class DQN(RLAlgorithm):
def __init__(self, env, do_render, num_threads, gamma, lr,
global_max_episode):
state_size, action_size = env.observation_space.shape[
0], env.action_space.n
self.qnetwork_global = QNetwork(state_size, action_size) #.to(device)
self.qnetwork_global.share_memory()
self.qnetwork_target = QNetwork(state_size, action_size) #.to(device)
self.qnetwork_target.share_memory()
self.agents = [
DQNAgent(id=id,
env=env,
do_render=do_render,
state_size=state_size,
action_size=action_size,
n_episodes=global_max_episode,
lr=lr,
gamma=gamma,
update_every=UPDATE_EVERY + num_threads,
global_network=self.qnetwork_global,
target_network=self.qnetwork_target)
for id in range(num_threads)
]
def train(self):
[agent.start() for agent in self.agents]
[agent.join() for agent in self.agents]
def __init__(self, state_size, action_size, num_agents, double_dqn=True):
self.action_size = action_size
self.double_dqn = double_dqn
# Q-Network
self.qnetwork_local = QNetwork(state_size, action_size).to(device)
self.qnetwork_target = copy.deepcopy(self.qnetwork_local)
self.optimizer = torch.optim.Adam(self.qnetwork_local.parameters(),
lr=LR)
# Replay memory
self.memory = ReplayBuffer(BUFFER_SIZE)
self.num_agents = num_agents
self.t_step = 0
def __init__(self, state_size, action_size, seed):
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = QNetwork(state_size, action_size, seed).to(device) #prediction net
self.qnetwork_target = QNetwork(state_size, action_size, seed).to(device) #target network
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# store SARS when reach the batch size train
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
self.t_step = 0 # time step
def __init__(self, state_size, action_size, seed, num_threads,
update_every, lr):
super(ParameterServer, self).__init__()
# TODO: Figure out how to just create a tensor that can do ASGD without the model needed
self.model = QNetwork(state_size, action_size, seed) # TODO remove
self.time = mp.Value('i', 0)
p = [p for p in self.model.parameters()]
# params = [torch.nn.Parameter(p[i].clone().detach()) for i in range(len(p))]
# [g.share_memory_() for g in params] # Store gradients in shared memory
# self.parameters = params
self.qs = [mp.Queue() for q in range(len(p))]
self.shard_mem = [p[i].share_memory_() for i in range(len(p))]
self.shards = [
ParameterServerShard(p[i], self.shard_mem[i], lr, self.qs[i])
for i in range(len(p))
]
[shard.start() for shard in self.shards]
class ParameterServer():
def __init__(self, state_size, action_size, seed, num_threads,
update_every, lr):
super(ParameterServer, self).__init__()
# TODO: Figure out how to just create a tensor that can do ASGD without the model needed
self.model = QNetwork(state_size, action_size, seed) # TODO remove
self.time = mp.Value('i', 0)
p = [p for p in self.model.parameters()]
# params = [torch.nn.Parameter(p[i].clone().detach()) for i in range(len(p))]
# [g.share_memory_() for g in params] # Store gradients in shared memory
# self.parameters = params
self.qs = [mp.Queue() for q in range(len(p))]
self.shard_mem = [p[i].share_memory_() for i in range(len(p))]
self.shards = [
ParameterServerShard(p[i], self.shard_mem[i], lr, self.qs[i])
for i in range(len(p))
]
[shard.start() for shard in self.shards]
# TODO: Readers / Writers lock on gradients
# self.gradients = SharedGradients(num_threads)
# def initialize_gradients(self, i, gradients):
# self.gradients.initialize(i, gradients)
#
# def apply_gradients(self):
# self.optimizer.zero_grad()
# self.model.set_gradients(self.gradients.sum())
# self.optimizer.step()
# Need to fixwith ASGD
# How to process asynchronously? Also do mini-batches
def record_gradients(self, gradients):
#with self.time.get_lock():
# TODO just create this as a tensor instead, it works better
self.time.value += 1
for q, g in zip(self.qs, gradients):
# shard.update(g, self.time.value)
q.put(torch.from_numpy(g).share_memory_())
# def set_gradients(self, gradients):
# for g, p in zip(gradients, self.parameters):
# if g is not None:
# p.grad = torch.tensor(g)
# # p.grad = torch.from_numpy(g)
def get(self):
return self.shard_mem
def __init__(self, state_size, action_size, seed=0):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = QNetwork(state_size, action_size, seed=seed).to(device)
self.qnetwork_target = QNetwork(state_size, action_size, seed=seed).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, seed=0):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = QNetwork(state_size, action_size, seed=seed).to(device)
self.qnetwork_target = QNetwork(state_size, action_size, seed=seed).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
self.memory.add(state, action, reward, next_state, done)
# Learn every UPDATE_EVERY time steps.
self.t_step = (self.t_step + 1) % UPDATE_EVERY
if self.t_step == 0:
# If enough samples are available in memory, get random subset and learn
if len(self.memory) > BATCH_SIZE:
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, eps=0.):
"""Returns actions for given state as per current policy.
Params
======
state (array_like): current state
eps (float): epsilon, for epsilon-greedy action selection
"""
state = torch.from_numpy(state).float().unsqueeze(0).to(device)
self.qnetwork_local.eval()
with torch.no_grad():
action_values = self.qnetwork_local(state)
self.qnetwork_local.train()
# Epsilon-greedy action selection
if random.random() > eps:
return np.argmax(action_values.cpu().data.numpy())
else:
return random.choice(np.arange(self.action_size))
def learn(self, experiences, gamma):
"""Update value parameters using given batch of experience tuples.
Params
======
experiences (Tuple[torch.Variable]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# compute and minimize the loss
states, actions, rewards, next_states, dones = experiences
# Get max predicted Q values (for next states) from target model
Q_targets_next = self.qnetwork_target(next_states).detach().max(1)[0].unsqueeze(1)
# Compute Q targets for current states
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Get expected Q values from local model
Q_expected = self.qnetwork_local(states).gather(1, actions)
# Compute loss
loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# ------------------- update target network ------------------- #
self.soft_update(self.qnetwork_local, self.qnetwork_target, TAU)
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model (PyTorch model): weights will be copied from
target_model (PyTorch model): weights will be copied to
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(), local_model.parameters()):
target_param.data.copy_(tau*local_param.data + (1.0-tau)*target_param.data)
class Agent:
def __init__(self, state_size, action_size, num_agents, double_dqn=True):
self.action_size = action_size
self.double_dqn = double_dqn
# Q-Network
self.qnetwork_local = QNetwork(state_size, action_size).to(device)
self.qnetwork_target = copy.deepcopy(self.qnetwork_local)
self.optimizer = torch.optim.Adam(self.qnetwork_local.parameters(),
lr=LR)
# Replay memory
self.memory = ReplayBuffer(BUFFER_SIZE)
self.num_agents = num_agents
self.t_step = 0
def reset(self):
self.finished = [False] * self.num_agents
# Decide on an action to take in the environment
def act(self, state, eps=0.):
state = torch.from_numpy(state).float().unsqueeze(0).to(device)
self.qnetwork_local.eval()
with torch.no_grad():
action_values = self.qnetwork_local(state)
# Epsilon-greedy action selection
if random.random() > eps:
return torch.argmax(action_values).item()
else:
return torch.randint(self.action_size, ()).item()
# Record the results of the agent's action and update the model
def step(self, handle, state, action, next_state, agent_done, episode_done,
collision):
if not self.finished[handle]:
if agent_done:
reward = 1
elif collision:
reward = -5
else:
reward = -.1
# Save experience in replay memory
self.memory.push(state, action, reward, next_state, agent_done
or episode_done)
self.finished[handle] = agent_done or episode_done
# Perform a gradient update every UPDATE_EVERY time steps
self.t_step = (self.t_step + 1) % UPDATE_EVERY
if self.t_step == 0 and len(self.memory) > BATCH_SIZE * 20:
self.learn(*self.memory.sample(BATCH_SIZE, device))
def learn(self, states, actions, rewards, next_states, dones):
self.qnetwork_local.train()
# Get expected Q values from local model
Q_expected = self.qnetwork_local(states).gather(1, actions)
if self.double_dqn:
Q_best_action = self.qnetwork_local(next_states).argmax(1)
Q_targets_next = self.qnetwork_target(next_states).gather(
1, Q_best_action.unsqueeze(-1))
else:
Q_targets_next = self.qnetwork_target(next_states).detach().max(
1)[0].unsqueeze(-1)
# Compute Q targets for current states
Q_targets = rewards + GAMMA * Q_targets_next * (1 - dones)
# Compute loss and perform a gradient step
self.optimizer.zero_grad()
loss = F.mse_loss(Q_expected, Q_targets)
loss.backward()
self.optimizer.step()
# Update the target network parameters to `tau * local.parameters() + (1 - tau) * target.parameters()`
for target_param, local_param in zip(self.qnetwork_target.parameters(),
self.qnetwork_local.parameters()):
target_param.data.copy_(TAU * local_param.data +
(1.0 - TAU) * target_param.data)
# Checkpointing methods
def save(self, path, *data):
torch.save(self.qnetwork_local.state_dict(),
path / 'dqn/model_checkpoint.local')
torch.save(self.qnetwork_target.state_dict(),
path / 'dqn/model_checkpoint.target')
torch.save(self.optimizer.state_dict(),
path / 'dqn/model_checkpoint.optimizer')
with open(path / 'dqn/model_checkpoint.meta', 'wb') as file:
pickle.dump(data, file)
def load(self, path, *defaults):
try:
print("Loading model from checkpoint...")
self.qnetwork_local.load_state_dict(
torch.load(path / 'dqn/model_checkpoint.local'))
self.qnetwork_target.load_state_dict(
torch.load(path / 'dqn/model_checkpoint.target'))
self.optimizer.load_state_dict(
torch.load(path / 'dqn/model_checkpoint.optimizer'))
with open(path / 'dqn/model_checkpoint.meta', 'rb') as file:
return pickle.load(file)
except:
print("No checkpoint file was found")
return defaults
class Agent():
def __init__(self, state_size, action_size, seed):
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = QNetwork(state_size, action_size, seed).to(device) #prediction net
self.qnetwork_target = QNetwork(state_size, action_size, seed).to(device) #target network
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# store SARS when reach the batch size train
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
self.t_step = 0 # time step
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
self.memory.add(state, action, reward, next_state, done)
# Learn every UPDATE_EVERY time steps.
self.t_step = (self.t_step + 1) % UPDATE_EVERY
if self.t_step == 0:
# learn
if len(self.memory) > BATCH_SIZE:
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, eps=0.):
"""Returns actions for given state as per current policy.
Params
======
state (array_like): current state
eps (float): epsilon, for epsilon-greedy action selection
"""
state = torch.from_numpy(state).float().unsqueeze(0).to(device)
self.qnetwork_local.eval()
with torch.no_grad():
action_values = self.qnetwork_local(state)
self.qnetwork_local.train()
# Epsilon-greedy action selection
if random.random() > eps:
return np.argmax(action_values.cpu().data.numpy())
else:
return random.choice(np.arange(self.action_size))
def learn(self, experiences, gamma):
states, actions, rewards, next_states, dones = experiences
# predictied Q value
Q_targets_next = self.qnetwork_target(next_states).detach().max(1)[0].unsqueeze(1)
# Compute Q targets for current states
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Get expected Q values from local model
Q_expected = self.qnetwork_local(states).gather(1, actions)
# Compute loss
loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# ------------------- update target network ------------------- #
self.soft_update(self.qnetwork_local, self.qnetwork_target, TAU)
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model (PyTorch model): weights will be copied from
target_model (PyTorch model): weights will be copied to
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(), local_model.parameters()):
target_param.data.copy_(tau*local_param.data + (1.0-tau)*target_param.data)