def test_rescale_torch_tensor(self):
rows, cols = 3, 5
original_tensor = torch.randint(low=10, high=40,
size=(rows, cols)).float()
prev_max_tensor = torch.ones(1, 5) * 40.0
prev_min_tensor = torch.ones(1, 5) * 10.0
new_min_tensor = torch.ones(1, 5) * -1.0
new_max_tensor = torch.ones(1, 5).float()
print("Original tensor: ", original_tensor)
rescaled_tensor = rescale_torch_tensor(
original_tensor,
new_min_tensor,
new_max_tensor,
prev_min_tensor,
prev_max_tensor,
)
print("Rescaled tensor: ", rescaled_tensor)
reconstructed_original_tensor = rescale_torch_tensor(
rescaled_tensor,
prev_min_tensor,
prev_max_tensor,
new_min_tensor,
new_max_tensor,
)
print("Reconstructed Original tensor: ", reconstructed_original_tensor)
comparison_tensor = torch.eq(original_tensor,
reconstructed_original_tensor)
self.assertTrue(torch.sum(comparison_tensor), rows * cols)
def internal_prediction(self, states, test=False):
""" Returns list of actions output from actor network
:param states states as list of states to produce actions for
"""
self.actor_network.eval()
with torch.no_grad():
actions = self.actor_network(
rlt.PreprocessedState.from_tensor(states)).action
if not test:
if self.minibatch < self.initial_exploration_ts:
actions = (torch.rand_like(actions) *
(self.max_action_range_tensor_training -
self.min_action_range_tensor_training) +
self.min_action_range_tensor_training)
else:
actions += torch.randn_like(actions) * self.exploration_noise
# clamp actions to make sure actions are in the range
clamped_actions = torch.max(
torch.min(actions, self.max_action_range_tensor_training),
self.min_action_range_tensor_training,
)
rescaled_actions = rescale_torch_tensor(
clamped_actions,
new_min=self.min_action_range_tensor_serving,
new_max=self.max_action_range_tensor_serving,
prev_min=self.min_action_range_tensor_training,
prev_max=self.max_action_range_tensor_training,
)
self.actor_network.train()
return rescaled_actions
def _maybe_scale_action_in_train(self, action):
if (self.min_action_range_tensor_training is not None
and self.max_action_range_tensor_training is not None
and self.min_action_range_tensor_serving is not None
and self.max_action_range_tensor_serving is not None):
action = rescale_torch_tensor(
action,
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,
)
return action
def sample_action(self, scores: GaussianSamplerScore) -> rlt.ActorOutput:
self.actor_network.eval()
action, log_prob = self._sample_action(scores.loc, scores.scale_log)
# clamp actions to make sure actions are in the range
clamped_actions = torch.max(
torch.min(action, self.max_training_action),
self.min_training_action)
rescaled_actions = rescale_torch_tensor(
clamped_actions,
new_min=self.min_serving_action,
new_max=self.max_serving_action,
prev_min=self.min_training_action,
prev_max=self.max_training_action,
)
self.actor_network.train()
action = rescaled_actions.squeeze(0)
log_prob = torch.tensor(log_prob.item())
return rlt.ActorOutput(action=action, log_prob=log_prob)
def internal_prediction(self, states, test=False):
""" Returns list of actions output from actor network
:param states states as list of states to produce actions for
"""
self.actor_network.eval()
with torch.no_grad():
actions = self.actor_network(rlt.PreprocessedState.from_tensor(states))
# clamp actions to make sure actions are in the range
clamped_actions = torch.max(
torch.min(actions.action, self.max_action_range_tensor_training),
self.min_action_range_tensor_training,
)
rescaled_actions = rescale_torch_tensor(
clamped_actions,
new_min=self.min_action_range_tensor_serving,
new_max=self.max_action_range_tensor_serving,
prev_min=self.min_action_range_tensor_training,
prev_max=self.max_action_range_tensor_training,
)
self.actor_network.train()
return rescaled_actions
def train(self, training_batch) -> 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_policy_network_training_batch"):
training_batch = training_batch.as_policy_network_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 = action._replace(
float_features=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,
)
)
with torch.enable_grad():
#
# First, optimize Q networks; minimizing MSE between
# Q(s, a) & r + discount * V'(next_s)
#
current_state_action = rlt.PreprocessedStateAction(
state=state, action=action
)
q1_value = self.q1_network(current_state_action).q_value
if self.q2_network:
q2_value = self.q2_network(current_state_action).q_value
actor_output = self.actor_network(rlt.PreprocessedState(state=state))
# Optimize Alpha
if self.alpha_optimizer is not None:
alpha_loss = -(
self.log_alpha
* (actor_output.log_prob + self.target_entropy).detach()
).mean()
self.alpha_optimizer.zero_grad()
alpha_loss.backward()
self.alpha_optimizer.step()
self.entropy_temperature = self.log_alpha.exp()
with torch.no_grad():
if self.value_network is not None:
next_state_value = self.value_network_target(
learning_input.next_state.float_features
)
else:
next_state_actor_output = self.actor_network(
rlt.PreprocessedState(state=learning_input.next_state)
)
next_state_actor_action = rlt.PreprocessedStateAction(
state=learning_input.next_state,
action=rlt.PreprocessedFeatureVector(
float_features=next_state_actor_output.action
),
)
next_state_value = self.q1_network_target(
next_state_actor_action
).q_value
if self.q2_network is not None:
target_q2_value = self.q2_network_target(
next_state_actor_action
).q_value
next_state_value = torch.min(next_state_value, target_q2_value)
log_prob_a = self.actor_network.get_log_prob(
learning_input.next_state, next_state_actor_output.action
)
log_prob_a = log_prob_a.clamp(-20.0, 20.0)
next_state_value -= self.entropy_temperature * log_prob_a
target_q_value = (
reward + discount * next_state_value * not_done_mask.float()
)
q1_loss = F.mse_loss(q1_value, target_q_value)
q1_loss.backward()
self._maybe_run_optimizer(
self.q1_network_optimizer, self.minibatches_per_step
)
if self.q2_network:
q2_loss = F.mse_loss(q2_value, target_q_value)
q2_loss.backward()
self._maybe_run_optimizer(
self.q2_network_optimizer, self.minibatches_per_step
)
#
# Second, 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)
#
state_actor_action = rlt.PreprocessedStateAction(
state=state,
action=rlt.PreprocessedFeatureVector(
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()
actor_loss_mean.backward()
self._maybe_run_optimizer(
self.actor_network_optimizer, self.minibatches_per_step
)
#
# Lastly, if applicable, optimize value network; minimizing MSE between
# V(s) & E_a~pi(s) [ Q(s,a) - log(pi(a|s)) ]
#
if self.value_network is not None:
state_value = self.value_network(state.float_features)
if self.logged_action_uniform_prior:
log_prob_a = torch.zeros_like(min_q_actor_value)
target_value = min_q_actor_value
else:
with torch.no_grad():
log_prob_a = actor_output.log_prob
log_prob_a = log_prob_a.clamp(-20.0, 20.0)
target_value = (
min_q_actor_value - self.entropy_temperature * log_prob_a
)
value_loss = F.mse_loss(state_value, target_value.detach())
value_loss.backward()
self._maybe_run_optimizer(
self.value_network_optimizer, self.minibatches_per_step
)
# Use the soft update rule to update the target networks
if self.value_network is not None:
self._maybe_soft_update(
self.value_network,
self.value_network_target,
self.tau,
self.minibatches_per_step,
)
else:
self._maybe_soft_update(
self.q1_network,
self.q1_network_target,
self.tau,
self.minibatches_per_step,
)
if self.q2_network is not None:
self._maybe_soft_update(
self.q2_network,
self.q2_network_target,
self.tau,
self.minibatches_per_step,
)
# 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)
if self.value_network:
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)
self.loss_reporter.report(
td_loss=float(q1_loss),
reward_loss=None,
logged_rewards=reward,
model_values_on_logged_actions=q1_value,
model_propensities=actor_output.log_prob.exp(),
model_values=min_q_actor_value,
)
