def kl(self, other):
"""
Args:
other: object of CategoricalDistribution
Returns:
kl: A float32 tensor with shape [BATCH_SIZE]
"""
assert isinstance(other, CategoricalDistribution)
logits = self.logits - layers.reduce_max(self.logits, dim=1)
other_logits = other.logits - layers.reduce_max(other.logits, dim=1)
e_logits = layers.exp(logits)
other_e_logits = layers.exp(other_logits)
z = layers.reduce_sum(e_logits, dim=1)
other_z = layers.reduce_sum(other_e_logits, dim=1)
prob = e_logits / z
kl = layers.reduce_sum(
prob *
(logits - layers.log(z) - other_logits + layers.log(other_z)),
dim=1)
return kl
def learn(self, obs, action, reward, next_obs, terminal, sample_weight):
# print("obs:",obs)
# raise NotImplementedError
# obs = layers.squeeze(input=obs,axes=[-1])
pred_value = self.model.value(obs)
action_onehot = layers.one_hot(action, self.act_dim)
pred_action_value = layers.reduce_sum(action_onehot * pred_value,
dim=1)
# calculate the target q value
next_action_value = self.model.value(next_obs)
greedy_action = layers.argmax(next_action_value, axis=-1)
greedy_action = layers.unsqueeze(greedy_action, axes=[1])
greedy_action_onehot = layers.one_hot(greedy_action, self.act_dim)
next_pred_value = self.target_model.value(next_obs)
max_v = layers.reduce_sum(greedy_action_onehot * next_pred_value,
dim=1)
max_v.stop_gradient = True
target = reward + (
1.0 - layers.cast(terminal, dtype='float32')) * self.gamma * max_v
delta = layers.abs(target - pred_action_value)
cost = sample_weight * layers.square_error_cost(
pred_action_value, target)
cost = layers.reduce_mean(cost)
optimizer = fluid.optimizer.Adam(learning_rate=self.lr, epsilon=1e-3)
optimizer.minimize(cost)
return cost, delta
def logp(self, actions, eps=1e-6):
"""
Args:
actions: An int64 tensor with shape [BATCH_SIZE]
eps: A small float constant that avoids underflows when computing the log probability
Returns:
actions_log_prob: A float32 tensor with shape [BATCH_SIZE]
"""
assert len(actions.shape) == 1
logits = self.logits - layers.reduce_max(self.logits, dim=1)
e_logits = layers.exp(logits)
z = layers.reduce_sum(e_logits, dim=1)
prob = e_logits / z
actions = layers.unsqueeze(actions, axes=[1])
actions_onehot = layers.one_hot(actions, prob.shape[1])
actions_onehot = layers.cast(actions_onehot, dtype='float32')
actions_prob = layers.reduce_sum(prob * actions_onehot, dim=1)
actions_prob = actions_prob + eps
actions_log_prob = layers.log(actions_prob)
return actions_log_prob
def __init__(self,
behaviour_actions_log_probs,
target_actions_log_probs,
policy_entropy,
dones,
discount,
rewards,
values,
bootstrap_value,
entropy_coeff=-0.01,
vf_loss_coeff=0.5,
clip_rho_threshold=1.0,
clip_pg_rho_threshold=1.0):
"""Policy gradient loss with vtrace importance weighting.
VTraceLoss takes tensors of shape [T, B, ...], where `B` is the
batch_size. The reason we need to know `B` is for V-trace to properly
handle episode cut boundaries.
Args:
behaviour_actions_log_probs: A float32 tensor of shape [T, B].
target_actions_log_probs: A float32 tensor of shape [T, B].
policy_entropy: A float32 tensor of shape [T, B].
dones: A float32 tensor of shape [T, B].
discount: A float32 scalar.
rewards: A float32 tensor of shape [T, B].
values: A float32 tensor of shape [T, B].
bootstrap_value: A float32 tensor of shape [B].
"""
self.vtrace_returns = from_importance_weights(
behaviour_actions_log_probs=behaviour_actions_log_probs,
target_actions_log_probs=target_actions_log_probs,
discounts=inverse(dones) * discount,
rewards=rewards,
values=values,
bootstrap_value=bootstrap_value,
clip_rho_threshold=clip_rho_threshold,
clip_pg_rho_threshold=clip_pg_rho_threshold)
# The policy gradients loss
self.pi_loss = -1.0 * layers.reduce_sum(
target_actions_log_probs * self.vtrace_returns.pg_advantages)
# The baseline loss
delta = values - self.vtrace_returns.vs
self.vf_loss = 0.5 * layers.reduce_sum(layers.square(delta))
# The entropy loss (We want to maximize entropy, so entropy_ceoff < 0)
self.entropy = layers.reduce_sum(policy_entropy)
# The summed weighted loss
self.total_loss = (self.pi_loss + self.vf_loss * vf_loss_coeff +
self.entropy * entropy_coeff)
def entropy(self):
"""
Returns:
entropy: A float32 tensor with shape [BATCH_SIZE] of entropy of self policy distribution.
"""
logits = self.logits - layers.reduce_max(self.logits, dim=1)
e_logits = layers.exp(logits)
z = layers.reduce_sum(e_logits, dim=1)
prob = e_logits / z
entropy = -1.0 * layers.reduce_sum(prob * (logits - layers.log(z)),
dim=1)
return entropy
def learn(self,
obs,
action,
reward,
next_obs,
terminal,
learning_rate=None):
""" update value model self.model with DQN algorithm
"""
# Support the modification of learning_rate
if learning_rate is None:
assert isinstance(
self.lr,
float), "Please set the learning rate of DQN in initializaion."
learning_rate = self.lr
pred_value = self.model.value(obs)
next_pred_value = self.target_model.value(next_obs)
best_v = layers.reduce_max(next_pred_value, dim=1)
best_v.stop_gradient = True
target = reward + (
1.0 - layers.cast(terminal, dtype='float32')) * self.gamma * best_v
action_onehot = layers.one_hot(action, self.act_dim)
action_onehot = layers.cast(action_onehot, dtype='float32')
pred_action_value = layers.reduce_sum(
layers.elementwise_mul(action_onehot, pred_value), dim=1)
cost = layers.square_error_cost(pred_action_value, target)
cost = layers.reduce_mean(cost)
optimizer = fluid.optimizer.Adam(
learning_rate=learning_rate, epsilon=1e-3)
optimizer.minimize(cost)
return cost
def _calc_kl(self, means, logvars, old_means, old_logvars):
""" Calculate KL divergence between old and new distributions
See: https://en.wikipedia.org/wiki/Multivariate_normal_distribution#Kullback.E2.80.93Leibler_divergence
Args:
means: shape (batch_size, act_dim)
logvars: shape (act_dim)
old_means: shape (batch_size, act_dim)
old_logvars: shape (act_dim)
Returns:
kl: shape (batch_size)
"""
log_det_cov_old = layers.reduce_sum(old_logvars)
log_det_cov_new = layers.reduce_sum(logvars)
tr_old_new = layers.reduce_sum(layers.exp(old_logvars - logvars))
kl = 0.5 * (layers.reduce_sum(
layers.square(means - old_means) / layers.exp(logvars), dim=1) +
(log_det_cov_new - log_det_cov_old) + tr_old_new -
self.act_dim)
return kl
def _calc_logprob(self, actions, means, logvars):
""" Calculate log probabilities of actions, when given means and logvars
of normal distribution.
The constant sqrt(2 * pi) is omitted, which will be eliminated in later.
Args:
actions: shape (batch_size, act_dim)
means: shape (batch_size, act_dim)
logvars: shape (act_dim)
Returns:
logprob: shape (batch_size)
"""
exp_item = layers.elementwise_div(layers.square(actions - means),
layers.exp(logvars),
axis=1)
exp_item = -0.5 * layers.reduce_sum(exp_item, dim=1)
vars_item = -0.5 * layers.reduce_sum(logvars)
logprob = exp_item + vars_item
return logprob
def sample(self, obs):
mean, log_std = self.actor.policy(obs)
std = layers.exp(log_std)
normal = Normal(mean, std)
x_t = normal.sample([1])[0]
y_t = layers.tanh(x_t)
action = y_t * self.max_action
log_prob = normal.log_prob(x_t)
log_prob -= layers.log(self.max_action * (1 - layers.pow(y_t, 2)) +
epsilon)
log_prob = layers.reduce_sum(log_prob, dim=1, keep_dim=True)
log_prob = layers.squeeze(log_prob, axes=[1])
return action, log_prob
def learn(self, obs, actions, advantages, target_values, learning_rate,
entropy_coeff):
"""
Args:
obs: An float32 tensor of shape ([B] + observation_space).
E.g. [B, C, H, W] in atari.
actions: An int64 tensor of shape [B].
advantages: A float32 tensor of shape [B].
target_values: A float32 tensor of shape [B].
learning_rate: float scalar of learning rate.
entropy_coeff: float scalar of entropy coefficient.
"""
logits = self.model.policy(obs)
policy_distribution = CategoricalDistribution(logits)
actions_log_probs = policy_distribution.logp(actions)
# The policy gradient loss
pi_loss = -1.0 * layers.reduce_sum(actions_log_probs * advantages)
# The value function loss
values = self.model.value(obs)
delta = values - target_values
vf_loss = 0.5 * layers.reduce_sum(layers.square(delta))
# The entropy loss (We want to maximize entropy, so entropy_ceoff < 0)
policy_entropy = policy_distribution.entropy()
entropy = layers.reduce_sum(policy_entropy)
total_loss = (pi_loss + vf_loss * self.vf_loss_coeff +
entropy * entropy_coeff)
fluid.clip.set_gradient_clip(clip=fluid.clip.GradientClipByGlobalNorm(
clip_norm=40.0))
optimizer = fluid.optimizer.AdamOptimizer(learning_rate)
optimizer.minimize(total_loss)
return total_loss, pi_loss, vf_loss, entropy
def cal_bellman_residual(self, obs, action, reward, next_obs, terminal):
""" use self.model to get squared Bellman residual with fed data
"""
pred_value = self.model.value(obs)
next_pred_value = self.target_model.value(next_obs)
best_v = layers.reduce_max(next_pred_value, dim=1)
best_v.stop_gradient = True
target = reward + (
1.0 - layers.cast(terminal, dtype='float32')) * self.gamma * best_v
action_onehot = layers.one_hot(action, self.act_dim)
action_onehot = layers.cast(action_onehot, dtype='float32')
pred_action_value = layers.reduce_sum(
layers.elementwise_mul(action_onehot, pred_value), dim=1)
cost = layers.square_error_cost(pred_action_value, target)
cost = layers.reduce_mean(cost)
return cost
def learn(self, obs, action, reward, next_obs, terminal, sample_weight):
""" update value model self.model with DQN algorithm
"""
pred_value = self.model.value(obs)
next_pred_value = self.target_model.value(next_obs)
best_v = layers.reduce_max(next_pred_value, dim=1)
best_v.stop_gradient = True
target = reward + (
1.0 - layers.cast(terminal, dtype='float32')) * self.gamma * best_v
action_onehot = layers.one_hot(action, self.act_dim)
action_onehot = layers.cast(action_onehot, dtype='float32')
pred_action_value = layers.reduce_sum(action_onehot * pred_value,
dim=1)
delta = layers.abs(target - pred_action_value)
cost = sample_weight * layers.square_error_cost(
pred_action_value, target)
cost = layers.reduce_mean(cost)
optimizer = fluid.optimizer.Adam(learning_rate=self.lr, epsilon=1e-3)
optimizer.minimize(cost)
return cost, delta # `delta` is the TD-error
def learn(self, obs, actions, means, log_std, rewards, dones,
learning_rate, entropy_coeff):
"""
Args:
obs: An float32 tensor of shape ([B] + observation_space).
E.g. [B, C, H, W] in atari.
actions: An int64 tensor of shape [B].
behaviour_logits: A float32 tensor of shape [B, NUM_ACTIONS].
rewards: A float32 tensor of shape [B].
dones: A float32 tensor of shape [B].
learning_rate: float scalar of learning rate.
entropy_coeff: float scalar of entropy coefficient.
"""
values = self.model.value(obs)
# pi
log_std = layers.exp(log_std)
normal_pi = Normal(means, log_std)
# x_t1 = normal_pi.sample([1])[0]
# x_t1.stop_gradient = True
y_t1 = actions / self.max_action
# action1 = y_t1 * self.max_action
log_prob1 = normal_pi.log_prob(actions)
log_prob1 -= layers.log(self.max_action * (1 - layers.pow(y_t1, 2)) +
epsilon)
log_prob1 = layers.reduce_sum(log_prob1, dim=1, keep_dim=True)
log_prob_pi = layers.squeeze(log_prob1, axes=[1])
# mu
actions_mu, log_std_mu = self.model.policy(obs)
log_std_mu = layers.exp(log_std_mu)
normal_mu = Normal(actions_mu, log_std_mu)
# x_t2 = normal_mu.sample([1])[0]
# x_t2.stop_gradient = True
# y_t2 = actions
# action2 = y_t2 * self.max_action
log_prob2 = normal_mu.log_prob(actions)
log_prob2 -= layers.log(self.max_action * (1 - layers.pow(y_t1, 2)) +
epsilon)
log_prob2 = layers.reduce_sum(log_prob2, dim=1, keep_dim=True)
log_prob_mu = layers.squeeze(log_prob2, axes=[1])
# target_policy_distribution = CategoricalDistribution(target_logits)
# behaviour_policy_distribution = CategoricalDistribution(
# behaviour_logits)
policy_entropy = normal_mu.entropy()
# policy_entropy = layers.reduce_mean(policy_entropy, dim=1)
target_actions_log_probs = log_prob_mu
behaviour_actions_log_probs = log_prob_pi
# Calculating kl for debug
# kl = target_policy_distribution.kl(behaviour_policy_distribution)
kl = normal_mu.kl_divergence(normal_pi)
kl = layers.reduce_mean(kl, dim=1)
# kl = layers.unsqueeze(kl, axes=[1])
"""
Split the tensor into batches at known episode cut boundaries.
[B * T] -> [T, B]
"""
T = self.sample_batch_steps
def split_batches(tensor):
B = tensor.shape[0] // T
splited_tensor = layers.reshape(tensor,
[B, T] + list(tensor.shape[1:]))
# transpose B and T
return layers.transpose(splited_tensor, [1, 0] +
list(range(2, 1 + len(tensor.shape))))
behaviour_actions_log_probs = split_batches(
behaviour_actions_log_probs)
target_actions_log_probs = split_batches(target_actions_log_probs)
policy_entropy = split_batches(policy_entropy)
dones = split_batches(dones)
rewards = split_batches(rewards)
values = split_batches(values)
# [T, B] -> [T - 1, B] for V-trace calc.
behaviour_actions_log_probs = layers.slice(behaviour_actions_log_probs,
axes=[0],
starts=[0],
ends=[-1])
target_actions_log_probs = layers.slice(target_actions_log_probs,
axes=[0],
starts=[0],
ends=[-1])
policy_entropy = layers.slice(policy_entropy,
axes=[0],
starts=[0],
ends=[-1])
dones = layers.slice(dones, axes=[0], starts=[0], ends=[-1])
rewards = layers.slice(rewards, axes=[0], starts=[0], ends=[-1])
bootstrap_value = layers.slice(values,
axes=[0],
starts=[T - 1],
ends=[T])
values = layers.slice(values, axes=[0], starts=[0], ends=[-1])
bootstrap_value = layers.squeeze(bootstrap_value, axes=[0])
vtrace_loss = VTraceLoss(
behaviour_actions_log_probs=behaviour_actions_log_probs,
target_actions_log_probs=target_actions_log_probs,
policy_entropy=policy_entropy,
dones=dones,
discount=self.gamma,
rewards=rewards,
values=values,
bootstrap_value=bootstrap_value,
entropy_coeff=entropy_coeff,
vf_loss_coeff=self.vf_loss_coeff,
clip_rho_threshold=self.clip_rho_threshold,
clip_pg_rho_threshold=self.clip_pg_rho_threshold)
fluid.clip.set_gradient_clip(clip=fluid.clip.GradientClipByGlobalNorm(
clip_norm=40.0))
optimizer = fluid.optimizer.AdamOptimizer(learning_rate)
optimizer.minimize(vtrace_loss.total_loss)
return vtrace_loss, kl
