def main():
experiment = 'model-builder-v0' #specify environments here
env = gym.make(experiment)
#steps= env.spec.timestep_limit #steps per episode
steps = 20
assert isinstance(env.observation_space,
Box), "observation space must be continuous"
assert isinstance(env.action_space, Box), "action space must be continuous"
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env, is_batch_norm)
exploration_noise = OUNoise(env.action_space.shape[0])
counter = 0
reward_per_episode = 0
total_reward = 0
num_states = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
print("Number of States:", num_states)
print("Number of Actions:", num_actions)
print("Number of Steps per episode:", steps)
#saving reward:
reward_st = np.array([0])
for i in range(episodes):
print("==== Starting episode no:", i, "====", "\n")
observation = env.reset()
reward_per_episode = 0
for t in range(steps):
#rendering environmet (optional)
env.render()
x = observation
action = agent.evaluate_actor(np.reshape(x, [1, 300, 300, 2]))
noise = exploration_noise.noise()
action = action[
0] + noise #Select action according to current policy and exploration noise
print("Action at step", t, " :", action, "\n")
observation, reward, done, info = env.step(action)
#add s_t,s_t+1,action,reward to experience memory
agent.add_experience(x, observation, action, reward, done)
#train critic and actor network
if counter > 64:
agent.train()
reward_per_episode += reward
counter += 1
#check if episode ends:
if (done or (t == steps - 1)):
print('EPISODE: ', i, ' Steps: ', t, ' Total Reward: ',
reward_per_episode)
print("Printing reward to file")
exploration_noise.reset(
) #reinitializing random noise for action exploration
reward_st = np.append(reward_st, reward_per_episode)
np.savetxt('episode_reward.txt', reward_st, newline="\n")
print('\n\n')
break
total_reward += reward_per_episode
print("Average reward per episode {}".format(total_reward / episodes))
def main():
experiment= 'InvertedPendulum-v1' #specify environments here
env= gym.make(experiment)
steps= env.spec.timestep_limit #steps per episode
assert isinstance(env.observation_space, Box), "observation space must be continuous"
assert isinstance(env.action_space, Box), "action space must be continuous"
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env, is_batch_norm)
exploration_noise = OUNoise(env.action_space.shape[0])
counter=0
reward_per_episode = 0
total_reward=0
num_states = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
print "Number of States:", num_states
print "Number of Actions:", num_actions
print "Number of Steps per episode:", steps
#saving reward:
reward_st = np.array([0])
for i in xrange(episodes):
print "==== Starting episode no:",i,"====","\n"
observation = env.reset()
reward_per_episode = 0
for t in xrange(steps):
#rendering environmet (optional)
env.render()
x = observation
action = agent.evaluate_actor(np.reshape(x,[1,num_states]))
noise = exploration_noise.noise()
action = action[0] + noise #Select action according to current policy and exploration noise
print "Action at step", t ," :",action,"\n"
observation,reward,done,info=env.step(action)
#add s_t,s_t+1,action,reward to experience memory
agent.add_experience(x,observation,action,reward,done)
#train critic and actor network
if counter > 64:
agent.train()
reward_per_episode+=reward
counter+=1
#check if episode ends:
if (done or (t == steps-1)):
print 'EPISODE: ',i,' Steps: ',t,' Total Reward: ',reward_per_episode
print "Printing reward to file"
exploration_noise.reset() #reinitializing random noise for action exploration
reward_st = np.append(reward_st,reward_per_episode)
np.savetxt('episode_reward.txt',reward_st, newline="\n")
print '\n\n'
break
total_reward+=reward_per_episode
print "Average reward per episode {}".format(total_reward / episodes)
def main():
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env, is_batch_norm, CA_OBS_SPACE, CA_ACTION_SPACE,
CA_ACTION_BOUND)
exploration_noise = OUNoise(CA_ACTION_SPACE)
counter = 0
reward_per_episode = 0
total_reward = 0
num_states = CA_OBS_SPACE
num_actions = CA_ACTION_SPACE
print "Number of States:", num_states
print "Number of Actions:", num_actions
print "Number of Steps per episode:", steps
#saving reward:
reward_st = np.array([0])
for i in xrange(episodes):
print "==== Starting episode no:", i, "====", "\n"
# observation = env.reset()
observation = ca_reset()
reward_per_episode = 0
for t in xrange(steps):
#rendering environmet (optional)
# env.render()
x = observation
action = agent.evaluate_actor(np.reshape(x, [1, num_states]))
noise = exploration_noise.noise()
action = action[
0] + noise #Select action according to current policy and exploration noise
print "Action at step", t, " :", action, "\n"
# observation,reward,done,info=env.step(action)
observation, reward, done, info = ca_step(action)
print x, observation, action, reward, done
#add s_t,s_t+1,action,reward to experience memory
agent.add_experience(x, observation, action, reward, done)
#train critic and actor network
if counter > 64:
agent.train()
reward_per_episode += reward
counter += 1
#check if episode ends:
if (done or (t == steps - 1)):
print 'EPISODE: ', i, ' Steps: ', t, ' Total Reward: ', reward_per_episode
print "Printing reward to file"
exploration_noise.reset(
) #reinitializing random noise for action exploration
reward_st = np.append(reward_st, reward_per_episode)
np.savetxt('episode_reward.txt', reward_st, newline="\n")
print '\n\n'
break
total_reward += reward_per_episode
print "Average reward per episode {}".format(total_reward / episodes)
def main():
'''main function'''
# Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(ACTORNET_PRE_TRAINED, STATENET_PRE_TRAINED)
exploration_noise = OUNoise(ACTION_DIM)
# saving reward:
reward_st = np.array([0])
img_server = ImgServer()
img_server.wait_for_connect()
observe_t_img = img_server.receive_img()
observe_t_data = convert_img2data(observe_t_img)
actor_t = agent.evaluate_actor(observe_t_data)
noise = exploration_noise.noise()
actor_t = actor_t[0] + noise
img_server.send_actor_cmd(actor_t)
observe_t_1_img = observe_t_img
actor_t_1 = actor_t
img_server.close_connect()
index = 1
while True:
observe_t_img = observe_t_1_img
actor_t = actor_t_1
img_server.wait_for_connect()
observe_t_1_img = img_server.receive_img()
observe_t_1_data = convert_img2data(observe_t_1_img)
actor_t_1 = agent.evaluate_actor(observe_t_1_data)
noise = exploration_noise.noise()
actor_t_1 = actor_t_1[0] + noise
cost = compute_cost(observe_t_img)
agent.add_experience(observe_t_img, observe_t_1_img, actor_t, cost, index)
if index > 32:
agent.train()
img_server.send_actor_cmd(actor_t_1)
img_server.close_connect()
index = index + 1
def main():
env = Env(19997)
steps= 10000
num_states = 59
num_actions = 3
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env, is_batch_norm)
exploration_noise = OUNoise(num_actions)
counter=0
reward_per_episode = 0
total_reward=0
reward_st = np.array([0])
agent.actor_net.load_actor(os.getcwd() + '/weights/actor/model.ckpt')
agent.critic_net.load_critic(os.getcwd() + '/weights/critic/model.ckpt')
for i in range(episodes):
# print "==== Starting episode no:",i,"====","\n"
observation = env.reset()
done =False
reward_per_episode = 0
for t in range(steps):
x = observation
action = agent.evaluate_actor(np.reshape(x,[1,num_states]))
noise = exploration_noise.noise()
action = action[0] + noise #Select action according to current policy and exploration noise
for i in range(num_actions):
if action[i] > 1.0:
action[i] = 1.0
if action[i] < -1.0:
action[i] = -1.0
observation,reward,done = env.step(action)
print("reward:", reward, "\n")
agent.add_experience(x,observation,action,reward,done)
#train critic and actor network
if counter > 64:
agent.train()
reward_per_episode+=reward
counter+=1
#check if episode ends:
if (done or (t == steps-1)):
print('Episode',i,'Steps: ',t,'Episode Reward:',reward_per_episode)
exploration_noise.reset()
reward_st = np.append(reward_st,reward_per_episode)
np.savetxt('episode_reward.txt', reward_st, newline="\n")
agent.actor_net.save_actor(os.getcwd() + '/weights/actor/model.ckpt')
agent.critic_net.save_critic(os.getcwd() + '/weights/critic/model.ckpt')
break
total_reward+=reward_per_episode
def main():
env= Env(19997)
steps = 300
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env, is_batch_norm)
exploration_noise = OUNoise(2)
counter = 0
reward_per_episode = 0.
num_states = 32*16
num_actions = 2
#saving reward:
reward_st = np.array([0])
for i in range(episodes):
print ("==== Starting episode no:",str(i),"====","\n")
observation = env.reset()
reward_per_episode = 0
for t in range(steps):
x = observation
action = agent.evaluate_actor(np.reshape(x,[1,num_states]))
noise = exploration_noise.noise()
action = action[0] + noise
observation,reward,done=env.step(action,t)
agent.add_experience(x,observation,action,reward,done)
if counter > 64:
agent.train()
reward_per_episode+=reward
counter+=1
#check if episode ends:
if (done):
print ('EPISODE: ',str(i),' Steps: ',str(t),' Total Reward: ',str(reward_per_episode))
exploration_noise.reset() #reinitializing random noise for action exploration
reward_st = np.append(reward_st,reward_per_episode)
np.savetxt('episode_reward.txt',reward_st, newline="\n")
agent.actor_net.save_actor('/home/lee/Projects/Tracking/RL/weights/actor/model.ckpt')
agent.critic_net.save_critic('/home/lee/Projects/Tracking/RL/weights/critic/model.ckpt')
print ('\n\n')
break
def main():
env = Env(20000)
steps = 50
agent = DDPG(env, is_batch_norm)
counter = 0
exploration_noise = OUNoise(2)
reward_per_episode = 0
num_states = 96 * 4 + 4
num_actions = 2
reward_st = np.array([0])
agent.actor_net.load_actor(
'/home/myounghoe/ddpgtf/norepeat_target_2action_scale2/weights/actor/model.ckpt'
)
agent.critic_net.load_critic(
'/home/myounghoe/ddpgtf/norepeat_target_2action_scale2/weights/critic/model.ckpt'
)
for i in range(episodes):
print("==== Starting episode no:", str(i), "====", "\n")
observation = env.reset()
reward_per_episode = 0
for t in range(steps):
x = observation
action = agent.evaluate_actor(np.reshape(x, [1, num_states]))
noise = exploration_noise.noise()
action = action[0] # + noise
action = np.array([-1.0, 0.0])
observation, reward, done = env.step(action, t)
reward_per_episode += reward
counter += 1
#check if episode ends:
if (done):
print('EPISODE: ', str(i), ' Steps: ', str(t),
' Total Reward: ', str(reward_per_episode))
# print "Printing reward to file"
reward_st = np.append(reward_st, reward_per_episode)
np.savetxt('test_reward.txt', reward_st, newline="\n")
print('\n\n')
break
class DDPG:
def __init__(self, pretrain=False):
# Make sure all the directories exist
if not tf.gfile.Exists(TFLOG_PATH):
tf.gfile.MakeDirs(TFLOG_PATH)
if not tf.gfile.Exists(EXPERIENCE_PATH):
tf.gfile.MakeDirs(EXPERIENCE_PATH)
if not tf.gfile.Exists(NET_SAVE_PATH):
tf.gfile.MakeDirs(NET_SAVE_PATH)
# Initialize our session
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
self.session = tf.Session(config=config)
# self.session = tf.Session()
self.graph = self.session.graph
with self.graph.as_default():
# View the state batches
# self.visualize_input = VISUALIZE_BUFFER
# if self.visualize_input:
# self.viewer = CostmapVisualizer()
# Hardcode input size and action size
self.height = 662
self.width = 1
self.depth = 4
self.action_dim = 2
# Initialize the current action and the old action and old state for setting experiences
self.old_state = np.zeros((self.width, self.height, self.depth),
dtype='float32')
self.old_action = np.ones(2, dtype='float32')
self.network_action = np.zeros(2, dtype='float32')
self.noise_action = np.zeros(2, dtype='float32')
self.action = np.zeros(2, dtype='float32')
# Initialize the grad inverter object to keep the action bounds
self.grad_inv = GradInverter(A0_BOUNDS, A1_BOUNDS, self.session)
# Make sure the directory for the data files exists
if not tf.gfile.Exists(DATA_PATH):
tf.gfile.MakeDirs(DATA_PATH)
# Initialize summary writers to plot variables during training
self.summary_op = tf.summary.merge_all()
self.summary_writer = tf.summary.FileWriter(TFLOG_PATH)
# Initialize actor and critic networks
self.actor_network = ActorNetwork(self.height, self.action_dim,
self.depth, self.session,
self.summary_writer)
self.critic_network = CriticNetwork(self.height, self.action_dim,
self.depth, self.session,
self.summary_writer)
# Initialize the saver to save the network params
self.saver = tf.train.Saver()
# initialize the experience data manger
self.data_manager = DataManager(BATCH_SIZE, EXPERIENCE_PATH,
self.session)
# Uncomment if collecting a buffer for the autoencoder
# self.buffer = deque()
# Should we load the pre-trained params?
# If so: Load the full pre-trained net
# Else: Initialize all variables the overwrite the conv layers with the pretrained filters
if PRE_TRAINED_NETS:
self.saver.restore(self.session, NET_LOAD_PATH)
else:
self.session.run(tf.initialize_all_variables())
tf.train.start_queue_runners(sess=self.session)
time.sleep(1)
# Initialize a random process the Ornstein-Uhlenbeck process for action exploration
self.exploration_noise = OUNoise(self.action_dim, MU, THETA, SIGMA)
self.noise_flag = True
# Initialize time step
self.training_step = 0
# Flag: don't learn the first experience
self.first_experience = True
# After the graph has been filled add it to the summary writer
self.summary_writer.add_graph(self.graph)
def train(self):
# Check if the buffer is big enough to start training
if self.data_manager.enough_data():
# start_ = time.time()
# get the next random batch from the data manger
state_batch, \
action_batch, \
reward_batch, \
next_state_batch, \
is_episode_finished_batch = self.data_manager.get_next_batch()
state_batch = np.divide(state_batch, 10.0)
next_state_batch = np.divide(next_state_batch, 10.0)
# Are we visualizing the first state batch for debugging?
# If so: We have to scale up the values for grey scale before plotting
# if self.visualize_input:
# state_batch_np = np.asarray(state_batch)
# state_batch_np = np.multiply(state_batch_np, -100.0)
# state_batch_np = np.add(state_batch_np, 100.0)
# self.viewer.set_data(state_batch_np)
# self.viewer.run()
# self.visualize_input = False
# Calculate y for the td_error of the critic
# start = time.time()
y_batch = []
next_action_batch = self.actor_network.target_evaluate(
next_state_batch, action_batch)
q_value_batch = self.critic_network.target_evaluate(
next_state_batch, next_action_batch)
# done = time.time()
# elapsed = done - start
# print "forward actor and critic time is: ", elapsed
for i in range(0, BATCH_SIZE):
if is_episode_finished_batch[i]:
y_batch.append([reward_batch[i]])
else:
y_batch.append(reward_batch[i] + GAMMA * q_value_batch[i])
# Now that we have the y batch lets train the critic
# start = time.time()
self.critic_network.train(y_batch, state_batch, action_batch)
# done = time.time()
# elapsed = done - start
# print "train critic time is: ", elapsed
# self.critic_network.train(y_batch, state_batch, action_batch)
# Get the action batch so we can calculate the action gradient with it
# Then get the action gradient batch and adapt the gradient with the gradient inverting method
# start = time.time()
action_batch_for_gradients = self.actor_network.evaluate(
state_batch, action_batch)
# done = time.time()
# elapsed = done - start
# print "forward action after critic training time is: ", elapsed
q_gradient_batch = self.critic_network.get_action_gradient(
state_batch, action_batch_for_gradients)
q_gradient_batch = self.grad_inv.invert(
q_gradient_batch, action_batch_for_gradients)
# Now we can train the actor
# start = time.time()
self.actor_network.train(q_gradient_batch, state_batch,
action_batch)
# done = time.time()
# elapsed = done - start
# print "train actor time is: ", elapsed
# done = time.time()
# elapsed = done - start_
# print "====== total time is: ", elapsed
# Save model if necessary
if self.training_step > 0 and self.training_step % SAVE_STEP == 0:
self.saver.save(self.session,
NET_SAVE_PATH,
global_step=self.training_step)
# Update time step
self.training_step += 1
if self.training_step % 400 == 0:
print "iter: ", self.training_step
# start_ = time.time()
self.data_manager.check_for_enqueue()
# done = time.time()
# elapsed = done - start_
# print "############ check enqueue time is: ", elapsed
def get_action(self, state, old_action):
# normalize the state
state = state.astype(float)
state = np.divide(state, 10.0)
# Get the action
self.action = self.actor_network.get_action(state, old_action)
self.action = self.action.reshape((2, ))
# Are we using noise?
if self.noise_flag:
# scale noise down to 0 at training step 3000000
self.action = 0.8 * self.exploration_noise.noise()
# if self.training_step < MAX_NOISE_STEP:
# self.action += (MAX_NOISE_STEP - self.training_step) / \
# MAX_NOISE_STEP * self.exploration_noise.noise()
# if action value lies outside of action bounds, rescale the action vector
# if self.action[0] < A0_BOUNDS[0] or self.action[0] > A0_BOUNDS[1]:
# self.action *= np.fabs(A0_BOUNDS[0] / self.action[0])
# if self.action[1] < A0_BOUNDS[0] or self.action[1] > A0_BOUNDS[1]:
# self.action *= np.fabs(A1_BOUNDS[0] / self.action[1])
# Life q value output for this action and state
self.print_q_value(state, self.action)
return self.action
def set_experience(self, state, reward, is_episode_finished):
# Make sure we're saving a new old_state for the first experience of every episode
if self.first_experience:
self.first_experience = False
else:
state.astype('float32')
self.old_action.astype('float32')
self.old_action.astype('float32')
self.data_manager.store_experience_to_file(self.old_state,
self.old_action, reward,
state,
is_episode_finished)
# Uncomment if collecting data for the auto_encoder
# experience = (self.old_state, self.old_action, reward, state, is_episode_finished)
# self.buffer.append(experience)
if is_episode_finished:
self.first_experience = True
self.exploration_noise.reset()
# Safe old state and old action for next experience
self.old_state = state
self.old_action = self.action
def print_q_value(self, state, action):
string = "-"
q_value = self.critic_network.evaluate([state], [action])
stroke_pos = 30 * q_value[0][0] + 30
if stroke_pos < 0:
stroke_pos = 0
elif stroke_pos > 60:
stroke_pos = 60
class DDPG:
"""docstring for DDPG"""
def __init__(self, env):
self.name = 'DDPG' # name for uploading results
self.environment = env
# Randomly initialize actor network and critic network
# with both their target networks
self.state_dim = env.observation_space.shape[0]
self.action_dim = env.action_space.shape[0]
self.sess = tf.InteractiveSession()
self.actor_network = ActorNetwork(self.sess,self.state_dim,self.action_dim)
self.critic_network = CriticNetwork(self.sess,self.state_dim,self.action_dim)
# initialize replay buffer
self.replay_buffer = ReplayBuffer(REPLAY_BUFFER_SIZE)
# Initialize a random process the Ornstein-Uhlenbeck process for action exploration
self.exploration_noise = OUNoise(self.action_dim)
def train(self):
#print "train step",self.time_step
# Sample a random minibatch of N transitions from replay buffer
minibatch = self.replay_buffer.get_batch(BATCH_SIZE)
state_batch = np.asarray([data[0] for data in minibatch])
action_batch = np.asarray([data[1] for data in minibatch])
reward_batch = np.asarray([data[2] for data in minibatch])
next_state_batch = np.asarray([data[3] for data in minibatch])
done_batch = np.asarray([data[4] for data in minibatch])
# for action_dim = 1
action_batch = np.resize(action_batch,[BATCH_SIZE,self.action_dim])
# Calculate y_batch
next_action_batch = self.actor_network.target_actions(next_state_batch)
q_value_batch = self.critic_network.target_q(next_state_batch,next_action_batch)
y_batch = []
for i in range(len(minibatch)):
if done_batch[i]:
y_batch.append(reward_batch[i])
else :
y_batch.append(reward_batch[i] + GAMMA * q_value_batch[i])
y_batch = np.resize(y_batch,[BATCH_SIZE,1])
# Update critic by minimizing the loss L
self.critic_network.train(y_batch,state_batch,action_batch)
# Update the actor policy using the sampled gradient:
action_batch_for_gradients = self.actor_network.actions(state_batch)
q_gradient_batch = self.critic_network.gradients(state_batch,action_batch_for_gradients)
self.actor_network.train(q_gradient_batch,state_batch)
# Update the target networks
self.actor_network.update_target()
self.critic_network.update_target()
def noise_action(self,state):
# Select action a_t according to the current policy and exploration noise
action = self.actor_network.action(state)
return action+self.exploration_noise.noise()
def action(self,state):
action = self.actor_network.action(state)
return action
def perceive(self,state,action,reward,next_state,done):
# Store transition (s_t,a_t,r_t,s_{t+1}) in replay buffer
self.replay_buffer.add(state,action,reward,next_state,done)
# Store transitions to replay start size then start training
if self.replay_buffer.count() > REPLAY_START_SIZE:
self.train()
#if self.time_step % 10000 == 0:
#self.actor_network.save_network(self.time_step)
#self.critic_network.save_network(self.time_step)
# Re-iniitialize the random process when an episode ends
if done:
self.exploration_noise.reset()
def main():
experiment = 'MountainCarContinuous-v0'
env = gym.make(experiment)
steps = env.spec.timestep_limit
assert isinstance(env.observation_space, Box)
assert isinstance(env.action_space, Box)
agent = DDPG(env, is_batch_norm) #这个在循环前面，所以所有的weight都有继承
#也就是说，整个过程只训练了一个模型出来。
exploration_noise = OUNoise(env.action_space.shape[0])
reward_per_episode = 0
total_reward = 0
counter = 0
num_states = env.observation_space.shape[0] - 1
num_actions = env.action_space.shape[0]
#这是state的维度和action的维度
print 'Number of States:', num_states
print 'Number of Actions:', num_actions
print 'Number of steps per episode:', steps
if is_exploration == True:
print("\nExploration phase for {} steps. ".format(exploration_steps))
e_steps = 0
while e_steps < exploration_steps:
s = env.reset()
one_step = 0
done = False
exploration_noise.reset()
exp = []
while not done:
a = exploration_noise.noise()
ss, r, done, _ = env.step(a)
exp.append((s[:-1], a, ss[:-1], r, done))
s = ss
one_step += 1
if one_step > 998:
break
agent.add_experience(exp)
e_steps += 1
reward_st = np.array([0]) #这个是用来存每一次的rewards的
for i in xrange(episodes): #一共要循环1000次
print '====starting episode no:', i, '====', '\n'
observation = env.reset() #每个情节初始化，但是模型参数不初始化
reward_per_episode = 0
LSTM_SIZE = 40
statec_t1 = np.zeros((BATCH_SIZE, LSTM_SIZE))
stateh_t1 = np.zeros((BATCH_SIZE, LSTM_SIZE))
exp = []
for t in xrange(steps):
#env.render()
x = [observation[0:num_states]]
x = np.reshape(x * BATCH_SIZE, [BATCH_SIZE, num_states])
actor, statec_t1, stateh_t1 = agent.evaluate_actor(
x, statec_t1, stateh_t1)
noise = exploration_noise.noise()
#ra = random.random()
if (i < 500):
action = actor[0] + noise
else:
action = actor[0]
observation, reward, done, info = env.step(action)
#print 'Action at step',t,':',action,'reward:',reward,'\n'
exp.append((x, action, observation[0:num_states], reward, done))
if counter > 64:
agent.train()
counter += 1
reward_per_episode += reward
if (done or (t == steps - 1)):
#一个情节结束了～
agent.add_experience(exp)
print 'EPISODE:', i, 'Steps', t, 'Total Reward:', reward_per_episode
print 'Printing reward to file'
exploration_noise.reset()
reward_st = np.append(reward_st, reward_per_episode)
np.savetxt('episode_reward.txt', reward_st, newline='\n')
print '\n\n'
break
total_reward += reward_per_episode
#这里是计算平均值的
print "Average reward per episode {}".format(total_reward / episodes)
class RDPG:
"""docstring for RDPG"""
def __init__(self, env):
self.name = 'RDPG' # name for uploading results
self.environment = env
# Randomly initialize actor network and critic network
# with both their target networks
self.state_dim = env.observation_space.shape[0]
self.action_dim = env.action_space.shape[0]
self.sess = tf.InteractiveSession()
self.actor_network = ActorNetwork(self.sess, self.state_dim,
self.action_dim)
self.critic_network = CriticNetwork(self.sess, self.state_dim,
self.action_dim)
# initialize replay buffer
self.replay_buffer = ReplayBuffer(REPLAY_BUFFER_SIZE)
# Initialize a random process the Ornstein-Uhlenbeck process for action exploration
self.exploration_noise = OUNoise(self.action_dim)
self.saver = tf.train.Saver()
def train(self):
# Sample a random minibatch of N sequences from replay buffer
minibatch = self.replay_buffer.get_batch(BATCH_SIZE)
# Construct histories
observations = []
next_observations = []
actions = []
rewards = []
dones = []
for each in minibatch:
for i in range(1, len(each.observations)):
observations.append(self.pad(each.observations[0:i]))
next_observations.append(self.pad(each.observations[1, i + 1]))
actions.append(each.actions[0:i - 1])
rewards.append(each.rewards[0:i])
if i == len(each.observations) - 1:
dones.append(True)
else:
dones.append(False)
# Calculate y_batch
next_action_batch = self.actor_network.target_action(observations)
q_value_batch = self.critic_network.target_q(
next_observations,
[self.pad(i + j) for (i, j) in zip(actions, next_action_batch)])
y_batch = []
for i in range(len(observations)):
if dones[i]:
y_batch.append(rewards[i][-1])
else:
y_batch.append(rewards[i][-1] + GAMMA * q_value_batch[i])
y_batch = np.resize(y_batch, [len(observations), 1])
# Update critic by minimizing the loss L
self.critic_network.train(y_batch, observations,
[self.pad(i) for i in actions])
# Update the actor policy using the sampled gradient:
action_batch_for_gradients = self.actor_network.actions(observations)
q_gradient_batch = self.critic_network.gradients(
observations, action_batch_for_gradients)
self.actor_network.train(q_gradient_batch, observations)
# Update the target networks
self.actor_network.update_target()
self.critic_network.update_target()
def save_model(self, path, episode):
self.saver.save(self.sess, path + "modle.ckpt", episode)
def noise_action(self, history):
# Select action a_t according to a sequence of observation and action
action = self.actor_network.action(history)
return action + self.exploration_noise.noise()
def action(self, history):
action = self.actor_network.action(history)
return action
def perceive(self, history):
# Store the history sequence in the replay buffer
self.replay_buffer.add(history)
# Store history to replay start size then start training
if self.replay_buffer.count() > REPLAY_START_SIZE:
self.train()
# Re-iniitialize the random process when an episode ends
if done:
self.exploration_noise.reset()
def pad(self, input):
dim = len(input[0])
return input + [[0] * dim] * (1000 - len(input))
class DDPG:
"""docstring for DDPG"""
def __init__(self, environment):
self.name = 'DDPG' # name for uploading results
self.environment = environment
# Randomly initialize actor network and critic network
# with both their target networks
self.actor_network = ActorNetwork(
state_size=environment.observation_space.shape[0],
action_size=environment.action_space.shape[0])
self.critic_network = CriticNetwork(
state_size=environment.observation_space.shape[0],
action_size=environment.action_space.shape[0])
# initialize replay buffer
self.replay_buffer = deque()
# Initialize a random process the Ornstein-Uhlenbeck process for action exploration
self.exploration_noise = OUNoise(environment.action_space.shape[0])
# Initialize time step
self.time_step = 0
def set_init_observation(self, observation):
# receive initial observation state
self.state = observation
def train(self):
# Sample a random minibatch of N transitions from replay buffer
minibatch = random.sample(self.replay_buffer, BATCH_SIZE)
state_batch = [data[0] for data in minibatch]
action_batch = [data[1] for data in minibatch]
reward_batch = [data[2] for data in minibatch]
next_state_batch = [data[3] for data in minibatch]
action_batch = np.resize(action_batch, [BATCH_SIZE, 1])
# Calculate y
y_batch = []
next_action_batch = self.actor_network.target_evaluate(
next_state_batch)
q_value_batch = self.critic_network.target_evaluate(
next_state_batch, next_action_batch)
for i in range(0, BATCH_SIZE):
done = minibatch[i][4]
if done:
y_batch.append(reward_batch[i])
else:
y_batch.append(reward_batch[i] + GAMMA * q_value_batch[i])
# Update critic by minimizing the loss L
self.critic_network.train(y_batch, state_batch, action_batch)
# Update the actor policy using the sampled gradient:
action_batch_for_gradients = self.actor_network.evaluate(state_batch)
q_gradient_batch = self.critic_network.gradients(
state_batch, action_batch_for_gradients) / BATCH_SIZE
self.actor_network.train(q_gradient_batch, state_batch)
# Update the target networks
self.actor_network.update_target()
self.critic_network.update_target()
def get_action(self):
# Select action a_t according to the current policy and exploration noise
action = self.actor_network.get_action(self.state)
return np.clip(action + self.exploration_noise.noise(),
self.environment.action_space.low,
self.environment.action_space.high)
def set_feedback(self, observation, action, reward, done):
# Store transition (s_t,a_t,r_t,s_{t+1}) in replay buffer
next_state = observation
self.replay_buffer.append(
(self.state, action, reward, next_state, done))
# Update current state
self.state = next_state
# Update time step
self.time_step += 1
# Limit the replay buffer size
if len(self.replay_buffer) > REPLAY_BUFFER_SIZE:
self.replay_buffer.popleft()
# Store transitions to replay start size then start training
if self.time_step > REPLAY_START_SIZE:
self.train()
if self.time_step % 10000 == 0:
self.actor_network.save_network(self.time_step)
self.critic_network.save_network(self.time_step)
# Re-iniitialize the random process when an episode ends
if done:
self.exploration_noise.reset()
class PilotNode(object):
def __init__(self, model, logfolder):
print('initialize pilot node')
np.random.seed(FLAGS.random_seed)
tf.set_random_seed(FLAGS.random_seed)
# Initialize replay memory
self.logfolder = logfolder
self.world_name = ''
self.logfile = logfolder + '/tensorflow_log'
self.run = 0
self.run_eva = 0
self.maxy = -10
self.speed = FLAGS.speed
self.accumlosses = {}
self.current_distance = 0
self.furthest_point = 0
self.average_distance = 0
self.average_distance_eva = 0
self.last_pose = []
self.model = model
self.ready = False
self.finished = True
self.target_control = []
self.target_depth = []
self.target_odom = []
self.aux_depth = []
self.aux_odom = []
self.odom_error = []
self.prev_control = [0]
self.nfc_images = [
] #used by n_fc networks for building up concatenated frames
self.nfc_poses = [] #used by n_fc networks for calculating odometry
rospy.init_node('pilot', anonymous=True)
self.exploration_noise = OUNoise(4, 0, FLAGS.ou_theta, 1)
self.state = []
# self.delay_evaluation = 5 #can't be set by ros because node is started before ros is started...
if FLAGS.show_depth:
self.depth_pub = rospy.Publisher('/depth_prediction',
numpy_msg(Floats),
queue_size=1)
if FLAGS.show_odom:
self.odom_pub = rospy.Publisher('/odom_prediction',
numpy_msg(Floats),
queue_size=1)
# if FLAGS.off_policy:
# self.action_pub = rospy.Publisher('/supervised_vel', Twist, queue_size=1)
# if rospy.has_param('control'):
# rospy.Subscriber(rospy.get_param('control'), Twist, self.supervised_callback)
if FLAGS.real or FLAGS.off_policy:
self.action_pub = rospy.Publisher('/pilot_vel',
Twist,
queue_size=1)
else:
rospy.Subscriber('/supervised_vel', Twist,
self.supervised_callback)
if rospy.has_param('control'):
self.action_pub = rospy.Publisher(rospy.get_param('control'),
Twist,
queue_size=1)
if rospy.has_param('ready'):
rospy.Subscriber(rospy.get_param('ready'), Empty,
self.ready_callback)
if rospy.has_param('finished'):
rospy.Subscriber(rospy.get_param('finished'), Empty,
self.finished_callback)
if rospy.has_param('rgb_image') and not FLAGS.depth_input:
rospy.Subscriber(rospy.get_param('rgb_image'), Image,
self.image_callback)
if rospy.has_param('depth_image'):
if FLAGS.depth_input or FLAGS.auxiliary_depth or FLAGS.rl:
rospy.Subscriber(rospy.get_param('depth_image'), Image,
self.depth_callback)
if FLAGS.recovery_cameras:
# callbacks={'left':{'30':image_callback_left_30,'60':image_callback_left_60},'right':{'30':image_callback_right_30,'60':image_callback_right_60}}
# callbacks_depth={'left':{'30':depth_callback_left_30,'60':depth_callback_left_60},'right':{'30':depth_callback_right_30,'60':depth_callback_right_60}}
self.recovery_images = {}
for d in ['left', 'right']:
self.recovery_images[d] = {}
for c in ['30', '60']:
self.recovery_images[d][c] = {}
self.recovery_images[d][c]['rgb'] = []
self.recovery_images[d][c]['depth'] = []
rospy.Subscriber(
re.sub(r"kinect", "kinect_" + d + "_" + c,
rospy.get_param('rgb_image')), Image,
self.image_callback_recovery, (d, c))
rospy.Subscriber(
re.sub(r"kinect", "kinect_" + d + "_" + c,
rospy.get_param('depth_image')), Image,
self.depth_callback_recovery, (d, c))
if not FLAGS.real: # in simulation
self.replay_buffer = ReplayBuffer(FLAGS.buffer_size,
FLAGS.random_seed)
self.accumloss = 0
rospy.Subscriber('/ground_truth/state', Odometry, self.gt_callback)
def ready_callback(self, msg):
if not self.ready and self.finished:
print('Neural control activated.')
self.ready = True
self.start_time = rospy.get_time()
self.finished = False
self.exploration_noise.reset()
self.speed = FLAGS.speed + (
not FLAGS.evaluate) * np.random.uniform(
-FLAGS.sigma_x, FLAGS.sigma_x)
if rospy.has_param('evaluate') and not FLAGS.real:
# FLAGS.evaluate = False
FLAGS.evaluate = rospy.get_param('evaluate')
print '--> set evaluate to: ', FLAGS.evaluate
# if FLAGS.lstm:
# self.state=self.model.get_init_state(True)
# print 'set state to: ', self.state
if rospy.has_param('world_name'):
self.world_name = os.path.basename(
rospy.get_param('world_name').split('.')[0])
if 'sandbox' in self.world_name: self.world_name = 'sandbox'
def gt_callback(self, data):
if not self.ready: return
# Keep track of positions for logging
current_pos = [
data.pose.pose.position.x, data.pose.pose.position.y,
data.pose.pose.position.z
]
if len(self.last_pose) != 0:
self.current_distance += np.sqrt(
(self.last_pose[0, 3] - current_pos[0])**2 +
(self.last_pose[1, 3] - current_pos[1])**2)
self.furthest_point = max([
self.furthest_point,
np.sqrt(current_pos[0]**2 + current_pos[1]**2)
])
# Get pose (rotation and translation) for odometry
quaternion = (data.pose.pose.orientation.x,
data.pose.pose.orientation.y,
data.pose.pose.orientation.z,
data.pose.pose.orientation.w)
self.last_pose = transformations.quaternion_matrix(
quaternion
) # orientation of current frame relative to global frame
self.last_pose[0:3, 3] = current_pos
def process_rgb(self, msg):
# self.time_1 = time.time()
# if not self.ready or self.finished or (rospy.get_time()-self.start_time) < self.delay_evaluation: return
if not self.ready or self.finished: return []
try:
# Convert your ROS Image message to OpenCV2
im = bridge.imgmsg_to_cv2(
msg, 'rgb8'
) # changed to normal RGB order as i ll use matplotlib and PIL instead of opencv
# an idea could be to swap these channels during online training as this shouldnt matter though this could
# explain the performance drop coming from a pretrained network.
# This does mean that online trained nets might be worth nothing...
# im = bridge.imgmsg_to_cv2(msg, 'bgr8')
except CvBridgeError as e:
print(e)
else:
# self.time_2 = time.time()
size = self.model.input_size[1:]
im = sm.imresize(im, tuple(size), 'nearest')
# im = im*1/255.
# Basic preprocessing: center + make 1 standard deviation
# im -= FLAGS.mean
# im = im*1/FLAGS.std
return im
def process_depth(self, msg):
# if not self.ready or self.finished or (rospy.get_time()-self.start_time) < self.delay_evaluation: return
if not self.ready or self.finished: return []
try:
# Convert your ROS Image message to OpenCV2
im = bridge.imgmsg_to_cv2(msg, desired_encoding='passthrough'
) #gets float of 32FC1 depth image
except CvBridgeError as e:
print(e)
else:
im = im[::8, ::8]
shp = im.shape
# assume that when value is not a number it is due to a too large distance
# values can be nan for when they are closer than 0.5m but than the evaluate node should
# kill the run anyway.
im = np.asarray([
e * 1.0 if not np.isnan(e) else 5 for e in im.flatten()
]).reshape(shp) # clipping nans: dur: 0.010
# print 'min: ',np.amin(im),' and max: ',np.amax(im)
# im=np.asarray([ e*1.0 if not np.isnan(e) else 0 for e in im.flatten()]).reshape(shp) # clipping nans: dur: 0.010
# Resize image
if FLAGS.auxiliary_depth or FLAGS.rl:
size = self.model.depth_input_size #(55,74)
im = sm.imresize(im, size, 'nearest') # dur: 0.002
# cv2.imshow('depth', im) # dur: 0.002
if FLAGS.depth_input:
size = (self.model.input_size[1], self.model.input_size[1])
im = sm.imresize(im, size, 'nearest') # dur: 0.009
im = im[im.shape[0] / 2, :]
# cv2.imshow('depth', im.reshape(1,im.shape[0])) # dur: 0.002
# cv2.waitKey(2)
im = im * 1 / 255. * 5. # dur: 0.00004
return im
def image_callback(self, msg):
im = self.process_rgb(msg)
if len(im) != 0:
if FLAGS.n_fc:
self.nfc_images.append(im)
self.nfc_poses.append(copy.deepcopy(self.last_pose))
if len(self.nfc_images) < FLAGS.n_frames:
# print('filling concatenated frames: ',len(self.nfc_images))
return
else:
# concatenate last n-frames
im = np.concatenate(np.asarray(
self.nfc_images[-FLAGS.n_frames:]),
axis=2)
self.nfc_images = self.nfc_images[
-FLAGS.n_frames + 1:] # concatenate last n-1-frames
self.nfc_poses.pop(0) #get rid of the first one
assert len(self.nfc_poses) == FLAGS.n_frames - 1
# # calculate target odometry from previous global pose and current global pose
# euler = transformations.euler_from_matrix(self.nfc_poses[1], 'rxyz')
# # print 'current: ',str(euler[2]),str(self.nfc_poses[1][0,3]),str(self.nfc_poses[1][1,3])
# i_T_pg = transformations.inverse_matrix(self.nfc_poses[0])
# euler = transformations.euler_from_matrix(i_T_pg, 'rxyz')
# # print 'inverse prev: ',str(euler[2]), str(i_T_pg[0,3]),str(i_T_pg[1,3])
# T_cp = transformations.concatenate_matrices(i_T_pg, self.nfc_poses[1])
# r,p,yw = transformations.euler_from_matrix(T_cp, 'rxyz')
# x,y,z = T_cp[0:3,3]
# self.target_odom = [x,y,z,r,p,yw]
# print 'odom: ',str(self.target_odom[5]),str(self.target_odom[0]),str(self.target_odom[1])
# self.target_odom = [self.nfc_poses[1][i]-self.nfc_poses[0][i] for i in range(len(self.nfc_poses[0]))]
# print 'Target odometry: ', self.target_odom
self.process_input(im)
def image_callback_recovery(self, msg, args):
im = self.process_rgb(msg)
if len(im) == 0: return
trgt = -100.
if FLAGS.auxiliary_depth and len(
self.recovery_images[args[0]][args[1]]['depth']) == 0:
print("No target depth: {0} {1}".format(args[0], args[1]))
return
else:
trgt_depth = copy.deepcopy(
self.recovery_images[args[0]][args[1]]['depth'])
if len(self.target_control) == 0:
print("No target control: {0} {1}".format(args[0], args[1]))
return
else:
# left ==> -1, right ==> +1, 30dg ==> 0.5, 60dg ==> 1.0
compensation = -(args[0] == 'left') * int(
args[1]) / 60. + (args[0] == 'right') * int(args[1]) / 60.
trgt = compensation + self.target_control[5]
if FLAGS.experience_replay and not FLAGS.evaluate and trgt != -100:
if FLAGS.auxiliary_depth:
print('added experience of camera: {0} {1} with control {2}'.
format(args[0], args[1], trgt))
self.replay_buffer.add(im, [trgt], [trgt_depth])
else:
self.replay_buffer.add(im, [trgt])
def depth_callback(self, msg):
im = self.process_depth(msg)
if len(im) != 0:
if FLAGS.auxiliary_depth or FLAGS.rl:
self.target_depth = im #(64,)
if FLAGS.depth_input:
if FLAGS.network == 'nfc_control':
self.nfc_images.append(im)
if len(self.nfc_images) < 4:
# print('filling concatenated frames: ',len(self.nfc_images))
return
else:
# print np.asarray(self.nfc_images).shape
im = np.concatenate(np.asarray(self.nfc_images))
# print im.shape
self.nfc_images.pop(0)
self.process_input(im)
def depth_callback_recovery(self, msg, args):
im = self.process_depth(msg)
self.recovery_images[args[0]][args[1]]['depth'] = im
def process_input(self, im):
self.time_3 = time.time()
trgt = -100.
# if self.target_control == None or FLAGS.evaluate:
if FLAGS.evaluate: ### EVALUATE
trgt_depth = []
trgt_odom = []
with_loss = False
if len(
self.target_control
) != 0 and not FLAGS.auxiliary_depth and not FLAGS.auxiliary_odom:
trgt = self.target_control[5]
with_loss = True
elif len(self.target_control
) != 0 and FLAGS.auxiliary_depth and len(
self.target_depth) != 0 and not FLAGS.auxiliary_odom:
trgt = self.target_control[5]
trgt_depth = [copy.deepcopy(self.target_depth)]
with_loss = True
elif len(
self.target_control
) != 0 and not FLAGS.auxiliary_depth and FLAGS.auxiliary_odom and len(
self.target_odom) != 0:
trgt = self.target_control[5]
trgt_odom = [copy.deepcopy(self.target_odom)]
with_loss = True
elif len(
self.target_control
) != 0 and FLAGS.auxiliary_depth and len(
self.target_depth) != 0 and FLAGS.auxiliary_odom and len(
self.target_odom) != 0:
trgt = self.target_control[5]
trgt_odom = [copy.deepcopy(self.target_odom)]
trgt_depth = [copy.deepcopy(self.target_depth)]
with_loss = True
if with_loss and False: # for now skip calculating accumulated loses.
prev_ctr = [[self.prev_control[0]]]
control, self.state, losses, aux_results = self.model.forward(
[[im]] if FLAGS.lstm else [im],
states=self.state,
auxdepth=FLAGS.show_depth,
auxodom=FLAGS.show_odom,
prev_action=prev_ctr,
targets=[[trgt]],
target_depth=trgt_depth,
target_odom=trgt_odom)
if len(self.accumlosses.keys()) == 0:
self.accumlosses = losses
else:
# self.accumlosses=[self.accumlosses[i]+losses[i] for i in range(len(losses))]
for v in losses.keys():
self.accumlosses[v] = self.accumlosses[v] + losses[v]
else:
prev_ctr = [[self.prev_control[0]]]
control, self.state, losses, aux_results = self.model.forward(
[[im]] if FLAGS.lstm else [im],
states=self.state,
auxdepth=FLAGS.show_depth,
auxodom=FLAGS.show_odom,
prev_action=prev_ctr)
if FLAGS.show_depth and FLAGS.auxiliary_depth and len(
aux_results) > 0:
self.aux_depth = aux_results['depth']
if FLAGS.show_odom and FLAGS.auxiliary_odom and len(
aux_results) > 0:
self.aux_odom = aux_results['odom']
else: ###TRAINING
# Get necessary labels, if label is missing wait...
if len(self.target_control) == 0:
print('No target control')
return
else:
trgt = self.target_control[5]
# print(trgt)
if (FLAGS.auxiliary_depth or FLAGS.rl) and len(
self.target_depth) == 0:
print('No target depth')
return
else:
trgt_depth = copy.deepcopy(self.target_depth)
# self.target_depth = []
if FLAGS.auxiliary_odom and (len(self.target_odom) == 0
or len(self.prev_control) == 0):
print('no target odometry or previous control')
return
else:
trgt_odom = copy.deepcopy(self.target_odom)
# check if depth image corresponds to rgb image
# cv2.imshow('rgb', im)
# cv2.waitKey(2)
# cv2.imshow('depth', trgt_depth*1/5.)
# cv2.waitKey(2)
# ---------------------------------------------------------- DEPRECATED
# if not FLAGS.experience_replay: ### TRAINING WITHOUT EXPERIENCE REPLAY
# if FLAGS.auxiliary_depth:
# control, losses = self.model.backward([im],[[trgt]], [[[trgt_depth]]])
# else:
# control, losses = self.model.backward([im],[[trgt]])
# print 'Difference: '+str(control[0,0])+' and '+str(trgt)+'='+str(abs(control[0,0]-trgt))
# self.accumlosses += losses[0]
# else: ### TRAINING WITH EXPERIENCE REPLAY
# wait for first target depth in case of auxiliary depth.
# in case the network can predict the depth
self.time_4 = time.time()
prev_ctr = [[self.prev_control[0]]]
control, self.state, losses, aux_results = self.model.forward(
[[im]] if FLAGS.lstm else [im],
states=self.state,
auxdepth=FLAGS.show_depth,
auxodom=FLAGS.show_odom,
prev_action=prev_ctr)
if FLAGS.show_depth and FLAGS.auxiliary_depth:
self.aux_depth = aux_results['depth']
if FLAGS.show_odom and FLAGS.auxiliary_odom:
self.aux_odom = aux_results['odom']
self.time_5 = time.time()
# print 'state: ', self.state
### SEND CONTROL
noise = self.exploration_noise.noise()
# yaw = control[0,0]
# if np.random.binomial(1,FLAGS.epsilon) and not FLAGS.evaluate:
# yaw = max(-1,min(1,np.random.normal()))
if trgt != 100 and not FLAGS.evaluate:
action = trgt if np.random.binomial(1, FLAGS.alpha**
self.run) else control[0, 0]
else:
action = control[0, 0]
msg = Twist()
if FLAGS.type_of_noise == 'ou':
msg.linear.x = self.speed #0.8 # 1.8 #
# msg.linear.x = FLAGS.speed+(not FLAGS.evaluate)*FLAGS.sigma_x*noise[0] #0.8 # 1.8 #
msg.linear.y = (not FLAGS.evaluate) * noise[1] * FLAGS.sigma_y
msg.linear.z = (not FLAGS.evaluate) * noise[2] * FLAGS.sigma_z
msg.angular.z = max(
-1,
min(1, action +
(not FLAGS.evaluate) * FLAGS.sigma_yaw * noise[3]))
elif FLAGS.type_of_noise == 'uni':
msg.linear.x = self.speed
# msg.linear.x = FLAGS.speed + (not FLAGS.evaluate)*np.random.uniform(-FLAGS.sigma_x, FLAGS.sigma_x)
msg.linear.y = (not FLAGS.evaluate) * np.random.uniform(
-FLAGS.sigma_y, FLAGS.sigma_y)
msg.linear.z = (not FLAGS.evaluate) * np.random.uniform(
-FLAGS.sigma_z, FLAGS.sigma_z)
msg.angular.z = max(
-1,
min(
1, action + (not FLAGS.evaluate) *
np.random.uniform(-FLAGS.sigma_yaw, FLAGS.sigma_yaw)))
else:
raise IOError('Type of noise is unknown: {}'.format(
FLAGS.type_of_noise))
self.action_pub.publish(msg)
self.prev_control = [msg.angular.z]
self.time_6 = time.time()
if FLAGS.show_depth and len(self.aux_depth) != 0 and not self.finished:
# print('shape aux depth: {}'.format(self.aux_depth.shape))
self.aux_depth = self.aux_depth.flatten()
self.depth_pub.publish(self.aux_depth)
self.aux_depth = []
if FLAGS.show_odom and len(self.aux_odom) != 0 and not self.finished:
# trgt_odom = [copy.deepcopy(self.target_odom)]
# final_img = cv2.hconcat((im[:,:,0:3], im[:,:,3:6],im[:,:,6:]))
# final_img = cv2.hconcat((im[:,:,[2,1,0]], im[:,:,[5,4,3]],im[:,:,[8,7,6]]))
# print trgt_odom
# cv2.imshow('Final', final_img)
# cv2.waitKey(100)
# cv2.destroyAllWindows()
concat_odoms = np.concatenate(
(self.aux_odom.astype(np.float32).flatten(),
np.array(trgt_odom).astype(np.float32).flatten()))
# self.odom_pub.publish(self.aux_odom.flatten())
# print concat_odoms[4:6],' and ',concat_odoms[0:2]
self.odom_pub.publish(concat_odoms.astype(np.float32))
# self.odom_error.append(np.abs(np.array(trgt_odom).flatten()-self.aux_odom.flatten()))
self.aux_odom = []
# ADD EXPERIENCE REPLAY
if FLAGS.experience_replay and not FLAGS.evaluate and trgt != -100:
aux_info = {}
if FLAGS.auxiliary_depth or FLAGS.rl:
aux_info['target_depth'] = trgt_depth
if FLAGS.auxiliary_odom:
# print trgt_odom
# print 'target odom ',trgt_odom
aux_info['target_odom'] = trgt_odom
aux_info['prev_action'] = prev_ctr
if FLAGS.lstm:
# aux_info['state']=(np.zeros(()))
# state type: <type 'tuple'> len: 2 len sub: 2 len subsub: 1 len subsubsub: 100
aux_info['state'] = self.state
# aux_info['state']=((np.zeros((1,100)),np.zeros((1,100))+10),(np.ones((1,100)),np.ones((1,100))+20))
# print aux_info['state']
# (state layer0,output layer0,state layer1,output layer1)
# print 'state type: ',type(aux_info['state']),' len: ', len(aux_info['state']),' len sub: ', len(aux_info['state'][0]),' len subsub: ', len(aux_info['state'][0][0]),' len subsubsub: ', len(self.state[0][0][0])
self.replay_buffer.add(im, [trgt], aux_info=aux_info)
self.time_7 = time.time()
if FLAGS.save_input:
self.depthfile = open(self.logfolder + '/depth_input', 'a')
np.set_printoptions(precision=5)
message = "{0} : {1} : {2:.4f} \n".format(
self.run,
' '.join('{0:.5f}'.format(k) for k in np.asarray(im)), trgt)
self.depthfile.write(message)
self.depthfile.close()
self.time_8 = time.time()
# print 'processed image @: {0:.2f}'.format(time.time())
# print("Time debugging: \n cvbridge: {0} , \n resize: {1}, \n copy: {2} , \n net pred: {3}, \n pub: {4},\n exp buf: {5},\n pos file: {6} s".format((self.time_2-self.time_1),
# (self.time_3-self.time_2),(self.time_4-self.time_3),(self.time_5-self.time_4),(self.time_6-self.time_5),(self.time_7-self.time_6),(self.time_8-self.time_7)))
# Delay values with auxiliary depth (at the beginning of training)
# cv bridge (RGB): 0.0003s
# resize (RGB): 0.0015s
# copy control+depth: 2.7e-5 s
# net prediction: 0.011s
# publication: 0.0002s
# fill experience buffer: 1.8e-5 s
# write position: 2.1e-6 s
def supervised_callback(self, data):
if not self.ready: return
self.target_control = [
data.linear.x, data.linear.y, data.linear.z, data.angular.x,
data.angular.y, data.angular.z
]
def finished_callback(self, msg):
if self.ready and not self.finished:
# self.depth_pub.publish(self.aux_depth)
print('neural control deactivated.')
self.ready = False
self.finished = True
# Train model from experience replay:
# Train the model with batchnormalization out of the image callback loop
activation_images = []
depth_predictions = []
endpoint_activations = []
tloss = [] #total loss
closs = [] #control loss
dloss = [] #depth loss
oloss = [] #odometry loss
qloss = [] #RL cost-to-go loss
tlossm, clossm, dlossm, olossm, qlossm, tlossm_eva, clossm_eva, dlossm_eva, olossm_eva, qlossm_eva = 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
#tot_batch_loss = []
if FLAGS.experience_replay and self.replay_buffer.size() > (
FLAGS.batch_size if not FLAGS.lstm else FLAGS.batch_size *
FLAGS.num_steps) and not FLAGS.evaluate:
for b in range(
min(int(self.replay_buffer.size() / FLAGS.batch_size),
10)):
inputs, targets, aux_info = self.replay_buffer.sample_batch(
FLAGS.batch_size)
# import pdb; pdb.set_trace()
#print('time to smaple batch of images: ',time.time()-st)
if b == 0:
if FLAGS.plot_activations:
activation_images = self.model.plot_activations(
inputs, targets.reshape((-1, 1)))
if FLAGS.plot_depth and FLAGS.auxiliary_depth:
depth_predictions = self.model.plot_depth(
inputs,
aux_info['target_depth'].reshape(-1, 55, 74))
if FLAGS.plot_histograms:
endpoint_activations = self.model.get_endpoint_activations(
inputs)
init_state = []
depth_targets = []
odom_targets = []
prev_action = []
if FLAGS.lstm:
init_state = (aux_info['state'][:, 0, 0, 0, 0, :],
aux_info['state'][:, 0, 0, 1, 0, :],
aux_info['state'][:, 0, 1, 0, 0, :],
aux_info['state'][:, 0, 1, 1, 0, :])
# if FLAGS.use_init_state:
# init_state=
assert init_state[0].shape[0] == FLAGS.batch_size
# print 'init_state sizes ',init_state[0].shape
if FLAGS.auxiliary_depth or FLAGS.rl:
depth_targets = aux_info['target_depth'].reshape(
-1, 55, 74)
# depth_targets=aux_info['target_depth'].reshape(-1,55,74) if not FLAGS.lstm else aux_info['target_depth'].reshape(-1,FLAGS.num_steps, 55,74)
if FLAGS.auxiliary_odom:
odom_targets = aux_info['target_odom'].reshape(
-1, 4) if not FLAGS.lstm else aux_info[
'target_odom'].reshape(-1, FLAGS.num_steps, 4)
# odom_targets=aux_info['target_odom'].reshape(-1,6) if not FLAGS.lstm else aux_info['target_odom'].reshape(-1,FLAGS.num_steps, 6)
prev_action = aux_info['prev_action'].reshape(
-1, 1
) #if not FLAGS.lstm else aux_info['prev_action'].reshape(-1,FLAGS.num_steps, 1)
# todo add initial state for each rollout in the batch
controls, losses = self.model.backward(
inputs, init_state, targets[:].reshape(-1, 1),
depth_targets, odom_targets, prev_action)
tloss = losses['t']
if not FLAGS.rl or FLAGS.auxiliary_ctr: closs = losses['c']
if FLAGS.auxiliary_depth: dloss.append(losses['d'])
if FLAGS.auxiliary_odom: oloss.append(losses['o'])
if FLAGS.rl: qloss.append(losses['q'])
tlossm = np.mean(tloss)
clossm = np.mean(
closs) if not FLAGS.rl or FLAGS.auxiliary_ctr else 0
dlossm = np.mean(dloss) if FLAGS.auxiliary_depth else 0
olossm = np.mean(oloss) if FLAGS.auxiliary_odom else 0
qlossm = np.mean(qloss) if FLAGS.rl else 0
else:
print('Evaluating or filling buffer or no experience_replay: ',
self.replay_buffer.size())
if 't' in self.accumlosses.keys():
tlossm_eva = self.accumlosses['t']
if 'c' in self.accumlosses.keys():
clossm_eva = self.accumlosses['c']
if 'd' in self.accumlosses.keys():
dlossm_eva = self.accumlosses['d']
if 'o' in self.accumlosses.keys():
olossm_eva = self.accumlosses['o']
if 'q' in self.accumlosses.keys():
qlossm_eva = self.accumlosses['q']
if not FLAGS.evaluate:
self.average_distance = self.average_distance - self.average_distance / (
self.run + 1)
self.average_distance = self.average_distance + self.current_distance / (
self.run + 1)
else:
self.average_distance_eva = self.average_distance_eva - self.average_distance_eva / (
self.run_eva + 1)
self.average_distance_eva = self.average_distance_eva + self.current_distance / (
self.run_eva + 1)
odom_errx, odom_erry, odom_errz, odom_erryaw = 0, 0, 0, 0
if len(self.odom_error) != 0:
odom_errx = np.mean([e[0] for e in self.odom_error])
odom_erry = np.mean([e[1] for e in self.odom_error])
odom_errz = np.mean([e[2] for e in self.odom_error])
odom_erryaw = np.mean([e[3] for e in self.odom_error])
try:
sumvar = {}
# sumvar={k : 0 for k in self.model.summary_vars.keys()}
sumvar["Distance_current_" +
self.world_name if len(self.world_name) != 0 else
"Distance_current"] = self.current_distance
sumvar["Distance_furthest_" +
self.world_name if len(self.world_name) != 0 else
"Distance_furthest"] = self.furthest_point
if FLAGS.evaluate:
sumvar["Distance_average_eva"] = self.average_distance_eva
else:
sumvar["Distance_average"] = self.average_distance
if tlossm != 0: sumvar["Loss_total"] = tlossm
if clossm != 0: sumvar["Loss_control"] = clossm
if dlossm != 0: sumvar["Loss_depth"] = dlossm
if olossm != 0: sumvar["Loss_odom"] = olossm
if qlossm != 0: sumvar["Loss_q"] = qlossm
if tlossm_eva != 0: sumvar["Loss_total_eva"] = tlossm_eva
if clossm_eva != 0: sumvar["Loss_control_eva"] = clossm_eva
if dlossm_eva != 0: sumvar["Loss_depth_eva"] = dlossm_eva
if olossm_eva != 0: sumvar["Loss_odom_eva"] = olossm_eva
if qlossm_eva != 0: sumvar["Loss_q_eva"] = qlossm_eva
if odom_errx != 0: sumvar["odom_errx"] = odom_errx
if odom_erry != 0: sumvar["odom_erry"] = odom_erry
if odom_errz != 0: sumvar["odom_errz"] = odom_errz
if odom_erryaw != 0: sumvar["odom_erryaw"] = odom_erryaw
if FLAGS.plot_activations and len(activation_images) != 0:
sumvar["conv_activations"] = activation_images
# sumvar.append(activation_images)
if FLAGS.plot_depth and FLAGS.auxiliary_depth:
sumvar["depth_predictions"] = depth_predictions
# sumvar.append(depth_predictions)
if FLAGS.plot_histograms:
for i, ep in enumerate(self.model.endpoints):
sumvar['activations_{}'.format(
ep)] = endpoint_activations[i]
# sumvar.extend(endpoint_activations)
self.model.summarize(sumvar)
except Exception as e:
print('failed to write', e)
pass
else:
print(
'{0}: control finished {1}:[ current_distance: {2:0.3f}, average_distance: {3:0.3f}, furthest point: {4:0.1f}, total loss: {5:0.3f}, control loss: {6:0.3e}, depth loss: {7:0.3e}, odom loss: {8:0.3e}, q loss: {9:0.3e}, world: {10}'
.format(
time.strftime('%H:%M'),
self.run if not FLAGS.evaluate else self.run_eva,
self.current_distance, self.average_distance
if not FLAGS.evaluate else self.average_distance_eva,
self.furthest_point,
tlossm if not FLAGS.evaluate else tlossm_eva,
clossm if not FLAGS.evaluate else clossm_eva,
dlossm if not FLAGS.evaluate else dlossm_eva,
olossm if not FLAGS.evaluate else olossm_eva,
qlossm if not FLAGS.evaluate else qlossm_eva,
self.world_name))
l_file = open(self.logfile, 'a')
tag = 'train'
if FLAGS.evaluate:
tag = 'val'
l_file.write(
'{0} {1} {2} {3} {4} {5} {6} {7} {8} {9} {10}\n'.format(
self.run if not FLAGS.evaluate else self.run_eva,
self.current_distance, self.average_distance
if not FLAGS.evaluate else self.average_distance_eva,
self.furthest_point, tlossm, clossm, dlossm, olossm,
qlossm, tag, self.world_name))
l_file.close()
self.accumlosses = {}
self.maxy = -10
self.current_distance = 0
self.last_pose = []
self.nfc_images = []
self.nfc_poses = []
self.furthest_point = 0
if FLAGS.lstm and not FLAGS.evaluate: self.replay_buffer.new_run()
self.world_name = ''
if self.run % 10 == 0 and not FLAGS.evaluate:
# Save a checkpoint every 20 runs.
self.model.save(self.logfolder)
self.state = []
if not FLAGS.evaluate:
self.run += 1
else:
self.run_eva += 1
# wait for gzserver to be killed
gzservercount = 1
while gzservercount > 0:
#print('gzserver: ',gzservercount)
gzservercount = os.popen("ps -Af").read().count('gzserver')
time.sleep(0.1)
sys.stdout.flush()
def s2l():
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
num_states = feature_size #num_states = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
print("Number of States:", num_states)
print("Number of Actions:", num_actions)
agent = Saved_Policy(num_states, num_actions)
exploration_noise = OUNoise(env.action_space.shape[0])
counter = 0
total_reward = 0
print("Number of Rollouts per episode:", num_rollouts)
print("Number of Steps per roll out:", steps)
reward_st = np.array([0]) #saving reward
reward_st_all = np.array([0]) #saving reward after every step
eval_metric_st = np.array([0]) #saving evalutation metric
activity_obj = Vid_Feature()
demo_vid_array = demo_array_extractor(demo_folder)
demo_features = activity_obj.feature_extractor(demo_vid_array)
frame_obj = Frame_Feature()
for episode in range(num_episodes):
print("==== Starting episode no:", episode, "====", "\n")
env.reset() # Reset env in the begining of each episode
env.render()
obs_img = env.render(mode='rgb_array') # Get the observation
obs_img = np.array(misc.imresize(obs_img, [112, 112, 3]))
observation = np.array(frame_obj.frame_feature_extractor(obs_img))
observation = observation.reshape(-1)
reward_per_episode = 0
for t in range(num_rollouts):
reward_per_rollout = 0
vid_robo_ = []
for i in range(steps):
x = observation
action = agent.get_action(np.reshape(x, [1, num_states]))
noise = exploration_noise.noise()
#action = action[0] + noise #Select action according to current policy and exploration noise
print('Action at episode-', episode, 'rollout-', t, 'step-', i,
" :", action)
_, _, done, info = env.step(action)
env.render()
obs_robo_ = env.render(mode='rgb_array') # Get the observation
obs_robo = misc.imresize(obs_robo_, [112, 112, 3])
vid_robo_.append(obs_robo)
observation = np.array(
frame_obj.frame_feature_extractor(np.array(obs_robo)))
observation = observation.reshape(-1)
#pasue()
# Printing eval_metric after every rollout
eval_metric = np.array(env.get_eval())
eval_metric = eval_metric.reshape(-1)
print('Distance to goal:', eval_metric)
eval_metric_st = np.append(eval_metric_st, eval_metric)
np.savetxt('test_eval_metric_per_rollout.txt',
eval_metric_st,
newline="\n")
class DDPG(object):
def __init__(self, env):
self.action_dim = env.action_space.shape[0]
self.state_dim = env.observation_space.shape[0]
self.h1_dim = 400
self.h2_dim = 300
self.actor_learning_rate = 1e-4
self.critic_learning_rate = 1e-3
self.gamma = 0.99
# Ornstein-Uhlenbeck noise parameters
self.noise_theta = 0.15
self.noise_sigma = 0.20
self.ou = OUNoise(self.action_dim,
theta=self.noise_theta,
sigma=self.noise_sigma)
self.replay_buffer_size = 1000000
self.replay_buffer = deque(maxlen=self.replay_buffer_size)
self.replay_start_size = 1000
self.batch_size = 64
self.target_update_rate = 0.001
self.total_parameters = 0
self.global_steps = 0
self.reg_param = 0.01
def construct_model(self, gpu):
if gpu == -1: # use CPU
device = '/cpu:0'
sess_config = tf.ConfigProto()
else: # use GPU
device = '/gpu:' + str(gpu)
sess_config = tf.ConfigProto(log_device_placement=True,
allow_soft_placement=True)
sess_config.gpu_options.allow_growth = True
self.sess = tf.Session(config=sess_config)
with tf.device(device):
# output action and q_value and compute gradients of q_val
# w.r.t. action
with tf.name_scope('predict_actions'):
self.states = tf.placeholder(tf.float32,
[None, self.state_dim],
name='states')
self.action = tf.placeholder(tf.float32,
[None, self.action_dim],
name='action')
self.is_training = tf.placeholder(tf.bool, name='is_training')
with tf.variable_scope('actor_network'):
self.action_outputs, self.actor_params = \
self.actor(self.states, bn=True)
with tf.variable_scope('critic_network'):
self.value_outputs, self.critic_params = \
self.critic(self.states, self.action, bn=False)
self.action_gradients = tf.gradients(
self.value_outputs, self.action)[0]
# estimate target_q for update critic
with tf.name_scope('estimate_target_q'):
self.next_states = tf.placeholder(tf.float32,
[None, self.state_dim],
name='next_states')
self.mask = tf.placeholder(tf.float32, [
None,
], name='mask')
self.rewards = tf.placeholder(tf.float32, [
None,
],
name='rewards')
with tf.variable_scope('target_actor_network'):
self.target_actor_outputs, self.target_actor_params = \
self.actor(self.next_states, bn=True)
with tf.variable_scope('target_critic_network'):
self.target_value_outputs, self.target_critic_params = \
self.critic(self.next_states,
self.target_actor_outputs,
bn=False)
self.target_q = self.rewards + self.gamma * \
(self.target_value_outputs[:, 0] * self.mask)
with tf.name_scope('compute_gradients'):
# optimizer
self.actor_optimizer = tf.train.AdamOptimizer(
self.actor_learning_rate)
self.critic_optimizer = tf.train.AdamOptimizer(
self.critic_learning_rate)
# critic gradients
td_error = self.target_q - self.value_outputs[:, 0]
critic_mse = tf.reduce_mean(tf.square(td_error))
critic_reg = tf.reduce_sum(
[tf.nn.l2_loss(v) for v in self.critic_params])
critic_loss = critic_mse + self.reg_param * critic_reg
self.critic_gradients = \
self.critic_optimizer.compute_gradients(
critic_loss, self.critic_params)
# actor gradients
self.q_action_grads = tf.placeholder(tf.float32,
[None, self.action_dim],
name='q_action_grads')
actor_gradients = tf.gradients(self.action_outputs,
self.actor_params,
-self.q_action_grads)
self.actor_gradients = zip(actor_gradients, self.actor_params)
# apply gradient to update model
self.train_actor = self.actor_optimizer.apply_gradients(
self.actor_gradients)
self.train_critic = self.critic_optimizer.apply_gradients(
self.critic_gradients)
with tf.name_scope('update_target_networks'):
# batch norm paramerters should not be included when updating!
target_networks_update = []
for v_source, v_target in zip(self.actor_params,
self.target_actor_params):
update_op = v_target.assign_sub(self.target_update_rate *
(v_target - v_source))
target_networks_update.append(update_op)
for v_source, v_target in zip(self.critic_params,
self.target_critic_params):
update_op = v_target.assign_sub(self.target_update_rate *
(v_target - v_source))
target_networks_update.append(update_op)
self.target_networks_update = tf.group(*target_networks_update)
with tf.name_scope('total_numbers_of_parameters'):
for v in tf.trainable_variables():
shape = v.get_shape()
param_num = 1
for d in shape:
param_num *= d.value
print(v.name, ' ', shape, ' param nums: ', param_num)
self.total_parameters += param_num
print('Total nums of parameters: ', self.total_parameters)
def sample_action(self, states, explore):
# is_training suppose to be False when sampling action!!!
action = self.sess.run(self.action_outputs, {
self.states: states,
self.is_training: False
})
ou_noise = self.ou.noise() if explore else 0
return action + ou_noise
def store_experience(self, s, a, r, next_s, done):
self.replay_buffer.append([s, a[0], r, next_s, done])
self.global_steps += 1
def update_model(self):
if len(self.replay_buffer) < self.replay_start_size:
return
# get batch
batch = random.sample(self.replay_buffer, self.batch_size)
s, _a, r, next_s, done = np.vstack(batch).T.tolist()
mask = -np.array(done)
# compute a = u(s)
a = self.sess.run(self.action_outputs, {
self.states: s,
self.is_training: True
})
# gradients of q_value w.r.t action a
dq_da = self.sess.run(self.action_gradients, {
self.states: s,
self.action: a,
self.is_training: True
})
# train
self.sess.run(
[self.train_actor, self.train_critic],
{
# train_actor feed
self.states: s,
self.is_training: True,
self.q_action_grads: dq_da,
# train_critic feed
self.next_states: next_s,
self.action: _a,
self.mask: mask,
self.rewards: r
})
# update target network
self.sess.run(self.target_networks_update)
def actor(self, states, bn=False):
init = tf.contrib.layers.variance_scaling_initializer(factor=1.0,
mode='FAN_IN',
uniform=True)
if bn:
states = self.batch_norm(states,
self.is_training,
tf.identity,
scope='actor_bn_states')
w1 = tf.get_variable('w1', [self.state_dim, self.h1_dim],
initializer=init)
b1 = tf.get_variable('b1', [self.h1_dim], initializer=init)
h1 = tf.matmul(states, w1) + b1
if bn:
h1 = self.batch_norm(h1,
self.is_training,
tf.nn.relu,
scope='actor_bn_h1')
w2 = tf.get_variable('w2', [self.h1_dim, self.h2_dim],
initializer=init)
b2 = tf.get_variable('b2', [self.h2_dim], initializer=init)
h2 = tf.matmul(h1, w2) + b2
if bn:
h2 = self.batch_norm(h2,
self.is_training,
tf.nn.relu,
scope='actor_bn_h2')
# use tanh to bound the action
w3 = tf.get_variable('w3', [self.h2_dim, self.action_dim],
initializer=tf.random_uniform_initializer(
-3e-3, 3e-3))
b3 = tf.get_variable('b3', [self.action_dim],
initializer=tf.random_uniform_initializer(
-3e-4, 3e-4))
a = tf.nn.tanh(tf.matmul(h2, w3) + b3)
return a, [w1, b1, w2, b2, w3, b3]
def critic(self, states, action, bn=False):
init = tf.contrib.layers.variance_scaling_initializer(factor=1.0,
mode='FAN_IN',
uniform=True)
if bn:
states = self.batch_norm(states,
self.is_training,
tf.identity,
scope='critic_bn_state')
w1 = tf.get_variable('w1', [self.state_dim, self.h1_dim],
initializer=init)
b1 = tf.get_variable('b1', [self.h1_dim], initializer=init)
h1 = tf.matmul(states, w1) + b1
if bn:
h1 = self.batch_norm(h1,
self.is_training,
tf.nn.relu,
scope='critic_bn_h1')
# skip action from the first layer
h1_concat = tf.concat([h1, action], 1)
w2 = tf.get_variable('w2',
[self.h1_dim + self.action_dim, self.h2_dim],
initializer=init)
b2 = tf.get_variable('b2', [self.h2_dim], initializer=init)
h2 = tf.nn.relu(tf.matmul(h1_concat, w2) + b2)
w3 = tf.get_variable('w3', [self.h2_dim, 1],
initializer=tf.random_uniform_initializer(
-3e-3, 3e-3))
b3 = tf.get_variable('b3', [1],
initializer=tf.random_uniform_initializer(
-3e-4, 3e-4))
q = tf.matmul(h2, w3) + b3
return q, [w1, b1, w2, b2, w3, b3]
def batch_norm(self, x, is_training, activation_fn, scope):
# switch the 'is_training' flag and 'reuse' flag
return tf.cond(
is_training,
lambda: tf.contrib.layers.batch_norm(x,
activation_fn=activation_fn,
center=True,
scale=True,
updates_collections=None,
is_training=True,
reuse=None,
scope=scope,
decay=0.9,
epsilon=1e-5),
lambda: tf.contrib.layers.batch_norm(
x,
activation_fn=activation_fn,
center=True,
scale=True,
updates_collections=None,
is_training=False,
reuse=True, # to be able to reuse scope must be given
scope=scope,
decay=0.9,
epsilon=1e-5))
class MaDDPG:
def __init__(self, num_agents, state_dim, action_dim):
# track training times
self.time_step = 0
# use set session use GPU
#self.sess = tf.InteractiveSession()
self.sess = tf.Session(config=tf.ConfigProto(
log_device_placement=True))
self.num_agents = num_agents
self.state_dim = state_dim
self.action_dim = action_dim
self.agents = self.create_multi_agents(self.sess, num_agents,
self.state_dim, self.action_dim)
# make sure create Criticnetwork later, summarise mean Q value inside
self.critic = CriticNetwork(self.sess, state_dim, action_dim)
self.exploration_noise = OUNoise((self.num_agents, action_dim))
self.replay_buffer = ReplayBuffer(REPLAY_BUFFER_SIZE)
# for store checkpoint
self.saver = tf.train.Saver()
def train(self):
minibatch = self.replay_buffer.get_batch(BATCH_SIZE)
state_batch = np.zeros((BATCH_SIZE, self.num_agents, self.state_dim))
action_batch = np.zeros((BATCH_SIZE, self.num_agents, self.action_dim))
reward_batch = np.zeros((BATCH_SIZE, self.num_agents))
next_state_batch = np.zeros(
(BATCH_SIZE, self.num_agents, self.state_dim))
done_batch = np.zeros((BATCH_SIZE))
for ii in range(BATCH_SIZE):
state_batch[ii, :, :] = minibatch[ii][0]
action_batch[ii, :, :] = minibatch[ii][1]
reward_batch[ii, :] = minibatch[ii][2]
next_state_batch[ii, :, :] = minibatch[ii][3]
done_batch[ii] = minibatch[ii][4]
# calculate Gt batch
next_action_batch = self.target_actions(next_state_batch)
q_value_batch = self.critic.target_q(next_state_batch,
next_action_batch)
gt = np.zeros((BATCH_SIZE, self.num_agents))
for ii in range(BATCH_SIZE):
if done_batch[ii]:
gt[ii, :] = reward_batch[ii, :]
else:
gt[ii, :] = reward_batch[ii, :] + GAMMA * q_value_batch[ii, :]
#update critic by minimizing the loss
self.critic.train(gt, state_batch, action_batch)
# update policy using the sampling gradients
actions_for_grad = self.actions(state_batch)
q_gradients_batch = self.critic.gradients(state_batch,
actions_for_grad)
self.train_agents(q_gradients_batch, state_batch)
# update critic target network
self.critic.update_target()
# update actor target
self.update_agents_target()
def summary(self, record_num):
if self.replay_buffer.count() > SUMMARY_BATCH_SIZE:
mini_batch = self.replay_buffer.popn(SUMMARY_BATCH_SIZE)
state_batch = np.zeros(
(SUMMARY_BATCH_SIZE, self.num_agents, self.state_dim))
for ii in range(SUMMARY_BATCH_SIZE):
state_batch[ii, :, :] = mini_batch[ii][0]
actions_for_summary = self.actions(state_batch)
self.critic.write_summaries(state_batch, actions_for_summary,
record_num)
def update_agents_target(self):
for agent in self.agents:
agent.update_target()
def train_agents(self, gradients_batch, state_batch):
# gradients_batch = [batchsize* agents* action_dim]
# state_batch = [batchsize* agents * state_dim ]
for ii in range(self.num_agents):
grad = gradients_batch[:, ii, :]
state = state_batch[:, ii, :]
self.agents[ii].train(grad, state)
def create_multi_agents(self, sess, num_agents, state_dim, action_dim):
agents = []
nets = None
for ii in range(num_agents):
agent_name = 'agent' + str(ii)
agents.append(
ActorNetwork(sess, state_dim, action_dim, agent_name, nets))
nets = agents[-1].nets
return agents
def add_agents(self, add_num):
for ii in range(add_num):
#self.num_agents+=1
agent_name = 'agent' + str(self.num_agents)
self.agents.append(
ActorNetwork(self.sess, self.state_dim, self.action_dim,
agent_name, self.agents[-1].nets))
# the agents' name is from 0-num_agents-1
self.num_agents += 1
# if add a new agent then reset the noise and replay buffer
self.exploration_noise = OUNoise((self.num_agents, self.action_dim))
#self.replay_buffer = ReplayBuffer(REPLAY_BUFFER_SIZE)
self.replay_buffer.erase()
# re-create a saver
# the new saver will contains all the savable variables.
# otherwise only contains the initially created agents
self.saver = tf.train.Saver()
# reset the time step
# self.time_step = 0
def action(
self, state
): # here is action, for one state on agent, not batch_sized actions
# state = [num_agents * state_dim]
# actions = [num_agents * action_dim]
action = np.zeros((self.num_agents, self.action_dim))
for ii in range(self.num_agents):
action[ii, :] = self.agents[ii].action(state[ii, :])
return action
def actions(self, state_batch):
#state = batch_size*numOfagents*state_dim
#actions = batch_size*numOfagents*action_dim
batch_size = state_batch.shape[0]
actions = np.zeros((batch_size, self.num_agents, self.action_dim))
for ii in range(self.num_agents):
actions[:, ii, :] = self.agents[ii].actions(state_batch[:, ii, :])
return actions
def target_actions(self, state_batch):
# the state size is batch_size* num_agents * state_dimension
actions = np.zeros(
(state_batch.shape[0], self.num_agents, self.action_dim))
for ii in range(self.num_agents):
actions[:,
ii, :] = self.agents[ii].target_actions(state_batch[:,
ii, :])
return actions
def noise_action(self, state):
action = self.action(state)
# clip the action, action \in [-1,+1]
return np.clip(action + self.exploration_noise.noise(), -1, 1)
def close_session(self):
self.sess.close()
def perceive(self, state, action, reward, next_state, done):
# store {st,at,Rt+1,st+1}
self.replay_buffer.add(state, action, reward, next_state, done)
if self.replay_buffer.count() > REPLAY_START_SIZE:
self.time_step += 1
self.train()
if self.time_step % SAVE_STEPS == 0:
self.save_network()
# if self.time_step % 10000 == 0:
# self.actor_network.save_network(self.time_step)
# self.critic_network.save_network(self.time_step)
# Re-initialize the random process when an episode ends
if done:
self.exploration_noise.reset()
def load_network(self):
checkpoint = tf.train.get_checkpoint_state("saved_network")
if checkpoint and checkpoint.model_checkpoint_path:
self.saver.restore(self.sess, checkpoint.model_checkpoint_path)
print("Successfully loaded:", checkpoint.model_checkpoint_path)
else:
print('Could not find old network weights')
def save_network(self):
# do not processing under Dropbox
# exit drop box then run
print('save network...', self.time_step)
self.saver.save(self.sess,
'saved_network/' + 'network',
global_step=self.time_step)
class DDPG:
"""docstring for DDPG"""
def __init__(self, env, DIRECTORY):
self.batch_size = BATCH_SIZE
self.replay_start_size = REPLAY_START_SIZE # self.sub_batch_size = BATCH_SIZE / n_gpu
self.name = 'DDPG' # name for uploading results
self.environment = env
# Randomly initialize actor network and critic network
# with both their target networks
self.state_dim = env.observation_space.shape[0]
self.action_dim = env.action_space.shape[0]
self.sess = tf.InteractiveSession(config=tf.ConfigProto(
allow_soft_placement=True, log_device_placement=False))
self.trace_length = TRACE_LENGTH
self.temp_abstract = TEMP_ABSTRACT
self.actor_network = ActorNetwork(self.sess, BATCH_SIZE,
self.state_dim, self.action_dim,
self.temp_abstract, DIRECTORY)
self.critic_network = CriticNetwork(self.sess, BATCH_SIZE,
self.state_dim, self.action_dim,
self.temp_abstract, DIRECTORY)
# initialize replay buffer
max_len_trajectory = self.environment.spec.timestep_limit + 1 # trace_length
self.replay_buffer = ReplayBuffer(REPLAY_BUFFER_SIZE, DIRECTORY,
max_len_trajectory,
self.actor_network.last_epi)
# Initialize a random process the Ornstein-Uhlenbeck process for action exploration
self.exploration_noise = OUNoise(self.action_dim)
###
self.diff = 0.
self.discounting_mat_dict = {}
###
def state_initialiser(self, shape, mode='g'):
if mode == 'z': #Zero
initial = np.zeros(shape=shape)
elif mode == 'g': #Gaussian
# initial = stats.truncnorm.rvs(a=-0.02/0.01,b=0.02/0.01,loc=0.,scale=0.01,size=shape)
initial = np.random.normal(loc=0.,
scale=1. / float(shape[1]),
size=shape)
else: # May do some adaptive initialiser can be built in later
raise NotImplementedError
return initial
def train(self, time_step): #,time_step):
###1) Get-batch data for opt
minibatch, trace_length = self.replay_buffer.get_batch(
self.batch_size, self.trace_length,
time_step) #, self.trace_length)
try:
state_trace_batch = np.stack(minibatch[:, :, 2].ravel()).reshape(
self.batch_size, trace_length, self.state_dim)
action_trace_batch = np.stack(minibatch[:, :, 3].ravel()).reshape(
self.batch_size, trace_length, self.action_dim)
next_state_batch = np.stack(minibatch[:, -1, 6].ravel()).reshape(
self.batch_size, 1, self.state_dim)
next_state_trace_batch = np.concatenate(
[state_trace_batch, next_state_batch], axis=1)
reward_trace_batch = np.stack(minibatch[:, :, 4].ravel()).reshape(
self.batch_size, trace_length, 1)
done_trace_batch = np.stack(minibatch[:, :, 7].ravel()).reshape(
self.batch_size, trace_length, 1)
except Exception as e:
print(str(e))
raise
###2) Painfully initialise initial memories of LSTMs: not super-efficient, but no error guaranteed from tf's None-type zero-state problem
init_actor_hidden1_cORm_batch = self.state_initialiser(
shape=(self.batch_size, self.actor_network.rnn_size), mode='z')
actor_init_h_batch = (
init_actor_hidden1_cORm_batch, init_actor_hidden1_cORm_batch
) #((init_hidden1_cORm_batch,init_hidden1_cORm_batch),(init_actor_hidden2_cORm_batch,init_actor_hidden2_cORm_batch))
init_critic_hidden1_cORm_batch = self.state_initialiser(
shape=(self.batch_size, self.critic_network.rnn_size), mode='z')
critic_init_h_batch = (
init_critic_hidden1_cORm_batch, init_critic_hidden1_cORm_batch
) #,(init_critic_hidden3_cORm_batch,init_critic_hidden3_cORm_batch))
###
self.dt_list = np.zeros(shape=(15, ))
self.dt_list[-1] = time.time()
if trace_length <= OPT_LENGTH:
target_actor_init_h_batch = actor_init_h_batch
target_critic_init_h_batch = critic_init_h_batch
pass
else:
### memory stuff
actor_init_h_batch = self.actor_network.action(
state_trace_batch[:, :-OPT_LENGTH, :],
actor_init_h_batch,
mode=1)
target_actor_init_h_batch = actor_init_h_batch
critic_init_h_batch = self.critic_network.evaluation(
state_trace_batch[:, :-OPT_LENGTH, :],
action_trace_batch[:, :-OPT_LENGTH, :],
critic_init_h_batch,
mode=1)
target_critic_init_h_batch = critic_init_h_batch
state_trace_batch = state_trace_batch[:, -OPT_LENGTH:, :]
next_state_trace_batch = next_state_trace_batch[:, -(OPT_LENGTH +
1):, :]
action_trace_batch = action_trace_batch[:, -OPT_LENGTH:, :]
reward_trace_batch = reward_trace_batch[:, -OPT_LENGTH:, :]
done_trace_batch = done_trace_batch[:, -OPT_LENGTH:, :]
self.dt_list[0] = time.time() - np.sum(self.dt_list)
###3) Obtain target output
next_action_batch = self.actor_network.target_action(
next_state_trace_batch,
init_temporal_hidden_cm_batch=target_actor_init_h_batch)
self.dt_list[1] = time.time() - np.sum(self.dt_list)
next_action_trace_batch = np.concatenate(
[action_trace_batch,
np.expand_dims(next_action_batch, axis=1)],
axis=1)
self.dt_list[2] = time.time() - np.sum(self.dt_list)
target_lastQ_batch = self.critic_network.target_q_trace(
next_state_trace_batch,
next_action_trace_batch,
init_temporal_hidden_cm_batch=target_critic_init_h_batch)
self.dt_list[3] = time.time() - np.sum(self.dt_list)
# Control the length of time-step for gradient
if trace_length <= OPT_LENGTH:
update_length = np.minimum(
trace_length,
OPT_LENGTH // 1) #//denom: 2(opt1) #1(opt0) #OPT_LENGTH(opt2)
else:
update_length = OPT_LENGTH // 1 #//denom: 2(opt1) #1(opt0) #OPT_LENGTH(opt2)
target_lastQ_batch_masked = target_lastQ_batch * (
1. - done_trace_batch[:, -1])
rQ = np.concatenate([
np.squeeze(reward_trace_batch[:, -update_length:], axis=-1),
target_lastQ_batch_masked
],
axis=1)
self.dt_list[4] = time.time() - np.sum(self.dt_list)
try:
discounting_mat = self.discounting_mat_dict[update_length]
except KeyError:
discounting_mat = np.zeros(shape=(update_length,
update_length + 1),
dtype=np.float)
for i in range(update_length):
discounting_mat[i, :i] = 0.
discounting_mat[i,
i:] = GAMMA**np.arange(0.,
-i + update_length + 1)
discounting_mat = np.transpose(discounting_mat)
self.discounting_mat_dict[update_length] = discounting_mat
try:
y_trace_batch = np.expand_dims(np.matmul(rQ, discounting_mat),
axis=-1)
except Exception as e:
print('?')
raise
self.dt_list[5] = time.time() - np.sum(self.dt_list)
###4)Train Critic: get next_action, target_q, then optimise
critic_grad = self.critic_network.train(
y_trace_batch,
update_length,
state_trace_batch,
action_trace_batch,
init_temporal_hidden_cm_batch=critic_init_h_batch)
self.dt_list[6] = time.time() - np.sum(self.dt_list)
###5) Train Actor: while updated critic, we declared the dQda. Hence sess,run(dQda*dadParam_actor), then optimise actor
for i in range(update_length):
actor_init_h_batch_trace = (np.expand_dims(actor_init_h_batch[0],
axis=1),
np.expand_dims(actor_init_h_batch[1],
axis=1))
critic_init_h_batch_trace = (np.expand_dims(critic_init_h_batch[0],
axis=1),
np.expand_dims(critic_init_h_batch[1],
axis=1))
if i == 0:
actor_init_h_batch_stack = actor_init_h_batch_trace
critic_init_h_batch_stack = critic_init_h_batch_trace
else:
actor_init_h_batch_stack = (np.concatenate(
(actor_init_h_batch_stack[0], actor_init_h_batch_trace[0]),
axis=1),
np.concatenate(
(actor_init_h_batch_stack[1],
actor_init_h_batch_trace[1]),
axis=1))
critic_init_h_batch_stack = (
np.concatenate((critic_init_h_batch_stack[0],
critic_init_h_batch_trace[0]),
axis=1),
np.concatenate((critic_init_h_batch_stack[1],
critic_init_h_batch_trace[1]),
axis=1))
action_trace_batch_for_gradients, actor_init_h_batch = self.actor_network.action_trace(
np.expand_dims(state_trace_batch[:, i], 1),
init_temporal_hidden_cm_batch=actor_init_h_batch)
critic_init_h_batch = self.critic_network.evaluation_trace(
np.expand_dims(state_trace_batch[:, i], 1),
np.expand_dims(action_trace_batch[:, i], 1),
init_temporal_hidden_cm_batch=critic_init_h_batch)
if i == 0:
action_trace_batch_for_gradients_stack = action_trace_batch_for_gradients
else:
action_trace_batch_for_gradients_stack = np.concatenate(
(action_trace_batch_for_gradients_stack,
action_trace_batch_for_gradients),
axis=1)
self.dt_list[7] = time.time() - np.sum(self.dt_list)
state_trace_batch_stack = np.reshape(
state_trace_batch,
(self.batch_size * update_length, 1, self.state_dim))
action_trace_batch_stack = np.reshape(
action_trace_batch,
(self.batch_size * update_length, 1, self.action_dim))
action_trace_batch_for_gradients_stack = np.reshape(
action_trace_batch_for_gradients_stack,
(self.batch_size * update_length, 1, self.action_dim))
actor_init_h_batch_stack = (np.reshape(
actor_init_h_batch_stack[0],
(self.batch_size * update_length, self.actor_network.rnn_size)),
np.reshape(
actor_init_h_batch_stack[1],
(self.batch_size * update_length,
self.actor_network.rnn_size)))
critic_init_h_batch_stack = (np.reshape(
critic_init_h_batch_stack[0],
(self.batch_size * update_length, self.critic_network.rnn_size)),
np.reshape(
critic_init_h_batch_stack[1],
(self.batch_size * update_length,
self.critic_network.rnn_size)))
q_gradient_trace_batch = self.critic_network.gradients(
1,
state_trace_batch_stack,
action_trace_batch_for_gradients_stack,
init_temporal_hidden_cm_batch=critic_init_h_batch_stack)
self.dt_list[8] = time.time() - np.sum(self.dt_list)
# Update the actor policy using the sampled gradient:
actor_grad = self.actor_network.train(
q_gradient_trace_batch,
1,
state_trace_batch_stack,
action_trace_batch_stack,
init_temporal_hidden_cm_batch=actor_init_h_batch_stack)
self.dt_list[9] = time.time() - np.sum(self.dt_list)
# Update the target networks via EMA & Indicators
# self.critic_network.update_target()
self.dt_list[10] = time.time() - np.sum(self.dt_list)
# self.actor_network.update_target()
self.dt_list[11] = time.time() - np.sum(self.dt_list)
# actor_diff = self.actor_network.get_diff()
self.dt_list[12] = time.time() - np.sum(self.dt_list)
# critic_diff = self.critic_network.get_diff()
self.dt_list[13] = time.time() - np.sum(self.dt_list)
self.dt_list = np.delete(self.dt_list, -1)
return actor_grad, critic_grad, # actor_diff, actor_grad, critic_diff, critic_grad
def action(self, state_trace, init_hidden_cm, epi, noisy=True):
# Select action a_t according to the current policy and exploration noise
action, last_hidden_cm = self.actor_network.action([state_trace],
init_hidden_cm,
mode=2)
if noisy:
noise = self.exploration_noise.noise() #epi)
return action + noise, last_hidden_cm #, dt#, np.linalg.norm(noise)
else:
return action, last_hidden_cm
def evaluation(self, state_trace, action_trace, action_last,
init_hidden_cm):
return self.critic_network.evaluation([state_trace], [action_trace],
action_last,
init_hidden_cm,
mode=2) #q_value, last_hidden_cm
# def perceive(self,actor_init_hidden_cm,critic_last_hidden_cm,state,action,reward,next_state,done,time_step,epi):
def perceive(self, state, action, reward, next_state, done, time_step,
epi):
# Store transition (s_t,a_t,r_t,s_{t+1}) in replay buffer
# self.replay_buffer.add(actor_init_hidden_cm,critic_last_hidden_cm,state,action,reward,next_state,done,epi)
done = float(done)
self.replay_buffer.add(state, action, reward, next_state, done, epi,
time_step)
# Store transitions to replay start size then start training
if (self.replay_buffer.num_experiences > REPLAY_START_SIZE):
# Non-zero diff should be found
self.actor_grad, self.critic_grad = self.train(time_step)
# self.actor_diff, self.actor_grad, self.critic_diff, self.critic_grad = self.train(time_step)
else:
# Zero diff as is not trained
# self.actor_diff = 0.
self.actor_grad = 0.
# self.critic_diff = 0.
self.critic_grad = 0.
# Re-iniitialize the random process when an episode ends
if done:
self.exploration_noise.reset()
def s2l():
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
num_states = feature_size #num_states = env.observation_space.shape[0]
num_actions = num_controls
print ("Number of States:", num_states)
print ("Number of Actions:", num_actions)
action_space_high=[1.5] #[0.0,0.0,0.0]
action_space_low=[0.03] #[0.5,0.5,0.5]
print ("Action space highest values", action_space_high)
print ("Action space lowest values:", action_space_low)
robot=RoboControl()
#while True:
# #robot.check()
# robot.publish_control([1])
# robot.reset()
agent = DDPG(is_batch_norm,num_states,num_actions,action_space_high,action_space_low)
exploration_noise = OUNoise(num_actions)
counter=0
total_reward=0
print ("Number of Rollouts per episode:", num_rollouts)
print ("Number of Steps per roll out:", steps)
reward_st = np.array([0]) #saving reward
eval_metric_st= np.array([0])
reward_st_all = np.array([0]) #saving reward after every step
activity_obj=Vid_Feature()
demo_vid_array=demo_array_extractor(demo_folder)
demo_features=activity_obj.feature_extractor(demo_vid_array)
frame_obj=Frame_Feature()
#camera_obj= Camera()
camera_obj= CameraSub()
for episode in range(num_episodes):
print ("==== Starting episode no:",episode,"====","\n")
robot.reset() # Reset env in the begining of each episode
obs_img=camera_obj.camera_subscribe() # Get the observation
#obs_img=np.array(misc.imresize(obs_img,[112,112,3]))
observation =np.array(frame_obj.frame_feature_extractor(obs_img))
observation=observation.reshape(-1)
reward_per_episode = 0
for t in range(num_rollouts):
reward_per_rollout=0
vid_robo_=[]
for i in range(steps):
x = observation
action = agent.evaluate_actor(np.reshape(x,[1,num_states]))
noise = exploration_noise.noise()
action = action[0] + noise #Select action according to current policy and exploration noise
print ('Action at episode-',episode,'rollout-',t, 'step-', i ," :",action)
robot.publish_control(action)
obs_robo=camera_obj.camera_subscribe() # Get the observation
#obs_robo=misc.imresize(obs_robo,[112,112,3])
vid_robo_.append(obs_robo)
observation=np.array(frame_obj.frame_feature_extractor(np.array(obs_robo)))
observation=observation.reshape(-1)
#pasue()
if(i==15):
vid_robo=np.array(vid_robo_)
robo_features=activity_obj.feature_extractor(vid_robo)
reward=-(distance(demo_features,robo_features))
reward=np.array(reward)
print('reward: ',reward)
else:
reward=0
reward=np.array(reward)
print('reward: ',reward)
# Storing reward after every rollout
reward_st_all = np.append(reward_st_all,reward)
np.savetxt('reward_all.txt',reward_st_all, newline="\n")
#add s_t,s_t+1,action,reward to experience memory
agent.add_experience(x,observation,action,reward,False)
reward_per_rollout+=reward
counter+=1
#train critic and actor network
if counter > start_training:
agent.train()
print ('\n\n')
#Saving policy
if ((episode%100)==0 and t==num_rollouts-1):
print('saving policy...........................!')
agent.save_actor(episode)
reward_per_episode+=reward_per_rollout
#check if episode ends:
print ('EPISODE: ',episode,' Total Reward: ',reward_per_episode)
print ("Printing reward to file")
exploration_noise.reset() #reinitializing random noise for action exploration
reward_st = np.append(reward_st,reward_per_episode)
np.savetxt('episode_reward.txt',reward_st, fmt='%f', newline="\n")
print ('\n\n')
total_reward+=reward_per_episode
print ("Average reward per episode {}".format(total_reward / num_episodes))
done = False
epsilon -= (epsilon / EXPLORE)
epsilon = np.maximum(min_epsilon, epsilon)
episode_r = 0.
step = 0
max_steps = 110
r_episodio_actual = []
while (not done):
step += 1
print('step =', step)
wait = input("PRESS ENTER TO CONTINUE.")
action = ddpg.predict_action(np.reshape(state, (1, state_dim)))
action1 = action
print('LA ACCION sin clipear ES', action1, action1.shape)
action = np.clip(action1, min_action, max_action)
action = action + max(epsilon, 0) * ruido.noise() * 2.
action = np.clip(action, min_action, max_action)
print('ruido =',
max(epsilon, 0) * ruido.noise() * 2., 'epsilon =',
epsilon)
print('LA ACCION clipeada ES', action, action.shape)
next_state, reward, done, info = env.step(action)
print('EL NEXT_ESTADO ES', next_state, next_state.shape)
#reward = np.clip(reward,-1.,1.)
print('instaneous r = ', reward)
replay_buffer.add(np.reshape(state, (state_dim, )),
np.reshape(action, (action_dim, )), reward,
done, np.reshape(next_state, (state_dim, )))
state = next_state
episode_r = episode_r + reward
class DDPG:
"""docstring for DDPG"""
def __init__(self, sess, data_fname):
self.name = 'DDPG'
# Randomly initialize actor network and critic network
# with both their target networks
self.name = 'DDPG' # name for uploading results
# Randomly initialize actor network and critic network
# with both their target networks
self.state_dim = Hp.state_dim
self.action_dim = Hp.action_dim
print(self.state_dim, self.action_dim)
self.sess = sess
self.state_input = [
tf.placeholder(tf.float32, shape=(None, None, Hp.n_coord))
for _ in xrange(Hp.categories)
]
#tf.placeholder("float",[None,self.state_dim])
self.target_state_input = [
tf.placeholder(tf.float32, shape=(None, None, Hp.n_coord))
for _ in xrange(Hp.categories)
]
#tf.placeholder("float",[None,self.state_dim])
self.state_network = StateEnc(self.sess, self.state_input,
self.target_state_input)
state_batch = self.state_network.encoding
next_state_batch = self.state_network.target_encoding
weights, biases, w_i2h0, w_h2h0, w_b0, w_i2h1, w_h2h1, w_b1, w_i2h2, w_h2h2, w_b2 = self.state_network.get_parameters(
)
state_network_params = weights + biases + [
w_i2h0, w_h2h0, w_b0, w_i2h1, w_h2h1, w_b1, w_i2h2, w_h2h2, w_b2
]
self.actor_network = ActorNetwork(self.sess, Hp.n_hidden,
self.action_dim, self.state_input,
state_batch, next_state_batch,
state_network_params)
self.critic_network = CriticNetwork(self.sess, Hp.n_hidden,
self.action_dim, state_batch,
next_state_batch)
# initialize replay buffer
self.replay_buffer = ReplayBuffer(Hp.REPLAY_BUFFER_SIZE, data_fname)
self.summary_str2 = None
# Initialize a random process the Ornstein-Uhlenbeck process for action exploration
self.exploration_noise = OUNoise(self.action_dim)
def train(self):
#print "train step",self.time_step
# Sample a random minibatch of N transitions from replay buffer
minibatches = self.replay_buffer.get_batch(Hp.batch_size * Hp.N_TRAIN)
print("######### TRAINING #############")
for k in range(Hp.N_TRAIN):
minibatch = minibatches[k * Hp.batch_size:(k + 1) * Hp.batch_size]
state_batch_r = np.asarray([data[0] for data in minibatch])
state_batch = []
for j in range(Hp.categories):
new_cat = np.stack(state_batch_r[:, j], axis=0)
state_batch.append(new_cat)
#state_batch = [np.expand_dims(state_batch, axis=1)]
action_batch = np.asarray([data[1] for data in minibatch])
reward_batch = np.asarray([data[2] for data in minibatch])
next_state_batch_r = np.asarray([data[3] for data in minibatch])
next_state_batch = []
for j in range(Hp.categories):
new_cat = np.stack(next_state_batch_r[:, j], axis=0)
next_state_batch.append(new_cat)
#next_state_batch = [np.expand_dims(next_state_batch, axis=1)]
done_batch = np.asarray([data[4] for data in minibatch])
# for action_dim = 1
action_batch = np.resize(action_batch,
[Hp.batch_size, self.action_dim])
next_action_batch = self.actor_network.target_actions(
self.target_state_input, next_state_batch)
q_value_batch = self.critic_network.target_q(
self.target_state_input, next_state_batch, next_action_batch)
y_batch = []
for i in range(len(minibatch)):
if done_batch[i]:
y_batch.append(reward_batch[i])
else:
y_batch.append(reward_batch[i] +
Hp.GAMMA * q_value_batch[i])
y_batch = np.resize(y_batch, [Hp.batch_size, 1])
# Update critic by minimizing the loss L
self.critic_network.train(y_batch, self.state_input, state_batch,
action_batch)
# Update the actor policy using the sampled gradient:
action_batch_for_gradients = self.actor_network.actions(
self.state_input, state_batch)
q_gradient_batch = self.critic_network.gradients(
self.state_input, state_batch, action_batch_for_gradients)
self.summary_str2 = self.actor_network.train(
q_gradient_batch, self.state_input, state_batch)
# Update the target networks
self.actor_network.update_target()
self.critic_network.update_target()
self.state_network.update_target()
def noise_action(self, state):
# Select action a_t according to the current policy and exploration noise
state = [np.expand_dims(el, axis=0) for el in state]
action = self.actor_network.action(state)
print("no noise ", action)
return np.clip(
action +
self.exploration_noise.noise() * np.array([-17.0, 17.0, 900.0]),
[-35.0, 0.0, 0.0], [0.0, 35.0, 2000.0])
def action(self, state):
state = [np.expand_dims(el, axis=0) for el in state]
action = self.actor_network.action(state)
return action
def perceive(self, state, action, reward, next_state, done):
# Store transition (s_t,a_t,r_t,s_{t+1}) in replay buffer
self.replay_buffer.add(state, action, reward, next_state, done)
# Store transitions to replay start size then start training
if self.replay_buffer.count() > Hp.REPLAY_START_SIZE:
self.train()
#if self.time_step % 10000 == 0:
#self.actor_network.save_network(self.time_step)
#self.critic_network.save_network(self.time_step)
# Re-iniitialize the random process when an episode ends
if done:
self.exploration_noise.reset()
class DDPG(object):
def __init__(self, env, args):
self.action_dim = env.action_space.shape[0]
self.state_dim = env.observation_space.shape[0]
self.actor_lr = args.a_lr
self.critic_lr = args.c_lr
self.gamma = args.gamma
# Ornstein-Uhlenbeck noise parameters
self.ou = OUNoise(
self.action_dim, theta=args.noise_theta, sigma=args.noise_sigma)
self.replay_buffer = deque(maxlen=args.buffer_size)
self.replay_start_size = args.replay_start_size
self.batch_size = args.batch_size
self.target_update_rate = args.target_update_rate
self.total_parameters = 0
self.global_steps = 0
self.reg_param = args.reg_param
def construct_model(self, gpu):
if gpu == -1: # use CPU
device = '/cpu:0'
sess_config = tf.ConfigProto()
else: # use GPU
device = '/gpu:' + str(gpu)
sess_config = tf.ConfigProto(
log_device_placement=True, allow_soft_placement=True)
sess_config.gpu_options.allow_growth = True
self.sess = tf.Session(config=sess_config)
with tf.device(device):
# output action, q_value and gradients of q_val w.r.t. action
with tf.name_scope('predict_actions'):
self.states = tf.placeholder(
tf.float32, [None, self.state_dim], name='states')
self.action = tf.placeholder(
tf.float32, [None, self.action_dim], name='action')
self.is_training = tf.placeholder(tf.bool, name='is_training')
self.action_outputs, self.actor_params = self._build_actor(
self.states, scope='actor_net', bn=True)
self.value_outputs, self.critic_params = self._build_critic(
self.states, self.action, scope='critic_net', bn=False)
self.action_gradients = tf.gradients(
self.value_outputs, self.action)[0]
# estimate target_q for update critic
with tf.name_scope('estimate_target_q'):
self.next_states = tf.placeholder(
tf.float32, [None, self.state_dim], name='next_states')
self.mask = tf.placeholder(tf.float32, [None], name='mask')
self.rewards = tf.placeholder(tf.float32, [None], name='rewards')
# target actor network
self.t_action_outputs, self.t_actor_params = self._build_actor(
self.next_states, scope='t_actor_net', bn=True,
trainable=False)
# target critic network
self.t_value_outputs, self.t_critic_params = self._build_critic(
self.next_states, self.t_action_outputs, bn=False,
scope='t_critic_net', trainable=False)
self.target_q = self.rewards + self.gamma * \
(self.t_value_outputs[:, 0] * self.mask)
with tf.name_scope('compute_gradients'):
self.actor_opt = tf.train.AdamOptimizer(self.actor_lr)
self.critic_opt = tf.train.AdamOptimizer(self.critic_lr)
# critic gradients
td_error = self.target_q - self.value_outputs[:, 0]
critic_mse = tf.reduce_mean(tf.square(td_error))
critic_reg = tf.reduce_sum(
[tf.nn.l2_loss(v) for v in self.critic_params])
critic_loss = critic_mse + self.reg_param * critic_reg
self.critic_gradients = \
self.critic_opt.compute_gradients(
critic_loss, self.critic_params)
# actor gradients
self.q_action_grads = tf.placeholder(
tf.float32, [None, self.action_dim], name='q_action_grads')
actor_gradients = tf.gradients(
self.action_outputs, self.actor_params,
-self.q_action_grads)
self.actor_gradients = zip(actor_gradients, self.actor_params)
# apply gradient to update model
self.train_actor = self.actor_opt.apply_gradients(
self.actor_gradients)
self.train_critic = self.critic_opt.apply_gradients(
self.critic_gradients)
with tf.name_scope('update_target_networks'):
# batch norm paramerters should not be included when updating!
target_networks_update = []
for v_source, v_target in zip(
self.actor_params, self.t_actor_params):
update_op = v_target.assign_sub(
0.001 * (v_target - v_source))
target_networks_update.append(update_op)
for v_source, v_target in zip(
self.critic_params, self.t_critic_params):
update_op = v_target.assign_sub(
0.01 * (v_target - v_source))
target_networks_update.append(update_op)
self.target_networks_update = tf.group(*target_networks_update)
with tf.name_scope('total_numbers_of_parameters'):
for v in tf.trainable_variables():
shape = v.get_shape()
param_num = 1
for d in shape:
param_num *= d.value
print(v.name, ' ', shape, ' param nums: ', param_num)
self.total_parameters += param_num
print('Total nums of parameters: ', self.total_parameters)
def sample_action(self, states, noise):
# is_training suppose to be False when sampling action.
action = self.sess.run(
self.action_outputs,
feed_dict={self.states: states, self.is_training: False})
ou_noise = self.ou.noise() if noise else 0
return action + ou_noise
def store_experience(self, s, a, r, next_s, done):
self.replay_buffer.append([s, a[0], r, next_s, done])
self.global_steps += 1
def update_model(self):
if len(self.replay_buffer) < self.replay_start_size:
return
# get batch
batch = random.sample(self.replay_buffer, self.batch_size)
s, _a, r, next_s, done = np.vstack(batch).T.tolist()
mask = ~np.array(done)
# compute a = u(s)
a = self.sess.run(self.action_outputs, {
self.states: s,
self.is_training: True
})
# gradients of q_value w.r.t action a
dq_da = self.sess.run(self.action_gradients, {
self.states: s,
self.action: a,
self.is_training: True
})
# train
self.sess.run([self.train_actor, self.train_critic], {
# train_actor feed
self.states: s,
self.is_training: True,
self.q_action_grads: dq_da,
# train_critic feed
self.next_states: next_s,
self.action: _a,
self.mask: mask,
self.rewards: r
})
# update target network
self.sess.run(self.target_networks_update)
def _build_actor(self, states, scope, bn=False, trainable=True):
h1_dim = 400
h2_dim = 300
init = tf.contrib.layers.variance_scaling_initializer(
factor=1.0, mode='FAN_IN', uniform=True)
with tf.variable_scope(scope):
if bn:
states = self.batch_norm(
states, self.is_training, tf.identity,
scope='actor_bn_states', trainable=trainable)
h1 = tcl.fully_connected(
states, h1_dim, activation_fn=None, weights_initializer=init,
biases_initializer=init, trainable=trainable, scope='actor_h1')
if bn:
h1 = self.batch_norm(
h1, self.is_training, tf.nn.relu, scope='actor_bn_h1',
trainable=trainable)
else:
h1 = tf.nn.relu(h1)
h2 = tcl.fully_connected(
h1, h2_dim, activation_fn=None, weights_initializer=init,
biases_initializer=init, trainable=trainable, scope='actor_h2')
if bn:
h2 = self.batch_norm(
h2, self.is_training, tf.nn.relu, scope='actor_bn_h2',
trainable=trainable)
else:
h2 = tf.nn.relu(h2)
# use tanh to bound the action
a = tcl.fully_connected(
h2, self.action_dim, activation_fn=tf.nn.tanh,
weights_initializer=tf.random_uniform_initializer(-3e-3, 3e-3),
biases_initializer=tf.random_uniform_initializer(-3e-4, 3e-4),
trainable=trainable, scope='actor_out')
params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope)
return a, params
def _build_critic(self, states, action, scope, bn=False, trainable=True):
h1_dim = 400
h2_dim = 300
init = tf.contrib.layers.variance_scaling_initializer(
factor=1.0, mode='FAN_IN', uniform=True)
with tf.variable_scope(scope):
if bn:
states = self.batch_norm(
states, self.is_training, tf.identity,
scope='critic_bn_state', trainable=trainable)
h1 = tcl.fully_connected(
states, h1_dim, activation_fn=None, weights_initializer=init,
biases_initializer=init, trainable=trainable, scope='critic_h1')
if bn:
h1 = self.batch_norm(
h1, self.is_training, tf.nn.relu, scope='critic_bn_h1',
trainable=trainable)
else:
h1 = tf.nn.relu(h1)
# skip action from the first layer
h1 = tf.concat([h1, action], 1)
h2 = tcl.fully_connected(
h1, h2_dim, activation_fn=None, weights_initializer=init,
biases_initializer=init, trainable=trainable,
scope='critic_h2')
if bn:
h2 = self.batch_norm(
h2, self.is_training, tf.nn.relu, scope='critic_bn_h2',
trainable=trainable)
else:
h2 = tf.nn.relu(h2)
q = tcl.fully_connected(
h2, 1, activation_fn=None,
weights_initializer=tf.random_uniform_initializer(-3e-3, 3e-3),
biases_initializer=tf.random_uniform_initializer(-3e-4, 3e-4),
trainable=trainable, scope='critic_out')
params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope)
return q, params
def batch_norm(self, x, is_training, activation_fn, scope, trainable=True):
# switch the 'is_training' flag and 'reuse' flag
return tf.cond(
is_training,
lambda: tf.contrib.layers.batch_norm(
x,
activation_fn=activation_fn,
center=True,
scale=True,
updates_collections=None,
is_training=True,
reuse=None,
scope=scope,
decay=0.9,
epsilon=1e-5,
trainable=trainable),
lambda: tf.contrib.layers.batch_norm(
x,
activation_fn=activation_fn,
center=True,
scale=True,
updates_collections=None,
is_training=False,
reuse=True, # to be able to reuse scope must be given
scope=scope,
decay=0.9,
epsilon=1e-5,
trainable=trainable))
class DDPG_REC:
def __init__(self, state_item_num, action_item_num, emb_dim, batch_size, tau, actor_lr, critic_lr,
gamma, buffer_size, item_space, summary_dir):
self.state_item_num = state_item_num
self.action_item_num = action_item_num
self.emb_dim = emb_dim
self.batch_size = batch_size
self.tau = tau
self.actor_lr = actor_lr
self.critic_lr = critic_lr
self.gamma = gamma
self.buffer_size = buffer_size
self.item_space = item_space
self.summary_dir = summary_dir
self.sess = tf.Session()
self.s_dim = emb_dim * state_item_num
self.a_dim = emb_dim * action_item_num
self.actor = Actor(self.sess, state_item_num, action_item_num, emb_dim, batch_size, tau, actor_lr)
self.critic = Critic(self.sess, state_item_num, action_item_num, emb_dim,
self.actor.get_num_trainable_vars(), gamma, tau, critic_lr)
self.exploration_noise = OUNoise(self.a_dim)
# set up summary operators
self.summary_ops, self.summary_vars = self.build_summaries()
self.sess.run(tf.global_variables_initializer())
self.writer = tf.summary.FileWriter(summary_dir, self.sess.graph)
# initialize target network weights
self.actor.hard_update_target_network()
self.critic.hard_update_target_network()
# initialize replay memory
self.replay_buffer = ReplayBuffer(buffer_size)
def gene_actions(self, weight_batch):
"""use output of actor network to calculate action list
Args:
weight_batch: actor network outputs
Returns:
recommendation list
"""
item_ids = list(self.item_space.keys())
item_weights = list(self.item_space.values())
max_ids = list()
for weight in weight_batch:
score = np.dot(item_weights, np.transpose(weight))
idx = np.argmax(score, 0)
max_ids.append([item_ids[_] for _ in idx])
return max_ids
# def gene_action(self, weight):
# """use output of actor network to calculate action list
# Args:
# weight: actor network outputs
#
# Returns:
# recommendation list
# """
# item_ids = list(self.item_space.keys())
# item_weights = list(self.item_space.values())
# score = np.dot(item_weights, np.transpose(weight))
# idx = np.argmax(score)
# return item_ids[idx]
@staticmethod
def build_summaries():
episode_reward = tf.Variable(0.)
tf.summary.scalar("reward", episode_reward)
episode_max_q = tf.Variable(0.)
tf.summary.scalar("max_q_value", episode_max_q)
critic_loss = tf.Variable(0.)
tf.summary.scalar("critic_loss", critic_loss)
summary_vars = [episode_reward, episode_max_q, critic_loss]
summary_ops = tf.summary.merge_all()
return summary_ops, summary_vars
def _train(self):
samples = self.replay_buffer.sample_batch(self.batch_size)
state_batch = np.asarray([_[0] for _ in samples])
action_batch = np.asarray([_[1] for _ in samples])
reward_batch = np.asarray([_[2] for _ in samples])
n_state_batch = np.asarray([_[3] for _ in samples])
done_batch = np.asarray([_[4] for _ in samples])
seq_len_batch = np.asarray([self.state_item_num] * self.batch_size)
# calculate predicted q value
action_weights = self.actor.predict_target(state_batch, seq_len_batch) # [batch_size,
n_action_batch = self.gene_actions(action_weights.reshape((-1, self.action_item_num, self.emb_dim)))
n_action_emb_batch = get_item_emb(n_action_batch, item_ids_emb_dict)
target_q_batch = self.critic.predict_target(n_state_batch.reshape((-1, self.s_dim)),
n_action_emb_batch.reshape((-1, self.a_dim)), seq_len_batch)
y_batch = []
for i in range(self.batch_size):
if done_batch[i]:
y_batch.append(reward_batch[i])
else:
y_batch.append(reward_batch[i] + self.critic.gamma * target_q_batch[i])
# train critic
q_value, critic_loss, _ = self.critic.train(state_batch, action_batch,
np.reshape(y_batch, (self.batch_size, 1)), seq_len_batch)
# train actor
action_weight_batch_for_gradients = self.actor.predict(state_batch, seq_len_batch)
action_batch_for_gradients = self.gene_actions(action_weight_batch_for_gradients)
action_emb_batch_for_gradients = get_item_emb(action_batch_for_gradients, item_ids_emb_dict)
a_gradient_batch = self.critic.action_gradients(state_batch,
action_emb_batch_for_gradients.reshape((-1, self.a_dim)),
seq_len_batch)
self.actor.train(state_batch, a_gradient_batch[0], seq_len_batch)
# update target networks
self.actor.update_target_network()
self.critic.update_target_network()
return np.amax(q_value), critic_loss
def action(self, state):
weight = self.actor.predict(np.reshape(state, [1, self.s_dim]), np.array([self.state_item_num])) + \
self.exploration_noise.noise().reshape(
(1, self.action_item_num, int(self.a_dim / self.action_item_num)))
action = self.gene_actions(weight)
return np.array(action[0])
def perceive_and_train(self, state, action, reward, n_state, done):
action_emb = get_item_emb(action, item_ids_emb_dict)
self.replay_buffer.add(list(state.reshape((self.s_dim,))),
list(action_emb.reshape((self.a_dim,))),
[reward],
list(n_state.reshape((self.s_dim,))),
[done])
# Store transitions to replay start size then start training
ep_q_value_, critic_loss = 0, 0
if self.replay_buffer.size() > self.batch_size:
ep_q_value_, critic_loss = self._train()
# if self.time_step % 10000 == 0:
# self.actor_network.save_network(self.time_step)
# self.critic_network.save_network(self.time_step)
# Re-iniitialize the random process when an episode ends
if done:
self.exploration_noise.reset()
return ep_q_value_, critic_loss
def write_summary(self, ep_reward, ep_q_value, loss, i):
summary_str = self.sess.run(self.summary_ops, feed_dict={self.summary_vars[0]: ep_reward,
self.summary_vars[1]: ep_q_value,
self.summary_vars[2]: loss})
self.writer.add_summary(summary_str, i)
def save(self):
self.writer.close()
saver = tf.train.Saver()
ckpt_path = os.path.join(os.path.dirname(__file__), "models")
saver.save(self.sess, ckpt_path, write_meta_graph=False)
class DDPG:
"""docstring for DDPG"""
def __init__(self, env):
mx.random.seed(seed)
np.random.seed(seed)
self.env = env
if flg_gpu:
self.ctx = mx.gpu(0)
else:
self.ctx = mx.cpu()
self.state_dim = env.observation_space.shape[0]
self.action_dim = env.action_space.shape[0]
self.ddpgnet = DDPGNet(self.state_dim, self.action_dim)
self.exploration_noise = OUNoise(self.action_dim)
self.replay_buffer = ReplayBuffer(memory_size)
self.batch_size = batch_size
self.ddpgnet.init()
self.train_step = 0
def train(self):
# print "train step",self.time_step
# Sample a random minibatch of N transitions from replay buffer
minibatch = self.replay_buffer.get_batch(self.batch_size)
state_batch = np.asarray([data[0] for data in minibatch])
action_batch = np.asarray([data[1] for data in minibatch])
reward_batch = np.asarray([data[2] for data in minibatch])
next_state_batch = np.asarray([data[3] for data in minibatch])
done_batch = np.asarray([data[4] for data in minibatch])
# for action_dim = 1
action_batch = np.resize(action_batch,
[self.batch_size, self.action_dim])
# Calculate y_batch
next_qvals = self.ddpgnet.get_target_q(next_state_batch).asnumpy()
y_batch = []
for i in range(len(minibatch)):
if done_batch[i]:
y_batch.append(reward_batch[i])
else:
y_batch.append(reward_batch[i] + GAMMA * next_qvals[i][0])
y_batch = np.resize(y_batch, [self.batch_size, 1])
# Update critic by minimizing the loss L
self.ddpgnet.update_critic(state_batch, action_batch, y_batch)
# Update actor by maxmizing Q
self.ddpgnet.update_actor(state_batch)
self.train_step += 1
# update target networks
self.ddpgnet.update_target()
def noise_action(self, state):
# Select action a_t according to the current policy and exploration noise
state = np.reshape(state, (1, self.state_dim))
action = self.ddpgnet.get_step_action(state)
return action + self.exploration_noise.noise()
def action(self, state):
state = np.reshape(state, (1, self.state_dim))
action = self.ddpgnet.get_step_action(state)
return action
def perceive(self, state, action, reward, next_state, done):
# Store transition (s_t,a_t,r_t,s_{t+1}) in replay buffer
self.replay_buffer.add(state, action, reward, next_state, done)
# Store transitions to replay start size then start training
if self.replay_buffer.count() > memory_start_size:
self.train()
# if self.time_step % 10000 == 0:
# self.actor_network.save_network(self.time_step)
# self.critic_network.save_network(self.time_step)
# Re-iniitialize the random process when an episode ends
if done:
self.exploration_noise.reset()
class DDPG:
"""docstring for DDPG"""
def __init__(self, env, time_steps, hidden_dim):
self.name = 'DDPG' # name for uploading results
self.scale = env.asset
self.unit = env.unit
self.seed = env.rd_seed
self.time_dim = time_steps
self.state_dim = env.observation_space.shape[1]
self.action_dim = env.action_space.shape[0]
self.batch_size = 64
self.memory_size = self.time_dim + self.batch_size * 10
self.start_size = self.time_dim + self.batch_size * 2
# Initialise actor & critic networks
self.actor_network = Actor(self.time_dim, self.state_dim,
self.action_dim, hidden_dim)
self.critic_network = Critic(self.time_dim, self.state_dim,
self.action_dim, hidden_dim)
# Initialize replay buffer
self.replay_state = torch.zeros(
(self.start_size - 1, 3, self.state_dim), device=cuda)
self.replay_next_state = torch.zeros(
(self.start_size - 1, 3, self.state_dim), device=cuda)
self.replay_action = torch.zeros(
(self.start_size - 1, 1, self.state_dim), device=cuda)
self.replay_reward = torch.zeros((self.start_size - 1, ), device=cuda)
# Initialize a random process the Ornstein-Uhlenbeck process for action exploration
self.exploration_noise = OUNoise(self.action_dim,
sigma=0.01 / self.action_dim)
self.initial()
def initial(self):
self.steps = 0
self.action = torch.zeros(self.action_dim, device=cuda)
self.replay_state = torch.zeros(
(self.start_size - 1, 3, self.state_dim), device=cuda)
self.replay_next_state = torch.zeros(
(self.start_size - 1, 3, self.state_dim), device=cuda)
self.replay_action = torch.zeros((self.start_size - 1, self.state_dim),
device=cuda)
self.replay_reward = torch.zeros((self.start_size - 1, ), device=cuda)
def train_on_batch(self):
# Sample a random minibatch of N transitions from replay buffer
sample = torch.randint(self.time_dim,
self.replay_reward.shape[0], [self.batch_size],
device=cuda)
index = torch.stack([sample - i for i in range(self.time_dim, 0, -1)
]).t().reshape(-1)
state_data = min_max_scale(self.replay_state[:, 0, :])
amount_data = min_max_scale(self.replay_state[:, 2, :])
next_state_data = min_max_scale(self.replay_next_state[:, 0, :])
next_amount_data = min_max_scale(self.replay_next_state[:, 2, :])
state_batch = torch.index_select(state_data, 0,
index).view(self.batch_size, -1)
amount_data = torch.index_select(amount_data, 0,
sample).view(self.batch_size, -1)
state_batch = torch.cat([state_batch, amount_data], dim=1)
next_state_batch = torch.index_select(next_state_data, 0,
index).view(self.batch_size, -1)
next_amount_data = torch.index_select(next_amount_data, 0,
sample).view(
self.batch_size, -1)
next_state_batch = torch.cat([next_state_batch, next_amount_data],
dim=1)
action_batch = torch.index_select(self.replay_action / self.unit, 0,
sample)
reward_batch = torch.index_select(self.replay_reward, 0, sample)
# Calculate y_batch
next_action_batch = self.actor_network.target_action(next_state_batch)
q_batch = self.critic_network.target_q(next_action_batch,
next_state_batch)
y_batch = torch.add(reward_batch, q_batch, alpha=GAMMA).view(-1, 1)
# train actor-critic by target loss
self.actor_network.train(
self.critic_network.train(y_batch, action_batch, state_batch))
# Update target networks by soft update
self.actor_network.update_target()
self.critic_network.update_target()
def perceive(self, state, action, reward, next_state, done):
if self.steps < self.start_size - 1:
self.replay_state[self.steps] = state
self.replay_next_state[self.steps] = next_state
self.replay_action[self.steps] = action
self.replay_reward[self.steps] = reward
else:
if self.steps >= self.memory_size:
self.replay_state = self.replay_state[1:]
self.replay_next_state = self.replay_next_state[1:]
self.replay_action = self.replay_action[1:]
self.replay_reward = self.replay_reward[1:]
self.replay_state = torch.cat(
(self.replay_state, state.unsqueeze(0)), dim=0)
self.replay_next_state = torch.cat(
(self.replay_next_state, next_state.unsqueeze(0)), dim=0)
self.replay_action = torch.cat(
(self.replay_action, action.unsqueeze(0)), dim=0)
self.replay_reward = torch.cat(
(self.replay_reward, reward.unsqueeze(0)), dim=0)
self.steps += 1
def act(self, next_state, portfolio):
if self.steps > self.start_size:
next_state_data = min_max_scale(
self.replay_next_state[:, 0, :])[-self.time_dim:].view(1, -1)
next_amount_data = min_max_scale(
self.replay_next_state[:, 2, :])[-1].view(1, -1)
next_state_data = torch.cat([next_state_data, next_amount_data],
dim=1)
self.train_on_batch()
allocation = self.actor_network.target_action(
next_state_data).data.view(-1)
allocation += torch.tensor(self.exploration_noise.noise().tolist(),
device=cuda)
allocation[allocation < 0] = 0
allocation /= sum(allocation)
allocation = torch.floor(portfolio * allocation /
next_state[1, :] / self.unit) * self.unit
self.action = allocation
return self.action.clone()
class DDPG:
def __init__(self, state_dim, state_channel, action_dim):
self.state_dim = state_dim
self.state_channel = state_channel
self.action_dim = action_dim
self.sess = tf.InteractiveSession()
self.state_input = tf.placeholder('float', [None, state_dim, state_dim, state_channel])
self.target_state_input = tf.placeholder('float', [None, state_dim, state_dim, state_channel])
self.action_input = tf.placeholder('float', [None, action_dim])
self.actor_network = ActorNetwork(self.sess, self.state_dim, self.state_channel, self.action_dim)
self.critic_network = CriticNetwork(self.sess, self.state_dim, self.state_channel, self.action_dim)
# create network
self.actor_network.create_network(self.state_input)
self.critic_network.create_q_network(self.state_input, self.actor_network.action_output)
# create target network
self.actor_network.create_target_network(self.target_state_input)
self.critic_network.create_target_q_network(self.target_state_input, self.actor_network.target_action_output)
# create training method
self.actor_network.create_training_method(self.critic_network.q_value_output)
self.critic_network.create_training_method()
self.sess.run(tf.initialize_all_variables())
self.actor_network.update_target()
self.critic_network.update_target()
self.replay_buffer = ReplayBuffer(REPLAY_BUFFER_SIZE)
self.exploration_noise = OUNoise(self.action_dim)
self.dir_path = os.path.dirname(os.path.realpath(__file__)) + '/models_ddpg'
if not os.path.exists(self.dir_path):
os.mkdir(self.dir_path)
# for log
self.reward_input = tf.placeholder(tf.float32)
tf.scalar_summary('reward', self.reward_input)
self.time_input = tf.placeholder(tf.float32)
tf.scalar_summary('living_time', self.time_input)
self.summary_op = tf.merge_all_summaries()
self.summary_writer = tf.train.SummaryWriter(self.dir_path + '/log', self.sess.graph)
self.episode_reward = 0.0
self.episode_start_time = 0.0
self.time_step = 1
self.saver = tf.train.Saver(tf.all_variables())
self.load_time_step()
self.load_network()
return
def train(self):
action_dim = self.action_dim
minibatch = self.replay_buffer.get_batch(BATCH_SIZE) # sample BATCH_SIZE from replay_buffer
state_batch = np.asarray([data[0] for data in minibatch])
action_batch = np.asarray([data[1] for data in minibatch])
reward_batch = np.asarray([data[2] for data in minibatch])
next_state_batch = np.asarray([data[3] for data in minibatch])
done_batch = np.asarray([data[4] for data in minibatch])
# if action_dim = 1, it's a number not a array
action_batch = np.resize(action_batch, [BATCH_SIZE, action_dim])
# calculate y_batch via target network
next_action_batch = self.actor_network.target_actions(next_state_batch)
q_value_batch = self.critic_network.target_q_value(next_state_batch, next_action_batch)
y_batch = []
for i in range(BATCH_SIZE):
if done_batch[i]:
y_batch.append(reward_batch[i])
else:
y_batch.append(reward_batch[i] + GAMMA * q_value_batch[i])
y_batch = np.resize(y_batch, [BATCH_SIZE, 1])
# print np.shape(reward_batch), np.shape(y_batch)
# train actor network
self.actor_network.train(state_batch)
# train critic network
self.critic_network.train(y_batch, state_batch, action_batch)
# update target network
self.actor_network.update_target()
self.critic_network.update_target()
return
def noise_action(self, state):
action = self.actor_network.action(state)
return action + self.exploration_noise.noise()
def action(self, state):
action = self.actor_network.action(state)
return action
def _record_log(self, reward, living_time):
summary_str = self.sess.run(self.summary_op, feed_dict={
self.reward_input: reward,
self.time_input: living_time
})
self.summary_writer.add_summary(summary_str, self.time_step)
return
def perceive(self, state, action, reward, next_state, done):
self.replay_buffer.add(state, action, reward, next_state, done)
if self.episode_start_time == 0.0:
self.episode_start_time = time.time()
# for testing
# self.time_step += 1
# if self.time_step == 100:
# print '--------------------------------'
# self.replay_buffer.save_to_pickle()
# return
self.episode_reward += reward
living_time = time.time() - self.episode_start_time
if self.time_step % 1000 == 0 or done:
self._record_log(self.episode_reward, living_time)
if self.replay_buffer.size() > REPLAY_START_SIZE:
self.train()
if self.time_step % 100000 == 0:
self.save_network()
if done:
print '===============reset noise========================='
self.exploration_noise.reset()
self.episode_reward = 0.0
self.episode_start_time = time.time()
self.time_step += 1
return
def load_time_step(self):
if not os.path.exists(self.dir_path):
return
files = os.listdir(self.dir_path)
step_list = []
for filename in files:
if ('meta' in filename) or ('-' not in filename):
continue
step_list.append(int(filename.split('-')[-1]))
step_list = sorted(step_list)
if len(step_list) == 0:
return
self.time_step = step_list[-1] + 1
return
def load_network(self):
checkpoint = tf.train.get_checkpoint_state(self.dir_path)
if checkpoint and checkpoint.model_checkpoint_path:
self.saver.restore(self.sess, checkpoint.model_checkpoint_path)
print 'Successfully loaded:', checkpoint.model_checkpoint_path
else:
print 'Could not find old network weights'
return
def save_network(self):
print 'save actor-critic network...', self.time_step
self.saver.save(self.sess, self.dir_path + '/ddpg', global_step=self.time_step)
return
class DDPG:
def __init__(self):
# Make sure all the directories exist
if not tf.gfile.Exists(TFLOG_PATH):
tf.gfile.MakeDirs(TFLOG_PATH)
if not tf.gfile.Exists(EXPERIENCE_PATH):
tf.gfile.MakeDirs(EXPERIENCE_PATH)
if not tf.gfile.Exists(NET_SAVE_PATH):
tf.gfile.MakeDirs(NET_SAVE_PATH)
# Initialize our session
self.session = tf.Session()
self.graph = self.session.graph
with self.graph.as_default():
# View the state batches
self.visualize_input = VISUALIZE_BUFFER
if self.visualize_input:
self.viewer = CostmapVisualizer()
# Hardcode input size and action size
self.height = 86
self.width = self.height
self.depth = 4
self.action_dim = 2
# Initialize the current action and the old action and old state for setting experiences
self.old_state = np.zeros((self.width, self.height, self.depth), dtype='int8')
self.old_action = np.ones(2, dtype='float')
self.network_action = np.zeros(2, dtype='float')
self.noise_action = np.zeros(2, dtype='float')
self.action = np.zeros(2, dtype='float')
# Initialize the grad inverter object to keep the action bounds
self.grad_inv = GradInverter(A0_BOUNDS, A1_BOUNDS, self.session)
# Make sure the directory for the data files exists
if not tf.gfile.Exists(DATA_PATH):
tf.gfile.MakeDirs(DATA_PATH)
# Initialize summary writers to plot variables during training
self.summary_op = tf.merge_all_summaries()
self.summary_writer = tf.train.SummaryWriter(TFLOG_PATH)
# Initialize actor and critic networks
self.actor_network = ActorNetwork(self.height, self.action_dim, self.depth, self.session,
self.summary_writer)
self.critic_network = CriticNetwork(self.height, self.action_dim, self.depth, self.session,
self.summary_writer)
# Initialize the saver to save the network params
self.saver = tf.train.Saver()
# initialize the experience data manger
self.data_manager = DataManager(BATCH_SIZE, EXPERIENCE_PATH, self.session)
# Uncomment if collecting a buffer for the autoencoder
# self.buffer = deque()
# Should we load the pre-trained params?
# If so: Load the full pre-trained net
# Else: Initialize all variables the overwrite the conv layers with the pretrained filters
if PRE_TRAINED_NETS:
self.saver.restore(self.session, NET_LOAD_PATH)
else:
self.session.run(tf.initialize_all_variables())
tf.train.start_queue_runners(sess=self.session)
time.sleep(1)
# Initialize a random process the Ornstein-Uhlenbeck process for action exploration
self.exploration_noise = OUNoise(self.action_dim, MU, THETA, SIGMA)
self.noise_flag = True
# Initialize time step
self.training_step = 0
# Flag: don't learn the first experience
self.first_experience = True
# After the graph has been filled add it to the summary writer
self.summary_writer.add_graph(self.graph)
def train(self):
# Check if the buffer is big enough to start training
if self.data_manager.enough_data():
# get the next random batch from the data manger
state_batch, \
action_batch, \
reward_batch, \
next_state_batch, \
is_episode_finished_batch = self.data_manager.get_next_batch()
state_batch = np.divide(state_batch, 100.0)
next_state_batch = np.divide(next_state_batch, 100.0)
# Are we visualizing the first state batch for debugging?
# If so: We have to scale up the values for grey scale before plotting
if self.visualize_input:
state_batch_np = np.asarray(state_batch)
state_batch_np = np.multiply(state_batch_np, -100.0)
state_batch_np = np.add(state_batch_np, 100.0)
self.viewer.set_data(state_batch_np)
self.viewer.run()
self.visualize_input = False
# Calculate y for the td_error of the critic
y_batch = []
next_action_batch = self.actor_network.target_evaluate(next_state_batch)
q_value_batch = self.critic_network.target_evaluate(next_state_batch, next_action_batch)
for i in range(0, BATCH_SIZE):
if is_episode_finished_batch[i]:
y_batch.append([reward_batch[i]])
else:
y_batch.append(reward_batch[i] + GAMMA * q_value_batch[i])
# Now that we have the y batch lets train the critic
self.critic_network.train(y_batch, state_batch, action_batch)
# Get the action batch so we can calculate the action gradient with it
# Then get the action gradient batch and adapt the gradient with the gradient inverting method
action_batch_for_gradients = self.actor_network.evaluate(state_batch)
q_gradient_batch = self.critic_network.get_action_gradient(state_batch, action_batch_for_gradients)
q_gradient_batch = self.grad_inv.invert(q_gradient_batch, action_batch_for_gradients)
# Now we can train the actor
self.actor_network.train(q_gradient_batch, state_batch)
# Save model if necessary
if self.training_step > 0 and self.training_step % SAVE_STEP == 0:
self.saver.save(self.session, NET_SAVE_PATH, global_step=self.training_step)
# Update time step
self.training_step += 1
self.data_manager.check_for_enqueue()
def get_action(self, state):
# normalize the state
state = state.astype(float)
state = np.divide(state, 100.0)
# Get the action
self.action = self.actor_network.get_action(state)
# Are we using noise?
if self.noise_flag:
# scale noise down to 0 at training step 3000000
if self.training_step < MAX_NOISE_STEP:
self.action += (MAX_NOISE_STEP - self.training_step) / MAX_NOISE_STEP * self.exploration_noise.noise()
# if action value lies outside of action bounds, rescale the action vector
if self.action[0] < A0_BOUNDS[0] or self.action[0] > A0_BOUNDS[1]:
self.action *= np.fabs(A0_BOUNDS[0]/self.action[0])
if self.action[1] < A0_BOUNDS[0] or self.action[1] > A0_BOUNDS[1]:
self.action *= np.fabs(A1_BOUNDS[0]/self.action[1])
# Life q value output for this action and state
self.print_q_value(state, self.action)
return self.action
def set_experience(self, state, reward, is_episode_finished):
# Make sure we're saving a new old_state for the first experience of every episode
if self.first_experience:
self.first_experience = False
else:
self.data_manager.store_experience_to_file(self.old_state, self.old_action, reward, state,
is_episode_finished)
# Uncomment if collecting data for the auto_encoder
# experience = (self.old_state, self.old_action, reward, state, is_episode_finished)
# self.buffer.append(experience)
if is_episode_finished:
self.first_experience = True
self.exploration_noise.reset()
# Safe old state and old action for next experience
self.old_state = state
self.old_action = self.action
def print_q_value(self, state, action):
string = "-"
q_value = self.critic_network.evaluate([state], [action])
stroke_pos = 30 * q_value[0][0] + 30
if stroke_pos < 0:
stroke_pos = 0
elif stroke_pos > 60:
stroke_pos = 60
print '[' + stroke_pos * string + '|' + (60-stroke_pos) * string + ']', "Q: ", q_value[0][0], \
"\tt: ", self.training_step
class DDPG:
"""docstring for DDPG"""
def __init__(self, state_space, action_dim):
self.name = 'DDPG' # name for uploading results
self.sess = tf.Session()
# Randomly initialize actor network and critic network
# with both their target networks
self.state_space = state_space
self.action_dim = action_dim # 1
self.ac_network = ActorCriticNetwork(self.sess, self.state_space,
self.action_dim)
# initialize replay buffer
self.replay_buffer = ReplayBuffer(REPLAY_BUFFER_SIZE)
# Initialize a random process the Ornstein-Uhlenbeck process for action exploration
self.exploration_noise = OUNoise(self.action_dim)
def train(self):
# Sample a random minibatch of N transitions from replay buffer
minibatch = self.replay_buffer.get_batch(BATCH_SIZE)
state_batch = np.asarray([data[0] for data in minibatch])
action_batch = np.asarray([data[1] for data in minibatch])
reward_batch = np.asarray([data[2] for data in minibatch])
next_state_batch = np.asarray([data[3] for data in minibatch])
done_batch = np.asarray([data[4] for data in minibatch])
# for action_dim = 1
action_batch = np.resize(action_batch, [BATCH_SIZE, self.action_dim])
# Get Q target label
# maxQ(s',a')
q_value_batch = self.ac_network.target_q(next_state_batch)
# Calculate target maxQ(s,a): y = reward + GAMMA * maxQ(s',a')
y_batch = []
batch_size = len(minibatch)
for i in range(batch_size):
if done_batch[i]:
y_batch.append(reward_batch[i])
else:
y_batch.append(reward_batch[i] + GAMMA * q_value_batch[i])
y_batch = np.resize(y_batch, [batch_size, 1])
# Update eval critic network by minimizing the loss L
cost = self.ac_network.train_critic(y_batch, state_batch, action_batch)
print('step_%d critic cost:' % self.ac_network.time_step, cost)
# Update eval actor policy using the sampled gradient:
self.ac_network.train_actor(state_batch)
# Update the target networks
self.ac_network.update_target()
def noise_action(self, state):
# Select action a_t according to the current policy and exploration noise
action = self.ac_network.actions(state)
return action[0] + self.exploration_noise.noise()
def action(self, state):
action = self.ac_network.actions([state])
return action[0]
def perceive(self, state, action, reward, next_state, done):
# Store transition (s_t,a_t,r_t,s_{t+1}) in replay buffer
self.replay_buffer.add(state, action, reward, next_state, done)
# Store transitions to replay start size then start training
if self.replay_buffer.count() > REPLAY_START_SIZE:
self.train()
#if self.time_step % 10000 == 0:
#self.actor_network.save_network(self.time_step)
#self.critic_network.save_network(self.time_step)
# Re-iniitialize the random process when an episode ends
if done:
self.exploration_noise.reset()
def sparse_tensor(self, state_batch, state_space):
row = len(state_batch)
indices = []
for r in range(row):
indices += [(r, c) for c in state_batch[r]]
values = [1.0 for i in range(len(indices))]
return tf.SparseTensorValue(indices=indices,
values=values,
dense_shape=[row, state_space])
N_STATES = env.observation_space.shape[0] #Number of observed states
N_ACTIONS = env.action_space.shape[0] #Number of input states
assert isinstance(env.observation_space, Box), "observation space must be continuous"
assert isinstance(env.action_space, Box), "action space must be continuous"
exploration_noise = OUNoise(1) #1 is number of input
agent = RDPG(env, N_STATES, N_ACTIONS,STEPS,BATCH_SIZE) #initialize critic, actor and corresponding target nets
exploration_noise = OUNoise(N_ACTIONS)
reward_st = np.array([0])
R = []
for i in xrange(EPISODES):
print("==== Starting episode no:",i,"====","\n")
o_t = env.reset()
reward_per_episode = 0
for t in xrange(STEPS):
a_t = agent.evaluate_actor(o_t,t) + exploration_noise.noise()
#a_t = random.uniform(U_MIN,U_MAX)
print("Action at step", t ," :",a_t,"\n")
o_t1, r_t, done, info = env.step(a_t)
#r_t = t #remove this
#print r_t
o_t = np.reshape(o_t,[1,N_STATES])
a_t = np.reshape(a_t,[1,N_ACTIONS])
r_t = np.reshape(r_t,[1,1])
if t == 0:
#initializing history at time, t = 0
h_t = np.hstack([o_t,a_t,r_t])
else:
h_t = np.append(h_t,np.hstack([o_t,a_t,r_t]),axis = 0)
reward_per_episode += r_t
def trainer(epochs=1000, MINIBATCH_SIZE=40, GAMMA = 0.99, epsilon=1.0, min_epsilon=0.01, BUFFER_SIZE=10000, train_indicator=True, render=False):
with tf.Session() as sess:
# configuring environment
env = gym.make(ENV_NAME)
# configuring the random processes
np.random.seed(RANDOM_SEED)
tf.set_random_seed(RANDOM_SEED)
env.seed(RANDOM_SEED)
# info of the environment to pass to the agent
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
action_bound = np.float64(10) # I choose this number since the mountain continuos does not have a boundary
# Creating agent
ruido = OUNoise(action_dim, mu = 0.4) # this is the Ornstein-Uhlenbeck Noise
actor = ActorNetwork(sess, state_dim, action_dim, action_bound, ACTOR_LEARNING_RATE, TAU, DEVICE)
critic = CriticNetwork(sess, state_dim, action_dim, CRITIC_LEARNING_RATE, TAU, actor.get_num_trainable_vars(), DEVICE)
sess.run(tf.global_variables_initializer())
# Initialize target network weights
actor.update_target_network()
critic.update_target_network()
# Initialize replay memory
replay_buffer = ReplayBuffer(BUFFER_SIZE, RANDOM_SEED)
goal = 0
max_state = -1.
try:
critic.recover_critic()
actor.recover_actor()
print('********************************')
print('models restored succesfully')
print('********************************')
except:
pass
# print('********************************')
# print('Failed to restore models')
# print('********************************')
for i in range(epochs):
state = env.reset()
state = np.hstack(state)
ep_reward = 0
ep_ave_max_q = 0
done = False
step = 0
max_state_episode = -1
epsilon -= (epsilon/EXPLORE)
epsilon = np.maximum(min_epsilon,epsilon)
while (not done):
if render:
env.render()
#print('step', step)
# 1. get action with actor, and add noise
action_original = actor.predict(np.reshape(state,(1,state_dim))) # + (10. / (10. + i))* np.random.randn(1)
action = action_original + max(epsilon,0)*ruido.noise()
# remove comment if you want to see a step by step update
# print(step,'a',action_original, action,'s', state[0], 'max state', max_state_episode)
# 2. take action, see next state and reward :
next_state, reward, done, info = env.step(action)
if train_indicator:
# 3. Save in replay buffer:
replay_buffer.add(np.reshape(state, (actor.s_dim,)), np.reshape(action, (actor.a_dim,)), reward,
done, np.reshape(next_state, (actor.s_dim,)))
# Keep adding experience to the memory until
# there are at least minibatch size samples
if replay_buffer.size() > MINIBATCH_SIZE:
# 4. sample random minibatch of transitions:
s_batch, a_batch, r_batch, t_batch, s2_batch = replay_buffer.sample_batch(MINIBATCH_SIZE)
# Calculate targets
# 5. Train critic Network (states,actions, R + gamma* V(s', a')):
# 5.1 Get critic prediction = V(s', a')
# the a' is obtained using the actor prediction! or in other words : a' = actor(s')
target_q = critic.predict_target(s2_batch, actor.predict_target(s2_batch))
# 5.2 get y_t where:
y_i = []
for k in range(MINIBATCH_SIZE):
if t_batch[k]:
y_i.append(r_batch[k])
else:
y_i.append(r_batch[k] + GAMMA * target_q[k])
# 5.3 Train Critic!
predicted_q_value, _ = critic.train(s_batch, a_batch, np.reshape(y_i, (MINIBATCH_SIZE, 1)))
ep_ave_max_q += np.amax(predicted_q_value)
# 6 Compute Critic gradient (depends on states and actions)
# 6.1 therefore I first need to calculate the actions the current actor would take.
a_outs = actor.predict(s_batch)
# 6.2 I calculate the gradients
grads = critic.action_gradients(s_batch, a_outs)
actor.train(s_batch, grads[0])
# Update target networks
actor.update_target_network()
critic.update_target_network()
state = next_state
if next_state[0] > max_state_episode:
max_state_episode = next_state[0]
ep_reward = ep_reward + reward
step +=1
if done:
ruido.reset()
if state[0] > 0.45:
#print('****************************************')
#print('got it!')
#print('****************************************')
goal += 1
if max_state_episode > max_state:
max_state = max_state_episode
print('th',i+1,'n steps', step,'R:', round(ep_reward,3),'Pos', round(epsilon,3),'Efficiency', round(100.*((goal)/(i+1.)),3) )
# print('Efficiency', 100.*((goal)/(i+1.)))
print('*************************')
print('now we save the model')
critic.save_critic()
actor.save_actor()
print('model saved succesfuly')
print('*************************')
class DDPG:
"""docstring for DDPG"""
def __init__(self, env, results_file):
self.name = 'DDPG' # name for uploading results
self.environment = env
# Randomly initialize actor network and critic network
# with both their target networks
self.state_dim = env.observation_space.shape[0]
self.action_dim = env.action_space.shape[0]
self.sess = tf.InteractiveSession()
self.actor_network = ActorNetwork(self.sess, self.state_dim,
self.action_dim)
self.critic_network = CriticNetwork(self.sess, self.state_dim,
self.action_dim)
# initialize replay buffer
self.replay_buffer = ReplayBuffer(REPLAY_BUFFER_SIZE)
# Initialize a random process the Ornstein-Uhlenbeck process for action exploration
self.exploration_noise = OUNoise(self.action_dim)
results_file.write(ActorNetwork.get_settings())
def train(self):
#print "train step",self.time_step
# Sample a random minibatch of N transitions from replay buffer
minibatch = self.replay_buffer.get_batch(BATCH_SIZE)
state_batch = np.asarray([data[0] for data in minibatch])
action_batch = np.asarray([data[1] for data in minibatch])
reward_batch = np.asarray([data[2] for data in minibatch])
next_state_batch = np.asarray([data[3] for data in minibatch])
done_batch = np.asarray([data[4] for data in minibatch])
# for action_dim = 1
action_batch = np.resize(action_batch, [BATCH_SIZE, self.action_dim])
# Calculate y_batch
next_action_batch = self.actor_network.target_actions(next_state_batch)
q_value_batch = self.critic_network.target_q(next_state_batch,
next_action_batch)
y_batch = []
for i in range(len(minibatch)):
if done_batch[i]:
y_batch.append(reward_batch[i])
else:
y_batch.append(reward_batch[i] + GAMMA * q_value_batch[i])
y_batch = np.resize(y_batch, [BATCH_SIZE, 1])
# Update critic by minimizing the loss L
self.critic_network.train(y_batch, state_batch, action_batch)
# Update the actor policy using the sampled gradient:
action_batch_for_gradients = self.actor_network.actions(state_batch)
q_gradient_batch = self.critic_network.gradients(
state_batch, action_batch_for_gradients)
self.actor_network.train(q_gradient_batch, state_batch)
# Update the target networks
self.actor_network.update_target()
self.critic_network.update_target()
def noise_action(self, state):
# Select action a_t according to the current policy and exploration noise
action = self.actor_network.action(state)
return action + self.exploration_noise.noise()
def action(self, state):
action = self.actor_network.action(state)
return action
def perceive(self, state, action, reward, next_state, done):
# Store transition (s_t,a_t,r_t,s_{t+1}) in replay buffer
self.replay_buffer.add(state, action, reward, next_state, done)
# Store transitions to replay start size then start training
if self.replay_buffer.count() > REPLAY_START_SIZE:
self.train()
#if self.time_step % 10000 == 0:
#self.actor_network.save_network(self.time_step)
#self.critic_network.save_network(self.time_step)
# Re-iniitialize the random process when an episode ends
if done:
self.exploration_noise.reset()
class PilotNode(object):
"""Node to listen to ROS topics like depth, rgb input and supervised control.
The node also publishes to pilot control and predicted depth for visualization.
"""
def __init__(self, FLAGS, model, logfolder):
print('initialize pilot node')
self.FLAGS=FLAGS
# Initialize fields
self.logfolder = logfolder
f=open(os.path.join(self.logfolder,'tf_log'),'a')
f.write(self.FLAGS.log_tag)
f.write('\n')
f.close()
self.model = model
self.ready=False
self.finished=True
self.training=False
self.last_pose=[] # previous pose, used for accumulative distance
self.world_name = ''
self.runs={'train':0, 'test':0} # number of online training run (used for averaging)
# self.accumlosses = {} # gather losses and info over the run in a dictionary
self.current_distance=0 # accumulative distance travelled from beginning of run used at evaluation
self.furthest_point=0 # furthest point reached from spawning point at the beginning of run
self.average_distances={'train':0, 'test':0} # running average over different runs
self.target_control = [] # field to keep the latest supervised control
self.target_depth = [] # field to keep the latest supervised depth
self.nfc_images =[] #used by n_fc networks for building up concatenated frames
self.exploration_noise = OUNoise(4, 0, self.FLAGS.ou_theta,1)
if not self.FLAGS.dont_show_depth: self.depth_pub = rospy.Publisher('/depth_prediction', numpy_msg(Floats), queue_size=1)
self.action_pub=rospy.Publisher('/nn_vel', Twist, queue_size=1)
self.model.reset_metrics()
rospy.Subscriber('/nn_start', Empty, self.ready_callback)
rospy.Subscriber('/nn_stop', Empty, self.finished_callback)
# extract imitation loss from supervised velocity
rospy.Subscriber('/supervised_vel', Twist, self.supervised_callback)
self.start_time = 0
self.imitation_loss=[]
self.depth_prediction=[]
self.depth_loss=[]
self.driving_duration=-1
self.skip_frames = 0
self.img_index = 0
self.fsm_index = 0
if rospy.has_param('rgb_image'):
image_topic=rospy.get_param('rgb_image')
if 'compressed' in image_topic:
rospy.Subscriber(image_topic, CompressedImage, self.compressed_image_callback)
else:
rospy.Subscriber(image_topic, Image, self.image_callback)
if rospy.has_param('depth_image'):
depth_topic = rospy.get_param('depth_image')
if 'scan' in depth_topic:
rospy.Subscriber(depth_topic, LaserScan, self.scan_depth_callback)
else:
rospy.Subscriber(depth_topic, Image, self.depth_callback)
if not self.FLAGS.real: # initialize the replay buffer
self.replay_buffer = ReplayBuffer(self.FLAGS, self.FLAGS.random_seed)
self.accumloss = 0
if rospy.has_param('gt_info'):
rospy.Subscriber(rospy.get_param('gt_info'), Odometry, self.gt_callback)
# Add some lines to debug delays:
self.time_im_received=[]
self.time_ctr_send=[]
rospy.init_node('pilot', anonymous=True)
def ready_callback(self,msg):
""" callback function that makes DNN policy starts the ready flag is set on 1 (for 3s)"""
if not self.ready and self.finished:
print('Neural control activated.')
self.ready = True
self.start_time = rospy.get_time()
self.finished = False
self.exploration_noise.reset()
# choose one speed for this flight
self.FLAGS.speed=self.FLAGS.speed + (not self.FLAGS.evaluate)*np.random.uniform(-self.FLAGS.sigma_x, self.FLAGS.sigma_x)
if rospy.has_param('evaluate'):
self.FLAGS.evaluate = rospy.get_param('evaluate')
print '--> set evaluate to: {0} with speed {1}'.format(self.FLAGS.evaluate, self.FLAGS.speed)
if rospy.has_param('skip_frames'):
self.skip_frames = rospy.get_param('skip_frames')
print '--> set skip_frames to: {0}'.format(self.skip_frames)
if rospy.has_param('world_name') :
self.world_name = rospy.get_param('world_name')
time.sleep(1) # wait one second, otherwise create_dataset can't follow...
def gt_callback(self, data):
"""Callback function that keeps track of positions for logging"""
if not self.ready or self.training: return
current_pos=[data.pose.pose.position.x,
data.pose.pose.position.y,
data.pose.pose.position.z]
if len(self.last_pose)!= 0:
self.current_distance += np.sqrt((self.last_pose[0,3]-current_pos[0])**2+(self.last_pose[1,3]-current_pos[1])**2)
self.furthest_point=max([self.furthest_point, np.sqrt(current_pos[0]**2+current_pos[1]**2)])
# Get pose (rotation and translation) [DEPRECATED: USED FOR ODOMETRY]
quaternion = (data.pose.pose.orientation.x,
data.pose.pose.orientation.y,
data.pose.pose.orientation.z,
data.pose.pose.orientation.w)
self.last_pose = transformations.quaternion_matrix(quaternion) # orientation of current frame relative to global frame
self.last_pose[0:3,3]=current_pos
def process_rgb(self, msg):
""" Convert RGB serial data to opencv image of correct size"""
try:
# Convert your ROS Image message to OpenCV2
# changed to normal RGB order as i ll use matplotlib and PIL instead of opencv
img =bridge.imgmsg_to_cv2(msg, 'rgb8')
except CvBridgeError as e:
print(e)
else:
img = img[::2,::5,:]
size = self.model.input_size[1:]
img = sm.resize(img,size,mode='constant').astype(float)
return img
def process_rgb_compressed(self, msg):
""" Convert RGB serial data to opencv image of correct size"""
# if not self.ready or self.finished: return []
try:
img = bridge.compressed_imgmsg_to_cv2(msg, desired_encoding='passthrough')
except CvBridgeError as e:
print(e)
else:
# 308x410 to 128x128
img = img[::2,::3,:]
size = self.model.input_size[1:]
img = sm.resize(img,size,mode='constant').astype(float)
return img
def process_depth(self, msg):
""" Convert depth serial data to opencv image of correct size"""
# if not self.ready or self.finished: return []
try:
# Convert your ROS Image message to OpenCV2
de = bridge.imgmsg_to_cv2(msg, desired_encoding='passthrough')#gets float of 32FC1 depth image
except CvBridgeError as e:
print(e)
else:
de = de[::6,::8]
shp=de.shape
# # assume that when value is not a number it is due to a too large distance (set to 5m)
# # values can be nan for when they are closer than 0.5m but than the evaluate node should
# # kill the run anyway.
de=np.asarray([ e*1.0 if not np.isnan(e) else 5 for e in de.flatten()]).reshape(shp) # clipping nans: dur: 0.010
size = (55,74)
# print 'DEPTH: min: ',np.amin(de),' and max: ',np.amax(de)
de = sm.resize(de,size,order=1,mode='constant', preserve_range=True)
return de
def process_scan(self, msg):
"""Preprocess serial scan: clip horizontal field of view, clip at 1's and ignore 0's, smooth over 4 bins."""
# field of view should follow camera:
# wide-angle camera: -60 to 60.
# normal camera: -35 to 35.
ranges=[1 if r > 1 or r==0 else r for r in msg.ranges]
# clip left 45degree range from 0:45 reversed with right 45degree range from the last 45:
ranges=list(reversed(ranges[:self.FLAGS.field_of_view/2]))+list(reversed(ranges[-self.FLAGS.field_of_view/2:]))
# add some smoothing by averaging over 4 neighboring bins
ranges = [sum(ranges[i*self.FLAGS.smooth_scan:i*self.FLAGS.smooth_scan+self.FLAGS.smooth_scan])/self.FLAGS.smooth_scan for i in range(int(len(ranges)/self.FLAGS.smooth_scan))]
# make it a numpy array
de = np.asarray(ranges).reshape((1,-1))
# if list(de.shape) != self.model.output_size: # reshape if necessary
# de = sm.resize(de,self.model.output_size,order=1,mode='constant', preserve_range=True)
return de
def compressed_image_callback(self, msg):
""" Process serial image data with process_rgb and concatenate frames if necessary"""
im = self.process_rgb_compressed(msg)
if len(im)!=0:
self.process_input(im)
def image_callback(self, msg):
""" Process serial image data with process_rgb and concatenate frames if necessary"""
self.time_im_received.append(time.time())
im = self.process_rgb(msg)
if len(im)!=0:
if 'nfc' in self.FLAGS.network: # when features are concatenated, multiple images should be kept.
self.nfc_images.append(im)
if len(self.nfc_images) < self.FLAGS.n_frames: return
else:
# concatenate last n-frames
im = np.concatenate(np.asarray(self.nfc_images[-self.FLAGS.n_frames:]),axis=2)
self.nfc_images = self.nfc_images[-self.FLAGS.n_frames+1:] # concatenate last n-1-frames
self.process_input(im)
def depth_callback(self, msg):
im = self.process_depth(msg)
if len(im)!=0 and self.FLAGS.auxiliary_depth:
self.target_depth = im
def scan_depth_callback(self, msg):
im = self.process_scan(msg)
if len(im)!=0:
self.depth = im
# calculate depth loss on the fly
if len(self.depth_prediction) != 0:
# print("pred: {0} trg: {1}".format(self.depth_prediction, self.depth))
self.depth_loss.append(np.mean((self.depth_prediction - self.depth.flatten())**2))
def process_input(self, im):
"""Process the inputs: images, targets, auxiliary tasks
Predict control based on the inputs.
Plot auxiliary predictions.
Fill replay buffer.
"""
# skip a number of frames to lower the actual control rate
# independently of the image frame rate
if self.skip_frames != 0:
self.img_index+=1
if self.img_index % (self.skip_frames+1) != 0:
return
aux_depth=[] # variable to keep predicted depth
trgt = -100.
inpt=im
if self.FLAGS.evaluate: ### EVALUATE
trgt=np.array([[self.target_control[5]]]) if len(self.target_control) != 0 else []
trgt_depth = np.array([copy.deepcopy(self.target_depth)]) if len(self.target_depth) !=0 and self.FLAGS.auxiliary_depth else []
control, aux_results = self.model.forward([inpt], auxdepth= not self.FLAGS.dont_show_depth,targets=trgt, depth_targets=trgt_depth)
if not self.FLAGS.dont_show_depth and self.FLAGS.auxiliary_depth and len(aux_results)>0: aux_depth = aux_results['d']
else: ###TRAINING
# Get necessary labels, if label is missing wait...
def check_field(target_name):
if len (target_name) == 0:
# print('Waiting for target {}'.format(target_name))
return False
else:
return True
if not check_field(self.target_control): return
else:
trgt = self.target_control[5]
if self.FLAGS.auxiliary_depth:
if not check_field(self.target_depth):
return
else:
trgt_depth = copy.deepcopy(self.target_depth)
control, aux_results = self.model.forward([inpt], auxdepth=not self.FLAGS.dont_show_depth)
if not self.FLAGS.dont_show_depth and self.FLAGS.auxiliary_depth: aux_depth = aux_results['d']
### SEND CONTROL
control = control[0]
# print control
if trgt != -100 and not self.FLAGS.evaluate: # policy mixing with self.FLAGS.alpha
action = trgt if np.random.binomial(1, self.FLAGS.alpha**(self.runs['train']+1)) else control
else:
action = control
msg = Twist()
msg.linear.x = self.FLAGS.speed
if self.FLAGS.noise == 'ou':
noise = self.exploration_noise.noise()
msg.linear.y = (not self.FLAGS.evaluate)*noise[1]*self.FLAGS.sigma_y
msg.linear.z = (not self.FLAGS.evaluate)*noise[2]*self.FLAGS.sigma_z
msg.angular.z = max(-1,min(1,action+(not self.FLAGS.evaluate)*self.FLAGS.sigma_yaw*noise[3]))
elif self.FLAGS.noise == 'uni':
# msg.linear.x = self.FLAGS.speed + (not self.FLAGS.evaluate)*np.random.uniform(-self.FLAGS.sigma_x, self.FLAGS.sigma_x)
msg.linear.y = (not self.FLAGS.evaluate)*np.random.uniform(-self.FLAGS.sigma_y, self.FLAGS.sigma_y)
msg.linear.z = (not self.FLAGS.evaluate)*np.random.uniform(-self.FLAGS.sigma_z, self.FLAGS.sigma_z)
msg.angular.z = max(-1,min(1,action+(not self.FLAGS.evaluate)*np.random.uniform(-self.FLAGS.sigma_yaw, self.FLAGS.sigma_yaw)))
else:
raise IOError( 'Type of noise is unknown: {}'.format(self.FLAGS.noise))
# if np.abs(msg.angular.z) > 0.3: msg.linear.x = 0.
if np.abs(msg.angular.z) > 0.3 and self.FLAGS.break_and_turn: msg.linear.x = 0. + np.random.binomial(1, 0.1)
self.action_pub.publish(msg)
self.time_ctr_send.append(time.time())
### keep track of imitation loss on the fly
if len(self.target_control) != 0:
self.imitation_loss.append((self.target_control[5]-action)**2)
if not self.FLAGS.dont_show_depth and len(aux_depth) != 0 and not self.finished:
aux_depth = aux_depth.flatten()
self.depth_pub.publish(aux_depth)
aux_depth = []
# ADD EXPERIENCE REPLAY
if not self.FLAGS.evaluate and trgt != -100:
experience={'state':im,
'action':action,
'trgt':trgt}
if self.FLAGS.auxiliary_depth: experience['target_depth']=trgt_depth
self.replay_buffer.add(experience)
# print("added experience: {0} vs {1}".format(action, trgt))
def supervised_callback(self, data):
"""Get target control from the /supervised_vel node"""
# print 'received control'
if not self.ready: return
self.target_control = [data.linear.x,
data.linear.y,
data.linear.z,
data.angular.x,
data.angular.y,
data.angular.z]
def finished_callback(self,msg):
"""When run is finished:
sample 10 batches from the replay buffer,
apply gradient descent on the model,
write log file and checkpoints away
"""
if self.ready and not self.finished:
print('neural control deactivated. @ time: {}'.format(time.time()))
self.ready=False
self.finished=True
if self.start_time!=0:
self.driving_duration = rospy.get_time() - self.start_time
# Train model from experience replay:
# Train the model with batchnormalization out of the image callback loop
depth_predictions = []
losses_train = {}
if self.replay_buffer.size()>self.FLAGS.batch_size and not self.FLAGS.evaluate:
for b in range(min(int(self.replay_buffer.size()/self.FLAGS.batch_size), 10)):
inputs, targets, aux_info = self.replay_buffer.sample_batch(self.FLAGS.batch_size)
if b==0:
if self.FLAGS.plot_depth and self.FLAGS.auxiliary_depth:
depth_predictions = tools.plot_depth(inputs, aux_info['target_depth'].reshape(-1,55,74))
depth_targets=[]
if self.FLAGS.auxiliary_depth:
depth_targets=aux_info['target_depth'].reshape(-1,55,74)
losses = self.model.backward(inputs,targets[:].reshape(-1,1),depth_targets)
for k in losses.keys():
try:
losses_train[k].append(losses[k])
except:
losses_train[k]=[losses[k]]
# Gather all info to build a proper summary and string of results
k='train' if not self.FLAGS.evaluate else 'test'
self.average_distances[k]= self.average_distances[k]-self.average_distances[k]/(self.runs[k]+1)
self.average_distances[k] = self.average_distances[k]+self.current_distance/(self.runs[k]+1)
self.runs[k]+=1
sumvar={}
result_string='{0}: run {1}'.format(time.strftime('%H:%M'),self.runs[k])
vals={'current':self.current_distance, 'furthest':self.furthest_point}
for d in ['current', 'furthest']:
name='Distance_{0}_{1}'.format(d,'train' if not self.FLAGS.evaluate else 'test')
if len(self.world_name)!=0: name='{0}_{1}'.format(name,self.world_name)
sumvar[name]=vals[d]
result_string='{0}, {1}:{2}'.format(result_string, name, vals[d])
for k in losses_train.keys():
name={'total':'Loss_train_total'}
sumvar[name[k]]=np.mean(losses_train[k])
result_string='{0}, {1}:{2}'.format(result_string, name[k], np.mean(losses_train[k]))
# get all metrics of this episode and add them to var
results = self.model.get_metrics()
for k in results.keys():
sumvar[k] = results[k]
result_string='{0}, {1}:{2}'.format(result_string, k, results[k])
if self.FLAGS.plot_depth and self.FLAGS.auxiliary_depth:
sumvar["depth_predictions"]=depth_predictions
# add driving duration (collision free)
if self.driving_duration != -1:
result_string='{0}, driving_duration: {1:0.3f}'.format(result_string, self.driving_duration)
sumvar['driving_time']=self.driving_duration
# add imitation loss
if len(self.imitation_loss)!=0:
result_string='{0}, imitation_loss: {1:0.3}'.format(result_string, np.mean(self.imitation_loss))
sumvar['imitation_loss']=np.mean(self.imitation_loss)
# add depth loss
if len(self.depth_loss)!=0:
result_string='{0}, depth_loss: {1:0.3f}, depth_loss_var: {2:0.3f}'.format(result_string, np.mean(self.depth_loss), np.var(self.depth_loss))
sumvar['depth_loss']=np.mean(self.depth_loss)
if len(self.time_ctr_send) > 10 and len(self.time_im_received) > 10:
# calculate control-rates and rgb-rates from differences
avg_ctr_rate = 1/np.mean([self.time_ctr_send[i+1]-self.time_ctr_send[i] for i in range(len(self.time_ctr_send)-1)])
std_ctr_delays = np.std([self.time_ctr_send[i+1]-self.time_ctr_send[i] for i in range(len(self.time_ctr_send)-1)])
avg_im_rate = 1/np.mean([self.time_im_received[i+1]-self.time_im_received[i] for i in range(1,len(self.time_im_received)-1)]) #skip first image delay as network still needs to 'startup'
std_im_delays = np.std([self.time_ctr_send[i+1]-self.time_ctr_send[i] for i in range(len(self.time_ctr_send)-1)])
result_string='{0}, control_rate: {1:0.3f}, image_rate: {2:0.3f} , control_delay_std: {1:0.3f}, image_delay_std: {2:0.3f} '.format(result_string, avg_ctr_rate, avg_im_rate, std_ctr_delays, std_im_delays)
try:
self.model.summarize(sumvar)
except Exception as e:
print('failed to write', e)
pass
else:
print(result_string)
# ! Note: tf_log is used by evaluate_model train_model and train_and_evaluate_model in simulation_supervised/scripts
# Script starts next run once this file is updated.
try:
f=open(os.path.join(self.logfolder,'tf_log'),'a')
f.write(result_string)
f.write('\n')
f.close()
except Exception as e:
print('failed to write txt tf_log {}'.format(e))
print('retry after sleep 60')
time.sleep(60)
f=open(os.path.join(self.logfolder,'tf_log'),'a')
f.write(result_string)
f.write('\n')
f.close()
# self.accumlosses = {}
self.current_distance = 0
self.last_pose = []
self.nfc_images = []
self.furthest_point = 0
self.world_name = ''
if self.runs['train']%20==0 and not self.FLAGS.evaluate:
# Save a checkpoint every 20 runs.
self.model.save(self.logfolder)
print('model saved [run {0}]'.format(self.runs['train']))
self.time_im_received=[]
self.time_ctr_send=[]
self.model.reset_metrics()
self.start_time=0
self.imitation_loss=[]
self.depth_loss=[]
self.driving_duration=-1
self.img_index=0
self.fsm_index = 0
def main():
experiment= 'InvertedPendulum-v1'
env= gym.make(experiment)
assert isinstance(env.observation_space, Box), "observation space must be continuous"
assert isinstance(env.action_space, Box), "action space must be continuous"
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env)
exploration_noise = OUNoise(env.action_space.shape[0])
counter=0
total_reward=0
num_states = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
#saving reward:
reward_st = np.array([0])
for i in xrange(episodes):
observation = env.reset()
reward_per_episode = 0
for t in xrange(steps):
#rendering environmet (optional)
#env.render()
x = observation
#select action using actor network model
action = agent.evaluate_actor(np.reshape(x,[num_actions,num_states]))
noise = exploration_noise.noise()
action = action[0] + noise
print 'Agent.Action :',action
print '\n'
print '\n'
observation,reward,done,[]=env.step(action)
#add s_t,s_t+1,action,reward to experience memeroy
agent.add_experience(x,observation,action,reward,done)
#train critic and actor network
if counter > 64:
agent.train()
reward_per_episode+=reward
counter+=1
#check if episode ends:
if done:
print 'EPISODE: ',i,' Steps: ',t,' Total Reward: ',reward_per_episode
exploration_noise.reset()
reward_st = np.append(reward_st,reward_per_episode)
np.savetxt('episode_reward.txt',reward_st, newline="\n")
print '\n'
print '\n'
break
total_reward+=reward_per_episode
print "Average reward per episode {}".format(total_reward / episodes)
def train_rdpg():
env = GazeboEnv()
object_name_list = env.GetObjectName()
map_table = env.RandomiseObject(object_name_list)
robot = GazeboRobot('robot1', flags.max_steps)
rate = rospy.Rate(robot.control_freq)
exploration_noise = OUNoise(action_dimension=flags.a_dim,
mu=flags.mu, theta=flags.theta, sigma=flags.sigma)
config = tf.ConfigProto()
config.gpu_options.allow_growth=True
with tf.Session(config=config) as sess:
agent = RDPG(flags, sess)
trainable_var = tf.trainable_variables()
model_dir = os.path.join(flags.model_dir, flags.model_name)
if not os.path.exists(model_dir):
os.makedirs(model_dir)
# summary
print " [*] printing trainable variables"
for idx, v in enumerate(trainable_var):
print " var {:3}: {:15} {}".format(idx, str(v.get_shape()), v.name)
with tf.name_scope(v.name.replace(':0', '')):
variable_summaries(v)
reward_ph = tf.placeholder(tf.float32, [], name='reward')
q_ph = tf.placeholder(tf.float32, [], name='q_pred')
tf.summary.scalar('reward', reward_ph)
tf.summary.scalar('q_estimate', q_ph)
merged = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(model_dir, sess.graph)
# model saver
saver = tf.train.Saver(trainable_var, max_to_keep=3)
sess.run(tf.global_variables_initializer())
# start training
T = 0
episode = 0
while T < flags.total_steps:
t = 0
init_pose = robot.SetInitialPose(map_table)
seq = []
total_reward = 0
loop_time = []
depth_img = robot.GetDepthImageObservation()
depth_stack = np.stack([depth_img, depth_img, depth_img], axis=-1)
while not rospy.is_shutdown():
start_time = time.time()
reward, terminate = robot.GetRewardAndTerminate(t)
if t > 0:
seq.append((depth_img, action, reward))
depth_img = robot.GetDepthImageObservation()
depth_stack = np.stack([depth_img, depth_stack[:, :, 0], depth_stack[:, :, 1]], axis=-1)
action = agent.ActorPredict([depth_stack], t)[0]
if episode < flags.noise_stop_episode:
action += exploration_noise.noise() * np.asarray(agent.action_range)
action = robot.SelfControl(action, agent.action_range)
total_reward += reward
if (T + 1) % flags.steps_per_checkpoint == 0:
saver.save(sess, os.path.join(model_dir, 'network') , global_step=episode)
if terminate:
if len(seq) > 3:
agent.Add2Mem(seq)
if episode > agent.batch_size:
for train_step in xrange(2):
q = agent.Train()
summary = sess.run(merged, feed_dict={reward_ph: total_reward,
q_ph: np.amax(q)})
summary_writer.add_summary(summary, episode)
print 'Episode:{:} | Steps:{:} | Reward:{:.2f} | T:{:} | Q:{:.2f}'.format(episode,
t,
total_reward,
T,
np.amax(q))
else:
print 'Episode:{:} | Steps:{:} | Reward:{:.2f} | T:{:}'.format(episode,
t,
total_reward,
T)
episode +=1
T += 1
break
t += 1
T += 1
rate.sleep()
class DDPG:
"""docstring for DDPG"""
def __init__(self, environment):
self.name = 'DDPG' # name for uploading results
self.environment = environment
# Randomly initialize actor network and critic network
# with both their target networks
self.actor_network = ActorNetwork(state_size = environment.observation_space.shape[0],action_size = environment.action_space.shape[0])
self.critic_network = CriticNetwork(state_size = environment.observation_space.shape[0],action_size = environment.action_space.shape[0])
# initialize replay buffer
self.replay_buffer = deque()
# Initialize a random process the Ornstein-Uhlenbeck process for action exploration
self.exploration_noise = OUNoise(environment.action_space.shape[0])
# Initialize time step
self.time_step = 0
def set_init_observation(self,observation):
# receive initial observation state
self.state = observation
def train(self):
# Sample a random minibatch of N transitions from replay buffer
minibatch = random.sample(self.replay_buffer,BATCH_SIZE)
state_batch = [data[0] for data in minibatch]
action_batch = [data[1] for data in minibatch]
reward_batch = [data[2] for data in minibatch]
next_state_batch = [data[3] for data in minibatch]
action_batch = np.resize(action_batch,[BATCH_SIZE,1])
# Calculate y
y_batch = []
next_action_batch = self.actor_network.target_evaluate(next_state_batch)
q_value_batch = self.critic_network.target_evaluate(next_state_batch,next_action_batch)
for i in range(0,BATCH_SIZE):
done = minibatch[i][4]
if done:
y_batch.append(reward_batch[i])
else:
y_batch.append(reward_batch[i] + GAMMA * q_value_batch[i])
# Update critic by minimizing the loss L
self.critic_network.train(y_batch,state_batch,action_batch)
# Update the actor policy using the sampled gradient:
action_batch_for_gradients = self.actor_network.evaluate(state_batch)
q_gradient_batch = self.critic_network.gradients(state_batch,action_batch_for_gradients)/BATCH_SIZE
self.actor_network.train(q_gradient_batch,state_batch)
# Update the target networks
self.actor_network.update_target()
self.critic_network.update_target()
def get_action(self):
# Select action a_t according to the current policy and exploration noise
action = self.actor_network.get_action(self.state)
return np.clip(action+self.exploration_noise.noise(),self.environment.action_space.low,self.environment.action_space.high)
def set_feedback(self,observation,action,reward,done):
# Store transition (s_t,a_t,r_t,s_{t+1}) in replay buffer
next_state = observation
self.replay_buffer.append((self.state,action,reward,next_state,done))
# Update current state
self.state = next_state
# Update time step
self.time_step += 1
# Limit the replay buffer size
if len(self.replay_buffer) > REPLAY_BUFFER_SIZE:
self.replay_buffer.popleft()
# Store transitions to replay start size then start training
if self.time_step > REPLAY_START_SIZE:
self.train()
if self.time_step % 10000 == 0:
self.actor_network.save_network(self.time_step)
self.critic_network.save_network(self.time_step)
# Re-iniitialize the random process when an episode ends
if done:
self.exploration_noise.reset()
def s2l():
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
num_states = feature_size #num_states = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
print("Number of States:", num_states)
print("Number of Actions:", num_actions)
agent = DDPG(env, is_batch_norm, num_states, num_actions)
exploration_noise = OUNoise(env.action_space.shape[0])
counter = 0
total_reward = 0
print("Number of Rollouts per episode:", num_rollouts)
print("Number of Steps per roll out:", steps)
reward_st = np.array([0]) #saving reward
eval_metric_st = np.array([0])
reward_st_all = np.array([0]) #saving reward after every step
frame_obj = Frame_Feature()
#activity_obj=Vid_Feature()
demo_vid_array = demo_array_extractor(demo_folder)
demo_features = frame_obj.video_feature_extractor(demo_vid_array)
for episode in range(num_episodes):
print("==== Starting episode no:", episode, "====", "\n")
env.reset() # Reset env in the begining of each episode
env.render()
obs_img = env.render(mode='rgb_array') # Get the observation
obs_img = np.array(misc.imresize(obs_img, [112, 112, 3]))
observation = np.array(frame_obj.frame_feature_extractor(obs_img))
observation = observation.reshape(-1)
reward_per_episode = 0
for t in range(num_rollouts):
reward_per_rollout = 0
vid_robo_ = []
for i in range(steps):
x = observation
action = agent.evaluate_actor(np.reshape(x, [1, num_states]))
noise = exploration_noise.noise()
action = action[
0] + noise #Select action according to current policy and exploration noise
print('Action at episode-', episode, 'rollout-', t, 'step-', i,
" :", action)
_, _, done, info = env.step(action)
env.render()
obs_robo_ = env.render(mode='rgb_array') # Get the observation
obs_robo = misc.imresize(obs_robo_, [112, 112, 3])
vid_robo_.append(obs_robo)
observation = np.array(
frame_obj.frame_feature_extractor(np.array(obs_robo)))
observation = observation.reshape(-1)
#pasue()
if (i == 15):
vid_robo = np.array(vid_robo_)
robo_features = frame_obj.video_feature_extractor(vid_robo)
reward = -(distance(demo_features, robo_features))
reward = np.array(reward)
print('reward: ', reward)
else:
reward = 0
reward = np.array(reward)
print('reward: ', reward)
# Printing eval_metric after every rollout
eval_metric = np.array(env.get_eval())
eval_metric = eval_metric.reshape(-1)
print('Distance to goal:', eval_metric)
eval_metric_st = np.append(eval_metric_st, eval_metric)
np.savetxt('eval_metric_per_step.txt',
eval_metric_st,
newline="\n")
# Storing reward after every rollout
reward_st_all = np.append(reward_st_all, reward)
np.savetxt('reward_all.txt', reward_st_all, newline="\n")
#add s_t,s_t+1,action,reward to experience memory
agent.add_experience(x, observation, action, reward, False)
reward_per_rollout += reward
counter += 1
#train critic and actor network
if counter > start_training:
agent.train()
print('\n\n')
#Saving policy
if ((episode % 50) == 0 and t == num_rollouts - 1):
print('saving policy...........................!')
agent.save_actor(episode)
reward_per_episode += reward_per_rollout
#check if episode ends:
print('EPISODE: ', episode, ' Total Reward: ', reward_per_episode)
print("Printing reward to file")
exploration_noise.reset(
) #reinitializing random noise for action exploration
reward_st = np.append(reward_st, reward_per_episode)
np.savetxt('episode_reward.txt', reward_st, fmt='%f', newline="\n")
print('\n\n')
total_reward += reward_per_episode
print("Average reward per episode {}".format(total_reward / num_episodes))
for i in range(MAX_EPISODES):
t_start = time.time()
sd = i * 3 + 100
m_sd, s, gm, loc = env.set_state_seed(sd)
lm = env.get_local_map(loc)
exploration_noise.reset()
ep_reward = 0
ave_dw = 0
j = 0
r = 0
for j in range(MAX_EP_STEPS):
# Add exploration noise
a = ddpg.choose_action(s, gm, lm)
ave_dw += np.linalg.norm(a)
a += exploration_noise.noise(
) # add randomness to action selection for exploration
a = np.minimum(a_bound1, np.maximum(-a_bound1, a))
a[0:4] /= max(np.linalg.norm(a[0:4]), 1e-8)
s_, loc_, r, done = env.step(a)
ddpg.memory.add(m_sd, s, loc, a, r, s_, loc_)
replay_num += 1
if ddpg.pointer > REPLAY_START:
ddpg.learn()
s = s_
loc = loc_
lm = env.get_local_map(loc_)
ep_reward += r
class AgentDDPG:
def __init__(self, env, state_size, action_size):
self.env = env
self.replay_memory = deque()
self.actor_network = actor_network.ActorNetwork(
state_size, action_size)
self.critic_network = critic_network.CriticNetwork(
state_size, action_size)
self.ou_noise = OUNoise(action_size)
self.time_step = 0
def set_state(self, obs):
self.state = obs
def get_action(self):
# Select action a_t according to the current policy and exploration noise
action = self.actor_network.get_action(self.state)
return np.clip(action + self.ou_noise.noise(),
self.env.action_space.low, self.env.action_space.high)
def set_feedback(self, obs, action, reward, done):
next_state = obs
self.replay_memory.append(
(self.state, action, reward, next_state, done))
self.state = next_state
self.time_step += 1
if len(self.replay_memory) > config.MEMORY_SIZE:
self.replay_memory.popleft()
# Store transitions to replay start size then start training
if self.time_step > config.OBSERVATION_STEPS:
self.train()
if self.time_step % config.SAVE_EVERY_X_STEPS == 0:
self.actor_network.save_network(self.time_step)
self.critic_network.save_network(self.time_step)
# reinit the random process when an episode ends
if done:
self.ou_noise.reset()
def train(self):
minibatch = random.sample(self.replay_memory, config.MINI_BATCH_SIZE)
state_batch = [data[0] for data in minibatch]
action_batch = [data[1] for data in minibatch]
reward_batch = [data[2] for data in minibatch]
next_state_batch = [data[3] for data in minibatch]
action_batch = np.resize(action_batch, [config.MINI_BATCH_SIZE, 1])
# Calculate y
y_batch = []
next_action_batch = self.actor_network.get_target_action_batch(
next_state_batch)
q_value_batch = self.critic_network.get_target_q_batch(
next_state_batch, next_action_batch)
for i in range(0, config.MINI_BATCH_SIZE):
done = minibatch[i][4]
if done:
y_batch.append(reward_batch[i])
else:
y_batch.append(reward_batch[i] +
config.FUTURE_REWARD_DISCOUNT *
q_value_batch[i])
y_batch = np.array(y_batch)
y_batch = np.reshape(y_batch, [len(y_batch), 1])
# Update critic by minimizing the loss
self.critic_network.train(y_batch, state_batch, action_batch)
# Update the actor policy using the sampled gradient:
action_batch = self.actor_network.get_action_batch(state_batch)
q_gradient_batch = self.critic_network.get_gradients(
state_batch, action_batch)
self.actor_network.train(q_gradient_batch, state_batch)
# Update the target networks
self.actor_network.update_target()
self.critic_network.update_target()
class DDPG:
def __init__(self, env):
self.name = 'DDPG' # name for uploading results
self.environment = env
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
# Initialize time step
self.time_step = 0
# initialize replay buffer
self.replay_buffer = deque()
# initialize networks
self.create_networks_and_training_method(state_dim,action_dim)
self.sess = tf.InteractiveSession()
self.sess.run(tf.initialize_all_variables())
# loading networks
self.saver = tf.train.Saver()
checkpoint = tf.train.get_checkpoint_state("saved_networks")
if checkpoint and checkpoint.model_checkpoint_path:
self.saver.restore(self.sess, checkpoint.model_checkpoint_path)
print "Successfully loaded:", checkpoint.model_checkpoint_path
else:
print "Could not find old network weights"
global summary_writer
summary_writer = tf.train.SummaryWriter('~/logs',graph=self.sess.graph)
def create_networks_and_training_method(self,state_dim,action_dim):
theta_p = networks.theta_p(state_dim,action_dim)
theta_q = networks.theta_q(state_dim,action_dim)
target_theta_p,target_update_p = self.exponential_moving_averages(theta_p,TAU)
target_theta_q,target_update_q = self.exponential_moving_averages(theta_q,TAU)
self.state = tf.placeholder(tf.float32,[None,state_dim],'state')
self.action_test = networks.policy_network(self.state,theta_p)
# Initialize a random process the Ornstein-Uhlenbeck process for action exploration
self.exploration = OUNoise(action_dim)
noise = self.exploration.noise()
self.action_exploration = self.action_test + noise
q = networks.q_network(self.state,self.action_test,theta_q)
# policy optimization
mean_q = tf.reduce_mean(q)
weight_decay_p = tf.add_n([L2_POLICY * tf.nn.l2_loss(var) for var in theta_p])
loss_p = -mean_q + weight_decay_p
optim_p = tf.train.AdamOptimizer(P_LEARNING_RATE)
grads_and_vars_p = optim_p.compute_gradients(loss_p, var_list=theta_p)
optimize_p = optim_p.apply_gradients(grads_and_vars_p)
with tf.control_dependencies([optimize_p]):
self.train_p = tf.group(target_update_p)
# q optimization
self.action_train = tf.placeholder(tf.float32,[None,action_dim],'action_train')
self.reward = tf.placeholder(tf.float32,[None],'reward')
self.next_state = tf.placeholder(tf.float32,[None,state_dim],'next_state')
self.done = tf.placeholder(tf.bool,[None],'done')
q_train = networks.q_network(self.state,self.action_train,theta_q)
next_action = networks.policy_network(self.next_state,theta=target_theta_p)
next_q = networks.q_network(self.next_state,next_action,theta=target_theta_q)
q_target = tf.stop_gradient(tf.select(self.done,self.reward,self.reward + GAMMA * next_q))
# q loss
q_error = tf.reduce_mean(tf.square(q_target - q_train))
weight_decay_q = tf.add_n([L2_Q * tf.nn.l2_loss(var) for var in theta_q])
loss_q = q_error + weight_decay_q
optim_q = tf.train.AdamOptimizer(Q_LEARNING_RATE)
grads_and_vars_q = optim_q.compute_gradients(loss_q, var_list=theta_q)
optimize_q = optim_q.apply_gradients(grads_and_vars_q)
with tf.control_dependencies([optimize_q]):
self.train_q = tf.group(target_update_q)
tf.scalar_summary("loss_q",loss_q)
tf.scalar_summary("loss_p",loss_p)
tf.scalar_summary("q_mean",mean_q)
global merged_summary_op
merged_summary_op = tf.merge_all_summaries()
def train(self):
#print "train step",self.time_step
# Sample a random minibatch of N transitions from replay buffer
minibatch = random.sample(self.replay_buffer,BATCH_SIZE)
state_batch = [data[0] for data in minibatch]
action_batch = [data[1] for data in minibatch]
reward_batch = [data[2] for data in minibatch]
next_state_batch = [data[3] for data in minibatch]
done_batch = [data[4] for data in minibatch]
_,_,summary_str = self.sess.run([self.train_p,self.train_q,merged_summary_op],feed_dict={
self.state:state_batch,
self.action_train:action_batch,
self.reward:reward_batch,
self.next_state:next_state_batch,
self.done:done_batch
})
summary_writer.add_summary(summary_str,self.time_step)
# save network every 1000 iteration
if self.time_step % 1000 == 0:
self.saver.save(self.sess, 'saved_networks/' + 'network' + '-ddpg', global_step = self.time_step)
def noise_action(self,state):
# Select action a_t according to the current policy and exploration noise
action = self.sess.run(self.action_exploration,feed_dict={
self.state:[state]
})[0]
return np.clip(action,self.environment.action_space.low,self.environment.action_space.high)
def action(self,state):
action = self.sess.run(self.action_test,feed_dict={
self.state:[state]
})[0]
return np.clip(action,self.environment.action_space.low,self.environment.action_space.high)
def perceive(self,state,action,reward,next_state,done):
# Store transition (s_t,a_t,r_t,s_{t+1}) in replay buffer
self.replay_buffer.append((state,action,reward,next_state,done))
# Update time step
self.time_step += 1
# Limit the replay buffer size
if len(self.replay_buffer) > REPLAY_BUFFER_SIZE:
self.replay_buffer.popleft()
# Store transitions to replay start size then start training
if self.time_step > REPLAY_START_SIZE:
self.train()
# Re-iniitialize the random process when an episode ends
if done:
self.exploration.reset()
# f fan-in size
def exponential_moving_averages(self,theta, tau=0.001):
ema = tf.train.ExponentialMovingAverage(decay=1 - tau)
update = ema.apply(theta) # also creates shadow vars
averages = [ema.average(x) for x in theta]
return averages, update