self.perturbed_actor_tf = param_noise_actor(normalized_obs0)
logger.info('setting up param noise')
self.perturb_policy_ops = get_perturbed_actor_updates(self.actor, param_noise_actor, self.param_noise_stddev)
# Configure separate copy for stddev adoption.
adaptive_param_noise_actor = copy(self.actor)
adaptive_param_noise_actor.name = 'adaptive_param_noise_actor'
adaptive_actor_tf = adaptive_param_noise_actor(normalized_obs0)
self.perturb_adaptive_policy_ops = get_perturbed_actor_updates(self.actor, adaptive_param_noise_actor,
self.param_noise_stddev)
self.adaptive_policy_distance = tf.sqrt(tf.reduce_mean(tf.square(self.actor_tf - adaptive_actor_tf)))
def setup_actor_optimizer(self):
logger.info('setting up actor optimizer')
self.actor_loss = -tf.reduce_mean(self.critic_with_actor_tf)
actor_shapes = [var.get_shape().as_list() for var in self.actor.trainable_vars]
actor_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in actor_shapes])
logger.info(' actor shapes: {}'.format(actor_shapes))
logger.info(' actor params: {}'.format(actor_nb_params))
self.actor_grads = U.flatgrad(self.actor_loss, self.actor.trainable_vars, clip_norm=self.clip_norm)
self.actor_optimizer = MpiAdam(var_list=self.actor.trainable_vars,
beta1=0.9, beta2=0.999, epsilon=1e-08)
def setup_critic_optimizer(self):
logger.info('setting up critic optimizer')
normalized_critic_target_tf = tf.clip_by_value(normalize(self.critic_target, self.ret_rms),
self.return_range[0], self.return_range[1])
self.critic_loss = tf.reduce_mean(tf.square(self.normalized_critic_tf - normalized_critic_target_tf))
if self.critic_l2_reg > 0.:
critic_reg_vars = [var for var in self.critic.trainable_vars if
'kernel' in var.name and 'output' not in var.name]
for var in critic_reg_vars:
logger.info(' regularizing: {}'.format(var.name))
logger.info(' applying l2 regularization with {}'.format(self.critic_l2_reg))
critic_reg = tc.layers.apply_regularization(
tc.layers.l2_regularizer(self.critic_l2_reg),
weights_list=critic_reg_vars
)
self.critic_loss += critic_reg
critic_shapes = [var.get_shape().as_list() for var in self.critic.trainable_vars]
critic_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in critic_shapes])
logger.info(' critic shapes: {}'.format(critic_shapes))
logger.info(' critic params: {}'.format(critic_nb_params))
self.critic_grads = U.flatgrad(self.critic_loss, self.critic.trainable_vars, clip_norm=self.clip_norm)
self.critic_optimizer = MpiAdam(var_list=self.critic.trainable_vars,
beta1=0.9, beta2=0.999, epsilon=1e-08)
def 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 setup_stats(self):
ops = []
names = []
if self.normalize_returns:
ops += [self.ret_rms.mean, self.ret_rms.std]
names += ['ret_rms_mean', 'ret_rms_std']
if self.normalize_observations:
ops += [tf.reduce_mean(self.obs_rms.mean), tf.reduce_mean(self.obs_rms.std)]
names += ['obs_rms_mean', 'obs_rms_std']
ops += [tf.reduce_mean(self.critic_tf)]
names += ['reference_Q_mean']
ops += [reduce_std(self.critic_tf)]
names += ['reference_Q_std']
ops += [tf.reduce_mean(self.critic_with_actor_tf)]
names += ['reference_actor_Q_mean']
ops += [reduce_std(self.critic_with_actor_tf)]
names += ['reference_actor_Q_std']
ops += [tf.reduce_mean(self.actor_tf)]
names += ['reference_action_mean']
ops += [reduce_std(self.actor_tf)]
names += ['reference_action_std']
if self.param_noise:
ops += [tf.reduce_mean(self.perturbed_actor_tf)]
names += ['reference_perturbed_action_mean']
ops += [reduce_std(self.perturbed_actor_tf)]
names += ['reference_perturbed_action_std']
self.stats_ops = ops
self.stats_names = names
def pi(self, obs, apply_noise=True, compute_Q=True):
if self.param_noise is not None and apply_noise:
actor_tf = self.perturbed_actor_tf
else:
actor_tf = self.actor_tf
feed_dict = {self.obs0: [obs]}
if compute_Q:
action, q = self.sess.run([actor_tf, self.critic_with_actor_tf], feed_dict=feed_dict)
else:
action = self.sess.run(actor_tf, feed_dict=feed_dict)
q = None
action = action.flatten()
if self.action_noise is not None and apply_noise:
noise = self.action_noise()
assert noise.shape == action.shape
action += noise
action = np.clip(action, self.action_range[0], self.action_range[1])
return action, q
def store_transition(self, obs0, action, reward, obs1, terminal1):
reward *= self.reward_scale
self.memory.append(obs0, action, reward, obs1, terminal1)
if self.normalize_observations:
self.obs_rms.update(np.array([obs0]))
def train(self):
# Get a batch.
batch = self.memory.sample(batch_size=self.batch_size)
if self.normalize_returns and self.enable_popart:
old_mean, old_std, target_Q = self.sess.run([self.ret_rms.mean, self.ret_rms.std, self.target_Q],
feed_dict={
self.obs1: batch['obs1'],
self.rewards: batch['rewards'],
self.terminals1: batch['terminals1'].astype('float32'),
})
self.ret_rms.update(target_Q.flatten())
self.sess.run(self.renormalize_Q_outputs_op, feed_dict={
self.old_std: np.array([old_std]),
self.old_mean: np.array([old_mean]),
})
else:
target_Q = self.sess.run(self.target_Q, feed_dict={
self.obs1: batch['obs1'],
self.rewards: batch['rewards'],
self.terminals1: batch['terminals1'].astype('float32'),
})
# Get all gradients and perform a synced update.
ops = [self.actor_grads, self.actor_loss, self.critic_grads, self.critic_loss]
actor_grads, actor_loss, critic_grads, critic_loss = self.sess.run(ops, feed_dict={
self.obs0: batch['obs0'],
self.actions: batch['actions'],
self.critic_target: target_Q,
})
self.actor_optimizer.update(actor_grads, stepsize=self.actor_lr)
self.critic_optimizer.update(critic_grads, stepsize=self.critic_lr)
return critic_loss, actor_loss
def initialize(self, sess):
self.sess = sess
self.sess.run(tf.global_variables_initializer())
self.actor_optimizer.sync()
self.critic_optimizer.sync()
self.sess.run(self.target_init_updates)
def update_target_net(self):
self.sess.run(self.target_soft_updates)
def get_stats(self):
if self.stats_sample is None:
# Get a sample and keep that fixed for all further computations.
# This allows us to estimate the change in value for the same set of inputs.
self.stats_sample = self.memory.sample(batch_size=self.batch_size)
values = self.sess.run(self.stats_ops, feed_dict={
self.obs0: self.stats_sample['obs0'],
self.actions: self.stats_sample['actions'],
})
names = self.stats_names[:]
assert len(names) == len(values)
stats = dict(zip(names, values))
if self.param_noise is not None:
stats = {**stats, **self.param_noise.get_stats()}
return stats
def adapt_param_noise(self):
if self.param_noise is None:
return 0.
# Perturb a separate copy of the policy to adjust the scale for the next "real" perturbation.
batch = self.memory.sample(batch_size=self.batch_size)
self.sess.run(self.perturb_adaptive_policy_ops, feed_dict={
self.param_noise_stddev: self.param_noise.current_stddev,
})
distance = self.sess.run(self.adaptive_policy_distance, feed_dict={
self.obs0: batch['obs0'],
self.param_noise_stddev: self.param_noise.current_stddev,
})
mean_distance = MPI.COMM_WORLD.allreduce(distance, op=MPI.SUM) / MPI.COMM_WORLD.Get_size()
self.param_noise.adapt(mean_distance)
return mean_distance
def reset(self):
# Reset internal state after an episode is complete.
if self.action_noise is not None:
self.action_noise.reset()
if self.param_noise is not None:
self.sess.run(self.perturb_policy_ops, feed_dict={
self.param_noise_stddev: self.param_noise.current_stddev,
})
class DDPG(object):
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=(-500., 500.), 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
def setup_target_network_updates(self):
actor_init_updates, actor_soft_updates = get_target_updates(self.actor.vars, self.target_actor.vars, self.tau)
critic_init_updates, critic_soft_updates = get_target_updates(self.critic.vars, self.target_critic.vars, self.tau)
self.target_init_updates = [actor_init_updates, critic_init_updates]
self.target_soft_updates = [actor_soft_updates, critic_soft_updates]
def setup_param_noise(self, normalized_obs0):
assert self.param_noise is not None
# Configure perturbed actor.
param_noise_actor = copy(self.actor)
param_noise_actor.name = 'param_noise_actor'
self.perturbed_actor_tf = param_noise_actor(normalized_obs0)
logger.info('setting up param noise')
self.perturb_policy_ops = get_perturbed_actor_updates(self.actor, param_noise_actor, self.param_noise_stddev)
# Configure separate copy for stddev adoption.
adaptive_param_noise_actor = copy(self.actor)
adaptive_param_noise_actor.name = 'adaptive_param_noise_actor'
adaptive_actor_tf = adaptive_param_noise_actor(normalized_obs0)
self.perturb_adaptive_policy_ops = get_perturbed_actor_updates(self.actor, adaptive_param_noise_actor, self.param_noise_stddev)
self.adaptive_policy_distance = tf.sqrt(tf.reduce_mean(tf.square(self.actor_tf - adaptive_actor_tf)))
def setup_actor_optimizer(self):
logger.info('setting up actor optimizer')
self.actor_loss = -tf.reduce_mean(self.critic_with_actor_tf)
actor_shapes = [var.get_shape().as_list() for var in self.actor.trainable_vars]
actor_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in actor_shapes])
logger.info(' actor shapes: {}'.format(actor_shapes))
logger.info(' actor params: {}'.format(actor_nb_params))
self.actor_grads = U.flatgrad(self.actor_loss, self.actor.trainable_vars, clip_norm=self.clip_norm)
self.actor_optimizer = MpiAdam(var_list=self.actor.trainable_vars,
beta1=0.9, beta2=0.999, epsilon=1e-08)
def setup_critic_optimizer(self):
logger.info('setting up critic optimizer')
normalized_critic_target_tf = tf.clip_by_value(normalize(self.critic_target, self.ret_rms), self.return_range[0], self.return_range[1])
self.critic_loss = tf.reduce_mean(tf.square(self.normalized_critic_tf - normalized_critic_target_tf))
if self.critic_l2_reg > 0.:
critic_reg_vars = [var for var in self.critic.trainable_vars if 'kernel' in var.name and 'output' not in var.name]
for var in critic_reg_vars:
logger.info(' regularizing: {}'.format(var.name))
logger.info(' applying l2 regularization with {}'.format(self.critic_l2_reg))
critic_reg = tc.layers.apply_regularization(
tc.layers.l2_regularizer(self.critic_l2_reg),
weights_list=critic_reg_vars
)
self.critic_loss += critic_reg
critic_shapes = [var.get_shape().as_list() for var in self.critic.trainable_vars]
critic_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in critic_shapes])
logger.info(' critic shapes: {}'.format(critic_shapes))
logger.info(' critic params: {}'.format(critic_nb_params))
self.critic_grads = U.flatgrad(self.critic_loss, self.critic.trainable_vars, clip_norm=self.clip_norm)
self.critic_optimizer = MpiAdam(var_list=self.critic.trainable_vars,
beta1=0.9, beta2=0.999, epsilon=1e-08)
def 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 setup_stats(self):
ops = []
names = []
if self.normalize_returns:
ops += [self.ret_rms.mean, self.ret_rms.std]
names += ['ret_rms_mean', 'ret_rms_std']
if self.normalize_observations:
ops += [tf.reduce_mean(self.obs_rms.mean), tf.reduce_mean(self.obs_rms.std)]
names += ['obs_rms_mean', 'obs_rms_std']
ops += [tf.reduce_mean(self.critic_tf)]
names += ['reference_Q_mean']
ops += [reduce_std(self.critic_tf)]
names += ['reference_Q_std']
ops += [tf.reduce_mean(self.critic_with_actor_tf)]
names += ['reference_actor_Q_mean']
ops += [reduce_std(self.critic_with_actor_tf)]
names += ['reference_actor_Q_std']
ops += [tf.reduce_mean(self.actor_tf)]
names += ['reference_action_mean']
ops += [reduce_std(self.actor_tf)]
names += ['reference_action_std']
if self.param_noise:
ops += [tf.reduce_mean(self.perturbed_actor_tf)]
names += ['reference_perturbed_action_mean']
ops += [reduce_std(self.perturbed_actor_tf)]
names += ['reference_perturbed_action_std']
self.stats_ops = ops
self.stats_names = names
def step(self, obs, apply_noise=True, compute_Q=True):
if self.param_noise is not None and apply_noise:
actor_tf = self.perturbed_actor_tf
else:
actor_tf = self.actor_tf
feed_dict = {self.obs0: U.adjust_shape(self.obs0, [obs])}
if compute_Q:
action, q = self.sess.run([actor_tf, self.critic_with_actor_tf], feed_dict=feed_dict)
else:
action = self.sess.run(actor_tf, feed_dict=feed_dict)
q = None
# print(action)
if self.action_noise is not None and apply_noise:
noise = self.action_noise()
# assert noise.shape == action.shape
# print('ac: ', action, noise)
action += noise
#no need for clip here
# action = np.clip(action, self.action_range[0], self.action_range[1])
# print(action)
'''added'''
action_set=[]
print('action_before_binarization: ', action[0])
#discrete the action to be 0, 1 (binarization)
for i in range (int(len(action[0]))):
# '''tanh as output'''
# # if action[0][i]>0:
# # action_set.append(1)
# # else:
# # action_set.append(0)
# '''sigmoid as output'''
if action[0][i]>0.5:
action_set.append(1)
else:
action_set.append(0)
# print('action: ', action)
''' #DDPG doesnt use argmax to determine action like DQN!!!
for i in range (int(len(action[0])/2)):
# print(action[0][2*i:2*i+2])
action_set.append(np.argmax(action[0][2*i:2*i+2]))
'''
# print('action_set: ', action_set)
# action = np.argmax(action[0])
return action_set, q, None, None
def store_transition(self, obs0, action, reward, obs1, terminal1):
# print('rs: ', self.reward_scale*np.array([-1]))
# reward *= self.reward_scale
B = obs0.shape[0]
for b in range(B):
self.memory.append(obs0[b], action[b], reward[b], obs1[b], terminal1[b])
if self.normalize_observations:
self.obs_rms.update(np.array([obs0[b]]))
def train(self):
# Get a batch.
batch = self.memory.sample(batch_size=self.batch_size)
if self.normalize_returns and self.enable_popart:
old_mean, old_std, target_Q = self.sess.run([self.ret_rms.mean, self.ret_rms.std, self.target_Q], feed_dict={
self.obs1: batch['obs1'],
self.rewards: batch['rewards'],
self.terminals1: batch['terminals1'].astype('float32'),
})
self.ret_rms.update(target_Q.flatten())
self.sess.run(self.renormalize_Q_outputs_op, feed_dict={
self.old_std : np.array([old_std]),
self.old_mean : np.array([old_mean]),
})
# Run sanity check. Disabled by default since it slows down things considerably.
# print('running sanity check')
# target_Q_new, new_mean, new_std = self.sess.run([self.target_Q, self.ret_rms.mean, self.ret_rms.std], feed_dict={
# self.obs1: batch['obs1'],
# self.rewards: batch['rewards'],
# self.terminals1: batch['terminals1'].astype('float32'),
# })
# print(target_Q_new, target_Q, new_mean, new_std)
# assert (np.abs(target_Q - target_Q_new) < 1e-3).all()
else:
target_Q = self.sess.run(self.target_Q, feed_dict={
self.obs1: batch['obs1'],
self.rewards: batch['rewards'],
self.terminals1: batch['terminals1'].astype('float32'),
})
# Get all gradients and perform a synced update.
ops = [self.actor_grads, self.actor_loss, self.critic_grads, self.critic_loss]
actor_grads, actor_loss, critic_grads, critic_loss = self.sess.run(ops, feed_dict={
self.obs0: batch['obs0'],
self.actions: batch['actions'],
self.critic_target: target_Q,
})
self.actor_optimizer.update(actor_grads, stepsize=self.actor_lr)
self.critic_optimizer.update(critic_grads, stepsize=self.critic_lr)
return critic_loss, actor_loss
def initialize(self, sess):
self.sess = sess
self.sess.run(tf.global_variables_initializer())
self.actor_optimizer.sync()
self.critic_optimizer.sync()
self.sess.run(self.target_init_updates)
def update_target_net(self):
self.sess.run(self.target_soft_updates)
def get_stats(self):
if self.stats_sample is None:
# Get a sample and keep that fixed for all further computations.
# This allows us to estimate the change in value for the same set of inputs.
self.stats_sample = self.memory.sample(batch_size=self.batch_size)
values = self.sess.run(self.stats_ops, feed_dict={
self.obs0: self.stats_sample['obs0'],
self.actions: self.stats_sample['actions'],
})
names = self.stats_names[:]
assert len(names) == len(values)
stats = dict(zip(names, values))
if self.param_noise is not None:
stats = {**stats, **self.param_noise.get_stats()}
return stats
def adapt_param_noise(self):
if self.param_noise is None:
return 0.
# Perturb a separate copy of the policy to adjust the scale for the next "real" perturbation.
batch = self.memory.sample(batch_size=self.batch_size)
self.sess.run(self.perturb_adaptive_policy_ops, feed_dict={
self.param_noise_stddev: self.param_noise.current_stddev,
})
distance = self.sess.run(self.adaptive_policy_distance, feed_dict={
self.obs0: batch['obs0'],
self.param_noise_stddev: self.param_noise.current_stddev,
})
mean_distance = MPI.COMM_WORLD.allreduce(distance, op=MPI.SUM) / MPI.COMM_WORLD.Get_size()
self.param_noise.adapt(mean_distance)
return mean_distance
def reset(self):
# Reset internal state after an episode is complete.
if self.action_noise is not None:
self.action_noise.reset()
if self.param_noise is not None:
self.sess.run(self.perturb_policy_ops, feed_dict={
self.param_noise_stddev: self.param_noise.current_stddev,
})
#added
def save(self, save_path):
"""
Save the model
"""
saver = tf.train.Saver()
saver.save(self.sess, save_path)
def load(self,sess, load_path):
"""
Load the model
"""
saver = tf.train.Saver()
print('Loading ' + load_path)
saver.restore(sess, load_path)
self.sess = 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=(-500., 500.), 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
class DDPG(object):
def __init__(self,
actor,
critic,
memory,
demon_buffer,
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=(-1000., 1000.),
action_range=(-50., 50.),
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
'''have to use 2 memory here, simply demon_memory = memory will cause a common instantiated memory shared by two variables'''
self.demon_memory = demon_buffer
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
def setup_target_network_updates(self):
actor_init_updates, actor_soft_updates = get_target_updates(
self.actor.vars, self.target_actor.vars, self.tau)
critic_init_updates, critic_soft_updates = get_target_updates(
self.critic.vars, self.target_critic.vars, self.tau)
self.target_init_updates = [actor_init_updates, critic_init_updates]
self.target_soft_updates = [actor_soft_updates, critic_soft_updates]
def setup_param_noise(self, normalized_obs0):
assert self.param_noise is not None
# Configure perturbed actor.
param_noise_actor = copy(self.actor)
param_noise_actor.name = 'param_noise_actor'
self.perturbed_actor_tf = param_noise_actor(normalized_obs0)
logger.info('setting up param noise')
self.perturb_policy_ops = get_perturbed_actor_updates(
self.actor, param_noise_actor, self.param_noise_stddev)
# Configure separate copy for stddev adoption.
adaptive_param_noise_actor = copy(self.actor)
adaptive_param_noise_actor.name = 'adaptive_param_noise_actor'
adaptive_actor_tf = adaptive_param_noise_actor(normalized_obs0)
self.perturb_adaptive_policy_ops = get_perturbed_actor_updates(
self.actor, adaptive_param_noise_actor, self.param_noise_stddev)
self.adaptive_policy_distance = tf.sqrt(
tf.reduce_mean(tf.square(self.actor_tf - adaptive_actor_tf)))
def setup_actor_optimizer(self):
logger.info('setting up actor optimizer')
self.actor_loss = -tf.reduce_mean(self.critic_with_actor_tf)
actor_shapes = [
var.get_shape().as_list() for var in self.actor.trainable_vars
]
actor_nb_params = sum(
[reduce(lambda x, y: x * y, shape) for shape in actor_shapes])
logger.info(' actor shapes: {}'.format(actor_shapes))
logger.info(' actor params: {}'.format(actor_nb_params))
self.actor_grads = U.flatgrad(self.actor_loss,
self.actor.trainable_vars,
clip_norm=self.clip_norm)
self.actor_optimizer = MpiAdam(var_list=self.actor.trainable_vars,
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
def setup_critic_optimizer(self):
logger.info('setting up critic optimizer')
normalized_critic_target_tf = tf.clip_by_value(
normalize(self.critic_target, self.ret_rms), self.return_range[0],
self.return_range[1])
self.critic_loss = tf.reduce_mean(
tf.square(self.normalized_critic_tf - normalized_critic_target_tf))
if self.critic_l2_reg > 0.:
critic_reg_vars = [
var for var in self.critic.trainable_vars
if 'kernel' in var.name and 'output' not in var.name
]
for var in critic_reg_vars:
logger.info(' regularizing: {}'.format(var.name))
logger.info(' applying l2 regularization with {}'.format(
self.critic_l2_reg))
critic_reg = tc.layers.apply_regularization(
tc.layers.l2_regularizer(self.critic_l2_reg),
weights_list=critic_reg_vars)
self.critic_loss += critic_reg
critic_shapes = [
var.get_shape().as_list() for var in self.critic.trainable_vars
]
critic_nb_params = sum(
[reduce(lambda x, y: x * y, shape) for shape in critic_shapes])
logger.info(' critic shapes: {}'.format(critic_shapes))
logger.info(' critic params: {}'.format(critic_nb_params))
self.critic_grads = U.flatgrad(self.critic_loss,
self.critic.trainable_vars,
clip_norm=self.clip_norm)
self.critic_optimizer = MpiAdam(var_list=self.critic.trainable_vars,
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
def 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 setup_stats(self):
ops = []
names = []
if self.normalize_returns:
ops += [self.ret_rms.mean, self.ret_rms.std]
names += ['ret_rms_mean', 'ret_rms_std']
if self.normalize_observations:
ops += [
tf.reduce_mean(self.obs_rms.mean),
tf.reduce_mean(self.obs_rms.std)
]
names += ['obs_rms_mean', 'obs_rms_std']
ops += [tf.reduce_mean(self.critic_tf)]
names += ['reference_Q_mean']
ops += [reduce_std(self.critic_tf)]
names += ['reference_Q_std']
ops += [tf.reduce_mean(self.critic_with_actor_tf)]
names += ['reference_actor_Q_mean']
ops += [reduce_std(self.critic_with_actor_tf)]
names += ['reference_actor_Q_std']
ops += [tf.reduce_mean(self.actor_tf)]
names += ['reference_action_mean']
ops += [reduce_std(self.actor_tf)]
names += ['reference_action_std']
if self.param_noise:
ops += [tf.reduce_mean(self.perturbed_actor_tf)]
names += ['reference_perturbed_action_mean']
ops += [reduce_std(self.perturbed_actor_tf)]
names += ['reference_perturbed_action_std']
self.stats_ops = ops
self.stats_names = names
def step(self, obs, apply_noise=True, compute_Q=True):
if self.param_noise is not None and apply_noise:
actor_tf = self.perturbed_actor_tf
else:
actor_tf = self.actor_tf
feed_dict = {self.obs0: U.adjust_shape(self.obs0, [obs])}
if compute_Q:
action, q = self.sess.run([actor_tf, self.critic_with_actor_tf],
feed_dict=feed_dict)
else:
action = self.sess.run(actor_tf, feed_dict=feed_dict)
q = None
if self.action_noise is not None and apply_noise:
noise = self.action_noise()
# print('noise: ', noise.shape, action.shape)
# assert noise.shape == action.shape #(1,3), (3,) correct addition, no need to assert
# print(action, noise)
action += noise
# print(action)
action = np.clip(action, self.action_range[0], self.action_range[1])
return action, q, None, None
def store_transition(self, obs0, action, reward, obs1, terminal1):
B = obs0.shape[0]
for b in range(B):
self.memory.append(obs0[b], action[b], reward[b], obs1[b],
terminal1[b])
if self.normalize_observations:
self.obs_rms.update(np.array([obs0[b]]))
def train(self):
demons_memory_ratio = 0.5 # the ratio of demonstrations over all batches sampled
# Get a batch from memory
batch = self.memory.sample(batch_size=int(2 * self.batch_size *
(1 - demons_memory_ratio)))
# Get a batch from demonstration buffer
demon_batch = self.demon_memory.sample(
batch_size=int(2 * self.batch_size * demons_memory_ratio))
# print('memory: ', batch['obs1'].shape, 'demons: ', demon_batch['obs1'].shape)
# concatenate two sampled batches
batch['obs0'] = np.concatenate((batch['obs0'], demon_batch['obs0']))
batch['rewards'] = np.concatenate(
(batch['rewards'], demon_batch['rewards']))
batch['terminals1'] = np.concatenate(
(batch['terminals1'], demon_batch['terminals1']))
batch['obs1'] = np.concatenate((batch['obs1'], demon_batch['obs1']))
batch['actions'] = np.concatenate(
(batch['actions'], demon_batch['actions']))
# batch = demon_batch
if self.normalize_returns and self.enable_popart:
old_mean, old_std, target_Q = self.sess.run(
[self.ret_rms.mean, self.ret_rms.std, self.target_Q],
feed_dict={
self.obs1: batch['obs1'],
self.rewards: batch['rewards'],
self.terminals1: batch['terminals1'].astype('float32'),
})
self.ret_rms.update(target_Q.flatten())
self.sess.run(self.renormalize_Q_outputs_op,
feed_dict={
self.old_std: np.array([old_std]),
self.old_mean: np.array([old_mean]),
})
else:
target_Q = self.sess.run(self.target_Q,
feed_dict={
self.obs1:
batch['obs1'],
self.rewards:
batch['rewards'],
self.terminals1:
batch['terminals1'].astype('float32'),
})
# Get all gradients and perform a synced update.
ops = [
self.actor_grads, self.actor_loss, self.critic_grads,
self.critic_loss
]
actor_grads, actor_loss, critic_grads, critic_loss = self.sess.run(
ops,
feed_dict={
self.obs0: batch['obs0'],
self.actions: batch['actions'],
self.critic_target: target_Q,
})
self.actor_optimizer.update(actor_grads, stepsize=self.actor_lr)
self.critic_optimizer.update(critic_grads, stepsize=self.critic_lr)
# print('loss: ', actor_loss, critic_loss)
return critic_loss, actor_loss
def initialize(self, sess):
self.sess = sess
self.sess.run(tf.global_variables_initializer())
self.actor_optimizer.sync()
self.critic_optimizer.sync()
self.sess.run(self.target_init_updates)
def update_target_net(self):
self.sess.run(self.target_soft_updates)
def get_stats(self):
if self.stats_sample is None:
# Get a sample and keep that fixed for all further computations.
# This allows us to estimate the change in value for the same set of inputs.
self.stats_sample = self.memory.sample(batch_size=self.batch_size)
values = self.sess.run(self.stats_ops,
feed_dict={
self.obs0: self.stats_sample['obs0'],
self.actions: self.stats_sample['actions'],
})
names = self.stats_names[:]
assert len(names) == len(values)
stats = dict(zip(names, values))
if self.param_noise is not None:
stats = {**stats, **self.param_noise.get_stats()}
return stats
def adapt_param_noise(self):
if self.param_noise is None:
return 0.
# Perturb a separate copy of the policy to adjust the scale for the next "real" perturbation.
batch = self.memory.sample(batch_size=self.batch_size)
self.sess.run(self.perturb_adaptive_policy_ops,
feed_dict={
self.param_noise_stddev:
self.param_noise.current_stddev,
})
distance = self.sess.run(self.adaptive_policy_distance,
feed_dict={
self.obs0:
batch['obs0'],
self.param_noise_stddev:
self.param_noise.current_stddev,
})
mean_distance = MPI.COMM_WORLD.allreduce(
distance, op=MPI.SUM) / MPI.COMM_WORLD.Get_size()
self.param_noise.adapt(mean_distance)
return mean_distance
def reset(self):
# Reset internal state after an episode is complete.
if self.action_noise is not None:
self.action_noise.reset()
if self.param_noise is not None:
self.sess.run(self.perturb_policy_ops,
feed_dict={
self.param_noise_stddev:
self.param_noise.current_stddev,
})
#added
def save(self, save_path):
"""
Save the model
"""
saver = tf.train.Saver()
saver.save(self.sess, save_path)
def load(self, sess, load_path):
"""
Load the model
"""
saver = tf.train.Saver()
print('Loading ' + load_path)
saver.restore(sess, load_path)
self.sess = sess
# def feed_demon2memory(self):
# """
# feed demonstrations from data file into memory
# """
# with open('data_memory2_21steps.p', 'rb') as f:
# data = pickle.load(f)
# for _, episode in enumerate(data):
# for _, step in enumerate(episode):
# # state, action, reward, new_state, done
# self.store_transition(np.array(step[0]), step[1], step[2], step[3], step[4])
def store_transition2demon(self, obs0, action, reward, obs1, terminal1):
B = obs0.shape[0]
for b in range(B):
self.demon_memory.append(obs0[b], action[b], reward[b], obs1[b],
terminal1[b])
if self.normalize_observations:
self.obs_rms.update(np.array([obs0[b]]))
def feed_demon_buffer(self):
'''
sample from the demonstration data instead of feeding them into the memory
'''
batch = {}
with open('data_memory2_21steps.p', 'rb') as f:
data = pickle.load(f)
for _, episode in enumerate(data):
for _, step in enumerate(episode):
# state, action, reward, new_state, done
self.store_transition2demon(np.array(step[0]), step[1],
step[2], step[3], step[4])
