class DDPGAgent:
def __init__(self, state_size, action_size, num_agents,
hidden_actor, hidden_critic, lr_actor, lr_critic,
buffer_size, agent_id, use_PER=False, seed=0):
super(DDPGAgent, self).__init__()
self.seed = torch.manual_seed(seed)
self.agent_id = agent_id
# num_agents*action_size
self.actor_local = ActorNet(state_size, hidden_actor, action_size, seed=seed).to(device)
self.critic_local = CriticNet(num_agents*state_size, num_agents*action_size, hidden_critic, 1, seed=seed).to(device)
self.actor_target = ActorNet(state_size, hidden_actor, action_size, seed=seed).to(device)
self.critic_target = CriticNet(num_agents*state_size, num_agents*action_size, hidden_critic, 1, seed=seed).to(device)
self.actor_optimizer = Adam(self.actor_local.parameters(), lr=lr_actor)
self.critic_optimizer = Adam(self.critic_local.parameters(), lr=lr_critic, weight_decay=0.) #weight_decay=1.e-5
self.memory = ReplayBuffer(buffer_size, num_agents, state_size, action_size, use_PER)
# initialize targets same as original networks
hard_update(self.actor_target, self.actor_local)
hard_update(self.critic_target, self.critic_local)
#self.actor_target.eval() #wont be training target net
#self.critic_target.eval()
def _act(self, obs):
obs = obs.to(device)
if len(obs.shape)==1: obs = obs.unsqueeze(0) #for batchnorm
self.actor_local.eval()
with torch.no_grad():
action_local = self.actor_local(obs).squeeze()
self.actor_local.train()
return action_local #tensor, action_size torch.Size([2])
def _target_act(self, obs):
obs = obs.to(device)
if len(obs.shape)==1: obs = obs.unsqueeze(0) #for batchnorm
with torch.no_grad():
action_target = self.actor_target(obs).squeeze()
return action_target
class Learner:
def __init__(self, learner_id, config, dev, shared_state, shared_queue):
self.action_size = config['action_space']
self.obs_size = config['obs_space']
self.shared_queue = shared_queue
self.shared_state = shared_state
self.dev = dev
self.id = learner_id
self.burn_in_length = config['burn_in_length'] # 40-80
self.learning_length = config['learning_length']
self.sequence_length = self.burn_in_length + self.learning_length
self.n_step = config['n_step']
self.sequence = []
self.recurrent_state = []
self.priority = []
self.td_loss = deque(maxlen=self.learning_length)
self.gamma = config['gamma']
# self.actor_parameter_update_interval = config['actor_parameter_update_interval']
self.actor = ActorNet(dev, config).to(self.dev)
self.target_actor = ActorNet(dev, config).to(self.dev)
self.critic = CriticNet(dev, config).to(self.dev)
self.target_critic = CriticNet(dev, config).to(self.dev)
self.actor.load_state_dict(self.shared_state["actor"].state_dict())
self.target_actor.load_state_dict(
self.shared_state["target_actor"].state_dict())
self.critic.load_state_dict(self.shared_state["critic"].state_dict())
self.target_critic.load_state_dict(
self.shared_state["target_critic"].state_dict())
# self.actor.load_state_dict(self.shared_state["actor"])
# self.target_actor.load_state_dict(self.shared_state["target_actor"])
# self.critic.load_state_dict(self.shared_state["critic"])
# self.target_critic.load_state_dict(self.shared_state["target_critic"])
self.learner_actor_rate = config['learner_actor_rate']
self.num_actors = learner_id
self.n_actions = 1
self.max_frame = config['learner_max_frame']
self.memory_sequence_size = config['memory_sequence_size']
self.batch_size = config['batch_size']
self.memory = LearnerReplayMemory(self.memory_sequence_size, config,
dev)
self.model_path = './'
# self.memory_path = './memory_data/'
# self.model_save_interval = 10 # 50
self.learner_parameter_update_interval = config[
'learner_parameter_update_interval'] # 50
self.target_update_inverval = config['target_update_interval'] # 100
self.gamma = config['gamma']
self.actor_lr = config['actor_lr']
self.critic_lr = config['critic_lr']
self.actor_optimizer = optim.Adam(self.actor.parameters(),
lr=self.actor_lr)
self.actor_criterion = nn.MSELoss()
self.critic_optimizer = optim.Adam(self.critic.parameters(),
lr=self.critic_lr)
self.critic_criterion = nn.MSELoss()
def __del__(self):
self.shared_queue.close()
self.shared_state.close()
# self.save_model()
def save_model(self):
model_dict = {
'actor': self.actor.state_dict(),
'target_actor': self.target_actor.state_dict(),
'critic': self.critic.state_dict(),
'target_critic': self.target_critic.state_dict()
}
torch.save(model_dict, self.model_path + 'model.pt')
def update_target_model(self):
self.target_actor.load_state_dict(self.actor.state_dict())
self.target_critic.load_state_dict(self.critic.state_dict())
def run(self):
time_check(-1)
while self.memory.size() < self.batch_size:
self.memory.append(self.shared_queue.get(block=True))
# self.memory.append(self.shared_queue.get())
print('\rmem size: ', self.memory.size(), end='\r')
time_check(1)
count_mem = 0
frame = 0
win_v = vis.line(Y=torch.Tensor([0]), opts=dict(title='V_loss'))
win_p = vis.line(Y=torch.Tensor([0]), opts=dict(title='P_loss'))
while frame < self.max_frame:
# sleep(0.0001)
# if self.shared_queue.qsize()==0 and count_mem <0:
# self.memory.append(self.shared_queue.get(block=True))
#
# for i in range(self.shared_queue.qsize()):
# self.memory.append(self.shared_queue.get(block=False))
# count_mem += self.learner_actor_rate
# print('waiting shared q {}/{}'.format(self.memory.size(),self.batch_size))
# self.shared_state['frame'][self.id]=frame
# while self.shared_state['sleep'][self.id] :
# sleep(0.5)
# if self.shared_queue.qsize()==0 and count_mem <0:
# self.memory.append(self.shared_queue.get(block=True))
# self.memory.append(self.shared_queue.get())
# for i in range(self.shared_queue.qsize()):
## global_buf.append(self.shared_queue.get())
# self.memory.append(self.shared_queue.get())
# count_mem += self.learner_actor_rate
if self.shared_queue.qsize() != 0:
self.memory.append(self.shared_queue.get(block=True))
frame += 1
count_mem -= 1
episode_index, sequence_index, obs_seq, action_seq, reward_seq, gamma_seq, a_state, ta_state, c_state, tc_state = self.memory.sample(
)
self.actor.set_state(a_state[0], a_state[1])
self.target_actor.set_state(ta_state[0], ta_state[1])
self.critic.set_state(c_state[0], c_state[1])
self.target_critic.set_state(tc_state[0], tc_state[1])
### burn-in step ###
_ = [self.actor(obs_seq[i]) for i in range(self.burn_in_length)]
_ = [
self.critic(obs_seq[i], action_seq[i])
for i in range(self.burn_in_length)
]
_ = [
self.target_actor(obs_seq[i])
for i in range(self.burn_in_length + self.n_step)
]
_ = [
self.target_critic(obs_seq[i], action_seq[i])
for i in range(self.burn_in_length + self.n_step)
]
### learning steps ###
# update ciritic
q_value = torch.zeros(self.learning_length * self.batch_size,
self.n_actions)
target_q_value = torch.zeros(
self.learning_length * self.batch_size, self.n_actions)
for i in range(self.learning_length):
obs_i = self.burn_in_length + i
next_obs_i = self.burn_in_length + i + self.n_step
q_value[i * self.batch_size:(i + 1) *
self.batch_size] = self.critic(obs_seq[obs_i],
action_seq[obs_i])
with torch.no_grad():
next_q_value = self.target_critic(
obs_seq[next_obs_i],
self.target_actor(obs_seq[next_obs_i]))
target_q_val = reward_seq[obs_i] + (
gamma_seq[next_obs_i]**self.n_step) * next_q_value
# target_q_val = invertical_vf(target_q_val)
target_q_value[i * self.batch_size:(i + 1) *
self.batch_size] = target_q_val
critic_loss = self.actor_criterion(q_value,
target_q_value.detach())
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
# update actor
self.actor.reset_state()
self.critic.reset_state()
actor_loss = torch.zeros(self.learning_length * self.batch_size,
self.n_actions).to(self.dev)
for i in range(self.learning_length):
obs_i = i + self.burn_in_length
action = self.actor(obs_seq[obs_i])
actor_loss[i * self.batch_size:(i + 1) *
self.batch_size] = -self.critic(
obs_seq[obs_i], self.actor(obs_seq[obs_i]))
actor_loss = actor_loss.mean()
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# update target networks
if frame % self.target_update_inverval == 0:
self.update_target_model()
print('#', frame, 'critic_loss:', critic_loss.item(),
' actor_loss:', actor_loss.item(), ' count:', count_mem)
win_p = vis.line(X=torch.Tensor([frame]),
Y=torch.Tensor([actor_loss.item()]),
win=win_p,
update='append')
win_v = vis.line(X=torch.Tensor([frame]),
Y=torch.Tensor([critic_loss.item()]),
win=win_v,
update='append')
# calc priority
average_td_loss = ((q_value - target_q_value)**2).detach().to(
self.dev)
# average_td_loss = np.mean(((q_value - target_q_value)**2).detach().cpu().numpy() , axis = 1)
for i in range(len(episode_index)):
td = average_td_loss[i:-1:self.batch_size]
self.memory.priority[episode_index[i]][
sequence_index[i]] = calc_priority(td).cpu().view(1, -1)
self.memory.total_priority[episode_index[i]] = torch.cat(
self.memory.priority[episode_index[i]]).sum(0).view(1, -1)
# self.memory.priority[episode_index[i]][sequence_index[i]] = calc_priority(td)
# self.memory.total_priority[episode_index[i]] = sum(self.memory.priority[episode_index[i]])
# if frame % self.model_save_interval == 0:
# self.save_model()
if frame % self.learner_parameter_update_interval == 0:
# print('learner update ')
# [self.shared_state["actor"][k] = v.cpu() for k,v in self.actor.state_dict().item() ]
# [self.shared_state["target_actor"][k] = v.cpu() for k,v in self.target_actor.state_dict().item() ]
# [self.shared_state["critic"][k] = v.cpu() for k,v in self.critic.state_dict().item() ]
# [self.shared_state["target_critic"][k] = v.cpu() for k,v in self.target_critic.state_dict().item() ]
#
# for k,v in self.actor.state_dict().items():
# self.shared_state["actor"][k] = v.cpu()
# for k,v in self.target_actor.state_dict().items():
# self.shared_state["target_actor"][k] = v.cpu()
# for k,v in self.critic.state_dict().items():
# self.shared_state["critic"][k] = v.cpu()
# for k,v in self.target_critic.state_dict().items():
# self.shared_state["target_critic"][k] = v.cpu()
# self.shared_state["actor"] = self.actor.state_dict()
# self.shared_state["target_actor"] = self.target_actor.state_dict()
# self.shared_state["critic"] = self.critic.state_dict()
# self.shared_state["target_critic"] = self.target_critic.state_dict()
self.shared_state["actor"].load_state_dict(
self.actor.state_dict())
self.shared_state["critic"].load_state_dict(
self.critic.state_dict())
self.shared_state["target_actor"].load_state_dict(
self.target_actor.state_dict())
self.shared_state["target_critic"].load_state_dict(
self.target_critic.state_dict())
for i in range(self.num_actors):
self.shared_state["update"][i] = True
print('learner_update', self.actor.policy_l0.weight.data[0][0])
self.actor.reset_state()
self.target_actor.reset_state()
self.critic.reset_state()
self.target_critic.reset_state()
class D4PG_Agent:
"""
PyTorch Implementation of D4PG:
"Distributed Distributional Deterministic Policy Gradients"
(Barth-Maron, Hoffman, et al., 2018)
As described in the paper at: https://arxiv.org/pdf/1804.08617.pdf
Much thanks also to the original DDPG paper:
"Continuous Control with Deep Reinforcement Learning"
(Lillicrap, Hunt, et al., 2016)
https://arxiv.org/pdf/1509.02971.pdf
And to:
"A Distributional Perspective on Reinforcement Learning"
(Bellemare, Dabney, et al., 2017)
https://arxiv.org/pdf/1707.06887.pdf
D4PG utilizes distributional value estimation, n-step returns,
prioritized experience replay (PER), distributed K-actor exploration,
and off-policy actor-critic learning to achieve very fast and stable
learning for continuous control tasks.
This version of the Agent is written to interact with Udacity's
Continuous Control robotic arm manipulation environment which provides
20 simultaneous actors, negating the need for K-actor implementation.
Thus, this code has no multiprocessing functionality. It could be easily
added as part of the main.py script.
In the original D4PG paper, it is suggested in the data that PER does
not have significant (or perhaps any at all) effect on the speed or
stability of learning. Thus, it too has been left out of this
implementation but may be added as a future TODO item.
"""
def __init__(self,
env,
args,
e_decay=1,
e_min=0.05,
l2_decay=0.0001,
update_type="hard"):
"""
Initialize a D4PG Agent.
"""
self.device = args.device
self.framework = "D4PG"
self.eval = args.eval
self.agent_count = env.agent_count
self.actor_learn_rate = args.actor_learn_rate
self.critic_learn_rate = args.critic_learn_rate
self.batch_size = args.batch_size
self.buffer_size = args.buffer_size
self.action_size = env.action_size
self.state_size = env.state_size
self.C = args.C
self._e = args.e
self.e_decay = e_decay
self.e_min = e_min
self.gamma = args.gamma
self.rollout = args.rollout
self.tau = args.tau
self.update_type = update_type
self.num_atoms = args.num_atoms
self.vmin = args.vmin
self.vmax = args.vmax
self.atoms = torch.linspace(self.vmin, self.vmax,
self.num_atoms).to(self.device)
self.t_step = 0
self.episode = 0
# Set up memory buffers, currently only standard replay is implemented #
self.memory = ReplayBuffer(self.device, self.buffer_size, self.gamma,
self.rollout)
# Initialize ACTOR networks #
self.actor = ActorNet(args.layer_sizes, self.state_size,
self.action_size).to(self.device)
self.actor_target = ActorNet(args.layer_sizes, self.state_size,
self.action_size).to(self.device)
self._hard_update(self.actor, self.actor_target)
self.actor_optim = optim.Adam(self.actor.parameters(),
lr=self.actor_learn_rate,
weight_decay=l2_decay)
# Initialize CRITIC networks #
self.critic = CriticNet(args.layer_sizes, self.state_size,
self.action_size,
self.num_atoms).to(self.device)
self.critic_target = CriticNet(args.layer_sizes, self.state_size,
self.action_size,
self.num_atoms).to(self.device)
self._hard_update(self.actor, self.actor_target)
self.critic_optim = optim.Adam(self.critic.parameters(),
lr=self.critic_learn_rate,
weight_decay=l2_decay)
self.new_episode()
def act(self, states, eval=False):
"""
Predict an action using a policy/ACTOR network π.
Scaled noise N (gaussian distribution) is added to all actions todo
encourage exploration.
"""
states = states.to(self.device)
with torch.no_grad():
actions = self.actor(states).detach().cpu().numpy()
if not eval:
noise = self._gauss_noise(actions.shape)
actions += noise
return np.clip(actions, -1, 1)
def step(self, states, actions, rewards, next_states, pretrain=False):
"""
Add the current SARS' tuple into the short term memory, then learn
"""
# Current SARS' stored in short term memory, then stacked for NStep
experience = list(zip(states, actions, rewards, next_states))
self.memory.store_experience(experience)
self.t_step += 1
# Learn after done pretraining
if not pretrain:
self.learn()
def learn(self):
"""
Performs a distributional Actor/Critic calculation and update.
Actor πθ and πθ'
Critic Zw and Zw' (categorical distribution)
"""
# Sample from replay buffer, REWARDS are sum of ROLLOUT timesteps
# Already calculated before storing in the replay buffer.
# NEXT_STATES are ROLLOUT steps ahead of STATES
batch = self.memory.sample(self.batch_size)
states, actions, rewards, next_states = batch
atoms = self.atoms.unsqueeze(0)
# Calculate Yᵢ from target networks using πθ' and Zw'
# These tensors are not needed for backpropogation, so detach from the
# calculation graph (literally doubles runtime if this is not detached)
target_dist = self._get_targets(rewards, next_states).detach()
# Calculate log probability DISTRIBUTION using Zw w.r.t. stored actions
log_probs = self.critic(states, actions, log=True)
# Calculate the critic network LOSS (Cross Entropy), CE-loss is ideal
# for categorical value distributions as utilized in D4PG.
# estimates distance between target and projected values
critic_loss = -(target_dist * log_probs).sum(-1).mean()
# Predict action for actor network loss calculation using πθ
predicted_action = self.actor(states)
# Predict value DISTRIBUTION using Zw w.r.t. action predicted by πθ
probs = self.critic(states, predicted_action)
# Multiply probabilities by atom values and sum across columns to get
# Q-Value
expected_reward = (probs * atoms).sum(-1)
# Calculate the actor network LOSS (Policy Gradient)
# Take the mean across the batch and multiply in the negative to
# perform gradient ascent
actor_loss = -expected_reward.mean()
# Perform gradient ascent
self.actor_optim.zero_grad()
actor_loss.backward()
self.actor_optim.step()
# Perform gradient descent
self.critic_optim.zero_grad()
critic_loss.backward()
self.critic_optim.step()
self._update_networks()
self.actor_loss = actor_loss.item()
self.critic_loss = critic_loss.item()
def initialize_memory(self, pretrain_length, env):
"""
Fills up the ReplayBuffer memory with PRETRAIN_LENGTH number of experiences
before training begins.
"""
if len(self.memory) >= pretrain_length:
print("Memory already filled, length: {}".format(len(self.memory)))
return
print("Initializing memory buffer.")
states = env.states
while len(self.memory) < pretrain_length:
actions = np.random.uniform(-1, 1,
(self.agent_count, self.action_size))
next_states, rewards, dones = env.step(actions)
self.step(states, actions, rewards, next_states, pretrain=True)
if self.t_step % 10 == 0 or len(self.memory) >= pretrain_length:
print("Taking pretrain step {}... memory filled: {}/{}\
".format(self.t_step, len(self.memory), pretrain_length))
states = next_states
print("Done!")
self.t_step = 0
def _get_targets(self, rewards, next_states):
"""
Calculate Yᵢ from target networks using πθ' and Zw'
"""
target_actions = self.actor_target(next_states)
target_probs = self.critic_target(next_states, target_actions)
# Project the categorical distribution onto the supports
projected_probs = self._categorical(rewards, target_probs)
return projected_probs
def _categorical(self, rewards, probs):
"""
Returns the projected value distribution for the input state/action pair
While there are several very similar implementations of this Categorical
Projection methodology around github, this function owes the most
inspiration to Zhang Shangtong and his excellent repository located at:
https://github.com/ShangtongZhang
"""
# Create local vars to keep code more concise
vmin = self.vmin
vmax = self.vmax
atoms = self.atoms
num_atoms = self.num_atoms
gamma = self.gamma
rollout = self.rollout
rewards = rewards.unsqueeze(-1)
delta_z = (vmax - vmin) / (num_atoms - 1)
# Rewards were stored with 0->(N-1) summed, take Reward and add it to
# the discounted expected reward at N (ROLLOUT) timesteps
projected_atoms = rewards + gamma**rollout * atoms.unsqueeze(0)
projected_atoms.clamp_(vmin, vmax)
b = (projected_atoms - vmin) / delta_z
# It seems that on professional level GPUs (for instance on AWS), the
# floating point math is accurate to the degree that a tensor printing
# as 99.00000 might in fact be 99.000000001 in the backend, perhaps due
# to binary imprecision, but resulting in 99.00000...ceil() evaluating
# to 100 instead of 99. Forcibly reducing the precision to the minimum
# seems to be the only solution to this problem, and presents no issues
# to the accuracy of calculating lower/upper_bound correctly.
precision = 1
b = torch.round(b * 10**precision) / 10**precision
lower_bound = b.floor()
upper_bound = b.ceil()
m_lower = (upper_bound +
(lower_bound == upper_bound).float() - b) * probs
m_upper = (b - lower_bound) * probs
projected_probs = torch.tensor(np.zeros(probs.size())).to(self.device)
for idx in range(probs.size(0)):
projected_probs[idx].index_add_(0, lower_bound[idx].long(),
m_lower[idx].double())
projected_probs[idx].index_add_(0, upper_bound[idx].long(),
m_upper[idx].double())
return projected_probs.float()
@property
def e(self):
"""
This property ensures that the annealing process is run every time that
E is called.
Anneals the epsilon rate down to a specified minimum to ensure there is
always some noisiness to the policy actions. Returns as a property.
"""
self._e = max(self.e_min, self._e * self.e_decay)
return self._e
def _gauss_noise(self, shape):
"""
Returns the epsilon scaled noise distribution for adding to Actor
calculated action policy.
"""
n = np.random.normal(0, 1, shape)
return self.e * n
def new_episode(self):
"""
Handle any cleanup or steps to begin a new episode of training.
"""
self.memory.init_n_step()
self.episode += 1
def _update_networks(self):
"""
Updates the network using either DDPG-style soft updates (w/ param \
TAU), or using a DQN/D4PG style hard update every C timesteps.
"""
if self.update_type == "soft":
self._soft_update(self.actor, self.actor_target)
self._soft_update(self.critic, self.critic_target)
elif self.t_step % self.C == 0:
self._hard_update(self.actor, self.actor_target)
self._hard_update(self.critic, self.critic_target)
def _soft_update(self, active, target):
"""
Slowly updated the network using every-step partial network copies
modulated by parameter TAU.
"""
for t_param, param in zip(target.parameters(), active.parameters()):
t_param.data.copy_(self.tau * param.data +
(1 - self.tau) * t_param.data)
def _hard_update(self, active, target):
"""
Fully copy parameters from active network to target network. To be used
in conjunction with a parameter "C" that modulated how many timesteps
between these hard updates.
"""
target.load_state_dict(active.state_dict())
class Agent:
def __init__(self, env):
"""Args:
env(gym.Core.env): environment
"""
with open('./configuration.json') as config_file:
self.config = json.load(config_file)['agent']
self.env = env
self.state = None
self.epsilon = None
self.epsilon_decay = None
self.state_size = None
self.actions_size = None
self.actor = None
self.actor_target = None
self.critic = None
self.critic_target = None
self.actor_optim = None
self.critic_optim = None
self.gamma = None
self.memory = None
self.batch_size = None
self.action_space = None
self.normalizer = None
def __str__(self):
return 'RL_Agent Object'
def reset(self):
self.action_space = self.env.action_space
obs_space = self.env.observation_space.spaces
obs_len = obs_space['observation'].shape[0]
goal_len = obs_space['desired_goal'].shape[0]
self.state_size = obs_len + goal_len
self.actions_size = self.action_space.shape[0]
max_action = float(self.env.action_space.high[0])
self.actor = ActorNet(self.state_size, *self.config['net_sizes'],
self.actions_size, max_action)
self.critic = CriticNet(self.state_size, *self.config['net_sizes'],
self.actions_size)
self.actor_target = ActorNet(self.state_size,
*self.config['net_sizes'],
self.actions_size, max_action)
self.critic_target = CriticNet(self.state_size,
*self.config['net_sizes'],
self.actions_size)
self.actor_optim = Adam(self.actor.parameters(),
lr=self.config['learning_rate'])
self.critic_optim = Adam(self.critic.parameters(),
lr=self.config['learning_rate'])
self.update(self.critic_target, self.critic, 1)
self.update(self.actor_target, self.actor, 1)
self.epsilon = self.config['epsilon']
self.epsilon_decay = self.config['epsilon_decay']
self.gamma = self.config['gamma']
if self.config['PER']:
self.memory = self.memory = PrioritizedMemory(
self.config['memory_size'], self.config["memory_alpha"],
self.config["memory_epsilon"], self.config["memory_beta"],
self.config["memory_beta_increment"])
else:
self.memory = ReplayBuffer(self.config['memory_size'])
self.batch_size = self.config['batch_size']
self.normalizer = Normalizer(obs_len, goal_len)
# warm up the normalizer
self.normalizer.observe(self.env.reset())
def run(self, train):
total_reward = 0
done = False
self.state = self.env.reset()
self.normalizer.observe(self.state)
self.state = self.normalizer.normalize(self.state)
ep_transitions = []
# start episode
while not done:
if self.config['render']:
self.env.render()
# act and observe
action = self._get_action_epsilon_greedy(self.state)
obs, reward, done, info = self.env.step(action)
total_reward += reward
# normalize the state
self.normalizer.observe(obs)
obs = self.normalizer.normalize(obs)
# save the transition for later HER processing
transition = [self.state, reward, action, obs, not done]
ep_transitions.append(transition)
# save to memory
self.append_sample_to_memory(
*copy.deepcopy((flatten_state_dict_for_model(self.state),
reward, action,
flatten_state_dict_for_model(obs), not done)))
self.state = obs
if random.random() < self.config["her-probability"]:
her_trs = generate_her_transitions(ep_transitions,
self.env.compute_reward,
self.config['her-type'],
self.config['her-k_value'])
for t in her_trs:
self.append_sample_to_memory(*t)
if len(self.memory) > self.batch_size * 5 and train:
for i in range(40):
self._train()
self.soft_update_networks()
if self.epsilon > self.config['epsilon_min']:
self.epsilon *= self.epsilon_decay
return total_reward
def _train(self):
batch, indexes, importance_sampling_weights = None, None, None
if self.config['PER']:
batch, indexes, importance_sampling_weights = \
self.sample_from_per_memory(self.batch_size)
importance_sampling_weights = torch.Tensor(
importance_sampling_weights)
else:
batch = self.memory.get_random_batch(self.batch_size)
state_batch = torch.Tensor(batch[0])
reward_batch = torch.Tensor(batch[1])
action_batch = torch.Tensor(batch[2])
next_state_batch = torch.Tensor(batch[3])
# unused - see additional info in the Readme
# mask_batch = torch.Tensor(batch[4] * 1)
next_q_values = self.critic_target(next_state_batch,
self.actor_target(next_state_batch))
expected_q_values = reward_batch + (self.gamma * next_q_values)
expected_q_values = expected_q_values.clamp_(-50., 0.).detach()
self.critic_optim.zero_grad()
q_values = self.critic(state_batch, action_batch)
if self.config['PER']:
errors = torch.abs(q_values - expected_q_values)
critic_loss = (importance_sampling_weights * errors**2).sum()
for i in range(self.batch_size):
index = indexes[i]
self.memory.update(index, errors[i].detach().numpy())
else:
critic_loss = mse_loss(q_values, expected_q_values)
critic_loss.backward()
self.critic_optim.step()
self.actor_optim.zero_grad()
policy_loss = self.critic(state_batch, self.actor(state_batch))
action_reg = (self.actor.action_preact**2).mean()
policy_loss = -policy_loss.mean() + action_reg
policy_loss.backward()
self.actor_optim.step()
def get_action_greedy(self, state):
"""Hey, actor - act!... plus detach().numpy() ..."""
return self.actor(flatten_state_dict_for_model(state)).detach().numpy()
def _get_action_epsilon_greedy(self, state):
"""Returns an action for given state by using the actor network.
With epsilon probability, it returns a fully random action.
In both cases, there is a OU noise added as well.
Parameters can be specified in the configuration file.
"""
if random.random() > self.epsilon:
action = self.get_action_greedy(state) + \
np.random.normal(scale=0.2, size=self.actions_size)
else:
action = self.env.action_space.sample()
return np.clip(action, -1., 1.)
def append_sample_to_memory(self, state, reward, action, next_state, done):
"""Adds given transition to the memory. In case of using Prioritized
Experience Replay, it calculates the TD error."""
if not self.config['PER']:
self.memory.append((state, reward, action, next_state, done))
else:
q = self.critic(
torch.Tensor(state).unsqueeze(0),
torch.Tensor(action).unsqueeze(0))
target_val = self.critic_target(
torch.Tensor(next_state).unsqueeze(0),
self.actor_target(torch.Tensor(next_state).unsqueeze(0)))
target = reward + (self.gamma * target_val * (done * 1)).detach()
error = abs(q - target).detach().numpy()
self.memory.add((state, reward, action, next_state, done), error)
def soft_update_networks(self):
self.update(self.critic_target, self.critic,
self.config['network_update_amount'])
self.update(self.actor_target, self.actor,
self.config['network_update_amount'])
def update(self, target, src, amount):
for target_param, param in zip(target.parameters(), src.parameters()):
target_param.data.copy_(target_param.data * (1.0 - amount) +
param.data * amount)
def sample_from_per_memory(self, batch_size):
transition_batch, indexes, importance_sampling_weights = \
self.memory.sample(batch_size)
x, r, u, y, d = [], [], [], [], []
for i in transition_batch:
X, R, U, Y, D = i
x.append(np.array(X, copy=False))
y.append(np.array(Y, copy=False))
u.append(np.array(U, copy=False))
r.append(np.array(R, copy=False))
d.append(np.array(D, copy=False))
return ((np.array(x), np.array(r).reshape(-1,
1), np.array(u), np.array(y),
np.array(d).reshape(-1,
1)), indexes, importance_sampling_weights)
class Learner:
def __init__(self, n_actors):
self.env = suite.load(domain_name="walker", task_name="run")
self.n_actions = self.env.action_spec().shape[0]
self.obs_size = get_obs(self.env.reset().observation).shape[1]
self.n_actors = n_actors
self.burn_in_length = 20 # 40-80
self.learning_length = 40
self.sequence_length = self.burn_in_length + self.learning_length
self.n_step = 5
self.memory_sequence_size = 5000000
self.batch_size = 32
self.memory = LearnerReplayMemory(
memory_sequence_size=self.memory_sequence_size,
batch_size=self.batch_size)
self.model_path = './model_data/'
self.memory_path = './memory_data/'
self.actor = ActorNet(self.obs_size, self.n_actions, 0).cuda()
self.target_actor = deepcopy(self.actor).eval()
self.critic = CriticNet(self.obs_size, self.n_actions, 0).cuda()
self.target_critic = deepcopy(self.critic).eval()
self.model_save_interval = 50 # 50
self.memory_update_interval = 50 # 50
self.target_update_inverval = 500 # 100
self.gamma = 0.997
self.actor_lr = 1e-4
self.critic_lr = 1e-3
self.actor_optimizer = optim.Adam(self.actor.parameters(),
lr=self.actor_lr)
self.actor_criterion = nn.MSELoss()
self.critic_optimizer = optim.Adam(self.critic.parameters(),
lr=self.critic_lr)
self.critic_criterion = nn.MSELoss()
self.save_model()
def save_model(self):
model_dict = {
'actor': self.actor.state_dict(),
'target_actor': self.target_actor.state_dict(),
'critic': self.critic.state_dict(),
'target_critic': self.target_critic.state_dict()
}
torch.save(model_dict, self.model_path + 'model.pt')
def update_target_model(self):
self.target_actor.load_state_dict(self.actor.state_dict())
self.target_critic.load_state_dict(self.critic.state_dict())
def run(self):
# memory not enough
while self.memory.sequence_counter < self.batch_size * 100:
for i in range(self.n_actors):
is_memory = os.path.isfile(self.memory_path +
'/memory{}.pt'.format(i))
if is_memory:
self.memory.load(i)
sleep(0.1)
print('learner memory sequence size:',
self.memory.sequence_counter)
step = 0
while True:
if step % 100 == 0:
print('learning step:', step)
start = time()
step += 1
episode_index, sequence_index, obs_seq, action_seq, reward_seq, terminal_seq, a_state, ta_state, c_state, tc_state = self.memory.sample(
)
self.actor.set_state(a_state[0], a_state[1])
self.target_actor.set_state(ta_state[0], ta_state[1])
self.critic.set_state(c_state[0], c_state[1])
self.target_critic.set_state(tc_state[0], tc_state[1])
### burn-in step ###
_ = [self.actor(obs) for obs in obs_seq[0:self.burn_in_length]]
_ = [
self.critic(obs, action)
for obs, action in zip(obs_seq[0:self.burn_in_length],
action_seq[0:self.burn_in_length])
]
_ = [
self.target_actor(obs)
for obs in obs_seq[0:self.burn_in_length + self.n_step]
]
_ = [
self.target_critic(obs, action) for obs, action in zip(
obs_seq[0:self.burn_in_length +
self.n_step], action_seq[0:self.burn_in_length +
self.n_step])
]
### learning steps ###
# update ciritic
q_value = torch.zeros(self.learning_length * self.batch_size,
self.n_actions).cuda()
target_q_value = torch.zeros(
self.learning_length * self.batch_size, self.n_actions).cuda()
for i in range(self.learning_length):
obs_i = self.burn_in_length + i
next_obs_i = self.burn_in_length + i + self.n_step
q_value[i * self.batch_size:(i + 1) *
self.batch_size] = self.critic(obs_seq[obs_i],
action_seq[obs_i])
next_q_value = self.target_critic(
obs_seq[next_obs_i],
self.target_actor(obs_seq[next_obs_i]))
target_q_val = reward_seq[obs_i] + (
self.gamma**self.n_step) * (
1. - terminal_seq[next_obs_i - 1]) * next_q_value
target_q_val = invertical_vf(target_q_val)
target_q_value[i * self.batch_size:(i + 1) *
self.batch_size] = target_q_val
critic_loss = self.actor_criterion(q_value,
target_q_value.detach())
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
# update actor
self.actor.reset_state()
self.critic.reset_state()
actor_loss = torch.zeros(self.learning_length * self.batch_size,
self.n_actions).cuda()
for i in range(self.learning_length):
obs_i = i + self.burn_in_length
action = self.actor(obs_seq[obs_i])
actor_loss[i * self.batch_size:(i + 1) *
self.batch_size] = -self.critic(
obs_seq[obs_i], self.actor(obs_seq[obs_i]))
actor_loss = actor_loss.mean()
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# update target networks
if step % self.target_update_inverval == 0:
self.update_target_model()
# calc priority
average_td_loss = np.mean(
(q_value - target_q_value).detach().cpu().numpy()**2., axis=1)
for i in range(len(episode_index)):
td = average_td_loss[i:-1:self.batch_size]
self.memory.priority[episode_index[i]][
sequence_index[i]] = calc_priority(td)
self.memory.total_priority[episode_index[i]] = sum(
self.memory.priority[episode_index[i]])
if step % self.model_save_interval == 0:
self.save_model()
if step % self.memory_update_interval == 0:
for i in range(self.n_actors):
is_memory = os.path.isfile(self.memory_path +
'/memory{}.pt'.format(i))
if is_memory:
self.memory.load(i)
sleep(0.1)
self.actor.reset_state()
self.target_actor.reset_state()
self.critic.reset_state()
self.target_critic.reset_state()