def evaluate(
self,
evaluator: Evaluator,
logged_actions: Optional[np.ndarray],
logged_propensities: Optional[np.ndarray],
):
workspace.RunNet(self.all_q_score_model.net)
all_action_scores = workspace.FetchBlob(self.all_q_score_output)
maxq_action_idxs = workspace.FetchBlob(self.maxq_action_idxs)
model_values_on_logged_actions = np.sum(
(logged_actions * all_action_scores), axis=1, keepdims=True
)
model_propensities = Evaluator.softmax(all_action_scores, self.rl_temperature)
logged_rewards = workspace.FetchBlob("rewards")
cpe_stats = BatchStatsForCPE(
td_loss=workspace.FetchBlob(self.loss_blob),
logged_actions=logged_actions,
logged_propensities=logged_propensities,
logged_rewards=logged_rewards,
logged_values=None,
model_propensities=model_propensities,
model_rewards=None,
model_values=all_action_scores,
model_values_on_logged_actions=model_values_on_logged_actions,
model_action_idxs=maxq_action_idxs,
)
evaluator.report(cpe_stats)
def train(self, training_samples: TrainingDataPage, evaluator=None) -> None:
if self.minibatch == 0:
# Assume that the tensors are the right shape after the first minibatch
assert (
training_samples.states.shape[0] == self.minibatch_size
), "Invalid shape: " + str(training_samples.states.shape)
assert (
training_samples.actions.shape[0] == self.minibatch_size
), "Invalid shape: " + str(training_samples.actions.shape)
assert training_samples.rewards.shape == torch.Size(
[self.minibatch_size, 1]
), "Invalid shape: " + str(training_samples.rewards.shape)
assert (
training_samples.next_states.shape == training_samples.states.shape
), "Invalid shape: " + str(training_samples.next_states.shape)
assert (
training_samples.not_terminals.shape == training_samples.rewards.shape
), "Invalid shape: " + str(training_samples.not_terminals.shape)
if self.use_seq_num_diff_as_time_diff:
assert (
training_samples.time_diffs.shape == training_samples.rewards.shape
), "Invalid shape: " + str(training_samples.time_diffs.shape)
self.minibatch += 1
states = training_samples.states.detach().requires_grad_(True)
actions = training_samples.actions.detach().requires_grad_(True)
# As far as ddpg is concerned all actions are [-1, 1] due to actor tanh
actions = rescale_torch_tensor(
actions,
new_min=self.min_action_range_tensor_training,
new_max=self.max_action_range_tensor_training,
prev_min=self.min_action_range_tensor_serving,
prev_max=self.max_action_range_tensor_serving,
)
rewards = training_samples.rewards
next_states = training_samples.next_states
time_diffs = training_samples.time_diffs
discount_tensor = torch.tensor(np.full(rewards.shape, self.gamma)).type(
self.dtype
)
not_done_mask = training_samples.not_terminals
# Optimize the critic network subject to mean squared error:
# L = ([r + gamma * Q(s2, a2)] - Q(s1, a1)) ^ 2
q_s1_a1 = self.critic(torch.cat((states, actions), dim=1))
next_actions = self.actor_target(next_states)
next_state_actions = torch.cat((next_states, next_actions), dim=1)
q_s2_a2 = self.critic_target(next_state_actions)
filtered_q_s2_a2 = not_done_mask * q_s2_a2
if self.use_seq_num_diff_as_time_diff:
discount_tensor = discount_tensor.pow(time_diffs)
if self.minibatch < self.reward_burnin:
target_q_values = rewards
else:
target_q_values = rewards + (discount_tensor * filtered_q_s2_a2)
# compute loss and update the critic network
critic_predictions = q_s1_a1
loss_critic = self.q_network_loss(critic_predictions, target_q_values.detach())
loss_critic_for_eval = loss_critic.detach()
self.critic_optimizer.zero_grad()
loss_critic.backward()
self.critic_optimizer.step()
# Optimize the actor network subject to the following:
# max mean(Q(s1, a1)) or min -mean(Q(s1, a1))
loss_actor = -self.critic(torch.cat((states, self.actor(states)), dim=1)).mean()
self.actor_optimizer.zero_grad()
loss_actor.backward()
self.actor_optimizer.step()
if self.minibatch < self.reward_burnin:
# Reward burnin: force target network
self._soft_update(self.actor, self.actor_target, 1.0)
self._soft_update(self.critic, self.critic_target, 1.0)
else:
# Use the soft update rule to update both target networks
self._soft_update(self.actor, self.actor_target, self.tau)
self._soft_update(self.critic, self.critic_target, self.tau)
if evaluator is not None:
cpe_stats = BatchStatsForCPE(td_loss=loss_critic_for_eval.cpu().numpy())
evaluator.report(cpe_stats)
def train(self, training_batch, evaluator=None) -> None:
if hasattr(training_batch, "as_parametric_sarsa_training_batch"):
training_batch = training_batch.as_parametric_sarsa_training_batch(
)
learning_input = training_batch.training_input
self.minibatch += 1
reward = learning_input.reward
discount_tensor = torch.full_like(reward, self.gamma)
not_done_mask = learning_input.not_terminal
if self.use_seq_num_diff_as_time_diff:
# TODO: Implement this in another diff
raise NotImplementedError
if self.maxq_learning:
# Compute max a' Q(s', a') over all possible actions using target network
next_q_values = self.get_max_q_values(
learning_input.tiled_next_state,
learning_input.possible_next_actions,
self.double_q_learning,
)
else:
# SARSA
next_q_values = self.get_next_action_q_values(
learning_input.next_state, learning_input.next_action)
filtered_max_q_vals = next_q_values.reshape(-1, 1) * not_done_mask
if self.minibatch < self.reward_burnin:
target_q_values = reward
else:
target_q_values = reward + (discount_tensor * filtered_max_q_vals)
# Get Q-value of action taken
current_state_action = rlt.StateAction(state=learning_input.state,
action=learning_input.action)
q_values = self.q_network(current_state_action).q_value
self.all_action_scores = q_values.detach()
value_loss = self.q_network_loss(q_values, target_q_values)
self.loss = value_loss.detach()
self.q_network_optimizer.zero_grad()
value_loss.backward()
if self.gradient_handler:
self.gradient_handler(self.q_network.parameters())
self.q_network_optimizer.step()
# TODO: Maybe soft_update should belong to the target network
if self.minibatch < self.reward_burnin:
# Reward burnin: force target network
self._soft_update(self.q_network, self.q_network_target, 1.0)
else:
# Use the soft update rule to update target network
self._soft_update(self.q_network, self.q_network_target, self.tau)
# get reward estimates
reward_estimates = self.reward_network(current_state_action).q_value
reward_loss = F.mse_loss(reward_estimates, reward)
self.reward_network_optimizer.zero_grad()
reward_loss.backward()
self.reward_network_optimizer.step()
self.loss_reporter.report(td_loss=float(self.loss),
reward_loss=float(reward_loss))
if evaluator is not None:
cpe_stats = BatchStatsForCPE(
model_values_on_logged_actions=self.all_action_scores)
evaluator.report(cpe_stats)
def train(self,
training_samples: TrainingDataPage,
evaluator=None) -> None:
if self.minibatch == 0:
# Assume that the tensors are the right shape after the first minibatch
assert (training_samples.states.shape[0] == self.minibatch_size
), "Invalid shape: " + str(training_samples.states.shape)
assert (training_samples.actions.shape[0] == self.minibatch_size
), "Invalid shape: " + str(training_samples.actions.shape)
assert training_samples.rewards.shape == torch.Size(
[self.minibatch_size,
1]), "Invalid shape: " + str(training_samples.rewards.shape)
assert (training_samples.next_states.shape ==
training_samples.states.shape), "Invalid shape: " + str(
training_samples.next_states.shape)
assert (training_samples.not_terminals.shape ==
training_samples.rewards.shape), "Invalid shape: " + str(
training_samples.not_terminals.shape)
assert training_samples.possible_next_actions_state_concat.shape[
1] == (
training_samples.states.shape[1] +
training_samples.actions.shape[1]
), ("Invalid shape: " + str(
training_samples.possible_next_actions_state_concat.shape))
assert training_samples.possible_next_actions_lengths.shape == torch.Size(
[
self.minibatch_size
]), ("Invalid shape: " +
str(training_samples.possible_next_actions_lengths.shape))
self.minibatch += 1
states = training_samples.states.detach().requires_grad_(True)
actions = training_samples.actions
state_action_pairs = torch.cat((states, actions), dim=1)
rewards = training_samples.rewards
discount_tensor = torch.full(training_samples.time_diffs.shape,
self.gamma).type(self.dtype)
not_done_mask = training_samples.not_terminals
if self.use_seq_num_diff_as_time_diff:
discount_tensor = discount_tensor.pow(training_samples.time_diffs)
if self.maxq_learning:
# Compute max a' Q(s', a') over all possible actions using target network
next_q_values = self.get_max_q_values(
training_samples.possible_next_actions_state_concat,
training_samples.possible_next_actions_lengths,
self.double_q_learning,
)
else:
# SARSA
next_state_action_pairs = torch.cat(
(training_samples.next_states, training_samples.next_actions),
dim=1)
next_q_values = self.get_next_action_q_values(
next_state_action_pairs)
filtered_max_q_vals = next_q_values.reshape(-1, 1) * not_done_mask
if self.minibatch < self.reward_burnin:
target_q_values = rewards
else:
target_q_values = rewards + (discount_tensor * filtered_max_q_vals)
# Get Q-value of action taken
q_values = self.q_network(state_action_pairs)
all_action_scores = q_values.detach()
self.model_values_on_logged_actions = q_values.detach()
value_loss = self.q_network_loss(q_values, target_q_values)
self.loss = value_loss.detach()
self.q_network_optimizer.zero_grad()
value_loss.backward()
if self.gradient_handler:
self.gradient_handler(self.q_network.parameters())
self.q_network_optimizer.step()
if self.minibatch < self.reward_burnin:
# Reward burnin: force target network
self._soft_update(self.q_network, self.q_network_target, 1.0)
else:
# Use the soft update rule to update target network
self._soft_update(self.q_network, self.q_network_target, self.tau)
# get reward estimates
reward_estimates = self.reward_network(state_action_pairs)
reward_loss = F.mse_loss(reward_estimates, rewards)
self.reward_network_optimizer.zero_grad()
reward_loss.backward()
self.reward_network_optimizer.step()
self.loss_reporter.report(td_loss=float(self.loss),
reward_loss=float(reward_loss))
if evaluator is not None:
cpe_stats = BatchStatsForCPE(
model_values_on_logged_actions=all_action_scores)
evaluator.report(cpe_stats)
def train(self,
training_samples: TrainingDataPage,
evaluator: Optional[Evaluator] = None):
if self.minibatch == 0:
# Assume that the tensors are the right shape after the first minibatch
assert (training_samples.states.shape[0] == self.minibatch_size
), "Invalid shape: " + str(training_samples.states.shape)
assert training_samples.actions.shape == torch.Size([
self.minibatch_size, len(self._actions)
]), "Invalid shape: " + str(training_samples.actions.shape)
assert training_samples.rewards.shape == torch.Size(
[self.minibatch_size,
1]), "Invalid shape: " + str(training_samples.rewards.shape)
assert (training_samples.next_states.shape ==
training_samples.states.shape), "Invalid shape: " + str(
training_samples.next_states.shape)
assert (training_samples.not_terminals.shape ==
training_samples.rewards.shape), "Invalid shape: " + str(
training_samples.not_terminals.shape)
if training_samples.possible_next_actions is not None:
assert (
training_samples.possible_next_actions.shape ==
training_samples.actions.shape), "Invalid shape: " + str(
training_samples.possible_next_actions.shape)
if training_samples.propensities is not None:
assert (training_samples.propensities.shape == training_samples
.rewards.shape), "Invalid shape: " + str(
training_samples.propensities.shape)
# Apply reward boost if specified
reward_boosts = torch.sum(training_samples.actions.float() *
self.reward_boosts,
dim=1,
keepdim=True)
boosted_rewards = training_samples.rewards + reward_boosts
self.minibatch += 1
states = training_samples.states.detach().requires_grad_(True)
actions = training_samples.actions
rewards = boosted_rewards
next_states = training_samples.next_states
discount_tensor = torch.full(training_samples.time_diffs.shape,
self.gamma).type(self.dtype)
not_done_mask = training_samples.not_terminals
if self.use_seq_num_diff_as_time_diff:
discount_tensor = discount_tensor.pow(training_samples.time_diffs)
if self.maxq_learning:
# Compute max a' Q(s', a') over all possible actions using target network
possible_next_actions = training_samples.possible_next_actions
next_q_values = self.get_max_q_values(next_states,
possible_next_actions,
self.double_q_learning)
else:
# SARSA
next_actions = training_samples.next_actions
next_q_values = self.get_next_action_q_values(
next_states, next_actions)
filtered_next_q_vals = next_q_values * not_done_mask
if self.minibatch < self.reward_burnin:
target_q_values = rewards
else:
target_q_values = rewards + (discount_tensor *
filtered_next_q_vals)
# Get Q-value of action taken
all_q_values = self.q_network(states)
self.all_action_scores = all_q_values.detach()
q_values = torch.sum(all_q_values * actions, 1, keepdim=True)
loss = self.q_network_loss(q_values, target_q_values)
self.loss = loss.detach()
self.q_network_optimizer.zero_grad()
loss.backward()
if self.gradient_handler:
self.gradient_handler(self.q_network.parameters())
self.q_network_optimizer.step()
if self.minibatch < self.reward_burnin:
# Reward burnin: force target network
self._soft_update(self.q_network, self.q_network_target, 1.0)
else:
# Use the soft update rule to update target network
self._soft_update(self.q_network, self.q_network_target, self.tau)
# get reward estimates
reward_estimates = self.reward_network(states)
self.reward_estimates = reward_estimates.detach()
reward_estimates_for_logged_actions = reward_estimates.gather(
1, actions.argmax(dim=1, keepdim=True))
reward_loss = F.mse_loss(reward_estimates_for_logged_actions, rewards)
self.reward_network_optimizer.zero_grad()
reward_loss.backward()
self.reward_network_optimizer.step()
self.loss_reporter.report(td_loss=float(self.loss),
reward_loss=float(reward_loss))
training_metadata = {}
if evaluator is not None:
model_propensities = torch.from_numpy(
Evaluator.softmax(self.all_action_scores.cpu().numpy(),
self.rl_temperature))
cpe_stats = BatchStatsForCPE(
logged_actions=training_samples.actions,
logged_propensities=training_samples.propensities,
logged_rewards=rewards,
logged_values=None, # Compute at end of each epoch for CPE
model_propensities=model_propensities,
model_rewards=self.reward_estimates,
model_values=self.all_action_scores,
model_values_on_logged_actions=
None, # Compute at end of each epoch for CPE
model_action_idxs=self.all_action_scores.argmax(dim=1,
keepdim=True),
)
evaluator.report(cpe_stats)
training_metadata["model_rewards"] = self.reward_estimates.cpu(
).numpy()
return training_metadata
def train(self, training_batch, evaluator=None) -> None:
"""
IMPORTANT: the input action here is assumed to be preprocessed to match the
range of the output of the actor.
"""
if hasattr(training_batch, "as_parametric_sarsa_training_batch"):
training_batch = training_batch.as_parametric_sarsa_training_batch(
)
learning_input = training_batch.training_input
self.minibatch += 1
state = learning_input.state
action = learning_input.action
reward = learning_input.reward
discount = torch.full_like(reward, self.gamma)
not_done_mask = learning_input.not_terminal
current_state_action = rlt.StateAction(state=state, action=action)
q1_value = self.q1_network(current_state_action).q_value
min_q_value = q1_value
if self.q2_network:
q2_value = self.q2_network(current_state_action).q_value
min_q_value = torch.min(q1_value, q2_value)
# Use the minimum as target, ensure no gradient going through
min_q_value = min_q_value.detach()
#
# First, optimize value network; minimizing MSE between
# V(s) & Q(s, a) - log(pi(a|s))
#
state_value = self.value_network(state.float_features) # .q_value
with torch.no_grad():
log_prob_a = self.actor_network.get_log_prob(
state, action.float_features)
target_value = min_q_value - self.entropy_temperature * log_prob_a
value_loss = F.mse_loss(state_value, target_value)
self.value_network_optimizer.zero_grad()
value_loss.backward()
self.value_network_optimizer.step()
#
# Second, optimize Q networks; minimizing MSE between
# Q(s, a) & r + discount * V'(next_s)
#
with torch.no_grad():
next_state_value = (self.value_network_target(
learning_input.next_state.float_features) * not_done_mask)
if self.minibatch < self.reward_burnin:
target_q_value = reward
else:
target_q_value = reward + discount * next_state_value
q1_loss = F.mse_loss(q1_value, target_q_value)
self.q1_network_optimizer.zero_grad()
q1_loss.backward()
self.q1_network_optimizer.step()
if self.q2_network:
q2_loss = F.mse_loss(q2_value, target_q_value)
self.q2_network_optimizer.zero_grad()
q2_loss.backward()
self.q2_network_optimizer.step()
#
# Lastly, optimize the actor; minimizing KL-divergence between action propensity
# & softmax of value. Due to reparameterization trick, it ends up being
# log_prob(actor_action) - Q(s, actor_action)
#
actor_output = self.actor_network(rlt.StateInput(state=state))
state_actor_action = rlt.StateAction(
state=state,
action=rlt.FeatureVector(float_features=actor_output.action))
q1_actor_value = self.q1_network(state_actor_action).q_value
min_q_actor_value = q1_actor_value
if self.q2_network:
q2_actor_value = self.q2_network(state_actor_action).q_value
min_q_actor_value = torch.min(q1_actor_value, q2_actor_value)
actor_loss = torch.mean(self.entropy_temperature *
actor_output.log_prob - min_q_actor_value)
self.actor_network_optimizer.zero_grad()
actor_loss.backward()
self.actor_network_optimizer.step()
if self.minibatch < self.reward_burnin:
# Reward burnin: force target network
self._soft_update(self.value_network, self.value_network_target,
1.0)
else:
# Use the soft update rule to update both target networks
self._soft_update(self.value_network, self.value_network_target,
self.tau)
if evaluator is not None:
cpe_stats = BatchStatsForCPE(
td_loss=q1_loss.detach().cpu().numpy(),
model_values_on_logged_actions=q1_value.detach().cpu().numpy(),
)
evaluator.report(cpe_stats)
def train(self, training_batch, evaluator=None) -> None:
"""
IMPORTANT: the input action here is assumed to be preprocessed to match the
range of the output of the actor.
"""
if hasattr(training_batch, "as_parametric_sarsa_training_batch"):
training_batch = training_batch.as_parametric_sarsa_training_batch(
)
learning_input = training_batch.training_input
self.minibatch += 1
state = learning_input.state
action = learning_input.action
reward = learning_input.reward
discount = torch.full_like(reward, self.gamma)
not_done_mask = learning_input.not_terminal
if self._should_scale_action_in_train():
action = rlt.FeatureVector(
rescale_torch_tensor(
action.float_features,
new_min=self.min_action_range_tensor_training,
new_max=self.max_action_range_tensor_training,
prev_min=self.min_action_range_tensor_serving,
prev_max=self.max_action_range_tensor_serving,
))
current_state_action = rlt.StateAction(state=state, action=action)
q1_value = self.q1_network(current_state_action).q_value
min_q_value = q1_value
if self.q2_network:
q2_value = self.q2_network(current_state_action).q_value
min_q_value = torch.min(q1_value, q2_value)
# Use the minimum as target, ensure no gradient going through
min_q_value = min_q_value.detach()
#
# First, optimize value network; minimizing MSE between
# V(s) & Q(s, a) - log(pi(a|s))
#
state_value = self.value_network(state.float_features) # .q_value
if self.logged_action_uniform_prior:
log_prob_a = torch.zeros_like(min_q_value)
target_value = min_q_value
else:
with torch.no_grad():
log_prob_a = self.actor_network.get_log_prob(
state, action.float_features)
log_prob_a = log_prob_a.clamp(-20.0, 20.0)
target_value = min_q_value - self.entropy_temperature * log_prob_a
value_loss = F.mse_loss(state_value, target_value)
self.value_network_optimizer.zero_grad()
value_loss.backward()
self.value_network_optimizer.step()
#
# Second, optimize Q networks; minimizing MSE between
# Q(s, a) & r + discount * V'(next_s)
#
with torch.no_grad():
next_state_value = (self.value_network_target(
learning_input.next_state.float_features) * not_done_mask)
if self.minibatch < self.reward_burnin:
target_q_value = reward
else:
target_q_value = reward + discount * next_state_value
q1_loss = F.mse_loss(q1_value, target_q_value)
self.q1_network_optimizer.zero_grad()
q1_loss.backward()
self.q1_network_optimizer.step()
if self.q2_network:
q2_loss = F.mse_loss(q2_value, target_q_value)
self.q2_network_optimizer.zero_grad()
q2_loss.backward()
self.q2_network_optimizer.step()
#
# Lastly, optimize the actor; minimizing KL-divergence between action propensity
# & softmax of value. Due to reparameterization trick, it ends up being
# log_prob(actor_action) - Q(s, actor_action)
#
actor_output = self.actor_network(rlt.StateInput(state=state))
state_actor_action = rlt.StateAction(
state=state,
action=rlt.FeatureVector(float_features=actor_output.action))
q1_actor_value = self.q1_network(state_actor_action).q_value
min_q_actor_value = q1_actor_value
if self.q2_network:
q2_actor_value = self.q2_network(state_actor_action).q_value
min_q_actor_value = torch.min(q1_actor_value, q2_actor_value)
actor_loss = (self.entropy_temperature * actor_output.log_prob -
min_q_actor_value)
# Do this in 2 steps so we can log histogram of actor loss
actor_loss_mean = actor_loss.mean()
self.actor_network_optimizer.zero_grad()
actor_loss_mean.backward()
self.actor_network_optimizer.step()
if self.minibatch < self.reward_burnin:
# Reward burnin: force target network
self._soft_update(self.value_network, self.value_network_target,
1.0)
else:
# Use the soft update rule to update both target networks
self._soft_update(self.value_network, self.value_network_target,
self.tau)
# Logging at the end to schedule all the cuda operations first
if (self.tensorboard_logging_freq is not None
and self.minibatch % self.tensorboard_logging_freq == 0):
SummaryWriterContext.add_histogram("q1/logged_state_value",
q1_value)
if self.q2_network:
SummaryWriterContext.add_histogram("q2/logged_state_value",
q2_value)
SummaryWriterContext.add_histogram("log_prob_a", log_prob_a)
SummaryWriterContext.add_histogram("value_network/target",
target_value)
SummaryWriterContext.add_histogram("q_network/next_state_value",
next_state_value)
SummaryWriterContext.add_histogram("q_network/target_q_value",
target_q_value)
SummaryWriterContext.add_histogram("actor/min_q_actor_value",
min_q_actor_value)
SummaryWriterContext.add_histogram("actor/action_log_prob",
actor_output.log_prob)
SummaryWriterContext.add_histogram("actor/loss", actor_loss)
if evaluator is not None:
cpe_stats = BatchStatsForCPE(
td_loss=q1_loss.detach().cpu().numpy(),
logged_rewards=reward.detach().cpu().numpy(),
model_values_on_logged_actions=q1_value.detach().cpu().numpy(),
model_propensities=actor_output.log_prob.exp().detach().cpu().
numpy(),
model_values=min_q_actor_value.detach().cpu().numpy(),
)
evaluator.report(cpe_stats)