def __init__(self, env, gamma, tau, v_lr, q_lr, policy_lr, buffer_maxlen):
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.env = env
self.action_range = [env.action_space.low, env.action_space.high]
self.obs_dim = env.observation_space.shape[0]
self.action_dim = env.action_space.shape[0]
# hyperparameters
self.gamma = gamma
self.tau = tau
# initialize networks
self.value_net = ValueNetwork(self.obs_dim, 1).to(self.device)
self.target_value_net = ValueNetwork(self.obs_dim, 1).to(self.device)
self.q_net1 = SoftQNetwork(self.obs_dim, self.action_dim).to(self.device)
self.q_net2 = SoftQNetwork(self.obs_dim, self.action_dim).to(self.device)
self.policy_net = GaussianPolicy(self.obs_dim, self.action_dim).to(self.device)
# copy params to target param
for target_param, param in zip(self.target_value_net.parameters(), self.value_net.parameters()):
target_param.data.copy_(param)
# initialize optimizers
self.value_optimizer = optim.Adam(self.value_net.parameters(), lr=v_lr)
self.q1_optimizer = optim.Adam(self.q_net1.parameters(), lr=q_lr)
self.q2_optimizer = optim.Adam(self.q_net2.parameters(), lr=q_lr)
self.policy_optimizer = optim.Adam(self.policy_net.parameters(), lr=policy_lr)
self.replay_buffer = Buffer(buffer_maxlen)
def __init__(self, env, gamma, tau, alpha, q_lr, policy_lr, a_lr,
buffer_maxlen):
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.env = env
self.action_range = [env.action_space.low, env.action_space.high]
self.obs_dim = env.observation_space.shape[0]
self.action_dim = env.action_space.shape[0]
# hyperparameters
self.gamma = gamma
self.tau = tau
self.update_step = 0
self.delay_step = 2
# initialize networks
self.q_net1 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.q_net2 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.target_q_net1 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.target_q_net2 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.policy_net = PolicyNetwork(self.obs_dim,
self.action_dim).to(self.device)
# copy params to target param
for target_param, param in zip(self.target_q_net1.parameters(),
self.q_net1.parameters()):
target_param.data.copy_(param)
for target_param, param in zip(self.target_q_net2.parameters(),
self.q_net2.parameters()):
target_param.data.copy_(param)
# initialize optimizers
self.q1_optimizer = optim.Adam(self.q_net1.parameters(), lr=q_lr)
self.q2_optimizer = optim.Adam(self.q_net2.parameters(), lr=q_lr)
self.policy_optimizer = optim.Adam(self.policy_net.parameters(),
lr=policy_lr)
# entropy temperature
self.alpha = alpha
self.target_entropy = -torch.prod(
torch.Tensor(self.env.action_space.shape).to(self.device)).item()
self.log_alpha = torch.zeros(1, requires_grad=True, device=self.device)
self.alpha_optim = optim.Adam([self.log_alpha], lr=a_lr)
self.replay_buffer = BasicBuffer(buffer_maxlen)
class SACAgent:
def __init__(self, env, gamma, tau, v_lr, q_lr, policy_lr, buffer_maxlen):
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.env = env
self.action_range = [env.action_space.low, env.action_space.high]
self.obs_dim = env.observation_space.shape[0]
self.action_dim = env.action_space.shape[0]
# hyperparameters
self.gamma = gamma
self.tau = tau
self.update_step = 0
self.delay_step = 2
# initialize networks
self.value_net = ValueNetwork(self.obs_dim, 1).to(self.device)
self.target_value_net = ValueNetwork(self.obs_dim, 1).to(self.device)
self.q_net1 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.q_net2 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.policy_net = PolicyNetwork(self.obs_dim,
self.action_dim).to(self.device)
# copy params to target param
for target_param, param in zip(self.target_value_net.parameters(),
self.value_net.parameters()):
target_param.data.copy_(param)
# initialize optimizers
self.value_optimizer = optim.Adam(self.value_net.parameters(), lr=v_lr)
self.q1_optimizer = optim.Adam(self.q_net1.parameters(), lr=q_lr)
self.q2_optimizer = optim.Adam(self.q_net2.parameters(), lr=q_lr)
self.policy_optimizer = optim.Adam(self.policy_net.parameters(),
lr=policy_lr)
self.replay_buffer = BasicBuffer(buffer_maxlen)
def get_action(self, state):
state = torch.FloatTensor(state).unsqueeze(0).to(self.device)
mean, log_std = self.policy_net.forward(state)
std = log_std.exp()
normal = Normal(mean, std)
z = normal.sample()
action = torch.tanh(z)
action = action.cpu().detach().squeeze(0).numpy()
return self.rescale_action(action)
def rescale_action(self, action):
return action * (self.action_range[1] - self.action_range[0]) / 2.0 +\
(self.action_range[1] + self.action_range[0]) / 2.0
def update(self, batch_size):
states, actions, rewards, next_states, dones = self.replay_buffer.sample(
batch_size)
states = torch.FloatTensor(states).to(self.device)
actions = torch.FloatTensor(actions).to(self.device)
rewards = torch.FloatTensor(rewards).to(self.device)
next_states = torch.FloatTensor(next_states).to(self.device)
dones = torch.FloatTensor(dones).to(self.device)
dones = dones.view(dones.size(0), -1)
next_actions, next_log_pi = self.policy_net.sample(next_states)
next_q1 = self.q_net1(next_states, next_actions)
next_q2 = self.q_net2(next_states, next_actions)
next_v = self.target_value_net(next_states)
# value Loss
next_v_target = torch.min(next_q1, next_q2) - next_log_pi
curr_v = self.value_net.forward(states)
v_loss = F.mse_loss(curr_v, next_v_target.detach())
# q loss
curr_q1 = self.q_net1.forward(states, actions)
curr_q2 = self.q_net2.forward(states, actions)
expected_q = rewards + (1 - dones) * self.gamma * next_v
q1_loss = F.mse_loss(curr_q1, expected_q.detach())
q2_loss = F.mse_loss(curr_q2, expected_q.detach())
# update value network and q networks
self.value_optimizer.zero_grad()
v_loss.backward()
self.value_optimizer.step()
self.q1_optimizer.zero_grad()
q1_loss.backward()
self.q1_optimizer.step()
self.q2_optimizer.zero_grad()
q2_loss.backward()
self.q2_optimizer.step()
#delayed update for policy net and target value nets
if self.update_step % self.delay_step == 0:
new_actions, log_pi = self.policy_net.sample(states)
min_q = torch.min(self.q_net1.forward(states, new_actions),
self.q_net2.forward(states, new_actions))
policy_loss = (log_pi - min_q).mean()
self.policy_optimizer.zero_grad()
policy_loss.backward()
self.policy_optimizer.step()
# target networks
for target_param, param in zip(self.target_value_net.parameters(),
self.value_net.parameters()):
target_param.data.copy_(self.tau * param +
(1 - self.tau) * target_param)
self.update_step += 1
def train_sac(args):
'''
Parse arguments and construct objects for training sac model
'''
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
token_tables = op.build_token_tables()
# initialize tensorboard for logging output
from os import path
train_logger = None
if args.log_dir is not None:
train_logger = tb.SummaryWriter(path.join(args.log_dir, 'train'),
flush_secs=1)
# Load Models
policy = RobustFill(string_size=len(op.CHARACTER),
string_embedding_size=args.embedding_size,
decoder_inp_size=128,
hidden_size=args.hidden_size,
program_size=len(token_tables.op_token_table),
device=device)
q_1 = SoftQNetwork(128, len(token_tables.op_token_table), args.hidden_size)
q_2 = SoftQNetwork(128, len(token_tables.op_token_table), args.hidden_size)
tgt_q_1 = SoftQNetwork(128, len(token_tables.op_token_table),
args.hidden_size).eval()
tgt_q_2 = SoftQNetwork(128, len(token_tables.op_token_table),
args.hidden_size).eval()
if args.continue_training_policy:
policy.load_state_dict(
torch.load(path.join(path.dirname(path.abspath(__file__)),
args.checkpoint_filename),
map_location=device))
elif args.continue_training:
policy.load_state_dict(
torch.load(path.join(path.dirname(path.abspath(__file__)),
args.checkpoint_filename),
map_location=device))
q_1.load_state_dict(
torch.load(path.join(path.dirname(path.abspath(__file__)),
args.q1_checkpoint_filename),
map_location=device))
q_2.load_state_dict(
torch.load(path.join(path.dirname(path.abspath(__file__)),
args.q2_checkpoint_filename),
map_location=device))
for target_param, param in zip(tgt_q_1.parameters(), q_1.parameters()):
target_param.data.copy_(param.data)
for target_param, param in zip(tgt_q_2.parameters(), q_2.parameters()):
target_param.data.copy_(param.data)
for param in tgt_q_1.parameters():
param.requires_grad = False
for param in tgt_q_2.parameters():
param.requires_grad = False
policy = policy.to(device)
policy.set_device(device)
q_1 = q_1.to(device)
q_2 = q_2.to(device)
tgt_q_1 = tgt_q_1.to(device)
tgt_q_2 = tgt_q_2.to(device)
# Initialize optimizers
if (args.optimizer == 'sgd'):
policy_opt = optim.SGD(policy.parameters(), lr=args.lr)
q_1_opt = optim.SGD(q_1.parameters(), lr=args.lr)
q_2_opt = optim.SGD(q_2.parameters(), lr=args.lr)
entropy_opt = optim.SGD([policy.log_alpha], lr=args.lr)
else:
policy_opt = optim.Adam(policy.parameters(), lr=args.lr)
q_1_opt = optim.Adam(q_1.parameters(), lr=args.lr)
q_2_opt = optim.Adam(q_2.parameters(), lr=args.lr)
entropy_opt = optim.Adam([policy.log_alpha], lr=args.lr)
# Other necessary objects
env = RobustFillEnv()
replay_buffer_size = 1_000_000
replay_buffer = Replay_Buffer(replay_buffer_size, args.batch_size)
her = HER()
train_sac_(args, policy, q_1, q_2, tgt_q_1, tgt_q_2, policy_opt, q_1_opt,
q_2_opt, entropy_opt, replay_buffer, her, env, train_logger,
args.checkpoint_filename, args.checkpoint_step_size,
args.print_tensors)
def __init__(self, env: object, gamma: float, tau: float,
buffer_maxlen: int, critic_lr: float, actor_lr: float,
reward_scale: int):
# Selecting the device to use, wheter CUDA (GPU) if available or CPU
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
# Creating the Gym environments for training and evaluation
self.env = env
# Get max and min values of the action of this environment
self.action_range = [
self.env.action_space.low, self.env.action_space.high
]
# Get dimension of of the state and the action
self.obs_dim = self.env.observation_space.shape[0]
self.action_dim = self.env.action_space.shape[0]
# hyperparameters
self.gamma = gamma
self.tau = tau
self.critic_lr = critic_lr
self.actor_lr = actor_lr
self.buffer_maxlen = buffer_maxlen
self.reward_scale = reward_scale
# Scaling and bias factor for the actions -> We need scaling of the actions because each environment has different min and max values of actions
self.scale = (self.action_range[1] - self.action_range[0]) / 2.0
self.bias = (self.action_range[1] + self.action_range[0]) / 2.0
# initialize networks
self.q_net1 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.target_q_net1 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.q_net2 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.target_q_net2 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.policy = PolicyNetwork(self.obs_dim,
self.action_dim).to(self.device)
# copy weight parameters to the target Q networks
for target_param, param in zip(self.target_q_net1.parameters(),
self.q_net1.parameters()):
target_param.data.copy_(param)
for target_param, param in zip(self.target_q_net2.parameters(),
self.q_net2.parameters()):
target_param.data.copy_(param)
# initialize optimizers
self.q1_optimizer = optim.Adam(self.q_net1.parameters(),
lr=self.critic_lr)
self.q2_optimizer = optim.Adam(self.q_net2.parameters(),
lr=self.critic_lr)
self.policy_optimizer = optim.Adam(self.policy.parameters(),
lr=self.actor_lr)
# Create a replay buffer
self.replay_buffer = BasicBuffer(self.buffer_maxlen)
class SACAgent():
def __init__(self, env: object, gamma: float, tau: float,
buffer_maxlen: int, critic_lr: float, actor_lr: float,
reward_scale: int):
# Selecting the device to use, wheter CUDA (GPU) if available or CPU
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
# Creating the Gym environments for training and evaluation
self.env = env
# Get max and min values of the action of this environment
self.action_range = [
self.env.action_space.low, self.env.action_space.high
]
# Get dimension of of the state and the action
self.obs_dim = self.env.observation_space.shape[0]
self.action_dim = self.env.action_space.shape[0]
# hyperparameters
self.gamma = gamma
self.tau = tau
self.critic_lr = critic_lr
self.actor_lr = actor_lr
self.buffer_maxlen = buffer_maxlen
self.reward_scale = reward_scale
# Scaling and bias factor for the actions -> We need scaling of the actions because each environment has different min and max values of actions
self.scale = (self.action_range[1] - self.action_range[0]) / 2.0
self.bias = (self.action_range[1] + self.action_range[0]) / 2.0
# initialize networks
self.q_net1 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.target_q_net1 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.q_net2 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.target_q_net2 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.policy = PolicyNetwork(self.obs_dim,
self.action_dim).to(self.device)
# copy weight parameters to the target Q networks
for target_param, param in zip(self.target_q_net1.parameters(),
self.q_net1.parameters()):
target_param.data.copy_(param)
for target_param, param in zip(self.target_q_net2.parameters(),
self.q_net2.parameters()):
target_param.data.copy_(param)
# initialize optimizers
self.q1_optimizer = optim.Adam(self.q_net1.parameters(),
lr=self.critic_lr)
self.q2_optimizer = optim.Adam(self.q_net2.parameters(),
lr=self.critic_lr)
self.policy_optimizer = optim.Adam(self.policy.parameters(),
lr=self.actor_lr)
# Create a replay buffer
self.replay_buffer = BasicBuffer(self.buffer_maxlen)
def update(self, batch_size: int):
# Sampling experiences from the replay buffer
states, actions, rewards, next_states, dones = self.replay_buffer.sample(
batch_size)
# Convert numpy arrays of experience tuples into pytorch tensors
states = torch.FloatTensor(states).to(self.device)
actions = torch.FloatTensor(actions).to(self.device)
rewards = self.reward_scale * torch.FloatTensor(rewards).to(
self.device) # in SAC we do reward scaling for the sampled rewards
next_states = torch.FloatTensor(next_states).to(self.device)
dones = torch.FloatTensor(dones).to(self.device)
dones = dones.view(dones.size(0), -1)
# Critic update (computing the loss)
# Please refer to equation (6) in the paper for details
# Sample actions for the next states (s_t+1) using the current policy
next_actions, next_log_pi, _, _ = self.policy.sample(
next_states, self.scale)
next_actions = self.rescale_action(next_actions)
# Compute Q(s_t+1,a_t+1) by giving the states and actions to the Q network and choose the minimum from 2 target Q networks
next_q1 = self.target_q_net1(next_states, next_actions)
next_q2 = self.target_q_net2(next_states, next_actions)
min_q = torch.min(next_q1,
next_q2) # find minimum between next_q1 and next_q2
# Compute the next Q_target (Q(s_t,a_t)-alpha(next_log_pi))
next_q_target = (min_q - next_log_pi)
# Compute the Q(s_t,a_t) using s_t and a_t from the replay buffer
curr_q1 = self.q_net1.forward(states, actions)
curr_q2 = self.q_net2.forward(states, actions)
# Find expected Q, i.e., r(t) + gamma*next_q_target
expected_q = rewards + (1 - dones) * self.gamma * next_q_target
# Compute loss between Q network and expected Q
q1_loss = F.mse_loss(curr_q1, expected_q.detach())
q2_loss = F.mse_loss(curr_q2, expected_q.detach())
# Backpropagate the losses and update Q network parameters
self.q1_optimizer.zero_grad()
q1_loss.backward()
self.q1_optimizer.step()
self.q2_optimizer.zero_grad()
q2_loss.backward()
self.q2_optimizer.step()
# Policy update (computing the loss)
# Sample new actions for the current states (s_t) using the current policy
new_actions, log_pi, _, _ = self.policy.sample(states, self.scale)
new_actions = self.rescale_action(new_actions)
# Compute Q(s_t,a_t) and choose the minimum from 2 Q networks
new_q1 = self.q_net1.forward(states, new_actions)
new_q2 = self.q_net2.forward(states, new_actions)
min_q = torch.min(new_q1, new_q2)
# Compute the next policy loss, i.e., alpha*log_pi - Q(s_t,a_t) eq. (7)
policy_loss = (log_pi - min_q).mean()
# Backpropagate the losses and update policy network parameters
self.policy_optimizer.zero_grad()
policy_loss.backward()
self.policy_optimizer.step()
# Updating target networks with soft update using update rate tau
for target_param, param in zip(self.target_q_net1.parameters(),
self.q_net1.parameters()):
target_param.data.copy_(self.tau * param +
(1 - self.tau) * target_param)
for target_param, param in zip(self.target_q_net2.parameters(),
self.q_net2.parameters()):
target_param.data.copy_(self.tau * param +
(1 - self.tau) * target_param)
def get_action(
self, state: np.ndarray,
stochastic: bool) -> Tuple[np.ndarray, torch.Tensor, torch.Tensor]:
# state: the state input to the pi network
# stochastic: boolean (True -> use noisy action, False -> use noiseless (deterministic action))
state = torch.FloatTensor(state).unsqueeze(0).to(self.device)
# Get mean and sigma from the policy network
mean, log_std = self.policy.forward(state)
std = log_std.exp()
# Stochastic mode is used for training, non-stochastic mode is used for evaluation
if stochastic:
normal = Normal(mean, std)
z = normal.sample()
action = torch.tanh(z)
action = action.cpu().detach().squeeze(0).numpy()
else:
normal = Normal(mean, 0)
z = normal.sample()
action = torch.tanh(z)
action = action.cpu().detach().squeeze(0).numpy()
# return a rescaled action, and also the mean and standar deviation of the action
# we use a rescaled action since the output of the policy network is [-1,1] and the mujoco environments could be ranging from [-n,n] where n is an arbitrary real value
return self.rescale_action(action), mean, std
def rescale_action(self, action: np.ndarray) -> np.ndarray:
# we use a rescaled action since the output of the policy network is [-1,1] and the mujoco environments could be ranging from [-n,n] where n is an arbitrary real value
# scale -> scalar multiplication
# bias -> scalar offset
return action * self.scale[0] + self.bias[0]
def Actor_save(self, WORKSPACE: str):
# save 각 node별 모델 저장
print("Save the torch model")
savePath = WORKSPACE + "./policy_model5_Hop_.pth"
torch.save(self.policy.state_dict(), savePath)
def Actor_load(self, WORKSPACE: str):
# save 각 node별 모델 로드
print("load the torch model")
savePath = WORKSPACE + "./policy_model5_Hop_.pth" # Best
self.policy = PolicyNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.policy.load_state_dict(torch.load(savePath))
class SACAgent:
def __init__(self, env, gamma, tau, alpha, q_lr, policy_lr, a_lr,
buffer_maxlen):
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.env = env
self.action_range = [0, 250]
self.obs_dim = env.state_dim
self.action_dim = env.action_dim
# hyperparameters
self.gamma = gamma
self.tau = tau
self.update_step = 0
self.delay_step = 2
# initialize networks
self.q_net1 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.q_net2 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.target_q_net1 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.target_q_net2 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.policy_net = PolicyNetwork(self.obs_dim,
self.action_dim).to(self.device)
# copy params to target param
for target_param, param in zip(self.target_q_net1.parameters(),
self.q_net1.parameters()):
target_param.data.copy_(param)
for target_param, param in zip(self.target_q_net2.parameters(),
self.q_net2.parameters()):
target_param.data.copy_(param)
# initialize optimizers
self.q1_optimizer = optim.Adam(self.q_net1.parameters(), lr=q_lr)
self.q2_optimizer = optim.Adam(self.q_net2.parameters(), lr=q_lr)
self.policy_optimizer = optim.Adam(self.policy_net.parameters(),
lr=policy_lr)
# entropy temperature
self.alpha = alpha
self.target_entropy = -torch.prod(
torch.Tensor([self.action_dim, 1]).to(self.device)).item()
self.log_alpha = torch.zeros(1, requires_grad=True, device=self.device)
self.alpha_optim = optim.Adam([self.log_alpha], lr=a_lr)
self.replay_buffer = BasicBuffer(buffer_maxlen)
def get_action(self, state):
state = torch.FloatTensor(state).unsqueeze(0).to(self.device)
mean, log_std = self.policy_net.forward(state)
std = log_std.exp()
normal = Normal(mean, std)
z = normal.sample()
action = torch.tanh(z)
action = action.cpu().detach().squeeze(0).numpy()
return self.rescale_action(action)
def rescale_action(self, action):
return action * (self.action_range[1] - self.action_range[0]) / 2.0 +\
(self.action_range[1] + self.action_range[0]) / 2.0
def update(self, batch_size):
states, actions, rewards, next_states, dones = self.replay_buffer.sample(
batch_size)
states = torch.FloatTensor(states).to(self.device)
actions = torch.FloatTensor(actions).to(self.device)
rewards = torch.FloatTensor(rewards).to(self.device)
next_states = torch.FloatTensor(next_states).to(self.device)
dones = torch.FloatTensor(dones).to(self.device)
dones = dones.view(dones.size(0), -1)
next_actions, next_log_pi = self.policy_net.sample(next_states)
next_q1 = self.target_q_net1(next_states, next_actions)
next_q2 = self.target_q_net2(next_states, next_actions)
next_q_target = torch.min(next_q1, next_q2) - self.alpha * next_log_pi
expected_q = rewards + (1 - dones) * self.gamma * next_q_target
# q loss
curr_q1 = self.q_net1.forward(states, actions)
curr_q2 = self.q_net2.forward(states, actions)
q1_loss = F.mse_loss(curr_q1, expected_q.detach())
q2_loss = F.mse_loss(curr_q2, expected_q.detach())
# update q networks
self.q1_optimizer.zero_grad()
q1_loss.backward()
self.q1_optimizer.step()
self.q2_optimizer.zero_grad()
q2_loss.backward()
self.q2_optimizer.step()
# delayed update for policy network and target q networks
new_actions, log_pi = self.policy_net.sample(states)
if self.update_step % self.delay_step == 0:
min_q = torch.min(self.q_net1.forward(states, new_actions),
self.q_net2.forward(states, new_actions))
policy_loss = (self.alpha * log_pi - min_q).mean()
self.policy_optimizer.zero_grad()
policy_loss.backward()
self.policy_optimizer.step()
# target networks
for target_param, param in zip(self.target_q_net1.parameters(),
self.q_net1.parameters()):
target_param.data.copy_(self.tau * param +
(1 - self.tau) * target_param)
for target_param, param in zip(self.target_q_net2.parameters(),
self.q_net2.parameters()):
target_param.data.copy_(self.tau * param +
(1 - self.tau) * target_param)
# update temperature
alpha_loss = (self.log_alpha *
(-log_pi - self.target_entropy).detach()).mean()
self.alpha_optim.zero_grad()
alpha_loss.backward()
self.alpha_optim.step()
self.alpha = self.log_alpha.exp()
self.update_step += 1
def __init__(self, env, gamma, tau, v_lr, q_lr, policy_lr, buffer_maxlen):
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.firsttime = 0
self.env = env
self.action_range = [env.action_space.low, env.action_space.high]
#self.obs_dim = env.observation_space.shape[0]
self.action_dim = env.action_space.shape[0] #1
self.conv_channels = 4
self.kernel_size = (3, 3)
self.img_size = (500, 500, 3)
print("Diagnostics:")
print(f"action_range: {self.action_range}")
#print(f"obs_dim: {self.obs_dim}")
print(f"action_dim: {self.action_dim}")
# hyperparameters
self.gamma = gamma
self.tau = tau
self.update_step = 0
self.delay_step = 2
# initialize networks
self.feature_net = FeatureExtractor(self.img_size[2],
self.conv_channels,
self.kernel_size).to(self.device)
print("Feature net init'd successfully")
input_dim = self.feature_net.get_output_size(self.img_size)
self.input_size = input_dim[0] * input_dim[1] * input_dim[2]
print(f"input_size: {self.input_size}")
self.value_net = ValueNetwork(self.input_size, 1).to(self.device)
self.target_value_net = ValueNetwork(self.input_size,
1).to(self.device)
self.q_net1 = SoftQNetwork(self.input_size,
self.action_dim).to(self.device)
self.q_net2 = SoftQNetwork(self.input_size,
self.action_dim).to(self.device)
self.policy_net = PolicyNetwork(self.input_size,
self.action_dim).to(self.device)
print("Finished initing all nets")
# copy params to target param
for target_param, param in zip(self.target_value_net.parameters(),
self.value_net.parameters()):
target_param.data.copy_(param)
print("Finished copying targets")
# initialize optimizers
self.value_optimizer = optim.Adam(self.value_net.parameters(), lr=v_lr)
self.q1_optimizer = optim.Adam(self.q_net1.parameters(), lr=q_lr)
self.q2_optimizer = optim.Adam(self.q_net2.parameters(), lr=q_lr)
self.policy_optimizer = optim.Adam(self.policy_net.parameters(),
lr=policy_lr)
print("Finished initing optimizers")
self.replay_buffer = BasicBuffer(buffer_maxlen)
print("End of init")
class SACAgent:
def __init__(self, env, gamma, tau, v_lr, q_lr, policy_lr, buffer_maxlen):
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.firsttime = 0
self.env = env
self.action_range = [env.action_space.low, env.action_space.high]
#self.obs_dim = env.observation_space.shape[0]
self.action_dim = env.action_space.shape[0] #1
self.conv_channels = 4
self.kernel_size = (3, 3)
self.img_size = (500, 500, 3)
print("Diagnostics:")
print(f"action_range: {self.action_range}")
#print(f"obs_dim: {self.obs_dim}")
print(f"action_dim: {self.action_dim}")
# hyperparameters
self.gamma = gamma
self.tau = tau
self.update_step = 0
self.delay_step = 2
# initialize networks
self.feature_net = FeatureExtractor(self.img_size[2],
self.conv_channels,
self.kernel_size).to(self.device)
print("Feature net init'd successfully")
input_dim = self.feature_net.get_output_size(self.img_size)
self.input_size = input_dim[0] * input_dim[1] * input_dim[2]
print(f"input_size: {self.input_size}")
self.value_net = ValueNetwork(self.input_size, 1).to(self.device)
self.target_value_net = ValueNetwork(self.input_size,
1).to(self.device)
self.q_net1 = SoftQNetwork(self.input_size,
self.action_dim).to(self.device)
self.q_net2 = SoftQNetwork(self.input_size,
self.action_dim).to(self.device)
self.policy_net = PolicyNetwork(self.input_size,
self.action_dim).to(self.device)
print("Finished initing all nets")
# copy params to target param
for target_param, param in zip(self.target_value_net.parameters(),
self.value_net.parameters()):
target_param.data.copy_(param)
print("Finished copying targets")
# initialize optimizers
self.value_optimizer = optim.Adam(self.value_net.parameters(), lr=v_lr)
self.q1_optimizer = optim.Adam(self.q_net1.parameters(), lr=q_lr)
self.q2_optimizer = optim.Adam(self.q_net2.parameters(), lr=q_lr)
self.policy_optimizer = optim.Adam(self.policy_net.parameters(),
lr=policy_lr)
print("Finished initing optimizers")
self.replay_buffer = BasicBuffer(buffer_maxlen)
print("End of init")
def get_action(self, state):
if state.shape != self.img_size:
print(
f"Invalid size, expected shape {self.img_size}, got {state.shape}"
)
return None
inp = torch.from_numpy(state).float().permute(2, 0, 1).unsqueeze(0).to(
self.device)
features = self.feature_net(inp)
features = features.view(-1, self.input_size)
mean, log_std = self.policy_net.forward(features)
std = log_std.exp()
normal = Normal(mean, std)
z = normal.sample()
action = torch.tanh(z)
action = action.cpu().detach().squeeze(0).numpy()
return self.rescale_action(action)
def rescale_action(self, action):
return action * (self.action_range[1] - self.action_range[0]) / 2.0 +\
(self.action_range[1] + self.action_range[0]) / 2.0
def update(self, batch_size):
states, actions, rewards, next_states, dones = self.replay_buffer.sample(
batch_size)
# states and next states are lists of ndarrays, np.stack converts them to
# ndarrays of shape (batch_size, height, width, num_channels)
states = np.stack(states)
next_states = np.stack(next_states)
states = torch.FloatTensor(states).permute(0, 3, 1, 2).to(self.device)
actions = torch.FloatTensor(actions).to(self.device)
rewards = torch.FloatTensor(rewards).to(self.device)
next_states = torch.FloatTensor(next_states).permute(0, 3, 1,
2).to(self.device)
dones = torch.FloatTensor(dones).to(self.device)
dones = dones.view(dones.size(0), -1)
# Process images
features = self.feature_net(
states) #.contiguous() # Properly shaped due to batching
next_features = self.feature_net(next_states) #.contiguous()
features = torch.reshape(features, (64, self.input_size))
next_features = torch.reshape(next_features, (64, self.input_size))
next_actions, next_log_pi = self.policy_net.sample(next_features)
next_q1 = self.q_net1(next_features, next_actions)
next_q2 = self.q_net2(next_features, next_actions)
next_v = self.target_value_net(next_features)
next_v_target = torch.min(next_q1, next_q2) - next_log_pi
curr_v = self.value_net.forward(features)
v_loss = F.mse_loss(curr_v, next_v_target.detach())
# q loss
expected_q = rewards + (1 - dones) * self.gamma * next_v
curr_q1 = self.q_net1.forward(features, actions)
curr_q2 = self.q_net2.forward(features, actions)
q1_loss = F.mse_loss(curr_q1, expected_q.detach())
q2_loss = F.mse_loss(curr_q2, expected_q.detach())
# update value and q networks
self.value_optimizer.zero_grad()
v_loss.backward(retain_graph=True)
self.value_optimizer.step()
self.q1_optimizer.zero_grad()
q1_loss.backward(retain_graph=True)
self.q1_optimizer.step()
self.q2_optimizer.zero_grad()
q2_loss.backward(retain_graph=True)
self.q2_optimizer.step()
# delayed update for policy network and target q networks
if self.update_step % self.delay_step == 0:
new_actions, log_pi = self.policy_net.sample(features)
min_q = torch.min(self.q_net1.forward(features, new_actions),
self.q_net2.forward(features, new_actions))
policy_loss = (log_pi - min_q).mean()
self.policy_optimizer.zero_grad()
policy_loss.backward(retain_graph=True)
self.policy_optimizer.step()
# target networks
for target_param, param in zip(self.target_value_net.parameters(),
self.value_net.parameters()):
target_param.data.copy_(self.tau * param +
(1 - self.tau) * target_param)
self.update_step += 1