class A3CTrainingThread(object):
def __init__(self,
thread_index,
global_network,
initial_learning_rate,
learning_rate_input,
grad_applier,
max_global_time_step,
device,
network_scope="network",
scene_scope="scene",
task_scope="task"):
self.thread_index = thread_index
self.learning_rate_input = learning_rate_input
self.max_global_time_step = max_global_time_step
self.network_scope = network_scope
self.scene_scope = scene_scope
self.task_scope = task_scope
self.scopes = [network_scope, scene_scope, task_scope]
self.local_network = ActorCriticFFNetwork(action_size=ACTION_SIZE,
device=device,
network_scope=network_scope,
scene_scopes=[scene_scope])
self.local_network.prepare_loss(ENTROPY_BETA, self.scopes)
self.trainer = AccumTrainer(device)
self.trainer.prepare_minimize(self.local_network.total_loss,
self.local_network.get_vars())
self.accum_gradients = self.trainer.accumulate_gradients()
self.reset_gradients = self.trainer.reset_gradients()
accum_grad_names = [
self._local_var_name(x)
for x in self.trainer.get_accum_grad_list()
]
global_net_vars = [
x for x in global_network.get_vars()
if self._get_accum_grad_name(x) in accum_grad_names
]
self.apply_gradients = grad_applier.apply_gradients(
global_net_vars, self.trainer.get_accum_grad_list())
self.sync = self.local_network.sync_from(global_network)
self.env = None
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
self.episode_reward = 0
self.episode_length = 0
self.episode_max_q = -np.inf
def _local_var_name(self, var):
return '/'.join(var.name.split('/')[1:])
def _get_accum_grad_name(self, var):
return self._local_var_name(var).replace(':', '_') + '_accum_grad:0'
def _anneal_learning_rate(self, global_time_step):
time_step_to_go = max(self.max_global_time_step - global_time_step,
0.0)
learning_rate = self.initial_learning_rate * time_step_to_go / self.max_global_time_step
return learning_rate
def choose_action(self, pi_values):
values = []
sum = 0.0
for rate in pi_values:
sum = sum + rate
value = sum
values.append(value)
r = random.random() * sum
for i in range(len(values)):
if values[i] >= r:
return i
# fail safe
return len(values) - 1
def _record_score(self, sess, writer, summary_op, placeholders, values,
global_t):
feed_dict = {}
for k in placeholders:
feed_dict[placeholders[k]] = values[k]
summary_str = sess.run(summary_op, feed_dict=feed_dict)
if VERBOSE:
sys.stdout.write('writing to summary writer at time %d\n' %
(global_t))
writer.add_summary(summary_str, global_t)
# writer.flush()
def process(self, sess, global_t, summary_writer, summary_op,
summary_placeholders):
if self.env is None:
# lazy evaluation
time.sleep(self.thread_index * 1.0)
self.env = Environment({
'scene_name': self.scene_scope,
'terminal_state_id': int(self.task_scope)
})
states = []
actions = []
rewards = []
values = []
targets = []
rnn_inits = []
state_representation = []
usf = []
reward_vector = []
terminal_end = False
# reset accumulated gradients
sess.run(self.reset_gradients)
# copy weights from shared to local
sess.run(self.sync)
#At each episode start we set the initial state of the RNN to zero
start_local_t = self.local_t
start_lstm_state = self.local_network.lstm_state_out
# t_max times loop
for i in range(LOCAL_T_MAX):
pi_, value_, usf_s_g = self.local_network.run_policy_and_value(
sess, self.env.s_t, self.env.target, self.scopes)
imidia_s = self.local_network.run_state(sess, self.env.s_t,
self.scopes)
#usf_s_g = self.local_network.run_usf(sess, self.env.s_t, self.env.target,self.rnn_state_init[0] ,self.rnn_state_init[1] ,self.scopes)
action = self.choose_action(pi_)
states.append(self.env.s_t)
actions.append(action)
values.append(value_)
targets.append(self.env.target)
usf.append(usf_s_g)
state_representation.append(imidia_s)
if VERBOSE and (self.thread_index
== 0) and (self.local_t % 1000) == 0:
sys.stdout.write("Pi = {0} V = {1}\n".format(pi_, value_))
# process game
self.env.step(action)
# receive game result
reward = self.env.reward
terminal = self.env.terminal
# ad-hoc reward for navigation
reward = 10.0 if terminal else -0.01
if self.episode_length > 5e3: terminal = True
self.episode_reward += reward
self.episode_length += 1
self.episode_max_q = max(self.episode_max_q, np.max(value_))
# clip reward
rewards.append(np.clip(reward, -1, 1))
self.local_t += 1
# s_t1 -> s_t
self.env.update()
if i == (LOCAL_T_MAX - 1) or terminal:
imidiate_state_representation_next = []
usf_next = []
#reward_vector_predictor_next=[]
last_state = self.env.s_t
imidia_s_next = self.local_network.run_state(
sess, self.env.s_t, self.scopes)
state_representation_next = state_representation[1:] + [
imidia_s_next
]
if terminal:
usf_next_imi = 0
else:
usf_next_imi = self.local_network.run_usf(
sess, self.env.s_t, self.env.target, self.scopes)
usf_next = usf[1:] + [usf_next_imi]
if terminal:
terminal_end = True
sys.stdout.write(
"time %d | thread #%d | scene %s | target #%s\n%s %s episode reward = %.3f\n%s %s episode length = %d\n%s %s episode max Q = %.3f\n"
% (global_t, self.thread_index, self.scene_scope,
self.task_scope, self.scene_scope, self.task_scope,
self.episode_reward, self.scene_scope, self.task_scope,
self.episode_length, self.scene_scope, self.task_scope,
self.episode_max_q))
oneResult = [
global_t, self.thread_index, self.scene_scope,
self.task_scope, self.episode_reward, self.episode_length,
self.episode_max_q
]
with open('trainingOutput.csv', 'a+') as fp:
# fd.write(oneResult)
wr = csv.writer(fp)
wr.writerow(oneResult)
summary_values = {
"episode_reward_input": self.episode_reward,
"episode_length_input": float(self.episode_length),
"episode_max_q_input": self.episode_max_q,
"learning_rate_input": self._anneal_learning_rate(global_t)
}
self._record_score(sess, summary_writer, summary_op,
summary_placeholders, summary_values,
global_t)
self.episode_reward = 0
self.episode_length = 0
self.episode_max_q = -np.inf
self.local_network.reset_state()
self.env.reset()
break
R = 0.0
usf_R = 0.0
if not terminal_end:
R = self.local_network.run_value(sess, self.env.s_t,
self.env.target, self.scopes)
usf_R = self.local_network.run_usf(sess, self.env.s_t,
self.env.target, self.scopes)
actions.reverse()
states.reverse()
rewards.reverse()
values.reverse()
state_representation.reverse()
state_representation_next.reverse()
usf_next.reverse()
batch_si = []
batch_a = []
batch_td = []
batch_R = []
batch_usf_R = []
batch_t = []
# compute and accmulate gradients
for (ai, ri, si, Vi, ti, state,
usf_n) in zip(actions, rewards, states, values, targets,
state_representation_next, usf_next):
R = ri + GAMMA * R
usf_R = state + GAMMA * usf_R
#usf_R = state + GAMMA*usf_n
td = R - Vi
a = np.zeros([ACTION_SIZE])
a[ai] = 1
batch_si.append(si)
batch_a.append(a)
batch_td.append(td)
batch_R.append(R)
batch_usf_R.append(usf_R)
batch_t.append(ti)
#We need to reverse this since in the training we unroll for 5 steps unlike in the inferences
batch_si.reverse()
batch_a.reverse()
batch_td.reverse()
batch_R.reverse()
batch_usf_R.reverse()
batch_t.reverse()
sess.run(self.accum_gradients,
feed_dict={
self.local_network.s: batch_si,
self.local_network.a: batch_a,
self.local_network.t: batch_t,
self.local_network.td: batch_td,
self.local_network.r: batch_R,
self.local_network.return_usf: batch_usf_R,
self.local_network.initial_lstm_state: start_lstm_state,
self.local_network.step_size: [len(batch_a)],
})
cur_learning_rate = self._anneal_learning_rate(global_t)
sess.run(self.apply_gradients,
feed_dict={self.learning_rate_input: cur_learning_rate})
if VERBOSE and (self.thread_index == 0) and (self.local_t % 100) == 0:
sys.stdout.write("Local timestep %d\n" % self.local_t)
# return advanced local step size
diff_local_t = self.local_t - start_local_t
return diff_local_t
class A3C_Thread(object):
def __init__(self,
thread_index,
global_network,
initial_learning_rate,
learning_rate_input,
grad_applier,
max_global_time_step,
device,
network_scope="network",
scene_scope="scene",
task_scope="task"):
self.thread_index = thread_index # Number the thread
self._set_local_network(device, network_scope, scene_scope,
task_scope) # Set local network
self.sync = self.local_network.sync_from(
global_network) # Synthesize from the global network
self.learning_rate_input = learning_rate_input # Set learning rate
self.max_global_time_step = max_global_time_step # Set maximum of global time step
self._set_trainer_optimizer(device, global_network,
grad_applier) # Set trainer
self._set_environment(initial_learning_rate) # Set environment
# Create local network
def _set_local_network(self, device, network_scope, scene_scope,
task_scope):
self.local_network = DRLNetwork(action_size=ACTION_SIZE,
device=device,
network_scope=network_scope,
scene_scopes=[scene_scope])
self.network_scope = network_scope
self.scene_scope = scene_scope
self.task_scope = task_scope
self.scopes = [network_scope, scene_scope, task_scope]
self.local_network.prepare_loss(ENTROPY_BETA, self.scopes)
# Set trainer and optimizer
# Set Actor-Critic gradient and optimizer
# Use the accumulated trainer from Zhu
def _set_trainer_optimizer(self, device, global_network, grad_applier):
self.trainer = AccumTrainer(device)
self.trainer.prepare_minimize(self.local_network.total_loss,
self.local_network.get_vars())
self.accum_gradients = self.trainer.accumulate_gradients()
self.reset_gradients = self.trainer.reset_gradients()
accum_grad_names = [
self._local_var_name(x)
for x in self.trainer.get_accum_grad_list()
]
global_net_vars = [
x for x in global_network.get_vars()
if self._get_accum_grad_name(x) in accum_grad_names
]
self.apply_gradients = grad_applier.apply_gradients(
global_net_vars, self.trainer.get_accum_grad_list())
def _local_var_name(self, var):
return '/'.join(var.name.split('/')[1:])
def _get_accum_grad_name(self, var):
return self._local_var_name(var).replace(':', '_') + '_accum_grad:0'
# Set environments
def _set_environment(self, initial_learning_rate):
self.episode_max_q = -np.inf
self.env = None
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
self.episode_reward = 0
self.episode_length = 0
# Choose one action according to the pi values
def choose_action(self, pi_values):
action = np.random.choice(np.arange(len(pi_values)), p=pi_values)
return action
# Take LOCAL_T_MAX step in one process
# And update the accumulated gradients
def process(self, sess, global_t, summary_writer, summary_op,
summary_placeholders):
if self.env is None:
# lazy evaluation
time.sleep(self.thread_index * 1.0)
self.env = Environment({
'scene_name': self.scene_scope,
'terminal_state_id': int(self.task_scope)
})
start_local_t = self.local_t
# Initialization
states = []
actions = []
rewards = []
values = []
targets = []
terminal_end = False
# Reset accmulated gradient variables
sess.run(self.reset_gradients)
# Obtain shared parameters from global
sess.run(self.sync)
# t_max times loop
for i in range(LOCAL_T_MAX):
pi_, value_ = self.local_network.run_policy_and_value(
sess, self.env.s_t, self.env.target, self.scopes)
pi_ = np.array(pi_) / np.sum(pi_)
action = self.choose_action(pi_)
states.append(self.env.s_t)
actions.append(action)
values.append(value_)
targets.append(self.env.target)
if VERBOSE and (self.thread_index
== 0) and (self.local_t % 1000) == 0:
sys.stdout.write("%s:" % self.scene_scope)
sys.stdout.write("Pi = {0} V = {1}\n".format(pi_, value_))
# process game
self.env.step(action)
# receive game result
reward = self.env.reward
terminal = self.env.terminal
# ad-hoc reward for navigation
# reward = 10.0 if terminal else -0.01
if self.episode_length > 5e3: terminal = True
self.episode_reward += reward
self.episode_length += 1
self.episode_max_q = max(self.episode_max_q, np.max(value_))
# clip reward
rewards.append(np.clip(reward, -1, 1))
self.local_t += 1
if terminal:
terminal_end = True
sys.stdout.write(
"#Thread: %d \n time %d | thread #%d | scene %s | target #%s\n%s %s episode reward = %.3f\n%s %s episode length = %d\n%s %s episode max Q = %.3f\n"
% (self.thread_index, global_t, self.thread_index,
self.scene_scope, self.task_scope, self.scene_scope,
self.task_scope, self.episode_reward, self.scene_scope,
self.task_scope, self.episode_length, self.scene_scope,
self.task_scope, self.episode_max_q))
summary_values = {
"episode_reward_input": self.episode_reward,
"episode_length_input": float(self.episode_length),
"episode_max_q_input": self.episode_max_q,
"learning_rate_input": self._anneal_learning_rate(global_t)
}
self._record_score(sess, summary_writer, summary_op,
summary_placeholders, summary_values,
global_t)
self.episode_reward = 0
self.episode_length = 0
self.episode_max_q = -np.inf
self.env.reset()
break
R = 0.0
if not terminal_end:
R = self.local_network.run_value(sess, self.env.s_t,
self.env.target, self.scopes)
actions.reverse()
states.reverse()
rewards.reverse()
values.reverse()
batch_si = []
batch_a = []
batch_td = []
batch_R = []
batch_t = []
# compute and accmulate gradients
for (ai, ri, si, Vi, ti) in zip(actions, rewards, states, values,
targets):
R = ri + GAMMA * R
td = R - Vi
a = np.zeros([ACTION_SIZE])
a[ai] = 1
batch_si.append(si)
batch_a.append(a)
batch_td.append(td)
batch_R.append(R)
batch_t.append(ti)
sess.run(self.accum_gradients,
feed_dict={
self.local_network.s: batch_si,
self.local_network.a: batch_a,
self.local_network.t: batch_t,
self.local_network.td: batch_td,
self.local_network.r: batch_R
})
cur_learning_rate = self._anneal_learning_rate(global_t)
sess.run(self.apply_gradients,
feed_dict={self.learning_rate_input: cur_learning_rate})
if VERBOSE and (self.thread_index == 0) and (self.local_t % 100) == 0:
sys.stdout.write(
"#Thread-%d-%s-Local timestep-%d\n" %
(self.thread_index, self.scene_scope, self.local_t))
# return advanced local step size
diff_local_t = self.local_t - start_local_t
return diff_local_t
def _record_score(self, sess, writer, summary_op, placeholders, values,
global_t):
feed_dict = {}
for k in placeholders:
feed_dict[placeholders[k]] = values[k]
summary_str = sess.run(summary_op, feed_dict=feed_dict)
if VERBOSE:
sys.stdout.write('writing to summary writer at time %d\n' %
(global_t))
writer.add_summary(summary_str, global_t)
# writer.flush()
def _anneal_learning_rate(self, global_time_step):
time_step_to_go = max(self.max_global_time_step - global_time_step,
0.0)
learning_rate = self.initial_learning_rate * time_step_to_go / self.max_global_time_step
return learning_rate
class A3CTrainingThread(object):
def __init__(self,
thread_index,
global_network,
initial_learning_rate,
learning_rate_input,
grad_applier,
max_global_time_step,
device,
network_scope="network",
scene_scope="scene",
training_scene="scene",
task_scope="task",
checkpoint_scope="checkpoint"):
self.thread_index = thread_index
self.learning_rate_input = learning_rate_input
self.max_global_time_step = max_global_time_step
self.network_scope = network_scope
self.scene_scope = scene_scope
self.training_scene = training_scene
self.task_scope = task_scope
self.checkpoint_scope = checkpoint_scope
self.scopes = [network_scope, scene_scope, task_scope]
self.local_network = ActorCriticFFNetwork(action_size=ACTION_SIZE,
device=device,
network_scope=network_scope,
scene_scopes=[scene_scope])
self.local_network.prepare_loss(ENTROPY_BETA, self.scopes)
self.trainer = AccumTrainer(device)
self.trainer.prepare_minimize(self.local_network.total_loss,
self.local_network.get_vars())
self.accum_gradients = self.trainer.accumulate_gradients()
self.reset_gradients = self.trainer.reset_gradients()
accum_grad_names = [
self._local_var_name(x)
for x in self.trainer.get_accum_grad_list()
]
global_net_vars = [
x for x in global_network.get_vars()
if self._get_accum_grad_name(x) in accum_grad_names
]
self.apply_gradients = grad_applier.apply_gradients(
global_net_vars, self.trainer.get_accum_grad_list())
self.sync = self.local_network.sync_from(global_network)
self.env = None
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
self.episode_reward = 0
self.episode_length = 0
self.episode_max_q = -np.inf
def _local_var_name(self, var):
return '/'.join(var.name.split('/')[1:])
def _get_accum_grad_name(self, var):
return self._local_var_name(var).replace(':', '_') + '_accum_grad:0'
def _anneal_learning_rate(self, global_time_step):
time_step_to_go = max(self.max_global_time_step - global_time_step,
0.0)
learning_rate = self.initial_learning_rate * time_step_to_go / self.max_global_time_step
return learning_rate
def choose_action(self, pi_values):
values = []
sum = 0.0
for rate in pi_values:
sum = sum + rate
value = sum
values.append(value)
r = random.random() * sum
for i in range(len(values)):
if values[i] >= r:
return i
# fail safe
return len(values) - 1
def _record_score(self, sess, writer, summary_op, placeholders, values,
global_t):
feed_dict = {}
for k in placeholders:
feed_dict[placeholders[k]] = values[k]
summary_str = sess.run(summary_op, feed_dict=feed_dict)
if VERBOSE:
sys.stdout.write('writing to summary writer at time %d\n' %
(global_t))
writer.add_summary(summary_str, global_t)
# writer.flush()
def process(self, sess, global_t, summary_writer, summary_op,
summary_placeholders):
if self.env is None:
# lazy evaluation
time.sleep(self.thread_index * 1.0)
self.env = Environment({
'scene_name': self.training_scene,
'terminal_state_id': self.task_scope,
'checkpoint_state_id': self.checkpoint_scope
})
self.env.reset()
states = []
actions = []
rewards = []
values = []
targets = []
checkpoints = []
positions = []
auxilaries = []
auxilaries_cl = []
aclists = []
colists = []
isCheckpointed = []
collision = []
terminal_end = False
# reset accumulated gradients
sess.run(self.reset_gradients)
# copy weights from shared to local
sess.run(self.sync)
start_local_t = self.local_t
# t_max times loop
for i in range(LOCAL_T_MAX):
pi_, value_ = self.local_network.run_policy_and_value(
sess, self.env.s_t, self.env.s_position, self.env.checkpoint,
self.env.s_a_t, self.env.s_c_t, self.env.isCheckpoint,
self.env.s_aux_cl, self.scopes)
action = self.choose_action(pi_)
states.append(self.env.s_t)
actions.append(action)
values.append(value_)
targets.append(self.env.target)
checkpoints.append(self.env.checkpoint)
positions.append(self.env.s_position)
aclists.append(self.env.s_a_t)
colists.append(self.env.s_c_t)
collision.append(self.env.s_aux_cl)
isCheckpointed.append(int(self.env.isCheckpoint))
if VERBOSE and (self.thread_index
== 0) and (self.local_t % 1000) == 0:
sys.stdout.write("Pi = {0} V = {1}\n".format(pi_, value_))
# process game
self.env.step(action)
# receive game result
reward = self.env.reward
terminal = self.env.terminal
#if self.env.isCheckpoint:
# sys.stdout.write("CHECKPOINT \n")
if self.episode_length > 5e3: terminal = True
self.episode_reward += reward
self.episode_length += 1
self.episode_max_q = max(self.episode_max_q, np.max(value_))
# clip reward
rewards.append(np.clip(reward, -1, 1))
self.local_t += 1
# s_t1 -> s_t
self.env.update()
auxilaries.append(self.env.s_aux)
auxilaries_cl.append(self.env.s_aux_cl)
if terminal:
terminal_end = True
sys.stdout.write(
"time %d | thread #%d | scene %s | target #%s\n%s %s episode reward = %.3f\n%s %s episode length = %d\n%s %s episode max Q = %.3f\n"
% (global_t, self.thread_index, self.scene_scope,
self.task_scope, self.scene_scope, self.task_scope,
self.episode_reward, self.scene_scope, self.task_scope,
self.episode_length, self.scene_scope, self.task_scope,
self.episode_max_q))
summary_values = {
"episode_reward_input": self.episode_reward,
"episode_length_input": float(self.episode_length),
"episode_max_q_input": self.episode_max_q,
"learning_rate_input": self._anneal_learning_rate(global_t)
}
self._record_score(sess, summary_writer, summary_op,
summary_placeholders, summary_values,
global_t)
self.episode_reward = 0
self.episode_length = 0
self.episode_max_q = -np.inf
self.env.reset()
break
R = 0.0
if not terminal_end:
R = self.local_network.run_value(sess, self.env.s_t,
self.env.s_position,
self.env.checkpoint,
self.env.s_a_t, self.env.s_c_t,
self.env.isCheckpoint,
self.env.s_aux_cl, self.scopes)
actions.reverse()
states.reverse()
rewards.reverse()
values.reverse()
positions.reverse()
auxilaries.reverse()
auxilaries_cl.reverse()
aclists.reverse()
colists.reverse()
isCheckpointed.reverse()
collision.reverse()
batch_si = []
batch_a = []
batch_td = []
batch_R = []
batch_t = []
batch_c = []
batch_p = []
batch_aux = []
batch_aux_cl = []
batch_al = []
batch_cl = []
batch_ic = []
batch_collision = []
# compute and accmulate gradients
for (ai, ri, si, Vi, ti, ci, pi, auxi, aux_cl_i, ali, cli, ici,
coli) in zip(actions, rewards, states, values, targets,
checkpoints, positions, auxilaries, auxilaries_cl,
aclists, colists, isCheckpointed, collision):
R = ri + GAMMA * R
td = R - Vi
a = np.zeros([ACTION_SIZE])
a[ai] = 1
batch_si.append(si)
batch_a.append(a)
batch_td.append(td)
batch_R.append(R)
batch_t.append(ti)
batch_c.append(ci)
batch_p.append(pi)
batch_aux.append(auxi)
batch_aux_cl.append(aux_cl_i)
batch_al.append(ali)
batch_cl.append(cli)
batch_ic.append(ici)
batch_collision.append(coli)
sess.run(self.accum_gradients,
feed_dict={
self.local_network.s: batch_si,
self.local_network.a: batch_a,
self.local_network.t: batch_p,
self.local_network.c: batch_c,
self.local_network.td: batch_td,
self.local_network.aux: batch_aux,
self.local_network.aux_cl: batch_aux_cl,
self.local_network.al: batch_al,
self.local_network.cl: batch_cl,
self.local_network.ic: batch_ic,
self.local_network.col: batch_collision,
self.local_network.r: batch_R
})
cur_learning_rate = self._anneal_learning_rate(global_t)
sess.run(self.apply_gradients,
feed_dict={self.learning_rate_input: cur_learning_rate})
if VERBOSE and (self.thread_index == 0) and (self.local_t % 100) == 0:
sys.stdout.write("Local timestep %d\n" % self.local_t)
# return advanced local step size
diff_local_t = self.local_t - start_local_t
return diff_local_t, terminal
class A3CTrainingThread(object):
def __init__(self,
thread_index,
global_network,
initial_learning_rate,
learning_rate_input,
grad_applier,
max_global_time_step,
device,
network_scope="network",
scene_scope="scene",
task_scope="task"):
self.thread_index = thread_index
self.learning_rate_input = learning_rate_input
self.max_global_time_step = max_global_time_step
self.network_scope = network_scope
self.scene_scope = scene_scope
self.task_scope = task_scope
self.scopes = [network_scope, scene_scope, task_scope]
if USE_LSTM:
self.local_network = ActorCriticLSTMNetwork(
action_size=ACTION_SIZE,
device=device,
network_scope=network_scope,
scene_scopes=[scene_scope])
else:
self.local_network = ActorCriticFFNetwork(
action_size=ACTION_SIZE,
device=device,
network_scope=network_scope,
scene_scopes=[scene_scope])
self.local_network.prepare_loss(ENTROPY_BETA, self.scopes)
self.trainer = AccumTrainer(device)
self.trainer.prepare_minimize(self.local_network.total_loss,
self.local_network.get_vars())
self.accum_gradients = self.trainer.accumulate_gradients()
self.reset_gradients = self.trainer.reset_gradients()
accum_grad_names = [
self._local_var_name(x)
for x in self.trainer.get_accum_grad_list()
]
global_net_vars = [
x for x in global_network.get_vars()
if self._get_accum_grad_name(x) in accum_grad_names
]
self.apply_gradients = grad_applier.apply_gradients(
global_net_vars, self.trainer.get_accum_grad_list())
self.sync = self.local_network.sync_from(global_network)
self.env = None
self.obs = None
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
self.episode_reward = 0
self.episode_length = 0
self.episode_max_q = -np.inf
self.entropy = np.zeros(20)
def _local_var_name(self, var):
return '/'.join(var.name.split('/')[1:])
def _get_accum_grad_name(self, var):
return self._local_var_name(var).replace(':', '_') + '_accum_grad:0'
def _anneal_learning_rate(self, global_time_step):
time_step_to_go = max(self.max_global_time_step - global_time_step,
0.0)
learning_rate = self.initial_learning_rate * time_step_to_go / self.max_global_time_step
return learning_rate
def choose_action(self, pi_values):
values = []
sum = 0.0
for rate in pi_values:
sum = sum + rate
value = sum
values.append(value)
r = random.random() * sum
for i in range(len(values)):
if values[i] >= r:
return i
# fail safe
return len(values) - 1
def _record_score(self, sess, writer, summary_op, placeholders, values,
global_t):
feed_dict = {}
for k in placeholders:
feed_dict[placeholders[k]] = values[k]
summary_str = sess.run(summary_op, feed_dict=feed_dict)
if VERBOSE:
sys.stdout.write('writing to summary writer at time %d\n' %
(global_t))
writer.add_summary(summary_str, global_t)
# writer.flush()
def process(self, sess, global_t, summary_writer, summary_op,
summary_placeholders):
if self.env is None:
# lazy evaluation
time.sleep(self.thread_index * 1.0)
self.env = gym.make('Pong-v0')
self.obs = self.env.reset()
states = []
actions = []
rewards = []
values = []
terminal_end = False
# reset accumulated gradients
sess.run(self.reset_gradients)
# copy weights from shared to local
sess.run(self.sync)
start_local_t = self.local_t
if USE_LSTM:
start_lstm_state = self.local_network.lstm_state_out
# t_max times loop
for i in range(LOCAL_T_MAX):
pi_, value_ = self.local_network.run_policy_and_value(
sess, self.obs, self.scopes)
action = self.choose_action(pi_)
states.append(self.obs)
actions.append(action)
values.append(value_)
if VERBOSE and (self.thread_index
== 0) and (self.local_t % 1000) == 0:
sys.stdout.write("Pi = {0} V = {1}\n".format(pi_, value_))
# process game
self.obs, reward, terminal, info = self.env.step(action)
# ad-hoc reward for navigation
# reward = 10.0 if terminal else -0.01
# if self.episode_length > 5e3: terminal = True
self.episode_reward += reward
self.episode_length += 1
self.episode_max_q = max(self.episode_max_q, np.max(value_))
# clip reward
rewards.append(np.clip(reward, -1, 1))
self.local_t += 1
# s_t1 -> s_t
# self.env.update()
if terminal:
terminal_end = True
sys.stdout.write(
"time %d | thread #%d | scene %s \n"
"%s %s episode reward = %.3f\n"
"%s %s episode length = %d\n"
"%s %s episode max Q = %.3f\n" %
(global_t, self.thread_index, self.scene_scope,
self.scene_scope, self.task_scope, self.episode_reward,
self.scene_scope, self.task_scope, self.episode_length,
self.scene_scope, self.task_scope, self.episode_max_q))
summary_values = {
"episode_reward_input": self.episode_reward,
"episode_length_input": float(self.episode_length),
"episode_max_q_input": self.episode_max_q,
"learning_rate_input":
self._anneal_learning_rate(global_t),
"episode_entropy": self.entropy[
0] # self.entropy here is a np.array([1, 20]) with same
# value for each element, don't know why
}
self._record_score(sess, summary_writer, summary_op,
summary_placeholders, summary_values,
global_t)
self.episode_reward = 0
self.episode_length = 0
self.episode_max_q = -np.inf
self.obs = self.env.reset()
if USE_LSTM:
self.local_network.reset_state()
break
R = 0.0
if not terminal_end:
R = self.local_network.run_value(sess, self.obs, self.scopes)
actions.reverse()
states.reverse()
rewards.reverse()
values.reverse()
batch_si = []
batch_a = []
batch_td = []
batch_R = []
# compute and accmulate gradients
for (ai, ri, si, Vi) in zip(actions, rewards, states, values):
R = ri + GAMMA * R
td = R - Vi
a = np.zeros([ACTION_SIZE])
a[ai] = 1
batch_si.append(si)
batch_a.append(a)
batch_td.append(td)
batch_R.append(R)
cur_learning_rate = self._anneal_learning_rate(global_t)
if USE_LSTM:
batch_si.reverse()
batch_a.reverse()
batch_td.reverse()
batch_R.reverse()
_, self.entropy = sess.run(
[self.apply_gradients, self.local_network.entropy],
feed_dict={
self.local_network.s: batch_si,
self.local_network.a: batch_a,
self.local_network.td: batch_td,
self.local_network.r: batch_R,
self.local_network.initial_lstm_state: start_lstm_state,
self.local_network.step_size: [len(batch_a)],
self.learning_rate_input: cur_learning_rate
})
# _, self.entropy = sess.run([self.accum_gradients, self.local_network.entropy],
# feed_dict={
# self.local_network.s: batch_si,
# self.local_network.a: batch_a,
# self.local_network.td: batch_td,
# self.local_network.r: batch_R,
# self.local_network.step_size: [len(batch_a)]
# })
else:
_, self.entropy = sess.run(
[self.accum_gradients, self.local_network.entropy],
feed_dict={
self.local_network.s: batch_si,
self.local_network.a: batch_a,
self.local_network.td: batch_td,
self.local_network.r: batch_R
})
sess.run(self.apply_gradients,
feed_dict={self.learning_rate_input: cur_learning_rate})
if VERBOSE and (self.thread_index == 0) and (self.local_t % 100) == 0:
sys.stdout.write("Local timestep %d\n" % self.local_t)
# return advanced local step size
diff_local_t = self.local_t - start_local_t
return diff_local_t
class A3CTrainingThread(object):
def __init__(self,
thread_index,
global_network,
global_discriminator,
initial_learning_rate,
learning_rate_input,
grad_applier,
grad_applier_discriminator,
max_global_time_step,
device,
device2,
network_scope="network",
scene_scope="scene",
task_scope="task"):
self.thread_index = thread_index
self.learning_rate_input = learning_rate_input
self.max_global_time_step = max_global_time_step
self.network_scope = network_scope
self.network_scope_D = network_scope + "_d"
self.scene_scope = scene_scope
self.task_scope = task_scope
self.scopes = [network_scope, scene_scope, task_scope]
self.scopes_d = [self.network_scope_D, task_scope]
self.local_network = ActorCriticFFNetwork(action_size=ACTION_SIZE,
device=device,
network_scope=network_scope,
scene_scopes=[scene_scope])
self.local_network.prepare_loss(ENTROPY_BETA, self.scopes)
self.trainer = AccumTrainer(device)
self.trainer.prepare_minimize(
self.local_network.
total_loss, #getting the gradients of for the local network variablkes
self.local_network.get_vars())
#This part is for the newly added PPO loss (we need to keep old and new update parameters)
new_variable_list = self.local_network.get_vars()
old_varaible_list = self.local_network.get_vars_old()
#For the ppo loss begining of the each iteration we need to sync old with current
self.old_new_sync = self.local_network.sync_curre_old()
self.accum_gradients = self.trainer.accumulate_gradients(
) #This is to assign gradients
self.reset_gradients = self.trainer.reset_gradients(
) #after applying the grads to variables we need to resent those variables
accum_grad_names = [
self._local_var_name(x)
for x in self.trainer.get_accum_grad_list()
] #get the name list of all the grad vars
global_net_vars = [
x for x in global_network.get_vars()
if self._get_accum_grad_name(x) in accum_grad_names
] #check whether the global_network vars are mentioned in gradiet computations for them
local_net_vars = [
x for x in self.local_network.get_vars()
if self._get_accum_grad_name(x) in accum_grad_names
]
self.apply_gradients = grad_applier.apply_gradients(
global_net_vars, self.trainer.get_accum_grad_list())
self.apply_gradients_local = grad_applier.apply_gradients_local_net(
local_net_vars, self.trainer.get_accum_grad_list())
#If this is unstable it is desireable to first apply the gradients on the local network and then clip and after that we apply
self.sync = self.local_network.sync_from(
global_network
) #this is to sync from the glocal network Apply updated global params to the local network
#This part is for the Discriminator
#########################################################################################
#
self.local_discriminator = Discriminator_WGAN( #
action_size=ACTION_SIZE, #
device=device, #
network_scope=network_scope, #
scene_scopes=[scene_scope]) #
#
self.local_discriminator.prepare_loss_D(ENTROPY_BETA, self.scopes_d) #
#
self.trainer_D = AccumTrainer_d(device=device,
name="AccumTrainer_d") #
#
self.trainer_D.prepare_minimize(
self.local_discriminator.total_loss_d, #
self.local_discriminator.get_vars()) #
#
#
self.accum_gradients_d = self.trainer_D.accumulate_gradients() #
self.reset_gradients_d = self.trainer_D.reset_gradients()
#
#
accum_grad_names_discrimi = [
self._local_var_name(x)
for x in self.trainer_D.get_accum_grad_list()
]
#
#
global_discri_vars = [
x for x in global_discriminator.get_vars()
if self._get_accum_grad_name(x) in accum_grad_names_discrimi
]
local_discri_vars = [
x for x in self.local_discriminator.get_vars()
if self._get_accum_grad_name(x) in accum_grad_names_discrimi
]
#
self.apply_gradients_discriminator = grad_applier_discriminator.apply_gradients(
local_discri_vars, self.trainer_D.get_accum_grad_list()
) #applying grad to the LOCAL network
#
self.clip_local_d_weights = self.local_discriminator.clip_weights(
) #here we are clipping the global net weights directly.
#
self.sync_discriminator_l_G = self.local_discriminator.sync_to(
global_discriminator) #
self.sync_discriminator_G_l = self.local_discriminator.sync_from(
global_discriminator)
#
self.D_var_G = global_discriminator.get_vars()
self.D_var_l = self.local_discriminator.get_vars() #
#
#
#########################################################################################
self.env = None
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
self.episode_reward = 0
self.episode_length = 0
self.episode_max_q = -np.inf
def _local_var_name(self, var):
return '/'.join(var.name.split('/')[1:])
def _get_accum_grad_name(self, var):
return self._local_var_name(var).replace(':', '_') + '_accum_grad:0'
def _anneal_learning_rate(self, global_time_step):
time_step_to_go = max(self.max_global_time_step - global_time_step,
0.0)
learning_rate = self.initial_learning_rate * time_step_to_go / self.max_global_time_step
return learning_rate
def choose_action(self, pi_values):
values = []
sum = 0.0
for rate in pi_values:
sum = sum + rate
value = sum
values.append(value)
r = random.random() * sum
for i in range(len(values)):
if values[i] >= r:
return i
# fail safe
return len(values) - 1
def _record_score(self, sess, writer, summary_op, placeholders, values,
global_t):
feed_dict = {}
for k in placeholders:
feed_dict[placeholders[k]] = values[k]
summary_str = sess.run(summary_op, feed_dict=feed_dict)
if VERBOSE:
print('writing to summary writer at time %d\n' % (global_t))
writer.add_summary(summary_str, global_t)
# writer.flush()
def process(self, sess, global_t, summary_writer, summary_op,
summary_placeholders):
if self.env is None:
# lazy evaluation
time.sleep(self.thread_index * 1.0)
self.env = Environment({
'scene_name': self.scene_scope,
'terminal_state_id': int(self.task_scope)
})
self.env.reset() #resetting the environment for each thread
self.env_Oracle = Environment(
{ #Every iteration in the thread the expert start with the current state of the agent
'scene_name': self.scene_scope,
'terminal_state_id': int(self.task_scope),
'initial_state': self.env.current_state_id
})
self.env_Oracle.reset()
states = [] #to keeep state ,actions ,targets and other stae
actions = []
rewards = []
values = []
targets = []
dones = []
states_oracle = []
actions_oracle = []
targets_oracle = []
terminal_end = False #in the start terminal state_end is false
sess.run(
self.reset_gradients
) #resetting the gradient positions when starting the process for each Iteration
sess.run(self.sync) # copy weights from shared to local
#dicriminator sync
##########################
sess.run(self.sync_discriminator_G_l
) #Copy the weights from the sharead to the local
sess.run(self.reset_gradients_d
) #resetting the gradients of the discriminator slosts
########################
start_local_t = self.local_t
self.oracle = ShortestPathOracle(self.env_Oracle, ACTION_SIZE)
#########################################################################################
#Sampling the Expert Trajectories
for i in range(100):
#We might need to use an for loop to finish the expert trajectory first
oracle_pi = self.oracle.run_policy(
self.env_Oracle.current_state_id
) #get the policy of the oracle which means the shotest path kind of action in the given state
oracle_action = self.choose_action(oracle_pi)
states_oracle.append(self.env_Oracle.s_t)
actions_oracle.append(oracle_action)
targets_oracle.append(self.env_Oracle.target)
self.env_Oracle.step(oracle_action)
terminal_o = self.env_Oracle.terminal
self.env_Oracle.update()
if terminal_o:
break
##############################################################################################
# t_max times loop
for i in range(
LOCAL_T_MAX
): #one thread will run for maximum amoound to 5 iterations then do a gradiet uodate
pi_, value_ = self.local_network.run_policy_and_value(
sess, self.env.s_t, self.env.target, self.scopes)
action = self.choose_action(pi_)
states.append(self.env.s_t)
actions.append(action)
values.append(value_)
targets.append(self.env.target)
if VERBOSE and (self.thread_index
== 0) and (self.local_t % 1000) == 0:
sys.stdout.write("Pi = {0} V = {1}\n".format(pi_, value_))
# process game
self.env.step(action)
# receive game result
reward = self.env.reward #getting the reward from the env
terminal = self.env.terminal #geting whether the agent went to a terminal state
# ad-hoc reward for navigation
reward = 10.0 if terminal else -0.01 #this is the normal reward here 10 if terminal all the others it is -0.01 (ollision donesst take in to the accout)
if self.episode_length > 5e3:
terminal = True #Here we do not let agent to run more that 5000 steps so we make it terminal
#but the above terminal thing has no effect on giving 10 as the rwaerd because we set the rweard above
self.episode_reward += reward
self.episode_length += 1
#this is what is the maximum value got in the episode
self.episode_max_q = max(self.episode_max_q, np.max(
value_)) #self.episode_max_q-This is -inf in the beggining
# clip reward
rewards.append(
np.clip(reward, -1, 1)
) #make sure the rewartds is between -1 and +1 even thore rtthere is a 10
self.local_t += 1
# s_t1 -> s_t
self.env.update()
if terminal: #if we go to the terminal state we will surely break the function
score = self.local_discriminator.run_critic(
sess, states, targets, actions, self.scopes_d)
sys.stdout.write("Critic_Score = {0}".format(score))
terminal_end = True
sys.stdout.write(
"time %d | thread #%d | scene %s | target #%s\n%s %s episode reward = %.3f\n%s %s episode length = %d\n%s %s episode max Q = %.3f\n"
% (global_t, self.thread_index, self.scene_scope,
self.task_scope, self.scene_scope, self.task_scope,
self.episode_reward, self.scene_scope, self.task_scope,
self.episode_length, self.scene_scope, self.task_scope,
self.episode_max_q))
summary_values = {
"episode_reward_input": self.episode_reward,
"episode_length_input": float(self.episode_length),
"episode_max_q_input": self.episode_max_q,
"learning_rate_input": self._anneal_learning_rate(global_t)
}
self._record_score(sess, summary_writer, summary_op,
summary_placeholders, summary_values,
global_t)
self.episode_reward = 0 #after terminal state we gonna make all these variables to zero
self.episode_length = 0 #Now the AI need to start from new position
self.episode_max_q = -np.inf #after a terminaltion we do this
self.env.reset()
break
R = 0.0 #In the terminal Return is nothing #If it's terminal end we do not have a return from the final state
if not terminal_end: #But if it's not the turminal Return is the next value function
R = self.local_network.run_value(sess, self.env.s_t,
self.env.target, self.scopes)
#Agent's Samples
actions.reverse()
states.reverse()
rewards.reverse()
values.reverse()
#Expert's Samples
states_oracle.reverse()
actions_oracle.reverse()
actions_oracle.reverse()
#Agent's batch
batch_si = []
batch_a = []
batch_actions = []
batch_td = []
batch_R = []
batch_t = []
#Expert's Batch
batch_si_ex = []
batch_a_ex = []
batch_t_ex = []
batch_si_d = []
batch_t_d = []
batch_actions_d = []
#This is for the
for (s_e, a_e, t_e) in zip(states_oracle, actions_oracle,
targets_oracle):
batch_si_ex.append(s_e)
batch_a_ex.append(a_e)
batch_t_ex.append(t_e)
for (ai, si, ti) in zip(actions, states, targets):
batch_actions_d.append(ai)
batch_si_d.append(si)
batch_t_d.append(ti)
cur_learning_rate = self._anneal_learning_rate(global_t)
for i in range(10):
#sess.run(self.reset_gradients_d)
sess.run(
self.
accum_gradients_d, #since we update the algorithm for given action ,given state, given advatns and given value and given reward we do not care about the sequence
feed_dict={
self.local_discriminator.s_e: batch_si_ex,
self.local_discriminator.Actions_e: batch_a_ex,
self.local_discriminator.s_a: batch_si_d,
self.local_discriminator.Actions_a: batch_actions_d,
self.local_discriminator.t_e: batch_t_ex,
self.local_discriminator.t_a: batch_t_d
})
sess.run(
self.
apply_gradients_discriminator, #directly gradients get apply on the global discri
feed_dict={self.learning_rate_input: 0.00005})
loss = sess.run(self.local_discriminator.total_loss_d,
feed_dict={
self.local_discriminator.s_e: batch_si_ex,
self.local_discriminator.Actions_e: batch_a_ex,
self.local_discriminator.s_a: batch_si_d,
self.local_discriminator.Actions_a:
batch_actions_d,
self.local_discriminator.t_e: batch_t_ex,
self.local_discriminator.t_a: batch_t_d
})
sess.run(self.clip_local_d_weights
) #every update make sure u clip weihtfs
critic_r = self.local_discriminator.run_critic(sess, batch_si_d,
batch_t_d,
batch_actions_d,
self.scopes_d)
critic_r = critic_r * 0.1
rewards = rewards + critic_r #We concatenate the rewrds function
# Compute the advantage function , return and stack them as batches in Agent
for (ai, ri, si, Vi, ti) in zip(actions, rewards, states, values,
targets):
R = ri + GAMMA * R #calculatung the adcantage function
td = R - Vi
a = np.zeros([ACTION_SIZE])
a[ai] = 1 #making the actions one hot
batch_actions.append(ai)
batch_si.append(si)
batch_a.append(a)
batch_td.append(td)
batch_R.append(R)
batch_t.append(ti)
#syncying the new paramters to the old network in the thread PPO
sess.run(self.old_new_sync)
for i in range(4):
#sess.run(self.reset_gradients) #reset the gradients
sess.run(
self.
accum_gradients, #since we update the algorithm for given action ,given state, given advatns and given value and given reward we do not care about the sequence
feed_dict={
self.local_network.s: batch_si,
self.local_network.a: batch_a,
self.local_network.t: batch_t,
self.local_network.td: batch_td,
self.local_network.r: batch_R,
})
sess.run(
self.
apply_gradients_local, #apply the gradients to the local networ
feed_dict={self.learning_rate_input: cur_learning_rate})
#theoritcally we can have one accume gradient operation here
sess.run(self.apply_gradients,
feed_dict={self.learning_rate_input: cur_learning_rate})
sess.run(
self.sync_discriminator_l_G
) #syncing the paramters from the local network to the global newok
if VERBOSE and (self.thread_index == 0) and (self.local_t % 100) == 0:
sys.stdout.write("Local timestep %d\n" % self.local_t)
# return advanced local step size
diff_local_t = self.local_t - start_local_t
return diff_local_t
class ADQN_Thread(object):
def __init__(self, thread_index, global_network, initial_learning_rate,
learning_rate_input, grad_applier, max_global_time_step,
device, network_scope="network", scene_scope="scene",
task_scope="task"):
self.thread_index = thread_index # Number the thread
self._set_local_network(device, network_scope, scene_scope, task_scope) # Set local network
self.sync = self.local_network.sync_from(global_network) # Synthesize from the global network
self.learning_rate_input = learning_rate_input # Set learning rate
self.max_global_time_step = max_global_time_step # Set maximum of global time step
self._set_trainer_optimizer(device, global_network, grad_applier) # Set trainer
self._set_environment(initial_learning_rate) # Set environment
self.memory_size = MEMORY_SIZE # memory size for replay buffer
self.memory = np.zeros((self.memory_size, 2048 * 4 * 2 + 2)) # initialize zero memory [s, a, r, s_]
self.replace_target_iter = DQN_REPLACE_TARGET_ITER
self.batch_size = DQN_BATCH_SIZE
self.gamma = REWARD_DECAY
# Create local network
def _set_local_network(self, device, network_scope, scene_scope, task_scope):
self.local_network = DRLNetwork(action_size=ACTION_SIZE, device=device, network_scope=network_scope,
scene_scopes=[scene_scope])
self.network_scope = network_scope
self.scene_scope = scene_scope
self.task_scope = task_scope
self.scopes = [network_scope, scene_scope, task_scope]
self.local_network.prepare_loss(ENTROPY_BETA, self.scopes)
# Set trainer and optimizer
# Set Actor-Critic gradient and optimizer
# Use the accumulated trainer from Zhu
def _set_trainer_optimizer(self, device, global_network, grad_applier):
self.trainer = AccumTrainer(device)
self.trainer.prepare_minimize(self.local_network.total_loss,
self.local_network.get_vars())
self.accum_gradients = self.trainer.accumulate_gradients()
self.reset_gradients = self.trainer.reset_gradients()
accum_grad_names = [self._local_var_name(x) for x in self.trainer.get_accum_grad_list()]
global_net_vars = [x for x in global_network.get_vars() if self._get_accum_grad_name(x) in accum_grad_names]
self.apply_gradients = grad_applier.apply_gradients(global_net_vars, self.trainer.get_accum_grad_list() )
def _local_var_name(self, var):
return '/'.join(var.name.split('/')[1:])
def _get_accum_grad_name(self, var):
return self._local_var_name(var).replace(':','_') + '_accum_grad:0'
# Set environments
def _set_environment(self, initial_learning_rate):
self.env = None
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
self.episode_reward = 0
self.episode_length = 0
def choose_action(self, actions_value):
# epsilon-greedy
if np.random.uniform() < EPSILON:
action = np.argmax(actions_value)
else:
action = np.random.randint(0, ACTION_SIZE)
return action
# Take LOCAL_T_MAX step in one process
def process(self, sess, global_t, summary_writer, summary_op, summary_placeholders):
#print("start process")
if self.env is None:
# lazy evaluation
time.sleep(self.thread_index*1.0)
self.env = Environment({
'scene_name': self.scene_scope,
'terminal_state_id': int(self.task_scope)
})
start_local_t = self.local_t
# Reset accmulated gradient variables
sess.run(self.reset_gradients)
# Obtain shared parameters from global
sess.run( self.sync )
# t_max times loop
for i in range(LOCAL_T_MAX):
old_s_t = self.env.s_t
actions_value = self.local_network.run_DQN(sess, self.env.s_t, self.env.target, self.scopes)
action = self.choose_action(actions_value)
if VERBOSE and (self.thread_index == 0) and (self.local_t % 1000) == 0:
sys.stdout.write("%s:" % self.scene_scope)
sys.stdout.write("Pi = {0} V = {1}\n".format(actions_value, action))
# process game
self.env.step(action)
# receive game result
reward = self.env.reward
terminal = self.env.terminal
# ad-hoc reward for navigation
# reward = 10.0 if terminal else -0.01
if self.episode_length > 5e3: terminal = True
self.episode_reward += reward
self.episode_length += 1
"""
print("Local t: {0:d}".format(self.local_t))
print("Reward: {0:f}".format(reward))
print("Episode reward: {0:f}".format(self.episode_reward))
print("Episode length: {0:d}".format(self.episode_length))
"""
self.local_t += 1
# store transition to replay buffer
self.store_transition(old_s_t, action, reward, self.env.s_t)
if terminal:
sys.stdout.write("#Thread: %d \n time %d | thread #%d | scene %s | target #%s\n%s %s episode reward = %.3f\n%s %s episode length = %d\n%s %s \n" % (self.thread_index, global_t, self.thread_index, self.scene_scope, self.task_scope, self.scene_scope, self.task_scope, self.episode_reward, self.scene_scope, self.task_scope, self.episode_length, self.scene_scope, self.task_scope))
summary_values = {
"episode_reward_input": self.episode_reward,
"episode_length_input": float(self.episode_length),
"learning_rate_input": self._anneal_learning_rate(global_t)
}
self._record_score(sess, summary_writer, summary_op, summary_placeholders,
summary_values, global_t)
self.episode_reward = 0
self.episode_length = 0
self.env.reset()
break
# update target network
if self.local_t % self.replace_target_iter == 0:
sess.run(self.local_network.replace_target_op)
# print('\ntarget_params_replaced\n')
# sample batch memory from all memory
if self.memory_counter > self.memory_size:
sample_index = np.random.choice(self.memory_size, size=self.batch_size)
else:
sample_index = np.random.choice(self.memory_counter, size=self.batch_size)
batch_memory = self.memory[sample_index, :]
batch_memory_s_ = np.reshape(batch_memory[:, -2048*4:], (-1, 2048, 4))
batch_memory_s = np.reshape(batch_memory[:, :2048*4], (-1, 2048, 4))
batch_memory_t = np.reshape(np.tile(self.env.target, [self.batch_size, 1]), (-1, 2048, 4))
q_next, q_eval = sess.run(
[self.local_network.q_next, self.local_network.q_eval],
feed_dict={
self.local_network.s_: batch_memory_s_, # fixed params
self.local_network.s: batch_memory_s, # newest params
self.local_network.t: batch_memory_t
})
# change q_target w.r.t q_eval's action
q_target = q_eval.copy()
batch_index = np.arange(self.batch_size, dtype=np.int32)
eval_act_index = batch_memory[:, 2048*4].astype(int)
reward = batch_memory[:, 2048*4 + 1]
key_eval = self.network_scope + '/' + self.scene_scope + '/eval'
if terminal:
q_target[key_eval][batch_index, eval_act_index] = reward
else:
key_target = self.network_scope + '/'+ self.scene_scope + '/target'
q_target[key_eval][batch_index, eval_act_index] = reward + self.gamma * np.max(q_next[key_target], axis=1)
for idx in batch_index:
# train eval network
sess.run(self.accum_gradients,
feed_dict={
self.local_network.s: [batch_memory_s[idx]],
self.local_network.t: [batch_memory_t[idx]],
self.local_network.q_target: [q_target[key_eval][idx]]})
cur_learning_rate = self._anneal_learning_rate(global_t)
# update global network
sess.run( self.apply_gradients,
feed_dict = { self.learning_rate_input: cur_learning_rate } )
if VERBOSE and (self.thread_index == 0) and (self.local_t % 100) == 0:
sys.stdout.write("#Thread-%d-%s-Local timestep-%d\n" % (self.thread_index, self.scene_scope, self.local_t))
# return advanced local step size
diff_local_t = self.local_t - start_local_t
return diff_local_t
def _record_score(self, sess, writer, summary_op, placeholders, values, global_t):
feed_dict = {}
for k in placeholders:
feed_dict[placeholders[k]] = values[k]
summary_str = sess.run(summary_op, feed_dict=feed_dict)
if VERBOSE: sys.stdout.write('writing to summary writer at time %d\n' % (global_t))
writer.add_summary(summary_str, global_t)
# writer.flush()
def _anneal_learning_rate(self, global_time_step):
time_step_to_go = max(self.max_global_time_step - global_time_step, 0.0)
learning_rate = self.initial_learning_rate * time_step_to_go / self.max_global_time_step
return learning_rate
def store_transition(self, s, a, r, s_):
if not hasattr(self, 'memory_counter'):
self.memory_counter = 0
transition = np.hstack((np.reshape(s, -1), [a, r], np.reshape(s_,-1)))
# replace the old memory with new memory
index = self.memory_counter % self.memory_size
self.memory[index, :] = transition
self.memory_counter += 1
class A3CTrainingThread(object):
def __init__(self,
thread_index,
global_network,
global_discriminator,
initial_learning_rate,
learning_rate_input,
grad_applier,
grad_applier_discriminator,
max_global_time_step,
device,
network_scope="network",
scene_scope="scene",
task_scope="task"):
self.thread_index = thread_index
self.learning_rate_input = learning_rate_input
self.max_global_time_step = max_global_time_step
self.network_scope = network_scope
self.network_scope_D = network_scope + "_d"
self.scene_scope = scene_scope
self.task_scope = task_scope
self.scopes = [network_scope, scene_scope, task_scope]
self.scopes_d = [self.network_scope_D, scene_scope, task_scope]
self.local_network = ActorCriticFFNetwork(action_size=ACTION_SIZE,
device=device,
network_scope=network_scope,
scene_scopes=[scene_scope])
self.local_discriminator = Discriminator_WGAN(
action_size=ACTION_SIZE,
device=device,
network_scope=network_scope,
scene_scopes=[scene_scope])
self.local_network.prepare_loss(ENTROPY_BETA, self.scopes)
self.local_discriminator.prepare_loss_D(ENTROPY_BETA, self.scopes_d)
self.trainer = AccumTrainer(device)
self.trainer_D = AccumTrainer(
device, name="AccumTrainer_d") #new instance for discrimninateor
self.trainer.prepare_minimize(
self.local_network.
total_loss, #getting the gradients of for the local network variablkes
self.local_network.get_vars())
self.trainer_D.prepare_minimize(self.local_discriminator.total_loss_d,
self.local_discriminator.get_vars())
new_variable_list = self.local_network.get_vars()
old_varaible_list = self.local_network.get_vars_old()
self.old_new_sync = self.local_network.sync_curre_old()
self.accum_gradients = self.trainer.accumulate_gradients(
) #This is to assign gradients
self.reset_gradients = self.trainer.reset_gradients(
) #after applying the grads to variables we need to resent those variables
#This is for the discriminatro
self.accum_gradients_d = self.trainer_D.accumulate_gradients()
self.reset_gradients_d = self.trainer_D.reset_gradients()
accum_grad_names = [
self._local_var_name(x)
for x in self.trainer.get_accum_grad_list()
] #get the name list of all the grad vars
accum_grad_names_discrimi = [
self._local_var_name(x)
for x in self.trainer_D.get_accum_grad_list()
]
global_net_vars = [
x for x in global_network.get_vars()
if self._get_accum_grad_name(x) in accum_grad_names
] #check whether the global_network vars are mentioned in gradiet computations for them
local_net_vars = [
x for x in self.local_network.get_vars()
if self._get_accum_grad_name(x) in accum_grad_names
]
#self.trainer.get_accum_grad_list() this is about gradients righjt now
global_discri_vars = [
x for x in global_discriminator.get_vars()
if self._get_accum_grad_name(x) in accum_grad_names_discrimi
]
self.apply_gradients = grad_applier.apply_gradients(
global_net_vars, self.trainer.get_accum_grad_list())
self.apply_gradients_local = grad_applier.apply_gradients_local_net(
local_net_vars, self.trainer.get_accum_grad_list())
#Discriminator
self.apply_gradients_discriminator = grad_applier_discriminator.apply_gradients(
global_discri_vars, self.trainer_D.get_accum_grad_list())
self.clip_global_d_weights = global_discriminator.clip_weights(
) #here we are clipping the global net weights directly.
#If this is unstable it is desireable to first apply the gradients on the local network and then clip and after that we apply
self.sync = self.local_network.sync_from(
global_network
) #this is to sync from the glocal network Apply updated global params to the local network
self.sync_discriminator = self.local_discriminator.sync_from(
global_discriminator)
self.env = None
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
self.episode_reward = 0
self.episode_length = 0
self.episode_max_q = -np.inf
def _local_var_name(self, var):
return '/'.join(var.name.split('/')[1:])
def _get_accum_grad_name(self, var):
return self._local_var_name(var).replace(':', '_') + '_accum_grad:0'
def _anneal_learning_rate(self, global_time_step):
time_step_to_go = max(self.max_global_time_step - global_time_step,
0.0)
learning_rate = self.initial_learning_rate * time_step_to_go / self.max_global_time_step
return learning_rate
def choose_action(self, pi_values):
values = []
sum = 0.0
for rate in pi_values:
sum = sum + rate
value = sum
values.append(value)
r = random.random() * sum
for i in range(len(values)):
if values[i] >= r:
return i
# fail safe
return len(values) - 1
def _record_score(self, sess, writer, summary_op, placeholders, values,
global_t):
feed_dict = {}
for k in placeholders:
feed_dict[placeholders[k]] = values[k]
summary_str = sess.run(summary_op, feed_dict=feed_dict)
if VERBOSE:
print('writing to summary writer at time %d\n' % (global_t))
writer.add_summary(summary_str, global_t)
# writer.flush()
def process(self, sess, global_t, summary_writer, summary_op,
summary_placeholders):
if self.env is None:
# lazy evaluation
time.sleep(self.thread_index * 1.0)
self.env = Environment({
'scene_name': self.scene_scope,
'terminal_state_id': int(self.task_scope)
})
self.env.reset() #resetting the environment for each thread
states = [] #to keeep state ,actions ,targets and other stae
actions = []
rewards = []
values = []
targets = []
dones = []
terminal_end = False #in the start terminal state_end is false
# reset accumulated gradients
sess.run(
self.reset_gradients
) #resetting the gradient positions when starting the process for each
# copy weights from shared to local
sess.run(self.sync)
start_local_t = self.local_t
# t_max times loop
for i in range(
LOCAL_T_MAX
): #one thread will run for maximum amoound to 5 iterations then do a gradiet uodate
pi_, value_ = self.local_network.run_policy_and_value(
sess, self.env.s_t, self.env.target, self.scopes)
#pi_Old, value_Old = self.local_network.run_policy_and_value_old(sess, self.env.s_t, self.env.target, self.scopes)
action = self.choose_action(pi_)
states.append(self.env.s_t)
actions.append(action)
values.append(value_)
targets.append(self.env.target)
if VERBOSE and (self.thread_index
== 0) and (self.local_t % 1000) == 0:
sys.stdout.write("Pi = {0} V = {1}\n".format(pi_, value_))
# process game
self.env.step(action)
# receive game result
reward = self.env.reward #getting the reward from the env
terminal = self.env.terminal #geting whether the agent went to a terminal state
# ad-hoc reward for navigation
reward = 10.0 if terminal else -0.01 #this is the normal reward here 10 if terminal all the others it is -0.01 (ollision donesst take in to the accout)
if self.episode_length > 5e3:
terminal = True #Here we do not let agent to run more that 5000 steps so we make it terminal
#but the above terminal thing has no effect on giving 10 as the rwaerd because we set the rweard above
self.episode_reward += reward
self.episode_length += 1
#this is what is the maximum value got in the episode
self.episode_max_q = max(self.episode_max_q, np.max(
value_)) #self.episode_max_q-This is -inf in the beggining
# clip reward
rewards.append(
np.clip(reward, -1, 1)
) #make sure the rewartds is between -1 and +1 even thore rtthere is a 10
self.local_t += 1
# s_t1 -> s_t
self.env.update()
if terminal: #if we go to the terminal state we will surely break the function
sys.stdout.write("Pi = {0} V = {1}\n".format(pi_, value_))
terminal_end = True
sys.stdout.write(
"time %d | thread #%d | scene %s | target #%s\n%s %s episode reward = %.3f\n%s %s episode length = %d\n%s %s episode max Q = %.3f\n"
% (global_t, self.thread_index, self.scene_scope,
self.task_scope, self.scene_scope, self.task_scope,
self.episode_reward, self.scene_scope, self.task_scope,
self.episode_length, self.scene_scope, self.task_scope,
self.episode_max_q))
summary_values = {
"episode_reward_input": self.episode_reward,
"episode_length_input": float(self.episode_length),
"episode_max_q_input": self.episode_max_q,
"learning_rate_input": self._anneal_learning_rate(global_t)
}
self._record_score(sess, summary_writer, summary_op,
summary_placeholders, summary_values,
global_t)
self.episode_reward = 0 #after terminal state we gonna make all these variables to zero
self.episode_length = 0 #Now the AI need to start from new position
self.episode_max_q = -np.inf #after a terminaltion we do this
self.env.reset()
break
'''
Here I should call the discriminator and get the reward signal from that
R_D=sess.run(D.get_reward(state,action))
'''
R = 0.0 #In the terminal Return is nothing #If it's terminal end we do not have a return from the final state
if not terminal_end: #But if it's not the turminal Return is the next value function
R = self.local_network.run_value(sess, self.env.s_t,
self.env.target, self.scopes)
Returns = np.zeros_like(rewards)
Advants = np.zeros_like(rewards)
lastgaelam = 0
LAMBDA = 0.9
GAM = 0.9
self.nsteps = len(rewards)
############################################################################# we should assined all params to the new params
#This will only has an effect on
#####################################################################
for t in reversed(range(self.nsteps)):
if t == self.nsteps - 1:
nextnonterminal = 1.0 - bool(
R == 0) #if R ==0 means the agent found the terminal stage
nextvalues = R
else:
nextnonterminal = 1.0 - bool(R == 0)
nextvalues = values[t + 1]
delta = rewards[t] + GAM * nextvalues * nextnonterminal - values[t]
Advants[
t] = lastgaelam = delta + GAM * LAMBDA * lastgaelam * nextnonterminal
Returns[t] = Advants[t] + values[t]
#Returns=Advants+values #This is more of the v_next
Advants = (Advants - Advants.mean()) / (Advants.std() + 1e-5)
#Returns=(Returns - Returns.mean()) / (Returns.std() + 1e-5)
Returns = Returns.tolist()
Advants = Advants.tolist()
actions.reverse()
states.reverse()
rewards.reverse()
values.reverse()
Returns.reverse()
Advants.reverse()
batch_si = []
batch_a = []
batch_td = []
batch_R = []
batch_t = []
batch_advant = []
batch_Return = []
# compute and accmulate gradients
for (ai, ri, si, Vi, ti, Re, Ad) in zip(actions, rewards, states,
values, targets, Returns,
Advants):
R = ri + GAMMA * R #calculatung the adcantage function
td = R - Vi
a = np.zeros([ACTION_SIZE])
a[ai] = 1 #making the actions one hot
batch_si.append(si)
batch_a.append(a)
batch_td.append(td)
batch_R.append(R)
batch_t.append(ti)
batch_advant.append(Ad)
batch_Return.append(Re)
sess.run(self.old_new_sync)
cur_learning_rate = self._anneal_learning_rate(global_t)
for i in range(3):
sess.run(
self.
accum_gradients, #since we update the algorithm for given action ,given state, given advatns and given value and given reward we do not care about the sequence
feed_dict={
self.local_network.s: batch_si,
self.local_network.a: batch_a,
self.local_network.t: batch_t,
self.local_network.td: batch_td,
self.local_network.r: batch_R,
self.local_network.Returns: batch_Return,
self.local_network.Advantages: batch_advant
})
sess.run(self.apply_gradients_local,
feed_dict={self.learning_rate_input: cur_learning_rate})
sess.run(self.apply_gradients,
feed_dict={self.learning_rate_input: cur_learning_rate})
if VERBOSE and (self.thread_index == 0) and (self.local_t % 100) == 0:
sys.stdout.write("Local timestep %d\n" % self.local_t)
# return advanced local step size
diff_local_t = self.local_t - start_local_t
return diff_local_t
class SmashNetTrainingThread(object): #Threading th training
def __init__(self,
thread_index,
global_network,
initial_learning_rate,
learning_rate_input,
grad_applier,
max_global_time_step,
device,
initial_diffidence_rate_seed,
mode="train",
network_scope="network",
scene_scope="scene",
task_scope="task",
encourage_symmetry=False):
self.thread_index = thread_index
self.learning_rate_input = learning_rate_input
self.max_global_time_step = max_global_time_step
self.network_scope = network_scope #assiciated with the thread number
self.scene_scope = scene_scope #Score
self.task_scope = task_scope #This the targe
self.scopes = [network_scope, scene_scope,
task_scope] # ["thread-n", "scene", "target"]
self.local_network = SmashNet( #locally smash net policy netwotk
action_size=ACTION_SIZE,
device=device,
network_scope=network_scope,
scene_scopes=[scene_scope])
self.local_network.prepare_loss(
self.scopes) #This is to calculate the loss for this thread
if mode is "train":
self.trainer = AccumTrainer(device)
self.trainer.prepare_minimize(self.local_network.loss,
self.local_network.get_vars())
self.accum_gradients = self.trainer.accumulate_gradients()
self.reset_gradients = self.trainer.reset_gradients()
accum_grad_names = [
self._local_var_name(x)
for x in self.trainer.get_accum_grad_list()
] #This is more of we apply gradients to globabl network
global_net_vars = [
x for x in global_network.get_vars()
if self._get_accum_grad_name(x) in accum_grad_names
]
self.apply_gradients = grad_applier.apply_gradients(
global_net_vars, self.trainer.get_accum_grad_list())
self.sync = self.local_network.sync_from(global_network)
self.env = None
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
# self.episode_reward = 0
self.episode_length = 0
# self.episode_max_q = -np.inf
self.episode_pi_sim = 0
self.episode_loss = 0
self.initial_diffidence_rate_seed = initial_diffidence_rate_seed
self.oracle = None
self.mode = mode
self.encourage_symmetry = encourage_symmetry
def _local_var_name(self, var):
return '/'.join(var.name.split('/')[1:])
def _get_accum_grad_name(self, var):
return self._local_var_name(var).replace(':', '_') + '_accum_grad:0'
def _anneal_rate(self, init_rate, global_time_step):
time_step_to_go = max(self.max_global_time_step - global_time_step,
0.0)
rate = init_rate * time_step_to_go / self.max_global_time_step
return rate
def _anneal_learning_rate(self, global_time_step):
learning_rate = self._anneal_rate(self.initial_learning_rate,
global_time_step)
return learning_rate
def _inverse_sigmoid_decay_rate(self, init_rate_seed, global_time_step):
rate = init_rate_seed * np.exp(-global_time_step / init_rate_seed)
rate = rate / (1. + rate)
return rate
def _anneal_diffidence_rate(self, global_time_step):
if self.initial_diffidence_rate_seed == 0: return 0
else:
return self._inverse_sigmoid_decay_rate(
self.initial_diffidence_rate_seed, global_time_step)
# TODO: check
def choose_action(self, smashnet_pi_values, oracle_pi_values,
confidence_rate): #can change the action to take
r = random.random()
if r < confidence_rate: pi_values = oracle_pi_values
else: pi_values = smashnet_pi_values
r = random.random() * np.sum(pi_values)
values = np.cumsum(pi_values)
for i in range(len(values)):
if values[i] >= r: return i
def _record_score(self, sess, writer, summary_op, placeholders, values,
global_t):
feed_dict = {}
for k in placeholders:
feed_dict[placeholders[k]] = values[k]
summary_str = sess.run(summary_op, feed_dict=feed_dict)
writer.add_summary(summary_str, global_t)
# writer.flush()
def _evaluate(self, sess, list_of_tasks, num_episodes, max_steps,
success_cutoff):
scene_scopes = list_of_tasks.keys()
results = {}
for scene_scope in scene_scopes:
for task_scope in list_of_tasks[scene_scope]:
env = Environment({
'scene_name': scene_scope,
'terminal_state_id': int(task_scope)
})
ep_lengths = []
ep_collisions = []
oracle_lengths = []
ep_successes = []
scopes = [self.network_scope, scene_scope, task_scope]
for i_episode in range(num_episodes):
env.reset()
oracle_lengths.append(env.shortest_path_distances[
env.current_state_id][env.terminal_state_id])
terminal = False
ep_length = 0
ep_collision = 0
while not terminal:
pi_values = self.local_network.run_policy(
sess, env.s_t, env.target, scopes)
action = sample_action(pi_values)
env.step(action)
env.update()
terminal = env.terminal
if ep_length == max_steps: break
if env.collided: ep_collision += 1
ep_length += 1
ep_lengths.append(ep_length)
ep_collisions.append(ep_collision)
ep_successes.append(int(ep_length < success_cutoff))
results[scene_scope + task_scope] = [
np.mean(ep_lengths),
np.mean(ep_collisions),
np.mean(oracle_lengths),
np.mean(ep_successes)
]
return results
def _flip_policy(self, policy):
flipped_policy = np.array([policy[3], policy[2], policy[1], policy[0]])
return flipped_policy
def process(self, sess, global_t, summary_writer, summary_op,
summary_placeholders): #This is to run the process
if self.env is None:
# lazy evaluation
time.sleep(self.thread_index * 1.0)
self.env = Environment(
{ #This is where you access in to the environment #scene_loader import THORDiscreteEnvironment
'scene_name': self.scene_scope,
'terminal_state_id': int(self.task_scope)
})
self.env.reset()
self.oracle = ShortestPathOracle(
self.env, ACTION_SIZE
) #Get the probabilities of the shortest paths to go to exat position
states = []
targets = []
oracle_pis = [] #expert policies
terminal_end = False
if self.mode is "train": #if the trainign is there
# reset accumulated gradients
sess.run(self.reset_gradients) #reet all the gradients
# copy weights from shared to local
sess.run(self.sync) #
start_local_t = self.local_t
# t_max times loop (5 steps)
for i in range(LOCAL_T_MAX): # This is for the training
flipped_run = self.encourage_symmetry and np.random.random() > 0.5
if flipped_run:
s_t = self.env.target
g = self.env.s_t
else:
s_t = self.env.s_t
g = self.env.target #first the initial state start with same state 4 times as the history stacked as frames 2048*5
smashnet_pi = self.local_network.run_policy(
sess, s_t, g,
self.scopes) #now gethe policy frmo the local network
if flipped_run: smashnet_pi = self._flip_policy(smashnet_pi)
oracle_pi = self.oracle.run_policy(
self.env.current_state_id
) #get the policy of the oracle which means the shotest path kind of action in the given state
diffidence_rate = self._anneal_diffidence_rate(global_t)
action = self.choose_action(smashnet_pi, oracle_pi,
diffidence_rate)
states.append(s_t) #stack action
targets.append(g) #stack target position
if flipped_run: oracle_pis.append(self._flip_policy(oracle_pi))
else: oracle_pis.append(oracle_pi) #get the expert's policies
# if VERBOSE and global_t % 10000 == 0:
# print("Thread %d" % (self.thread_index))
# sys.stdout.write("SmashNet Pi = {}, Oracle Pi = {}\n".format(["{:0.2f}".format(i) for i in smashnet_pi], ["{:0.2f}".format(i) for i in oracle_pi]))
if VALIDATE and global_t % VALIDATE_FREQUENCY == 0 and global_t > 0 and self.thread_index == 0: #This is for the alidation of the results
results = self._evaluate(sess,
list_of_tasks=VALID_TASK_LIST,
num_episodes=NUM_VAL_EPISODES,
max_steps=MAX_VALID_STEPS,
success_cutoff=SUCCESS_CUTOFF)
print("Thread %d" % (self.thread_index))
print("Validation results: %s" % (results))
self.env.step(action) #here we change the next step
is_terminal = self.env.terminal or self.episode_length > 5e3
if self.mode is "val" and self.episode_length > 1e3:
is_terminal = True
self.episode_length += 1
self.episode_pi_sim += 1. - cosine(smashnet_pi, oracle_pi)
self.local_t += 1
# s_t1 -> s_t
self.env.update() #update the new state
if is_terminal:
terminal_end = True
if self.mode is "val":
sess.run(self.sync)
sys.stdout.write(
"time %d | thread #%d | scene %s | target %s | episode length = %d\n"
% (global_t, self.thread_index, self.scene_scope,
self.task_scope, self.episode_length))
summary_values = {
"episode_length_input":
float(self.episode_length),
"episode_pi_sim_input":
self.episode_pi_sim / float(self.episode_length),
"episode_loss_input":
float(self.episode_loss)
}
self._record_score(sess, summary_writer, summary_op,
summary_placeholders, summary_values,
global_t)
self.episode_length = 0
self.episode_pi_sim = 0
self.episode_loss = 0
self.env.reset()
break
if self.mode is "train":
states.reverse()
oracle_pis.reverse()
batch_si = []
batch_ti = []
batch_opi = []
# compute and accmulate gradients
for (si, ti, opi) in zip(states, targets, oracle_pis):
batch_si.append(si)
batch_ti.append(ti)
batch_opi.append(opi)
sess.run(self.accum_gradients,
feed_dict={
self.local_network.s: batch_si,
self.local_network.t: batch_ti,
self.local_network.opi: batch_opi
})
self.episode_loss += sum(
sess.run(self.local_network.loss,
feed_dict={
self.local_network.s: batch_si,
self.local_network.t: batch_ti,
self.local_network.opi: batch_opi
}))
cur_learning_rate = self._anneal_learning_rate(global_t)
sess.run(self.apply_gradients,
feed_dict={self.learning_rate_input: cur_learning_rate})
# if VERBOSE and (self.thread_index == 0) and (self.local_t % 100) == 0:
# sys.stdout.write("Local timestep %d\n" % self.local_t)
# return advanced local step size
diff_local_t = self.local_t - start_local_t
return diff_local_t
class A3CTrainingThread(object):
def __init__(self,
thread_index,
global_network,
initial_learning_rate,
learning_rate_input,
grad_applier,
max_global_time_step,
device,
network_scope="network",
scene_scope="scene",
task_scope="task"):
self.thread_index = thread_index
self.learning_rate_input = learning_rate_input
self.max_global_time_step = max_global_time_step
self.network_scope = network_scope
self.scene_scope = scene_scope
self.task_scope = task_scope
self.scopes = [network_scope, scene_scope, task_scope]
self.task_scope_name=1
self.local_network = ActorCriticFFNetwork(
action_size=ACTION_SIZE,
device=device,
network_scope=network_scope,
scene_scopes=[scene_scope])
self.local_network.prepare_loss(ENTROPY_BETA, self.scopes)
self.trainer = AccumTrainer(device)
self.trainer.prepare_minimize(self.local_network.total_loss,
self.local_network.get_vars())
self.accum_gradients = self.trainer.accumulate_gradients()
self.reset_gradients = self.trainer.reset_gradients()
accum_grad_names = [self._local_var_name(x) for x in self.trainer.get_accum_grad_list()]
global_net_vars = [x for x in global_network.get_vars() if self._get_accum_grad_name(x) in accum_grad_names]
self.apply_gradients = grad_applier.apply_gradients(
global_net_vars, self.trainer.get_accum_grad_list() )
self.sync = self.local_network.sync_from(global_network)
self.env = None
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
self.episode_reward = 0
self.episode_length = 0
self.episode_max_q = -np.inf
def _local_var_name(self, var):
return '/'.join(var.name.split('/')[1:])
def _get_accum_grad_name(self, var):
return self._local_var_name(var).replace(':','_') + '_accum_grad:0'
def _anneal_learning_rate(self, global_time_step):
time_step_to_go = max(self.max_global_time_step - global_time_step, 0.0)
learning_rate = self.initial_learning_rate * time_step_to_go / self.max_global_time_step
return learning_rate
def choose_action(self, pi_values):
values = []
sum = 0.0
for rate in pi_values:
sum = sum + rate
value = sum
values.append(value)
r = random.random() * sum
for i in range(len(values)):
if values[i] >= r:
return i
# fail safe
return len(values) - 1
def _record_score(self, sess, writer, summary_op, placeholders, values, global_t):
feed_dict = {}
for k in placeholders:
feed_dict[placeholders[k]] = values[k]
summary_str = sess.run(summary_op, feed_dict=feed_dict)
if VERBOSE: print('writing to summary writer at time %d\n' % (global_t))
writer.add_summary(summary_str, global_t)
# writer.flush()
def process(self, sess, global_t, summary_writer, summary_op, summary_placeholders):
if self.env is None:
# lazy evaluation
time.sleep(self.thread_index*1.0)
self.task_scope_name = random.randint(1, 468) - 1
self.env = Environment({
'scene_name': self.scene_scope,
'terminal_state_id': self.task_scope_name
})
self.env.reset()
states = []
actions = []
rewards = []
values = []
targets = []
terminal_end = False
# reset accumulated gradients
sess.run( self.reset_gradients )
# copy weights from shared to local
sess.run( self.sync )
start_local_t = self.local_t
# t_max times loop
for i in range(LOCAL_T_MAX):
pi_, value_ = self.local_network.run_policy_and_value(sess, self.env.s_t, self.env.target, self.scopes)
action = self.choose_action(pi_)
states.append(self.env.s_t)
actions.append(action)
values.append(value_)
targets.append(self.env.target)
if VERBOSE and (self.thread_index == 0) and (self.local_t % 1000) == 0:
sys.stdout.write("Pi = {0} V = {1}\n".format(pi_, value_))
# process game
self.env.step(action)
# receive game result
reward = self.env.reward
terminal = self.env.terminal
# ad-hoc reward for navigation
reward = 10.0 if terminal else -0.01
if self.episode_length > 5e3: terminal = True
self.episode_reward += reward
self.episode_length += 1
self.episode_max_q = max(self.episode_max_q, np.max(value_))
# clip reward
rewards.append( np.clip(reward, -1, 1) )
self.local_t += 1
# s_t1 -> s_t
self.env.update()
if terminal:
terminal_end = True
print'----------'
print('real terminal id is {}'.format(self.task_scope_name))
sys.stdout.write("time %d | thread #%d | scene %s | target #%s\n%s %s episode reward = %.3f\n%s %s episode length = %d\n%s %s episode max Q = %.3f\n" % (global_t, self.thread_index, self.scene_scope, self.task_scope, self.scene_scope, self.task_scope, self.episode_reward, self.scene_scope, self.task_scope, self.episode_length, self.scene_scope, self.task_scope, self.episode_max_q))
summary_values = {
"episode_reward_input": self.episode_reward,
"episode_length_input": float(self.episode_length),
"episode_max_q_input": self.episode_max_q,
"learning_rate_input": self._anneal_learning_rate(global_t)
}
self._record_score(sess, summary_writer, summary_op, summary_placeholders,
summary_values, global_t)
self.episode_reward = 0
self.episode_length = 0
self.episode_max_q = -np.inf
self.task_scope_name = random.randint(1, 468) - 1
self.env = Environment({
'scene_name': self.scene_scope,
'terminal_state_id': self.task_scope_name
})
self.env.reset()
print ('init id is {}'.format(self.env.current_state_id))
print'----------'
break
R = 0.0
if not terminal_end:
R = self.local_network.run_value(sess, self.env.s_t, self.env.target, self.scopes)
actions.reverse()
states.reverse()
rewards.reverse()
values.reverse()
batch_si = []
batch_a = []
batch_td = []
batch_R = []
batch_t = []
# compute and accmulate gradients
for(ai, ri, si, Vi, ti) in zip(actions, rewards, states, values, targets):
R = ri + GAMMA * R
td = R - Vi
a = np.zeros([ACTION_SIZE])
a[ai] = 1
batch_si.append(si)
batch_a.append(a)
batch_td.append(td)
batch_R.append(R)
batch_t.append(ti)
if USE_LSTM:
batch_si.reverse()
batch_a.reverse()
batch_td.reverse()
batch_R.reverse()
batch_t.reverse()
sess.run(self.accum_gradients,
feed_dict={
self.local_network.s: batch_si,
self.local_network.a: batch_a,
self.local_network.t: batch_t,
self.local_network.td: batch_td,
self.local_network.r: batch_R,
self.local_network.initial_lstm_state: start_lstm_state,})
# self.local_network.step_size: [len(batch_a)]
else:
sess.run( self.accum_gradients,
feed_dict = {
self.local_network.s: batch_si,
self.local_network.a: batch_a,
self.local_network.t: batch_t,
self.local_network.td: batch_td,
self.local_network.r: batch_R} )
cur_learning_rate = self._anneal_learning_rate(global_t)
sess.run( self.apply_gradients,
feed_dict = { self.learning_rate_input: cur_learning_rate } )
if VERBOSE and (self.thread_index == 0) and (self.local_t % 100) == 0:
sys.stdout.write("Local timestep %d\n" % self.local_t)
# return advanced local step size
diff_local_t = self.local_t - start_local_t
return diff_local_t
class SmashNetTrainingThread(object):
def __init__(self,
thread_index,
global_network,
initial_learning_rate,
learning_rate_input,
grad_applier,
max_global_time_step,
device,
initial_diffidence_rate_seed,
mode="train",
network_scope="network",
scene_scope="scene",
task_scope="task",
encourage_symmetry=False):
self.thread_index = thread_index
self.learning_rate_input = learning_rate_input
self.max_global_time_step = max_global_time_step
self.network_scope = network_scope
self.scene_scope = scene_scope
self.task_scope = task_scope
self.scopes = [network_scope, scene_scope, task_scope] # ["thread-n", "scene", "target"]
self.local_network = SmashNet(
action_size=ACTION_SIZE,
device=device,
network_scope=network_scope,
scene_scopes=[scene_scope])
self.local_network.prepare_loss(self.scopes)
if mode is "train":
self.trainer = AccumTrainer(device)
self.trainer.prepare_minimize(self.local_network.loss,
self.local_network.get_vars())
self.accum_gradients = self.trainer.accumulate_gradients()
self.reset_gradients = self.trainer.reset_gradients()
accum_grad_names = [self._local_var_name(x) for x in self.trainer.get_accum_grad_list()]
global_net_vars = [x for x in global_network.get_vars() if self._get_accum_grad_name(x) in accum_grad_names]
self.apply_gradients = grad_applier.apply_gradients( global_net_vars, self.trainer.get_accum_grad_list() )
self.sync = self.local_network.sync_from(global_network)
self.env = None
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
# self.episode_reward = 0
self.episode_length = 0
# self.episode_max_q = -np.inf
self.episode_pi_sim = 0
self.episode_loss = 0
self.initial_diffidence_rate_seed = initial_diffidence_rate_seed
self.oracle = None
self.mode = mode
self.encourage_symmetry = encourage_symmetry
def _local_var_name(self, var):
return '/'.join(var.name.split('/')[1:])
def _get_accum_grad_name(self, var):
return self._local_var_name(var).replace(':','_') + '_accum_grad:0'
def _anneal_rate(self, init_rate, global_time_step):
time_step_to_go = max(self.max_global_time_step - global_time_step, 0.0)
rate = init_rate * time_step_to_go / self.max_global_time_step
return rate
def _anneal_learning_rate(self, global_time_step):
learning_rate = self._anneal_rate(self.initial_learning_rate, global_time_step)
return learning_rate
def _inverse_sigmoid_decay_rate(self, init_rate_seed, global_time_step):
rate = init_rate_seed*np.exp(-global_time_step/init_rate_seed)
rate = rate / (1. + rate)
return rate
def _anneal_diffidence_rate(self, global_time_step):
if self.initial_diffidence_rate_seed == 0: return 0
else: return self._inverse_sigmoid_decay_rate(self.initial_diffidence_rate_seed, global_time_step)
# TODO: check
def choose_action(self, smashnet_pi_values, oracle_pi_values, confidence_rate):
r = random.random()
if r < confidence_rate: pi_values = oracle_pi_values
else: pi_values = smashnet_pi_values
r = random.random() * np.sum(pi_values)
values = np.cumsum(pi_values)
for i in range(len(values)):
if values[i] >= r: return i
def _record_score(self, sess, writer, summary_op, placeholders, values, global_t):
feed_dict = {}
for k in placeholders:
feed_dict[placeholders[k]] = values[k]
summary_str = sess.run(summary_op, feed_dict=feed_dict)
writer.add_summary(summary_str, global_t)
# writer.flush()
def _evaluate(self, sess, list_of_tasks, num_episodes, max_steps, success_cutoff):
scene_scopes = list_of_tasks.keys()
results = {}
for scene_scope in scene_scopes:
for task_scope in list_of_tasks[scene_scope]:
env = Environment({
'scene_name': scene_scope,
'terminal_state_id': int(task_scope)
})
ep_lengths = []
ep_collisions = []
oracle_lengths = []
ep_successes = []
scopes = [self.network_scope, scene_scope, task_scope]
for i_episode in range(num_episodes):
env.reset()
oracle_lengths.append(env.shortest_path_distances[env.current_state_id][env.terminal_state_id])
terminal = False
ep_length = 0
ep_collision = 0
while not terminal:
pi_values = self.local_network.run_policy(sess, env.s_t, env.target, scopes)
action = sample_action(pi_values)
env.step(action)
env.update()
terminal = env.terminal
if ep_length == max_steps: break
if env.collided: ep_collision += 1
ep_length += 1
ep_lengths.append(ep_length)
ep_collisions.append(ep_collision)
ep_successes.append(int(ep_length < success_cutoff))
results[scene_scope + task_scope] = [np.mean(ep_lengths), np.mean(ep_collisions), np.mean(oracle_lengths), np.mean(ep_successes)]
return results
def _flip_policy(self, policy):
flipped_policy = np.array([policy[3],
policy[2],
policy[1],
policy[0]])
return flipped_policy
def process(self, sess, global_t, summary_writer, summary_op, summary_placeholders):
if self.env is None:
# lazy evaluation
time.sleep(self.thread_index*1.0)
self.env = Environment({
'scene_name': self.scene_scope,
'terminal_state_id': int(self.task_scope)
})
self.env.reset()
self.oracle = ShortestPathOracle(self.env, ACTION_SIZE)
states = []
targets = []
oracle_pis = []
terminal_end = False
if self.mode is "train":
# reset accumulated gradients
sess.run( self.reset_gradients )
# copy weights from shared to local
sess.run( self.sync )
start_local_t = self.local_t
# t_max times loop (5 steps)
for i in range(LOCAL_T_MAX):
flipped_run = self.encourage_symmetry and np.random.random() > 0.5
if flipped_run: s_t = self.env.target; g = self.env.s_t
else: s_t = self.env.s_t; g = self.env.target
smashnet_pi = self.local_network.run_policy(sess, s_t, g, self.scopes)
if flipped_run: smashnet_pi = self._flip_policy(smashnet_pi)
oracle_pi = self.oracle.run_policy(self.env.current_state_id)
diffidence_rate = self._anneal_diffidence_rate(global_t)
action = self.choose_action(smashnet_pi, oracle_pi, diffidence_rate)
states.append(s_t)
targets.append(g)
if flipped_run: oracle_pis.append(self._flip_policy(oracle_pi))
else: oracle_pis.append(oracle_pi)
# if VERBOSE and global_t % 10000 == 0:
# print("Thread %d" % (self.thread_index))
# sys.stdout.write("SmashNet Pi = {}, Oracle Pi = {}\n".format(["{:0.2f}".format(i) for i in smashnet_pi], ["{:0.2f}".format(i) for i in oracle_pi]))
if VALIDATE and global_t % VALIDATE_FREQUENCY == 0 and global_t > 0 and self.thread_index == 0:
results = self._evaluate(sess, list_of_tasks=VALID_TASK_LIST, num_episodes=NUM_VAL_EPISODES, max_steps=MAX_VALID_STEPS, success_cutoff=SUCCESS_CUTOFF)
print("Thread %d" % (self.thread_index))
print("Validation results: %s" % (results))
self.env.step(action)
is_terminal = self.env.terminal or self.episode_length > 5e3
if self.mode is "val" and self.episode_length > 1e3:
is_terminal = True
self.episode_length += 1
self.episode_pi_sim += 1. - cosine(smashnet_pi, oracle_pi)
self.local_t += 1
# s_t1 -> s_t
self.env.update()
if is_terminal:
terminal_end = True
if self.mode is "val":
sess.run(self.sync)
sys.stdout.write("time %d | thread #%d | scene %s | target %s | episode length = %d\n" % (global_t, self.thread_index, self.scene_scope, self.task_scope, self.episode_length))
summary_values = {
"episode_length_input": float(self.episode_length),
"episode_pi_sim_input": self.episode_pi_sim / float(self.episode_length),
"episode_loss_input": float(self.episode_loss)
}
self._record_score(sess, summary_writer, summary_op, summary_placeholders,
summary_values, global_t)
self.episode_length = 0
self.episode_pi_sim = 0
self.episode_loss = 0
self.env.reset()
break
if self.mode is "train":
states.reverse()
oracle_pis.reverse()
batch_si = []
batch_ti = []
batch_opi = []
# compute and accmulate gradients
for(si, ti, opi) in zip(states, targets, oracle_pis):
batch_si.append(si)
batch_ti.append(ti)
batch_opi.append(opi)
sess.run( self.accum_gradients,
feed_dict = {
self.local_network.s: batch_si,
self.local_network.t: batch_ti,
self.local_network.opi: batch_opi} )
self.episode_loss += sum(sess.run(self.local_network.loss,
feed_dict={
self.local_network.s: batch_si,
self.local_network.t: batch_ti,
self.local_network.opi: batch_opi}))
cur_learning_rate = self._anneal_learning_rate(global_t)
sess.run( self.apply_gradients, feed_dict = { self.learning_rate_input: cur_learning_rate } )
# if VERBOSE and (self.thread_index == 0) and (self.local_t % 100) == 0:
# sys.stdout.write("Local timestep %d\n" % self.local_t)
# return advanced local step size
diff_local_t = self.local_t - start_local_t
return diff_local_t