class SAC_v2:
def __init__(self, state_dim, action_dim, args):
self.buffer = Buffer(args.buffer_size)
self.actor_optimizer = tf.keras.optimizers.Adam(args.actor_lr)
self.critic1_optimizer = tf.keras.optimizers.Adam(args.critic_lr)
self.critic2_optimizer = tf.keras.optimizers.Adam(args.critic_lr)
self.state_dim = state_dim
self.action_dim = action_dim
self.batch_size = args.batch_size
self.tau = args.tau
self.gamma = args.gamma
self.training_start = args.training_start
self.training_step = args.training_step
self.current_step = 0
self.critic_update = args.critic_update
self.log_alpha = tf.Variable(np.log(args.alpha),
dtype=tf.float32,
trainable=True)
self.target_entropy = -action_dim
self.alpha_optimizer = tf.keras.optimizers.Adam(args.alpha_lr)
self.train_alpha = args.train_alpha
self.actor = Squashed_Gaussian_Actor(self.state_dim, self.action_dim,
args.hidden_dim, args.log_std_min,
args.log_std_max)
self.critic1 = Q_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.target_critic1 = Q_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.critic2 = Q_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.target_critic2 = Q_network(self.state_dim, self.action_dim,
args.hidden_dim)
copy_weight(self.critic1, self.target_critic1)
copy_weight(self.critic2, self.target_critic2)
self.network_list = {
'Actor': self.actor,
'Critic1': self.critic1,
'Critic2': self.critic2,
'Target_Critic1': self.target_critic1,
'Target_Critic2': self.target_critic2
}
self.name = 'SAC_v2'
@property
def alpha(self):
return tf.exp(self.log_alpha)
def get_action(self, state):
state = np.expand_dims(np.array(state), axis=0)
action, _ = self.actor(state)
action = np.clip(action.numpy()[0], -1, 1)
return action
def eval_action(self, state):
state = np.expand_dims(np.array(state), axis=0)
action, _ = self.actor(state, deterministic=True)
action = np.clip(action.numpy()[0], -1, 1)
return action
def train(self, training_num):
total_a_loss = 0
total_c1_loss, total_c2_loss = 0, 0
total_alpha_loss = 0
for i in range(training_num):
self.current_step += 1
s, a, r, ns, d = self.buffer.sample(self.batch_size)
ns_action, ns_logpi = self.actor(ns)
target_min_aq = tf.minimum(self.target_critic1(ns, ns_action),
self.target_critic2(ns, ns_action))
target_q = tf.stop_gradient(
r + self.gamma * (1 - d) *
(target_min_aq - self.alpha.numpy() * ns_logpi))
with tf.GradientTape(persistent=True) as tape1:
critic1_loss = 0.5 * tf.reduce_mean(
tf.square(self.critic1(s, a) - target_q))
critic2_loss = 0.5 * tf.reduce_mean(
tf.square(self.critic2(s, a) - target_q))
critic1_gradients = tape1.gradient(
critic1_loss, self.critic1.trainable_variables)
self.critic1_optimizer.apply_gradients(
zip(critic1_gradients, self.critic1.trainable_variables))
critic2_gradients = tape1.gradient(
critic2_loss, self.critic2.trainable_variables)
self.critic2_optimizer.apply_gradients(
zip(critic2_gradients, self.critic2.trainable_variables))
del tape1
with tf.GradientTape() as tape2:
s_action, s_logpi = self.actor(s)
min_aq_rep = tf.minimum(self.critic1(s, s_action),
self.critic2(s, s_action))
actor_loss = 0.5 * tf.reduce_mean(self.alpha.numpy() *
s_logpi - min_aq_rep)
actor_gradients = tape2.gradient(actor_loss,
self.actor.trainable_variables)
self.actor_optimizer.apply_gradients(
zip(actor_gradients, self.actor.trainable_variables))
del tape2
if self.train_alpha == True:
with tf.GradientTape() as tape3:
_, s_logpi = self.actor(s)
alpha_loss = -(
tf.exp(self.log_alpha) *
(tf.stop_gradient(s_logpi + self.target_entropy)))
alpha_loss = tf.nn.compute_average_loss(
alpha_loss) #from softlearning package
alpha_grad = tape3.gradient(alpha_loss, [self.log_alpha])
self.alpha_optimizer.apply_gradients(
zip(alpha_grad, [self.log_alpha]))
del tape3
if self.current_step % self.critic_update == 0:
soft_update(self.critic1, self.target_critic1, self.tau)
soft_update(self.critic2, self.target_critic2, self.tau)
total_a_loss += actor_loss.numpy()
total_c1_loss += critic1_loss.numpy()
total_c2_loss += critic2_loss.numpy()
if self.train_alpha == True:
total_alpha_loss += alpha_loss.numpy()
return [['Loss/Actor', total_a_loss], ['Loss/Critic1', total_c1_loss],
['Loss/Critic2', total_c2_loss],
['Loss/alpha', total_alpha_loss],
['Alpha', tf.exp(self.log_alpha).numpy()]]
class DBC_SACv2:
def __init__(self, obs_dim, action_dim, args):
self.buffer = Buffer(args.buffer_size)
self.obs_dim = obs_dim
self.action_dim = action_dim
self.log_alpha = tf.Variable(initial_value=tf.math.log(args.alpha),
trainable=True)
self.target_entropy = -action_dim
self.gamma = args.gamma
self.batch_size = args.batch_size
self.feature_dim = args.feature_dim
self.layer_num = args.layer_num
self.filter_num = args.filter_num
self.tau = args.tau
self.encoder_tau = args.encoder_tau
self.actor_update = args.actor_update
self.critic_update = args.critic_update
self.training_start = args.training_start
self.training_step = args.training_step
self.train_alpha = args.train_alpha
self.actor = Squashed_Gaussian_Actor(self.feature_dim, self.action_dim,
args.hidden_dim, args.log_std_min,
args.log_std_max)
self.critic1 = Q_network(self.feature_dim, self.action_dim,
args.hidden_dim)
self.critic2 = Q_network(self.feature_dim, self.action_dim,
args.hidden_dim)
self.target_critic1 = Q_network(self.feature_dim, self.action_dim,
args.hidden_dim)
self.target_critic2 = Q_network(self.feature_dim, self.action_dim,
args.hidden_dim)
self.encoder = PixelEncoder(self.obs_dim, self.feature_dim,
self.layer_num, self.filter_num)
self.target_encoder = PixelEncoder(self.obs_dim, self.feature_dim,
self.layer_num, self.filter_num)
self.dynamics_model = Transition_Network(self.feature_dim,
action_dim,
deterministic=False)
self.reward_model = Reward_Network(self.feature_dim)
copy_weight(self.critic1, self.target_critic1)
copy_weight(self.critic2, self.target_critic2)
copy_weight(self.encoder, self.target_encoder)
self.actor_optimizer = tf.keras.optimizers.Adam(args.actor_lr)
self.critic1_optimizer = tf.keras.optimizers.Adam(args.critic_lr)
self.critic2_optimizer = tf.keras.optimizers.Adam(args.critic_lr)
self.encoder_optimizer = tf.keras.optimizers.Adam(args.encoder_lr)
self.log_alpha_optimizer = tf.keras.optimizers.Adam(args.alpha_lr)
self.dynamics_optimizer = tf.keras.optimizers.Adam(args.decoder_lr)
self.reward_optimizer = tf.keras.optimizers.Adam(args.decoder_lr)
self.current_step = 0
self.network_list = {
'Actor': self.actor,
'Critic1': self.critic1,
'Critic2': self.critic2,
'Target_Critic1': self.target_critic1,
'Target_Critic2': self.target_critic2,
'Encoder': self.encoder,
'Target_Encoder': self.target_encoder,
'Dynamics': self.dynamics_model,
'Reward': self.reward_model
}
self.name = 'DBC_SACv2'
@property
def alpha(self):
return tf.exp(self.log_alpha)
def get_action(self, obs):
obs = np.expand_dims(np.array(obs), axis=0)
feature = self.encoder(obs)
action, _ = self.actor(feature)
action = action.numpy()[0]
return action
def eval_action(self, obs):
obs = np.expand_dims(np.array(obs), axis=0)
feature = self.encoder(obs)
action, _ = self.actor(feature, deterministic=True)
action = action.numpy()[0]
return action
def train(self, local_step):
self.current_step += 1
total_a_loss = 0
total_c1_loss, total_c2_loss = 0, 0
total_alpha_loss = 0
total_encoder_loss = 0
total_dynamics_loss = 0
total_reward_loss = 0
loss_list = []
s, a, r, ns, d = self.buffer.sample(self.batch_size)
ns_action, ns_logpi = self.actor(self.encoder(ns))
target_min_aq = tf.minimum(
self.target_critic1(self.target_encoder(ns), ns_action),
self.target_critic2(self.target_encoder(ns), ns_action))
target_q = tf.stop_gradient(
r + self.gamma * (1 - d) *
(target_min_aq - self.alpha.numpy() * ns_logpi))
with tf.GradientTape(persistent=True) as tape1:
critic1_loss = tf.reduce_mean(
tf.square(self.critic1(self.encoder(s), a) - target_q))
critic2_loss = tf.reduce_mean(
tf.square(self.critic2(self.encoder(s), a) - target_q))
critic1_gradients = tape1.gradient(
critic1_loss, self.encoder.trainable_variables +
self.critic1.trainable_variables)
self.critic1_optimizer.apply_gradients(
zip(
critic1_gradients, self.encoder.trainable_variables +
self.critic1.trainable_variables))
critic2_gradients = tape1.gradient(
critic2_loss, self.encoder.trainable_variables +
self.critic2.trainable_variables)
self.critic2_optimizer.apply_gradients(
zip(
critic2_gradients, self.encoder.trainable_variables +
self.critic2.trainable_variables))
del tape1
if self.current_step % self.actor_update == 0:
with tf.GradientTape() as tape2:
s_action, s_logpi = self.actor(
tf.stop_gradient(self.encoder(s)))
min_aq_rep = tf.minimum(
self.critic1(tf.stop_gradient(self.encoder(s)), s_action),
self.critic2(tf.stop_gradient(self.encoder(s)), s_action))
actor_loss = tf.reduce_mean(self.alpha.numpy() * s_logpi -
min_aq_rep)
actor_gradients = tape2.gradient(actor_loss,
self.actor.trainable_variables)
self.actor_optimizer.apply_gradients(
zip(actor_gradients, self.actor.trainable_variables))
del tape2
if self.train_alpha == True:
with tf.GradientTape() as tape3:
_, s_logpi = self.actor(self.encoder(s))
alpha_loss = -(
tf.exp(self.log_alpha) *
tf.stop_gradient(s_logpi + self.target_entropy))
alpha_loss = tf.nn.compute_average_loss(alpha_loss)
#alpha_loss = tf.reduce_mean(alpha_loss)
log_alpha_gradients = tape3.gradient(alpha_loss,
[self.log_alpha])
self.log_alpha_optimizer.apply_gradients(
zip(log_alpha_gradients, [self.log_alpha]))
del tape3
if self.current_step % self.critic_update == 0:
soft_update(self.critic1, self.target_critic1, self.tau)
soft_update(self.critic2, self.target_critic2, self.tau)
soft_update(self.encoder, self.target_encoder, self.encoder_tau)
#train encoder
with tf.GradientTape() as tape4:
new_ids = np.arange(len(s))
np.random.shuffle(new_ids)
s2 = tf.gather(s, new_ids)
feature = self.encoder(s)
#feature2 = tf.gather(feature, new_ids)
feature2 = self.encoder(s2)
reward = self.reward_model(tf.stop_gradient(feature))
#reward2 = tf.gather(reward, new_ids)
reward2 = self.reward_model(tf.stop_gradient(feature2))
feature_action, _ = self.actor(tf.stop_gradient(feature), True)
feature2_action, _ = self.actor(tf.stop_gradient(feature2), True)
mu, sigma = self.dynamics_model(tf.stop_gradient(feature),
feature_action)
mu2, sigma2 = self.dynamics_model(tf.stop_gradient(feature2),
feature2_action)
z_dist = tf.reshape(tf.keras.losses.huber(feature, feature2),
shape=[-1, 1])
r_dist = tf.reshape(tf.keras.losses.huber(reward, reward2),
shape=[-1, 1])
transition_dist = tf.sqrt(
tf.square(mu - mu2) + tf.square(sigma - sigma2))
bisimilarity = r_dist + self.gamma * transition_dist
encoder_loss = tf.reduce_mean(tf.square(z_dist - bisimilarity))
encoder_gradients = tape4.gradient(encoder_loss,
self.encoder.trainable_variables)
self.encoder_optimizer.apply_gradients(
zip(encoder_gradients, self.encoder.trainable_variables))
#train dynamics
with tf.GradientTape() as tape5:
feature = self.encoder(s)
mu, sigma = self.dynamics_model(feature, a)
if (sigma[0][0].numpy() == 0):
if self.dynamics_model.deterministic == False:
print("error")
sigma = tf.ones_like(mu)
next_feature = self.encoder(ns)
diff = (mu - tf.stop_gradient(next_feature)) / sigma
dynamics_loss = tf.reduce_mean(0.5 * tf.square(diff) +
tf.math.log(sigma))
dynamics_gradients = tape5.gradient(
dynamics_loss, self.encoder.trainable_variables +
self.dynamics_model.trainable_variables)
self.dynamics_optimizer.apply_gradients(
zip(
dynamics_gradients, self.encoder.trainable_variables +
self.dynamics_model.trainable_variables))
#train reward
with tf.GradientTape() as tape6:
feature = self.encoder(s)
sample_dynamics = self.dynamics_model.sample(feature, a)
reward_prediction = self.reward_model(sample_dynamics)
reward_loss = tf.reduce_mean(tf.square(reward_prediction - r))
reward_gradients = tape6.gradient(
reward_loss, self.encoder.trainable_variables +
self.reward_model.trainable_variables)
self.reward_optimizer.apply_gradients(
zip(
reward_gradients, self.encoder.trainable_variables +
self.reward_model.trainable_variables))
total_c1_loss += critic1_loss.numpy()
total_c2_loss += critic2_loss.numpy()
loss_list.append(['Loss/Critic1', total_c1_loss])
loss_list.append(['Loss/Critic2', total_c2_loss])
if self.current_step % self.actor_update == 0:
total_a_loss += actor_loss.numpy()
loss_list.append(['Loss/Actor', total_a_loss])
total_encoder_loss += encoder_loss.numpy()
loss_list.append(['Loss/Encoder', total_encoder_loss])
total_dynamics_loss += dynamics_loss.numpy()
loss_list.append(['Loss/Dynamics', total_dynamics_loss])
total_reward_loss += reward_loss.numpy()
loss_list.append(['Loss/Reward', total_reward_loss])
if self.current_step % self.actor_update == 0 and self.train_alpha == True:
total_alpha_loss += alpha_loss.numpy()
loss_list.append(['Loss/Alpha', total_alpha_loss])
loss_list.append(['Alpha', tf.exp(self.log_alpha).numpy()])
return loss_list
class TD3:
def __init__(self, state_dim, action_dim, args):
self.buffer = Buffer(args.buffer_size)
self.actor_optimizer = tf.keras.optimizers.Adam(args.actor_lr)
self.critic1_optimizer = tf.keras.optimizers.Adam(args.critic_lr)
self.critic2_optimizer = tf.keras.optimizers.Adam(args.critic_lr)
self.state_dim = state_dim
self.action_dim = action_dim
self.batch_size = args.batch_size
self.gamma = args.gamma
self.tau = args.tau
self.actor_lr = args.actor_lr
self.critic_lr = args.critic_lr
self.policy_delay = args.policy_delay
self.actor_noise = args.actor_noise
self.target_noise = args.target_noise
self.noise_clip = args.noise_clip
self.training_start = args.training_start
self.training_step = args.training_step
self.current_step = 0
self.actor = Policy_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.target_actor = Policy_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.critic1 = Q_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.target_critic1 = Q_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.critic2 = Q_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.target_critic2 = Q_network(self.state_dim, self.action_dim,
args.hidden_dim)
copy_weight(self.actor, self.target_actor)
copy_weight(self.critic1, self.target_critic1)
copy_weight(self.critic2, self.target_critic2)
self.network_list = {
'Actor': self.actor,
'Critic1': self.critic1,
'Critic2': self.critic2,
'Target_Critic1': self.target_critic1,
'Target_Critic2': self.target_critic2
}
self.name = 'TD3'
def get_action(self, state):
state = np.expand_dims(np.array(state), axis=0)
noise = np.random.normal(loc=0,
scale=self.actor_noise,
size=self.action_dim)
action = self.actor(state).numpy()[0] + noise
action = np.clip(action, -1, 1)
return action
def eval_action(self, state):
state = np.expand_dims(np.array(state), axis=0)
action = self.actor(state).numpy()[0]
action = np.clip(action, -1, 1)
return action
def train(self, training_num):
total_a_loss = 0
total_c1_loss, total_c2_loss = 0, 0
for i in range(training_num):
self.current_step += 1
s, a, r, ns, d = self.buffer.sample(self.batch_size)
target_action = tf.clip_by_value(
self.target_actor(ns) + tf.clip_by_value(
tf.random.normal(shape=self.target_actor(ns).shape,
mean=0,
stddev=self.target_noise),
-self.noise_clip, self.noise_clip), -1, 1)
target_value = tf.stop_gradient(
r + self.gamma *
(1 - d) * tf.minimum(self.target_critic1(ns, target_action),
self.target_critic2(ns, target_action)))
with tf.GradientTape(persistent=True) as tape:
critic1_loss = 0.5 * tf.reduce_mean(
tf.square(target_value - self.critic1(s, a)))
critic2_loss = 0.5 * tf.reduce_mean(
tf.square(target_value - self.critic2(s, a)))
critic1_grad = tape.gradient(critic1_loss,
self.critic1.trainable_variables)
self.critic1_optimizer.apply_gradients(
zip(critic1_grad, self.critic1.trainable_variables))
critic2_grad = tape.gradient(critic2_loss,
self.critic2.trainable_variables)
self.critic2_optimizer.apply_gradients(
zip(critic2_grad, self.critic2.trainable_variables))
if self.current_step % self.policy_delay == 0:
with tf.GradientTape() as tape2:
actor_loss = -tf.reduce_mean(self.critic1(
s, self.actor(s)))
actor_grad = tape2.gradient(actor_loss,
self.actor.trainable_variables)
self.actor_optimizer.apply_gradients(
zip(actor_grad, self.actor.trainable_variables))
soft_update(self.actor, self.target_actor, self.tau)
soft_update(self.critic1, self.target_critic1, self.tau)
soft_update(self.critic2, self.target_critic2, self.tau)
del tape, tape2
total_a_loss += actor_loss.numpy()
total_c1_loss += critic1_loss.numpy()
total_c2_loss += critic2_loss.numpy()
return [['Loss/Actor', total_a_loss], ['Loss/Critic1', total_c1_loss],
['Loss/Critic2', total_c2_loss]]
class DDQN:
def __init__(self, state_dim, action_dim, args):
self.buffer = Buffer(args.buffer_size)
self.optimizer = tf.keras.optimizers.Adam(args.learning_rate)
self.state_dim = state_dim
self.action_dim = action_dim
self.batch_size = args.batch_size
self.gamma = args.gamma
self.lr = args.learning_rate
self.epsilon = args.epsilon
self.training_start = args.training_start
self.training_step = args.training_step
self.current_step = 0
self.copy_iter = args.copy_iter
self.network = Policy_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.target_network = Policy_network(self.state_dim, self.action_dim,
args.hidden_dim)
copy_weight(self.network, self.target_network)
self.network_list = {
'Network': self.network,
'Target_Network': self.target_network
}
self.name = 'Double DQN'
def get_action(self, state):
state = np.expand_dims(np.array(state), axis=0)
q_value = self.network(state, activation='linear').numpy()
best_action = np.argmax(q_value, axis=1)[0]
if np.random.random() < self.epsilon:
return np.random.randint(low=0, high=self.action_dim)
else:
return best_action
def eval_action(self, state):
state = np.expand_dims(np.array(state), axis=0)
q_value = self.network(state, activation='linear').numpy()
best_action = np.argmax(q_value, axis=1)[0]
return best_action
def train(self, training_num):
total_loss = 0
for i in range(training_num):
self.current_step += 1
s, a, r, ns, d = self.buffer.sample(self.batch_size)
q_value = tf.expand_dims(tf.argmax(self.network(
ns, activation='linear'),
axis=1,
output_type=tf.int32),
axis=1)
q_value_one = tf.squeeze(tf.one_hot(q_value,
depth=self.action_dim),
axis=1)
target_value = r + self.gamma * (1 - d) * tf.reduce_sum(
self.target_network(ns, activation='linear') * q_value_one,
axis=1,
keepdims=True)
target_value = tf.stop_gradient(target_value)
with tf.GradientTape() as tape:
selected_values = tf.reduce_sum(
self.network(s, activation='linear') * tf.squeeze(
tf.one_hot(tf.cast(a, tf.int32), self.action_dim),
axis=1),
axis=1,
keepdims=True)
loss = 0.5 * tf.math.reduce_mean(
tf.square(target_value - selected_values))
variables = self.network.trainable_variables
gradients = tape.gradient(loss, variables)
self.optimizer.apply_gradients(zip(gradients, variables))
if self.current_step % self.copy_iter == 0:
copy_weight(self.network, self.target_network)
total_loss += loss.numpy()
return [['Loss/Loss', total_loss]]
class ImageDQN:
def __init__(self, obs_dim, action_dim, args):
self.buffer = Buffer(args.buffer_size)
self.optimizer = tf.keras.optimizers.Adam(args.learning_rate)
self.obs_dim = obs_dim
self.action_dim = action_dim
self.feature_dim = args.feature_dim
self.batch_size = args.batch_size
self.gamma = args.gamma
self.learning_rate = args.learning_rate
self.epsilon = args.epsilon
self.training_start = args.training_start
self.training_step = args.training_step
self.current_step = 0
self.copy_iter = args.copy_iter
self.layer_num = args.layer_num
self.filter_num = args.filter_num
self.network = Policy_network(self.feature_dim, self.action_dim,
args.hidden_dim)
self.target_network = Policy_network(self.feature_dim, self.action_dim,
args.hidden_dim)
self.encoder = PixelEncoder(self.obs_dim, self.feature_dim,
self.layer_num, self.filter_num,
'channels_last')
self.target_encoder = PixelEncoder(self.obs_dim, self.feature_dim,
self.layer_num, self.filter_num,
'channels_last')
copy_weight(self.network, self.target_network)
copy_weight(self.encoder, self.target_encoder)
self.network_list = {
'Network': self.network,
'Target_Network': self.target_network
}
self.name = 'ImageDQN'
def get_action(self, obs):
if np.random.random() < self.epsilon:
return np.random.randint(low=0, high=self.action_dim)
else:
obs = np.expand_dims(np.array(obs), axis=0)
feature = self.encoder(obs)
q_value = self.network(feature, activation='linear').numpy()
best_action = np.argmax(q_value, axis=1)[0]
return best_action
def eval_action(self, obs):
obs = np.expand_dims(np.array(obs), axis=0)
feature = self.encoder(obs)
q_value = self.network(feature, activation='linear').numpy()
best_action = np.argmax(q_value, axis=1)[0]
return best_action
def train(self, training_num):
total_loss = 0
for i in range(training_num):
self.current_step += 1
s, a, r, ns, d = self.buffer.sample(self.batch_size)
target_q = tf.reduce_max(self.target_network(
self.target_encoder(ns), activation='linear'),
axis=1,
keepdims=True)
target_value = r + self.gamma * (1 - d) * target_q
target_value = tf.stop_gradient(target_value)
with tf.GradientTape() as tape:
selected_values = tf.reduce_sum(
self.network(self.encoder(s), activation='linear') *
tf.squeeze(tf.one_hot(tf.cast(a, tf.int32),
self.action_dim),
axis=1),
axis=1,
keepdims=True)
loss = 0.5 * tf.reduce_mean(
tf.square(target_value - selected_values))
gradients = tape.gradient(
loss, self.encoder.trainable_variables +
self.network.trainable_variables)
self.optimizer.apply_gradients(
zip(
gradients, self.encoder.trainable_variables +
self.network.trainable_variables))
if self.current_step % self.copy_iter == 0:
copy_weight(self.network, self.target_network)
copy_weight(self.encoder, self.target_encoder)
total_loss += loss.numpy()
del tape
return [['Loss/Loss', total_loss]]
class SACv2_AE:
def __init__(self, obs_dim, action_dim, args):
self.buffer = Buffer(args.buffer_size)
self.obs_dim = obs_dim
self.action_dim = action_dim
self.image_size = args.image_size
self.current_step = 0
self.log_alpha = tf.Variable(initial_value=tf.math.log(args.alpha),
trainable=True)
self.target_entropy = -action_dim
self.gamma = args.gamma
self.batch_size = args.batch_size
self.feature_dim = args.feature_dim
self.layer_num = args.layer_num
self.filter_num = args.filter_num
self.tau = args.tau
self.encoder_tau = args.encoder_tau
self.actor_update = args.actor_update
self.critic_update = args.critic_update
self.decoder_update = args.decoder_update
self.decoder_latent_lambda = args.decoder_latent_lambda
self.decoder_weight_lambda = args.decoder_weight_lambda
self.training_start = args.training_start
self.training_step = args.training_step
self.train_alpha = args.train_alpha
self.actor = Squashed_Gaussian_Actor(self.feature_dim, self.action_dim,
args.hidden_dim, args.log_std_min,
args.log_std_max)
self.critic1 = Q_network(self.feature_dim, self.action_dim,
args.hidden_dim)
self.critic2 = Q_network(self.feature_dim, self.action_dim,
args.hidden_dim)
self.target_critic1 = Q_network(self.feature_dim, self.action_dim,
args.hidden_dim)
self.target_critic2 = Q_network(self.feature_dim, self.action_dim,
args.hidden_dim)
self.encoder = PixelEncoder(self.obs_dim, self.feature_dim,
self.layer_num, self.filter_num)
self.target_encoder = PixelEncoder(self.obs_dim, self.feature_dim,
self.layer_num, self.filter_num)
self.decoder = PixelDecoder(self.obs_dim, self.feature_dim,
self.layer_num, self.filter_num)
copy_weight(self.critic1, self.target_critic1)
copy_weight(self.critic2, self.target_critic2)
copy_weight(self.encoder, self.target_encoder)
self.actor_optimizer = tf.keras.optimizers.Adam(args.actor_lr)
self.critic1_optimizer = tf.keras.optimizers.Adam(args.critic_lr)
self.critic2_optimizer = tf.keras.optimizers.Adam(args.critic_lr)
self.encoder_optimizer = tf.keras.optimizers.Adam(args.encoder_lr)
self.decoder_optimizer = tfa.optimizers.AdamW(
weight_decay=self.decoder_weight_lambda,
learning_rate=args.decoder_lr)
self.log_alpha_optimizer = tf.keras.optimizers.Adam(args.alpha_lr,
beta_1=0.5)
self.network_list = {
'Actor': self.actor,
'Critic1': self.critic1,
'Critic2': self.critic2,
'Target_Critic1': self.target_critic1,
'Target_Critic2': self.target_critic2,
'Encoder': self.encoder,
'Target_Encoder': self.target_encoder,
'Decoder': self.decoder
}
self.name = 'SACv2_AE'
@property
def alpha(self):
return tf.exp(self.log_alpha)
def get_action(self, obs):
obs = np.expand_dims(np.array(obs), axis=0)
feature = self.encoder(obs)
action, _ = self.actor(feature)
action = action.numpy()[0]
return action
def eval_action(self, obs):
obs = np.expand_dims(np.array(obs), axis=0)
feature = self.encoder(obs)
action, _ = self.actor(feature, deterministic=True)
action = action.numpy()[0]
return action
def train(self, local_step):
self.current_step += 1
total_a_loss = 0
total_c1_loss, total_c2_loss = 0, 0
total_alpha_loss = 0
total_ae_loss = 0
loss_list = []
s, a, r, ns, d = self.buffer.sample(self.batch_size)
ns_action, ns_logpi = self.actor(self.encoder(ns))
target_min_aq = tf.minimum(
self.target_critic1(self.target_encoder(ns), ns_action),
self.target_critic2(self.target_encoder(ns), ns_action))
target_q = tf.stop_gradient(
r + self.gamma * (1 - d) *
(target_min_aq - self.alpha.numpy() * ns_logpi))
#critic update
with tf.GradientTape(persistent=True) as tape1:
critic1_loss = tf.reduce_mean(
tf.square(self.critic1(self.encoder(s), a) - target_q))
critic2_loss = tf.reduce_mean(
tf.square(self.critic2(self.encoder(s), a) - target_q))
critic1_gradients = tape1.gradient(
critic1_loss, self.encoder.trainable_variables +
self.critic1.trainable_variables)
self.critic1_optimizer.apply_gradients(
zip(
critic1_gradients, self.encoder.trainable_variables +
self.critic1.trainable_variables))
critic2_gradients = tape1.gradient(
critic2_loss, self.encoder.trainable_variables +
self.critic2.trainable_variables)
self.critic2_optimizer.apply_gradients(
zip(
critic2_gradients, self.encoder.trainable_variables +
self.critic2.trainable_variables))
del tape1
#actor update
if self.current_step % self.actor_update == 0:
with tf.GradientTape() as tape2:
s_action, s_logpi = self.actor(
tf.stop_gradient(self.encoder(s)))
min_aq_rep = tf.minimum(
self.critic1(tf.stop_gradient(self.encoder(s)), s_action),
self.critic2(tf.stop_gradient(self.encoder(s)), s_action))
actor_loss = tf.reduce_mean(self.alpha.numpy() * s_logpi -
min_aq_rep)
actor_gradients = tape2.gradient(actor_loss,
self.actor.trainable_variables)
self.actor_optimizer.apply_gradients(
zip(actor_gradients, self.actor.trainable_variables))
del tape2
#alpha update
if self.train_alpha == True:
with tf.GradientTape() as tape3:
_, s_logpi = self.actor(self.encoder(s))
alpha_loss = -(
tf.exp(self.log_alpha) *
tf.stop_gradient(s_logpi + self.target_entropy))
alpha_loss = tf.nn.compute_average_loss(alpha_loss)
log_alpha_gradients = tape3.gradient(alpha_loss,
[self.log_alpha])
self.log_alpha_optimizer.apply_gradients(
zip(log_alpha_gradients, [self.log_alpha]))
del tape3
if self.current_step % self.decoder_update == 0:
#encoder, decoder update
with tf.GradientTape(persistent=True) as tape4:
feature = self.encoder(s)
recovered_s = self.decoder(feature)
real_s = preprocess_obs(s)
rec_loss = tf.reduce_mean(tf.square(recovered_s - real_s))
latent_loss = tf.reduce_mean(
0.5 * tf.reduce_sum(tf.square(feature), axis=1))
ae_loss = rec_loss + self.decoder_latent_lambda * latent_loss
encoder_gradients = tape4.gradient(
ae_loss, self.encoder.trainable_variables)
decoder_gradients = tape4.gradient(
ae_loss, self.decoder.trainable_variables)
self.encoder_optimizer.apply_gradients(
zip(encoder_gradients, self.encoder.trainable_variables))
self.decoder_optimizer.apply_gradients(
zip(decoder_gradients, self.decoder.trainable_variables))
if self.current_step % self.critic_update == 0:
soft_update(self.critic1, self.target_critic1, self.tau)
soft_update(self.critic2, self.target_critic2, self.tau)
soft_update(self.encoder, self.target_encoder, self.encoder_tau)
del tape4
total_c1_loss += critic1_loss.numpy()
total_c2_loss += critic2_loss.numpy()
loss_list.append(['Loss/Critic1', total_c1_loss])
loss_list.append(['Loss/Critic2', total_c2_loss])
if self.current_step % self.decoder_update == 0:
total_ae_loss += ae_loss.numpy()
loss_list.append(['Loss/AutoEncoder', total_ae_loss])
if self.current_step % self.actor_update == 0:
total_a_loss += actor_loss.numpy()
loss_list.append(['Loss/Actor', total_a_loss])
if self.train_alpha == True:
total_alpha_loss += alpha_loss.numpy()
loss_list.append(['Loss/Alpha', total_alpha_loss])
loss_list.append(['Alpha', tf.exp(self.log_alpha).numpy()])
return loss_list
class SAC_v2:
def __init__(self,
state_dim,
action_dim,
hidden_dim=256,
training_step=1,
alpha=0.1,
train_alpha=True,
batch_size=128,
buffer_size=1e6,
tau=0.005,
learning_rate=0.0003,
gamma=0.99,
reward_scale=1,
training_start=500):
self.buffer = Buffer(buffer_size)
self.actor_optimizer = tf.keras.optimizers.Adam(learning_rate)
self.critic1_optimizer = tf.keras.optimizers.Adam(learning_rate)
self.critic2_optimizer = tf.keras.optimizers.Adam(learning_rate)
self.state_dim = state_dim
self.action_dim = action_dim
self.batch_size = batch_size
self.tau = tau
self.gamma = gamma
self.reward_scale = reward_scale
self.training_start = training_start
self.training_step = training_step
self.log_alpha = tf.Variable(np.log(alpha),
dtype=tf.float32,
trainable=True)
self.target_entropy = -action_dim
self.alpha_optimizer = tf.keras.optimizers.Adam(learning_rate)
self.train_alpha = train_alpha
self.actor = Squashed_Gaussian_Actor(self.state_dim, self.action_dim,
(hidden_dim, hidden_dim))
self.critic1 = Q_network(self.state_dim, self.action_dim,
(hidden_dim, hidden_dim))
self.target_critic1 = Q_network(self.state_dim, self.action_dim,
(hidden_dim, hidden_dim))
self.critic2 = Q_network(self.state_dim, self.action_dim,
(hidden_dim, hidden_dim))
self.target_critic2 = Q_network(self.state_dim, self.action_dim,
(hidden_dim, hidden_dim))
copy_weight(self.critic1, self.target_critic1)
copy_weight(self.critic2, self.target_critic2)
self.network_list = {
'Actor': self.actor,
'Critic1': self.critic1,
'Critic2': self.critic2,
'Target_Critic1': self.target_critic1,
'Target_Critic2': self.target_critic2
}
self.name = 'SAC_v2'
@property
def alpha(self):
return tf.exp(self.log_alpha)
def get_action(self, state):
state = np.expand_dims(np.array(state), axis=0)
action = self.actor(state).numpy()[0]
return action
def train(self, training_num):
for i in range(training_num):
s, a, r, ns, d = self.buffer.sample(self.batch_size)
target_min_aq = tf.minimum(self.target_critic1(ns, self.actor(ns)),
self.target_critic2(ns, self.actor(ns)))
target_q = tf.stop_gradient(
r + self.gamma * (1 - d) *
(target_min_aq - self.alpha.numpy() * self.actor.log_pi(ns)))
#critic training
with tf.GradientTape(persistent=True) as tape1:
critic1_loss = tf.reduce_mean(
tf.square(self.critic1(s, a) - target_q))
critic2_loss = tf.reduce_mean(
tf.square(self.critic2(s, a) - target_q))
critic1_gradients = tape1.gradient(
critic1_loss, self.critic1.trainable_variables)
self.critic1_optimizer.apply_gradients(
zip(critic1_gradients, self.critic1.trainable_variables))
critic2_gradients = tape1.gradient(
critic2_loss, self.critic2.trainable_variables)
self.critic2_optimizer.apply_gradients(
zip(critic2_gradients, self.critic2.trainable_variables))
del tape1
#actor training
with tf.GradientTape() as tape2:
mu, sigma = self.actor.mu_sigma(s)
output = mu + tf.random.normal(shape=mu.shape) * sigma
min_aq_rep = tf.minimum(self.critic1(s, output),
self.critic2(s, output))
actor_loss = tf.reduce_mean(self.alpha.numpy() *
self.actor.log_pi(s) - min_aq_rep)
actor_gradients = tape2.gradient(actor_loss,
self.actor.trainable_variables)
self.actor_optimizer.apply_gradients(
zip(actor_gradients, self.actor.trainable_variables))
del tape2
#alpha(temperature) training
if self.train_alpha == True:
with tf.GradientTape() as tape3:
alpha_loss = -(tf.exp(self.log_alpha) * (tf.stop_gradient(
self.actor.log_pi(s) + self.target_entropy)))
alpha_loss = tf.nn.compute_average_loss(
alpha_loss) #from softlearning package
alpha_grad = tape3.gradient(alpha_loss, [self.log_alpha])
self.alpha_optimizer.apply_gradients(
zip(alpha_grad, [self.log_alpha]))
del tape3
soft_update(self.critic1, self.target_critic1, self.tau)
soft_update(self.critic2, self.target_critic2, self.tau)
class SAC_v1:
def __init__(self,
state_dim,
action_dim,
hidden_dim=256,
training_step=1,
batch_size=128,
buffer_size=1e6,
tau=0.005,
learning_rate=0.0003,
gamma=0.99,
alpha=0.2,
reward_scale=1,
training_start=500):
self.buffer = Buffer(buffer_size)
self.actor_optimizer = tf.keras.optimizers.Adam(learning_rate)
self.critic1_optimizer = tf.keras.optimizers.Adam(learning_rate)
self.critic2_optimizer = tf.keras.optimizers.Adam(learning_rate)
self.v_network_optimizer = tf.keras.optimizers.Adam(learning_rate)
self.state_dim = state_dim
self.action_dim = action_dim
self.batch_size = batch_size
self.tau = tau
self.gamma = gamma
self.alpha = alpha
self.reward_scale = reward_scale
self.training_start = training_start
self.training_step = training_step
self.actor = Squashed_Gaussian_Actor(self.state_dim, self.action_dim,
(hidden_dim, hidden_dim))
self.critic1 = Q_network(self.state_dim, self.action_dim,
(hidden_dim, hidden_dim))
self.critic2 = Q_network(self.state_dim, self.action_dim,
(hidden_dim, hidden_dim))
self.v_network = V_network(self.state_dim, (hidden_dim, hidden_dim))
self.target_v_network = V_network(self.state_dim,
(hidden_dim, hidden_dim))
copy_weight(self.v_network, self.target_v_network)
self.network_list = {
'Actor': self.actor,
'Critic1': self.critic1,
'Critic2': self.critic2,
'V_network': self.v_network,
'Target_V_network': self.target_v_network
}
self.name = 'SAC_v1'
def get_action(self, state):
state = np.expand_dims(np.array(state), axis=0)
action = self.actor(state).numpy()[0]
return action
def train(self, training_num):
for i in range(training_num):
s, a, r, ns, d = self.buffer.sample(self.batch_size)
min_aq = tf.minimum(self.critic1(s, self.actor(s)),
self.critic2(s, self.actor(s)))
target_v = tf.stop_gradient(min_aq -
self.alpha * self.actor.log_pi(s))
#v_network training
with tf.GradientTape(persistent=True) as tape1:
v_loss = 0.5 * tf.reduce_mean(
tf.square(self.v_network(s) - target_v))
v_gradients = tape1.gradient(v_loss,
self.v_network.trainable_variables)
self.v_network_optimizer.apply_gradients(
zip(v_gradients, self.v_network.trainable_variables))
del tape1
target_q = tf.stop_gradient(r + self.gamma *
(1 - d) * self.target_v_network(ns))
#critic training
with tf.GradientTape(persistent=True) as tape2:
critic1_loss = 0.5 * tf.reduce_mean(
tf.square(self.critic1(s, a) - target_q))
critic2_loss = 0.5 * tf.reduce_mean(
tf.square(self.critic2(s, a) - target_q))
critic1_gradients = tape2.gradient(
critic1_loss, self.critic1.trainable_variables)
self.critic1_optimizer.apply_gradients(
zip(critic1_gradients, self.critic1.trainable_variables))
critic2_gradients = tape2.gradient(
critic2_loss, self.critic2.trainable_variables)
self.critic2_optimizer.apply_gradients(
zip(critic2_gradients, self.critic2.trainable_variables))
del tape2
#actor training
with tf.GradientTape() as tape3:
mu, sigma = self.actor.mu_sigma(s)
output = mu + tf.random.normal(shape=sigma.shape) * sigma
min_aq_rep = tf.minimum(self.critic1(s, output),
self.critic2(s, output))
actor_loss = tf.reduce_mean(self.alpha * self.actor.log_pi(s) -
min_aq_rep)
actor_grad = tape3.gradient(actor_loss,
self.actor.trainable_variables)
self.actor_optimizer.apply_gradients(
zip(actor_grad, self.actor.trainable_variables))
del tape3
soft_update(self.v_network, self.target_v_network, self.tau)
class SAC_v1:
def __init__(self, state_dim, action_dim, args):
self.buffer = Buffer(args.buffer_size)
self.actor_optimizer = tf.keras.optimizers.Adam(args.actor_lr)
self.critic1_optimizer = tf.keras.optimizers.Adam(args.critic_lr)
self.critic2_optimizer = tf.keras.optimizers.Adam(args.critic_lr)
self.v_network_optimizer = tf.keras.optimizers.Adam(args.v_lr)
self.state_dim = state_dim
self.action_dim = action_dim
self.batch_size = args.batch_size
self.tau = args.tau
self.gamma = args.gamma
self.alpha = args.alpha
self.training_start = args.training_start
self.training_step = args.training_step
self.current_step = 0
self.actor = Squashed_Gaussian_Actor(self.state_dim, self.action_dim,
args.hidden_dim, args.log_std_min,
args.log_std_max)
self.critic1 = Q_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.critic2 = Q_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.v_network = V_network(self.state_dim, args.hidden_dim)
self.target_v_network = V_network(self.state_dim, args.hidden_dim)
copy_weight(self.v_network, self.target_v_network)
self.network_list = {
'Actor': self.actor,
'Critic1': self.critic1,
'Critic2': self.critic2,
'V_network': self.v_network,
'Target_V_network': self.target_v_network
}
self.name = 'SAC_v1'
def get_action(self, state):
state = np.expand_dims(np.array(state), axis=0)
action, _ = self.actor(state)
action = np.clip(action.numpy()[0], -1, 1)
return action
def eval_action(self, state):
state = np.expand_dims(np.array(state), axis=0)
action, _ = self.actor(state, deterministic=True)
action = np.clip(action.numpy()[0], -1, 1)
return action
def train(self, training_num):
total_a_loss = 0
total_c1_loss, total_c2_loss = 0, 0
total_v_loss = 0
for i in range(training_num):
self.current_step += 1
s, a, r, ns, d = self.buffer.sample(self.batch_size)
s_action, s_logpi = self.actor(s)
min_aq = tf.minimum(self.critic1(s, s_action),
self.critic2(s, s_action))
target_v = tf.stop_gradient(min_aq - self.alpha * s_logpi)
with tf.GradientTape() as tape1:
v_loss = 0.5 * tf.reduce_mean(
tf.square(self.v_network(s) - target_v))
v_gradients = tape1.gradient(v_loss,
self.v_network.trainable_variables)
self.v_network_optimizer.apply_gradients(
zip(v_gradients, self.v_network.trainable_variables))
target_q = tf.stop_gradient(r + self.gamma *
(1 - d) * self.target_v_network(ns))
with tf.GradientTape(persistent=True) as tape2:
critic1_loss = 0.5 * tf.reduce_mean(
tf.square(self.critic1(s, a) - target_q))
critic2_loss = 0.5 * tf.reduce_mean(
tf.square(self.critic2(s, a) - target_q))
critic1_gradients = tape2.gradient(
critic1_loss, self.critic1.trainable_variables)
self.critic1_optimizer.apply_gradients(
zip(critic1_gradients, self.critic1.trainable_variables))
critic2_gradients = tape2.gradient(
critic2_loss, self.critic2.trainable_variables)
self.critic2_optimizer.apply_gradients(
zip(critic2_gradients, self.critic2.trainable_variables))
with tf.GradientTape() as tape3:
s_action, s_logpi = self.actor(s)
min_aq_rep = tf.minimum(self.critic1(s, s_action),
self.critic2(s, s_action))
actor_loss = tf.reduce_mean(self.alpha * s_logpi - min_aq_rep)
actor_grad = tape3.gradient(actor_loss,
self.actor.trainable_variables)
self.actor_optimizer.apply_gradients(
zip(actor_grad, self.actor.trainable_variables))
soft_update(self.v_network, self.target_v_network, self.tau)
del tape1, tape2, tape3
total_a_loss += actor_loss.numpy()
total_c1_loss += critic1_loss.numpy()
total_c2_loss += critic2_loss.numpy()
total_v_loss += v_loss.numpy()
return [['Loss/Actor', total_a_loss], ['Loss/Critic1', total_c1_loss],
['Loss/Critic2', total_c2_loss], ['Loss/V', total_v_loss]]
class DDPG:
def __init__(self, state_dim, action_dim, args):
self.buffer = Buffer(args.buffer_size)
self.actor_optimizer = tf.keras.optimizers.Adam(args.actor_lr)
self.critic_optimizer = tf.keras.optimizers.Adam(args.critic_lr)
self.state_dim = state_dim
self.action_dim = action_dim
self.batch_size = args.batch_size
self.gamma = args.gamma
self.tau = args.tau
self.noise_scale = args.noise_scale
self.training_start = args.training_start
self.training_step = args.training_step
self.current_step = 0
self.actor = Policy_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.target_actor = Policy_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.critic = Q_network(self.state_dim, self.action_dim,
args.hidden_dim)
self.target_critic = Q_network(self.state_dim, self.action_dim,
args.hidden_dim)
copy_weight(self.actor, self.target_actor)
copy_weight(self.critic, self.target_critic)
self.network_list = {
'Actor': self.actor,
'Target_Actor': self.target_actor,
'Critic': self.critic,
'Target_Critic': self.target_critic
}
self.name = 'DDPG'
def get_action(self, state):
state = np.expand_dims(np.array(state), axis=0)
noise = np.random.normal(loc=0,
scale=self.noise_scale,
size=self.action_dim)
action = self.actor(state).numpy()[0] + noise
action = np.clip(action, -1, 1)
return action
def eval_action(self, state):
state = np.expand_dims(np.array(state), axis=0)
action = self.actor(state).numpy()[0]
action = np.clip(action, -1, 1)
return action
def train(self, training_num):
total_a_loss = 0
total_c_loss = 0
for i in range(training_num):
self.current_step += 1
s, a, r, ns, d = self.buffer.sample(self.batch_size)
value_next = tf.stop_gradient(
self.target_critic(ns, self.target_actor(ns)))
target_value = r + (1 - d) * self.gamma * value_next
with tf.GradientTape(persistent=True) as tape:
critic_loss = 0.5 * tf.reduce_mean(
tf.square(target_value - self.critic(s, a)))
actor_loss = -tf.reduce_mean(self.critic(s, self.actor(s)))
critic_grad = tape.gradient(critic_loss,
self.critic.trainable_variables)
self.critic_optimizer.apply_gradients(
zip(critic_grad, self.critic.trainable_variables))
actor_grad = tape.gradient(actor_loss,
self.actor.trainable_variables)
self.actor_optimizer.apply_gradients(
(zip(actor_grad, self.actor.trainable_variables)))
soft_update(self.actor, self.target_actor, self.tau)
soft_update(self.critic, self.target_critic, self.tau)
del tape
total_a_loss += actor_loss.numpy()
total_c_loss += critic_loss.numpy()
return [['Loss/Actor', total_a_loss], ['Loss/Critic', total_c_loss]]