class Agent(object):
def __init__(self):
self.network, self.target_network = AtariNet(ACTIONS_SIZE), AtariNet(
ACTIONS_SIZE)
self.memory = Memory(MEMORY_SIZE)
self.learning_count = 0
self.optimizer = torch.optim.Adam(self.network.parameters(), lr=LR)
self.loss_func = nn.MSELoss()
def action(self, state, israndom):
if israndom and random.random() < EPSILON:
return np.random.randint(0, ACTIONS_SIZE)
state = torch.unsqueeze(torch.FloatTensor(state), 0)
actions_value = self.network.forward(state)
return torch.max(actions_value, 1)[1].data.numpy()[0]
def learn(self, state, action, reward, next_state, done):
old_val = self.network.forward(torch.FloatTensor([state])).gather(
1, torch.LongTensor([[action]]))[0]
target_val = self.network.forward(torch.FloatTensor([state]))
if done:
done = 0
target = reward
else:
done = 1
target = reward + GAMMA * torch.max(target_val)
error = abs(old_val[0] - target)
self.memory.add(error.data, (state, action, reward, next_state, done))
if self.memory.tree.n_entries < MEMORY_THRESHOLD:
return
if self.learning_count % UPDATE_TIME == 0:
self.target_network.load_state_dict(self.network.state_dict())
self.learning_count += 1
batch, idxs, is_weights = self.memory.sample(BATCH_SIZE)
state = torch.FloatTensor([x[0] for x in batch])
action = torch.LongTensor([[x[1]] for x in batch])
reward = torch.FloatTensor([[x[2]] for x in batch])
next_state = torch.FloatTensor([x[3] for x in batch])
done = torch.FloatTensor([[x[4]] for x in batch])
eval_q = self.network.forward(state).gather(1, action)
next_q = self.target_network(next_state).detach()
target_q = reward + GAMMA * next_q.max(1)[0].view(BATCH_SIZE, 1) * done
errors = torch.abs(eval_q - target_q).data.numpy().flatten()
loss = self.loss_func(eval_q, target_q)
for i in range(BATCH_SIZE):
idx = idxs[i]
self.memory.update(idx, errors[i])
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
class PERAgent(OffPolicyAgent):
# construct agent's model separately, so it can be sized according to problem
def __init__(self, n_replay, env, target_policy, behavior_policy, lr, discount, type = 'BC'):
super().__init__(n_replay, env, target_policy, behavior_policy, lr, discount, type)
# reseed numpy, reset weights of network
# Reset must be performed before every episode
def reset(self,seed):
# Reset time
self.t=0
# Set seed value
np.random.seed(seed)
# Reset replay buffer
self.replay_buffer = Memory(self.n_replay)
# Rebuild model
self.build_model(self.n_features,self.env.nA)
def generate_action(self, s, target_policy_sel = True):
pval = self.target_policy[s] if target_policy_sel else self.behavior_policy[s]
return np.random.choice(a=self.actions, p=pval)
def generate_all_actions(self,target_policy_sel = True):
return np.array([self.generate_action(item, target_policy_sel) for item in range(self.target_policy.shape[0])])
# Generate steps of experience
def generate_experience(self, k=16):
# Initialize environment
s = self.env.reset()
done = False
steps = 0
# For each step
while steps < k:
# choose action according to behavior policy
a = self.generate_action(s,False)
# Take a step in environment based on chosen action
(s2,r,done,_) = self.env.step(a)
# Compute importance ratios
ratio = self.target_policy[s,a] / self.behavior_policy[s,a]
# states and target action for Computing TD Error
current_state = self.construct_features([s])
next_state = self.construct_features([s2])
target_policy_action = self.generate_action(s,True)
# Get bootstrap estimate of next state action values
value_s = self.model.predict([current_state,np.zeros(current_state.shape[0])])
value_next_s = self.model.predict([next_state,np.zeros(next_state.shape[0])])
updated_val = r if done else (r + self.discount*value_next_s[0][target_policy_action])
# Compute TD error
td_error = np.abs(updated_val - value_s[0][a])
# Stop execution if weights blow up - not converged
if td_error > 10**5:
return 1
# Add experience to IR replay buffer
self.replay_buffer.add_per(td_error, (s,a,r,s2))
# Set for next step
s=s2
self.t += 1
steps += 1
# If episode ends, reset environment
if done:
done = False
s = self.env.reset()
return 0
# do batch of training using replay buffer
def train_batch(self, n_samples, batch_size):
# Sample a minibatch from replay buffer
data_samples, idxs, ratios, buffer_total = self.replay_buffer.sample(n_samples)
# Extract rewards, states, next states, actions from samples
rewards = extract_transition_components(data_samples, TransitionComponent.reward)
next_states = extract_transition_components(data_samples, TransitionComponent.next_state)
next_state_features = self.construct_features(next_states)
states = extract_transition_components(data_samples, TransitionComponent.state)
state_features = self.construct_features(states)
actions = extract_transition_components(data_samples, TransitionComponent.action)
# Calculate Target policy actions
target_policy_actions = np.array([self.generate_action(state, True) for state in states])
# Calculate state values for TD error
next_values_sa = self.model.predict([next_state_features, np.zeros(next_state_features.shape[0])])
next_values = np.choose(target_policy_actions,next_values_sa.T)
# v(s') is zero for terminal state, so need to fix model prediction
for i in range(n_samples):
# if experience ends in terminal state, value function returns 0
if next_states[i] == -1 or next_states[i] == 10: #TODO this only works for randomwalk of size 10
next_values[i] = 0.0
# Compute targets by bootstrap estimates
targets = (rewards + self.discount*next_values)
# Compute error for updating priorities
pred_values = self.model.predict([state_features, np.zeros(state_features.shape[0])])
final_targets = np.copy(pred_values)
np.put_along_axis(final_targets, np.expand_dims(actions,axis = 1),targets[:,np.newaxis],axis = 1)
pred = np.choose(actions, pred_values.T)
error = np.abs(pred - targets)
# Priority update
for i in range(batch_size):
self.replay_buffer.update(idxs[i], error[i])
# train on samples
self.model.fit([state_features, ratios], final_targets, batch_size=batch_size, verbose=0)
class MAD4PG:
"""Interacts with and learns from the environment."""
def __init__(self,
state_size,
action_size,
seed,
buffer_size=int(1e6),
batch_size=64,
gamma=0.99,
tau=1e-3,
update_every=3,
num_mc_steps=5,
num_agents=2):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.BATCH_SIZE = batch_size
self.GAMMA = gamma
self.TAU = tau
self.UPDATE_EVERY = update_every
self.num_mc_steps = num_mc_steps
self.experiences = [
ExperienceQueue(num_mc_steps) for _ in range(num_agents)
]
self.memory = Memory(buffer_size)
self.t_step = 0
self.train_start = batch_size
self.mad4pg_agent = [
D4PG(state_size,
action_size,
seed,
device,
num_atoms=N_ATOMS,
q_min=Vmin,
q_max=Vmax),
D4PG(state_size,
action_size,
seed,
device,
num_atoms=N_ATOMS,
q_min=Vmin,
q_max=Vmax)
]
def acts(self, states, add_noise=0.0):
acts = []
for s, a in zip(states, self.mad4pg_agent):
acts.append(a.act(np.expand_dims(s, 0), add_noise))
return np.vstack(acts)
# borrow from https://github.com/PacktPublishing/Deep-Reinforcement-Learning-Hands-On/tree/master/Chapter14
def distr_projection(self, next_distr_v, rewards_v, dones_mask_t, gamma):
next_distr = next_distr_v.data.cpu().numpy()
rewards = rewards_v.data.cpu().numpy()
dones_mask = dones_mask_t.cpu().numpy().astype(np.bool)
batch_size = len(rewards)
proj_distr = np.zeros((batch_size, N_ATOMS), dtype=np.float32)
dones_mask = np.squeeze(dones_mask)
rewards = rewards.reshape(-1)
for atom in range(N_ATOMS):
tz_j = np.minimum(
Vmax,
np.maximum(Vmin, rewards + (Vmin + atom * DELTA_Z) * gamma))
b_j = (tz_j - Vmin) / DELTA_Z
l = np.floor(b_j).astype(np.int64)
u = np.ceil(b_j).astype(np.int64)
eq_mask = u == l
proj_distr[eq_mask, l[eq_mask]] += next_distr[eq_mask, atom]
ne_mask = u != l
proj_distr[ne_mask,
l[ne_mask]] += next_distr[ne_mask,
atom] * (u - b_j)[ne_mask]
proj_distr[ne_mask,
u[ne_mask]] += next_distr[ne_mask,
atom] * (b_j - l)[ne_mask]
if dones_mask.any():
proj_distr[dones_mask] = 0.0
tz_j = np.minimum(Vmax, np.maximum(Vmin, rewards[dones_mask]))
b_j = (tz_j - Vmin) / DELTA_Z
l = np.floor(b_j).astype(np.int64)
u = np.ceil(b_j).astype(np.int64)
eq_mask = u == l
if dones_mask.shape == ():
if dones_mask:
proj_distr[0, l] = 1.0
else:
ne_mask = u != l
proj_distr[0, l] = (u - b_j)[ne_mask]
proj_distr[0, u] = (b_j - l)[ne_mask]
else:
eq_dones = dones_mask.copy()
eq_dones[dones_mask] = eq_mask
if eq_dones.any():
proj_distr[eq_dones, l[eq_mask]] = 1.0
ne_mask = u != l
ne_dones = dones_mask.copy()
ne_dones[dones_mask] = ne_mask
if ne_dones.any():
proj_distr[ne_dones, l[ne_mask]] = (u - b_j)[ne_mask]
proj_distr[ne_dones, u[ne_mask]] = (b_j - l)[ne_mask]
return torch.FloatTensor(proj_distr).to(device)
def step(self, states, actions, rewards, next_states, dones):
for agent_index in range(len(self.mad4pg_agent)):
agent_experiences = self.experiences[agent_index]
agent_experiences.states.appendleft(states[agent_index])
agent_experiences.rewards.appendleft(rewards[agent_index] *
self.GAMMA**self.num_mc_steps)
agent_experiences.actions.appendleft(actions[agent_index])
if len(agent_experiences.rewards) == self.num_mc_steps or dones[
agent_index]: # N-steps return: r= r1+gamma*r2+..+gamma^(t-1)*rt
done_tensor = torch.tensor(
dones[agent_index]).float().to(device)
condition = True
while condition:
for i in range(len(agent_experiences.rewards)):
agent_experiences.rewards[i] /= self.GAMMA
state = torch.tensor(
agent_experiences.states[-1]).float().unsqueeze(0).to(
device)
next_state = torch.tensor(
next_states[agent_index]).float().unsqueeze(0).to(
device)
action = torch.tensor(
agent_experiences.actions[-1]).float().unsqueeze(0).to(
device)
sum_reward = torch.tensor(sum(
agent_experiences.rewards)).float().unsqueeze(0).to(
device)
with evaluating(
self.mad4pg_agent[agent_index]) as cur_agent:
q_logits_expected = cur_agent.critic_local(
state, action)
action_next = cur_agent.actor_target(next_state)
q_target_logits_next = cur_agent.critic_target(
next_state, action_next)
q_target_distr_next = F.softmax(q_target_logits_next,
dim=1)
q_target_distr_next_projected = self.distr_projection(
q_target_distr_next, sum_reward, done_tensor,
self.GAMMA**self.num_mc_steps)
cross_entropy = -F.log_softmax(
q_logits_expected,
dim=1) * q_target_distr_next_projected
error = cross_entropy.sum(dim=1).mean().cpu().data
self.memory.add(
error,
(states[agent_index], actions[agent_index], sum_reward,
next_states[agent_index], dones[agent_index]))
agent_experiences.states.pop()
agent_experiences.rewards.pop()
agent_experiences.actions.pop()
condition = False and dones[agent_index] and len(
agent_experiences.states) > 0
if dones[agent_index]:
agent_experiences.states.clear()
agent_experiences.rewards.clear()
agent_experiences.actions.clear()
self.t_step = (self.t_step + 1) % self.UPDATE_EVERY
if self.t_step == 0:
# If enough samples are available in memory, get random subset and learn
# print(self.memory.tree.n_entries)
if self.memory.tree.n_entries > self.train_start:
for agent_index in range(len(self.mad4pg_agent)):
sampled_experiences, idxs = self.sample()
self.learn(self.mad4pg_agent[agent_index],
sampled_experiences, idxs)
def sample(self):
# prioritized experience replay
mini_batch, idxs, is_weights = self.memory.sample(self.BATCH_SIZE)
mini_batch = np.array(mini_batch).transpose()
statess = np.vstack([m for m in mini_batch[0] if m is not None])
actionss = np.vstack([m for m in mini_batch[1] if m is not None])
rewardss = np.vstack([m for m in mini_batch[2] if m is not None])
next_statess = np.vstack([m for m in mini_batch[3] if m is not None])
doness = np.vstack([m for m in mini_batch[4] if m is not None])
# bool to binary
doness = doness.astype(int)
statess = torch.from_numpy(statess).float().to(device)
actionss = torch.from_numpy(actionss).float().to(device)
rewardss = torch.from_numpy(rewardss).float().to(device)
next_statess = torch.from_numpy(next_statess).float().to(device)
doness = torch.from_numpy(doness).float().to(device)
return (statess, actionss, rewardss, next_statess, doness), idxs
def learn(self, agent, experiences, idxs):
"""Update policy and value parameters using given batch of experience tuples.
Q_targets = r + γ * critic_target(next_state, actor_target(next_state))
where:
actor_target(state) -> action
critic_target(state, action) -> Q-value
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# ---------------------------- update critic ---------------------------- #
# Get predicted next-state actions and Q values from target models
# Compute critic loss
q_logits_expected = agent.critic_local(states, actions)
actions_next = agent.actor_target(next_states)
q_targets_logits_next = agent.critic_target(next_states, actions_next)
q_targets_distr_next = F.softmax(q_targets_logits_next, dim=1)
q_targets_distr_projected_next = self.distr_projection(
q_targets_distr_next, rewards, dones,
self.GAMMA**self.num_mc_steps)
cross_entropy = -F.log_softmax(q_logits_expected,
dim=1) * q_targets_distr_projected_next
critic_loss = cross_entropy.sum(dim=1).mean()
with torch.no_grad():
errors = cross_entropy.sum(dim=1).cpu().data.numpy()
# update priority
for i in range(self.BATCH_SIZE):
idx = idxs[i]
self.memory.update(idx, errors[i])
# Minimize the loss
agent.critic_optimizer.zero_grad()
critic_loss.backward()
agent.critic_optimizer.step()
# Compute actor loss
actions_pred = agent.actor_local(states)
crt_distr_v = agent.critic_local(states, actions_pred)
actor_loss = -agent.critic_local.distr_to_q(crt_distr_v)
actor_loss = actor_loss.mean()
# Minimize the loss
agent.actor_optimizer.zero_grad()
actor_loss.backward()
agent.actor_optimizer.step()
# ----------------------- update target networks ----------------------- #
agent.soft_update(agent.critic_local, agent.critic_target, self.TAU)
agent.soft_update(agent.actor_local, agent.actor_target, self.TAU)
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self,
state_size,
action_size,
random_seed,
buffer_size=BUFFER_SIZE,
batch_size=BATCH_SIZE):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
buffer_size (int): maximum size of buffer
batch_size (int): size of each training batch
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.buffer_size = buffer_size
self.memory = Memory(
capacity=self.buffer_size) # internal memory using SumTree
self.batch_size = batch_size
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size, action_size).to(device)
self.actor_target = Actor(state_size, action_size).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=LR_ACTOR)
# Critic Network (w/ Target Network)
self.critic_local = Critic(state_size, action_size).to(device)
self.critic_target = Critic(state_size, action_size).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
# Noise process
self.noise = OUNoise(action_size, random_seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
def step(self,
state,
action,
reward,
next_state,
done,
batch_size=BATCH_SIZE,
update_every=UPDATE_EVERY):
"""Save experience in replay memory, and use random sample from buffer to learn."""
# Save experience / reward
self.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 enough samples are available in memory
if self.memory.tree.n_entries >= batch_size:
experiences, idxs, is_weights = self.sample()
self.learn(experiences, idxs, is_weights)
def act(self, state, add_noise=True):
"""Returns actions for given state as per current policy."""
state = torch.from_numpy(state).float().to(device)
self.actor_local.eval()
with torch.no_grad():
action = self.actor_local(state).cpu().data.numpy()
self.actor_local.train()
if add_noise:
#action = [act + self.noise.sample() for act in action]
action += self.noise.sample()
return np.clip(action, -1, 1)
def reset(self):
self.noise.reset()
def learn(self,
experiences,
idxs,
is_weights,
batch_size=BATCH_SIZE,
gamma=GAMMA):
"""Update policy and value parameters using given batch of experience tuples.
Q_targets = r + γ * critic_target(next_state, actor_target(next_state))
where:
actor_target(state) -> action
critic_target(state, action) -> Q-value
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# ---------------------------- update critic ---------------------------- #
# Get predicted next-state actions and Q values from target models
actions_next = self.actor_target(next_states)
Q_targets_next = self.critic_target(next_states, actions_next)
# Compute Q targets for current states (y_i)
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Compute critic loss
Q_expected = self.critic_local(states, actions)
#Loss calculation
critic_loss = (torch.from_numpy(is_weights).float().to(device) *
F.mse_loss(Q_expected, Q_targets)).mean()
# Minimize the loss
self.critic_optimizer.zero_grad()
critic_loss.backward()
#Introducing gradient clipping
torch.nn.utils.clip_grad_norm_(self.critic_local.parameters(), 1)
self.critic_optimizer.step()
# ---------------------------- update actor ---------------------------- #
# Compute actor loss
actions_pred = self.actor_local(states)
actor_loss = -self.critic_local(states, actions_pred).mean()
# Minimize the loss
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# ----------------------- update target networks ----------------------- #
self.soft_update(self.critic_local, self.critic_target, TAU)
self.soft_update(self.actor_local, self.actor_target, TAU)
#.......................update priorities in prioritized replay buffer.......#
#Calculate errors used in prioritized replay buffer
errors = (Q_expected - Q_targets).squeeze().cpu().data.numpy()
# update priority
for i in range(batch_size):
idx = idxs[i]
self.memory.update(idx, errors[i])
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)
def add(self, state, action, reward, next_state, done, gamma=GAMMA):
"""Add a new experience to memory."""
next_state_torch = torch.from_numpy(next_state).float().to(device)
reward_torch = torch.unsqueeze(
torch.from_numpy(np.array(reward)).float().to(device), 1)
done_torch = torch.unsqueeze(
torch.from_numpy(np.array(done).astype(
np.uint8)).float().to(device), 1)
state_torch = torch.from_numpy(state).float().to(device)
action_torch = torch.from_numpy(action).float().to(device)
self.actor_target.eval()
self.critic_target.eval()
self.critic_local.eval()
with torch.no_grad():
action_next = self.actor_target(next_state_torch)
Q_target_next = self.critic_target(next_state_torch, action_next)
Q_target = reward_torch + (gamma * Q_target_next *
(1 - done_torch))
Q_expected = self.critic_local(state_torch, action_torch)
self.actor_local.train()
self.critic_target.train()
self.critic_local.train()
#Error used in prioritized replay buffer
error = (Q_expected - Q_target).squeeze().cpu().data.numpy()
#Adding experiences to prioritized replay buffer
for i in np.arange(len(reward)):
self.memory.add(
error[i],
(state[i], action[i], reward[i], next_state[i], done[i]))
def sample(self):
"""Randomly sample a batch of experiences from memory."""
experiences, idxs, is_weights = self.memory.sample(self.batch_size)
states = np.vstack([e[0] for e in experiences])
states = torch.from_numpy(states).float().to(device)
actions = np.vstack([e[1] for e in experiences])
actions = torch.from_numpy(actions).float().to(device)
rewards = np.vstack([e[2] for e in experiences])
rewards = torch.from_numpy(rewards).float().to(device)
next_states = np.vstack([e[3] for e in experiences])
next_states = torch.from_numpy(next_states).float().to(device)
dones = np.vstack([e[4] for e in experiences]).astype(np.uint8)
dones = torch.from_numpy(dones).float().to(device)
return (states, actions, rewards, next_states, dones), idxs, is_weights
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, seed):
"""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).to(device)
self.qnetwork_target = QNetwork(state_size, action_size, seed).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# Replay memory
self.memory = Memory(BUFFER_SIZE)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
def step(self, state, action, reward, next_state, done):
state = torch.from_numpy(state).float().unsqueeze(0).to(device)
next_state = torch.from_numpy(next_state).float().unsqueeze(0).to(device)
self.qnetwork_local.eval()
self.qnetwork_target.eval()
with torch.no_grad():
target_action_values = self.qnetwork_target(next_state)
expected_action_values = self.qnetwork_local(state)
self.qnetwork_local.train()
self.qnetwork_target.train()
old_val = expected_action_values[0][action]
new_val = reward
if not done:
new_val += GAMMA * torch.max(target_action_values)
error = abs(old_val - new_val)
# Save experience in replay memory
self.memory.add(error, (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 self.memory.tree.n_entries > BATCH_SIZE:
experiences = self.memory.sample(BATCH_SIZE)
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()).astype(int)
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
"""
mini_batches, idxs, is_weights = experiences
states = torch.from_numpy(np.vstack([mini_batch[0] for mini_batch in mini_batches])).float().to(device)
actions = torch.from_numpy(np.vstack([mini_batch[1] for mini_batch in mini_batches])).long().to(device)
rewards = torch.from_numpy(np.vstack([mini_batch[2] for mini_batch in mini_batches])).float().to(device)
next_states = torch.from_numpy(np.vstack([mini_batch[3] for mini_batch in mini_batches])).float().to(device)
dones = torch.from_numpy(np.vstack([int(mini_batch[4]) for mini_batch in mini_batches])).float().to(device)
## TODO: compute and minimize the loss
"*** YOUR CODE HERE ***"
Q_source_next = self.qnetwork_local(next_states).detach().max(1)[1].unsqueeze(1)
Q_target = self.qnetwork_target(next_states)
Q_double_target = torch.tensor([Q_target[i][max_index] for i, max_index in enumerate(Q_source_next)]).detach().unsqueeze(1)
Q_observed = rewards + (gamma * Q_double_target * (1 - dones))
Q_expected = self.qnetwork_local(states).gather(1, actions)
errors = torch.abs(Q_expected - Q_observed).data.numpy()
# update priority
for i in range(BATCH_SIZE):
idx = idxs[i]
self.memory.update(idx, errors[i])
loss = (torch.FloatTensor(is_weights) * F.mse_loss(Q_expected, Q_observed)).mean()
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 DQNAgent():
def __init__(self, state_size, action_size):
self.render = False
self.load_model = False
# get size of state and action
self.state_size = state_size
self.action_size = action_size
self.discount_factor = 0.99
self.learning_rate = 0.001
self.lr_step_size = 10
self.lr_gamma = 0.9
self.memory_size = 2**15
self.epsilon = 1.0
self.epsilon_min = 0.05
self.explore_step = 1000
self.epsilon_decay = 0.99995
self.batch_size = 64
self.train_start = 10000
# create prioritized replay memory using SumTree
self.memory = Memory(self.memory_size)
# create main model and target model
self.model = DQN(state_size, action_size)
self.model.apply(self.weights_init)
self.target_model = DQN(state_size, action_size)
self.optimizer = optim.Adam(self.model.parameters(),
lr=self.learning_rate)
self.scheduler = StepLR(self.optimizer,
step_size=self.lr_step_size,
gamma=self.lr_gamma)
# initialize target model
self.update_target_model()
if self.load_model:
self.model = torch.load('save_model/per_dqn')
self.model.train()
# weight xavier initialize
def weights_init(self, m):
classname = m.__class__.__name__
if classname.find('Linear') != -1:
torch.nn.init.xavier_uniform_(m.weight)
# after some time interval update the target model to be same with model
def update_target_model(self):
self.target_model.load_state_dict(self.model.state_dict())
# get action from model using epsilon-greedy policy
def get_action(self, state):
if np.random.rand() <= self.epsilon:
return random.randrange(self.action_size)
else:
state = torch.from_numpy(state).float()
q_value = self.model(state)
_, action = torch.max(q_value, 1)
return int(action)
# save sample (error,<s,a,r,s'>) to the replay memory
def append_sample(self, state, action, reward, next_state, done):
target = self.model(torch.tensor(state).float()).data
old_val = target[0][action]
target_val = self.target_model(torch.tensor(next_state).float()).data
if done:
target[0][action] = reward
else:
target[0][action] = reward + \
self.discount_factor * torch.max(target_val)
error = abs(old_val - target[0][action])
self.memory.add(error, (state, action, reward, next_state, done))
# pick samples from prioritized replay memory (with batch_size)
def train_model(self):
if self.epsilon > self.epsilon_min:
self.epsilon *= self.epsilon_decay
self.epsilon = max(self.epsilon, self.epsilon_min)
mini_batch, idxs, is_weights = self.memory.sample(self.batch_size)
mini_batch = np.array(mini_batch).transpose()
states = np.vstack(mini_batch[0])
actions = list(mini_batch[1])
rewards = list(mini_batch[2])
next_states = np.vstack(mini_batch[3])
dones = mini_batch[4]
# bool to binary
dones = dones.astype(int)
# Q function of current state
states = torch.tensor(states).float()
pred = self.model(states)
# one-hot encoding
a = torch.tensor(actions, dtype=torch.long).view(-1, 1)
one_hot_action = torch.zeros(self.batch_size, self.action_size)
one_hot_action.scatter_(1, a, 1)
pred = torch.sum(pred.mul(one_hot_action), dim=1)
# Q function of next state
next_states = torch.tensor(next_states, dtype=torch.float)
next_pred = self.target_model(next_states.float()).data
rewards = torch.tensor(rewards, dtype=torch.float)
dones = torch.tensor(dones, dtype=torch.float)
# Q Learning: get maximum Q value at s' from target model
target = rewards + (1 - dones) * \
self.discount_factor * next_pred.max(1)[0]
errors = torch.abs(pred - target).data.numpy()
# update priority
for i in range(self.batch_size):
idx = idxs[i]
self.memory.update(idx, errors[i])
self.optimizer.zero_grad()
# MSE Loss function
loss = (torch.tensor(is_weights).float() *
F.mse_loss(pred, target)).mean()
loss.backward()
# and train
self.optimizer.step()
return loss.item()
class DDQN_Agent:
def __init__(self, useDepth=False):
self.useDepth = useDepth
self.eps_start = 0.9
self.eps_end = 0.05
self.eps_decay = 30000
self.gamma = 0.8
self.learning_rate = 0.001
self.batch_size = 512
self.memory = Memory(10000)
self.max_episodes = 10000
self.save_interval = 2
self.test_interval = 10
self.network_update_interval = 10
self.episode = -1
self.steps_done = 0
self.max_steps = 34
self.policy = DQN()
self.target = DQN()
self.test_network = DQN()
self.target.eval()
self.test_network.eval()
self.updateNetworks()
self.env = DroneEnv(useDepth)
self.optimizer = optim.Adam(self.policy.parameters(), self.learning_rate)
if torch.cuda.is_available():
print('Using device:', device)
print(torch.cuda.get_device_name(0))
else:
print("Using CPU")
# LOGGING
cwd = os.getcwd()
self.save_dir = os.path.join(cwd, "saved models")
if not os.path.exists(self.save_dir):
os.mkdir("saved models")
if not os.path.exists(os.path.join(cwd, "videos")):
os.mkdir("videos")
if torch.cuda.is_available():
self.policy = self.policy.to(device) # to use GPU
self.target = self.target.to(device) # to use GPU
self.test_network = self.test_network.to(device) # to use GPU
# model backup
files = glob.glob(self.save_dir + '\\*.pt')
if len(files) > 0:
files.sort(key=os.path.getmtime)
file = files[-1]
checkpoint = torch.load(file)
self.policy.load_state_dict(checkpoint['state_dict'])
self.episode = checkpoint['episode']
self.steps_done = checkpoint['steps_done']
self.updateNetworks()
print("Saved parameters loaded"
"\nModel: ", file,
"\nSteps done: ", self.steps_done,
"\nEpisode: ", self.episode)
else:
if os.path.exists("log.txt"):
open('log.txt', 'w').close()
if os.path.exists("last_episode.txt"):
open('last_episode.txt', 'w').close()
if os.path.exists("last_episode.txt"):
open('saved_model_params.txt', 'w').close()
self.optimizer = optim.Adam(self.policy.parameters(), self.learning_rate)
obs, _ = self.env.reset()
tensor = self.transformToTensor(obs)
writer.add_graph(self.policy, tensor)
def updateNetworks(self):
self.target.load_state_dict(self.policy.state_dict())
def transformToTensor(self, img):
tensor = torch.FloatTensor(img).to(device)
tensor = tensor.unsqueeze(0)
tensor = tensor.unsqueeze(0)
tensor = tensor.float()
return tensor
def convert_size(self, size_bytes):
if size_bytes == 0:
return "0B"
size_name = ("B", "KB", "MB", "GB", "TB", "PB", "EB", "ZB", "YB")
i = int(math.floor(math.log(size_bytes, 1024)))
p = math.pow(1024, i)
s = round(size_bytes / p, 2)
return "%s %s" % (s, size_name[i])
def act(self, state):
self.eps_threshold = self.eps_end + (self.eps_start - self.eps_end) * math.exp(
-1.0 * self.steps_done / self.eps_decay
)
self.steps_done += 1
if random.random() > self.eps_threshold:
# print("greedy")
if torch.cuda.is_available():
action = np.argmax(self.policy(state).cpu().data.squeeze().numpy())
else:
action = np.argmax(self.policy(state).data.squeeze().numpy())
else:
action = random.randrange(0, 4)
return int(action)
def append_sample(self, state, action, reward, next_state):
next_state = self.transformToTensor(next_state)
current_q = self.policy(state).squeeze().cpu().detach().numpy()[action]
next_q = self.target(next_state).squeeze().cpu().detach().numpy()[action]
expected_q = reward + (self.gamma * next_q)
error = abs(current_q - expected_q),
self.memory.add(error, state, action, reward, next_state)
def learn(self):
if self.memory.tree.n_entries < self.batch_size:
return
states, actions, rewards, next_states, idxs, is_weights = self.memory.sample(self.batch_size)
states = tuple(states)
next_states = tuple(next_states)
states = torch.cat(states)
actions = np.asarray(actions)
rewards = np.asarray(rewards)
next_states = torch.cat(next_states)
current_q = self.policy(states)[[range(0, self.batch_size)], [actions]]
next_q =self.target(next_states).cpu().detach().numpy()[[range(0, self.batch_size)], [actions]]
expected_q = torch.FloatTensor(rewards + (self.gamma * next_q)).to(device)
errors = torch.abs(current_q.squeeze() - expected_q.squeeze()).cpu().detach().numpy()
# update priority
for i in range(self.batch_size):
idx = idxs[i]
self.memory.update(idx, errors[i])
loss = F.smooth_l1_loss(current_q.squeeze(), expected_q.squeeze())
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
def train(self):
print("Starting...")
score_history = []
reward_history = []
if self.episode == -1:
self.episode = 1
for e in range(1, self.max_episodes + 1):
start = time.time()
state, _ = self.env.reset()
steps = 0
score = 0
while True:
state = self.transformToTensor(state)
action = self.act(state)
next_state, reward, done, _ = self.env.step(action)
if steps == self.max_steps:
done = 1
#self.memorize(state, action, reward, next_state)
self.append_sample(state, action, reward, next_state)
self.learn()
state = next_state
steps += 1
score += reward
if done:
print("----------------------------------------------------------------------------------------")
if self.memory.tree.n_entries < self.batch_size:
print("Training will start after ", self.batch_size - self.memory.tree.n_entries, " steps.")
break
print(
"episode:{0}, reward: {1}, mean reward: {2}, score: {3}, epsilon: {4}, total steps: {5}".format(
self.episode, reward, round(score / steps, 2), score, self.eps_threshold, self.steps_done))
score_history.append(score)
reward_history.append(reward)
with open('log.txt', 'a') as file:
file.write(
"episode:{0}, reward: {1}, mean reward: {2}, score: {3}, epsilon: {4}, total steps: {5}\n".format(
self.episode, reward, round(score / steps, 2), score, self.eps_threshold,
self.steps_done))
if torch.cuda.is_available():
print('Total Memory:', self.convert_size(torch.cuda.get_device_properties(0).total_memory))
print('Allocated Memory:', self.convert_size(torch.cuda.memory_allocated(0)))
print('Cached Memory:', self.convert_size(torch.cuda.memory_reserved(0)))
print('Free Memory:', self.convert_size(torch.cuda.get_device_properties(0).total_memory - (
torch.cuda.max_memory_allocated() + torch.cuda.max_memory_reserved())))
# tensorboard --logdir=runs
memory_usage_allocated = np.float64(round(torch.cuda.memory_allocated(0) / 1024 ** 3, 1))
memory_usage_cached = np.float64(round(torch.cuda.memory_reserved(0) / 1024 ** 3, 1))
writer.add_scalar("memory_usage_allocated", memory_usage_allocated, self.episode)
writer.add_scalar("memory_usage_cached", memory_usage_cached, self.episode)
writer.add_scalar('epsilon_value', self.eps_threshold, self.episode)
writer.add_scalar('score_history', score, self.episode)
writer.add_scalar('reward_history', reward, self.episode)
writer.add_scalar('Total steps', self.steps_done, self.episode)
writer.add_scalars('General Look', {'score_history': score,
'reward_history': reward}, self.episode)
# save checkpoint
if self.episode % self.save_interval == 0:
checkpoint = {
'episode': self.episode,
'steps_done': self.steps_done,
'state_dict': self.policy.state_dict()
}
torch.save(checkpoint, self.save_dir + '//EPISODE{}.pt'.format(self.episode))
if self.episode % self.network_update_interval == 0:
self.updateNetworks()
self.episode += 1
end = time.time()
stopWatch = end - start
print("Episode is done, episode time: ", stopWatch)
if self.episode % self.test_interval == 0:
self.test()
break
writer.close()
def test(self):
self.test_network.load_state_dict(self.target.state_dict())
start = time.time()
steps = 0
score = 0
image_array = []
state, next_state_image = self.env.reset()
image_array.append(next_state_image)
while True:
state = self.transformToTensor(state)
action = int(np.argmax(self.test_network(state).cpu().data.squeeze().numpy()))
next_state, reward, done, next_state_image = self.env.step(action)
image_array.append(next_state_image)
if steps == self.max_steps:
done = 1
state = next_state
steps += 1
score += reward
if done:
print("----------------------------------------------------------------------------------------")
print("TEST, reward: {}, score: {}, total steps: {}".format(
reward, score, self.steps_done))
with open('tests.txt', 'a') as file:
file.write("TEST, reward: {}, score: {}, total steps: {}\n".format(
reward, score, self.steps_done))
writer.add_scalars('Test', {'score': score, 'reward': reward}, self.episode)
end = time.time()
stopWatch = end - start
print("Test is done, test time: ", stopWatch)
# Convert images to video
frameSize = (256, 144)
import cv2
video = cv2.VideoWriter("videos\\test_video_episode_{}_score_{}.avi".format(self.episode, score), cv2.VideoWriter_fourcc(*'DIVX'), 7, frameSize)
for img in image_array:
video.write(img)
video.release()
break
class Agent: # todo change name to Agent
def __init__(self, eps, lr, gamma, batch_size, tau, max_memory, lambda_1,
lambda_2, lambda_3, n_steps, l_margin):
# Input Parameters
self.eps = eps # eps-greedy
self.gamma = gamma # discount factor
self.batch_size = batch_size
self.tau = tau # frequency of target replacement
self.ed = 0.005 # bonus for demonstration # todo they aren't used
self.ea = 0.001 # todo they aren't used
self.l_margin = l_margin
self.n_steps = n_steps
self.lambda1 = lambda_1 # n-step return
self.lambda2 = lambda_2 # supervised loss
self.lambda3 = lambda_3 # L2
self.counter = 0 # target replacement counter # todo change to iter_counter
self.replay = Memory(capacity=max_memory)
self.loss = nn.MSELoss()
self.policy = Policy() # todo change not have to pass architecture
self.opt = optim.Adam(self.policy.predictNet.parameters(),
lr=lr,
weight_decay=lambda_3)
self.replay.e = 0
self.demoReplay = ddict(list)
self.noisy = hasattr(self.policy.predictNet, "sample")
def choose_action(self, state):
state = torch.Tensor(state)
A = self.policy.sortedA(state)
if self.noisy:
self.policy.predictNet.sample()
return A[0]
if np.random.random() < self.eps:
return random.sample(A, 1)[0]
return A[0]
def sample(self):
return self.replay.sample(self.batch_size)
def store_demonstration(self, s, a, r, s_, done, episode):
s = torch.Tensor(s)
s_ = torch.Tensor(s_)
episodeReplay = self.demoReplay[
episode] # replay of certain demo episode
index = len(episodeReplay)
data = (s, a, r, s_, done, (episode, index))
episodeReplay.append(data)
self.replay.add(transition=data, demonstration=True)
def store_transition(self, s, a, r, s_, done):
s = torch.Tensor(s)
s_ = torch.Tensor(s_)
data = (s, a, r, s_, done, None)
self.replay.add(transition=data, demonstration=False)
def calculate_td_errors(self, samples):
if self.noisy:
self.policy.predictNet.sample() # for choosing action
alls, alla, allr, alls_, alldone, *_ = zip(*samples)
maxA = [self.policy.sortedA(s_)[0] for s_ in alls_]
if self.noisy:
self.policy.predictNet.sample() # for prediction
self.policy.targetNet.sample() # for target
Qtarget = torch.Tensor(allr)
Qtarget[torch.tensor(alldone) != 1] += self.gamma * self.policy.calcQ(
self.policy.targetNet, alls_, maxA)[torch.tensor(alldone) != 1]
Qpredict = self.policy.calcQ(self.policy.predictNet, alls, alla)
return Qpredict, Qtarget
def JE(self, samples):
loss = torch.tensor(0.0)
count = 0 # number of demo
for s, aE, *_, isdemo in samples:
if isdemo is None:
continue
A = self.policy.sortedA(s)
if len(A) == 1:
continue
QE = self.policy.calcQ(self.policy.predictNet, s, aE)
A1, A2 = np.array(A)[:
2] # action with largest and second largest Q
maxA = A2 if (A1 == aE).all() else A1
Q = self.policy.calcQ(self.policy.predictNet, s, maxA)
if (Q + self.l_margin) < QE:
continue
else:
loss += (Q - QE)
count += 1
return loss / count if count != 0 else loss
def Jn(self, samples, Qpredict):
# wait for refactoring, can't use with noisy layer
loss = torch.tensor(0.0)
count = 0
for i, (s, a, r, s_, done, isdemo) in enumerate(samples):
if isdemo is None:
continue
episode, idx = isdemo
nidx = idx + self.n_steps
lepoch = len(self.demoReplay[episode])
if nidx > lepoch:
continue
count += 1
ns, na, nr, ns_, ndone, _ = zip(
*self.demoReplay[episode][idx:nidx])
ns, na, ns_, ndone = ns[-1], na[-1], ns_[-1], ndone[-1]
discountedR = reduce(
lambda x, y: (x[0] + self.gamma**x[1] * y, x[1] + 1), nr,
(0, 0))[0]
maxA = self.policy.sortedA(ns_)[0]
target = discountedR if ndone else discountedR + self.gamma**self.n_steps * self.policy.calcQ(
self.policy.targetNet, ns_, maxA)
predict = Qpredict[i]
loss += (target - predict)**2
return loss / count
def L2(self, parameters):
loss = 0
for p in parameters:
loss += (p**2).sum()
return loss
def learn(self):
self.opt.zero_grad()
samples, idxs, = self.sample()
Qpredict, Qtarget = self.calculate_td_errors(samples)
for i in range(self.batch_size):
error = math.fabs(float(Qpredict[i] - Qtarget[i]))
self.replay.update(idxs[i], error)
JDQ = self.loss(Qpredict, Qtarget)
JE = self.JE(samples)
Jn = self.Jn(samples, Qpredict)
L2 = self.L2(self.policy.predictNet.parameters())
J = JDQ + self.lambda2 * JE + self.lambda1 * Jn + self.lambda3 * L2
J.backward()
self.opt.step()
self.counter += 1
if self.counter % self.tau == 0:
self.policy.updateTargetNet()
action_str, rotate_idx = utils.get_action_info(
pixel_index)
old_value = trainer.forward(
color,
depth,
action_str,
False,
rotate_idx,
clear_grad=True)[0, pixel_index[1],
pixel_index[2]]
print "New Q value: {:03f} -> {:03f} | TD Target: {:03f}".format(
old_q[i], old_value, td_target_list[i])
print "========================================================================================"
#update_tree[i/5].update(idxs[i], td_target-old_value)
if i / mini_batch_size == 0 or specific_tool == 0:
suction_1_memory_buffer.update(
idxs[i], td_target - old_value)
elif i / mini_batch_size == 1 or specific_tool == 1:
suction_2_memory_buffer.update(
idxs[i], td_target - old_value)
else:
gripper_memory_buffer.update(
idxs[i], td_target - old_value)
back_t = time.time() - back_ts
if arduino: arduino.write("b 1000")
print "Backpropagation& Updating: {} seconds \t|\t Avg. {} seconds".format(
back_t, back_t / (3 * mini_batch_size))
if learned_times % updating_freq == 0:
print "[%f] Replace target network to behavior network" % (
program_time + time.time() - program_ts)
trainer.target_net.load_state_dict(
trainer.behavior_net.state_dict())