def test_masked_softmax(self):
# Postive value case
x = torch.tensor([[15.0, 6.0, 9.0], [3.0, 2.0, 1.0]])
temperature = 1
mask = torch.tensor([[1.0, 0.0, 1.0], [0.0, 1.0, 1.0]])
out = masked_softmax(x, mask, temperature)
expected_out = torch.tensor([[0.9975, 0.0000, 0.0025],
[0, 0.7311, 0.2689]])
npt.assert_array_almost_equal(out, expected_out, 4)
# Postive value case (masked value goes to inf)
x = torch.tensor([[150.0, 2.0]])
temperature = 0.01
mask = torch.tensor([[0.0, 1.0]])
out = masked_softmax(x, mask, temperature)
expected_out = torch.tensor([[0.0, 1.0]])
npt.assert_array_almost_equal(out, expected_out, 4)
# Negative value case
x = torch.tensor([[-10.0, -1.0, -5.0]])
temperature = 0.01
mask = torch.tensor([[1.0, 1.0, 0.0]])
out = masked_softmax(x, mask, temperature)
expected_out = torch.tensor([[0.0, 1.0, 0.0]])
npt.assert_array_almost_equal(out, expected_out, 4)
# All values in a row are masked case
x = torch.tensor([[-5.0, 4.0, 3.0], [2.0, 1.0, 2.0]])
temperature = 1
mask = torch.tensor([[0.0, 0.0, 0.0], [1.0, 1.0, 1.0]])
out = masked_softmax(x, mask, temperature)
expected_out = torch.tensor([[0.0, 0.0, 0.0], [0.4223, 0.1554,
0.4223]])
npt.assert_array_almost_equal(out, expected_out, 4)
def policy_given_q_values(
q_scores: torch.Tensor,
action_names: List[str],
softmax_temperature: float,
possible_actions_presence: Optional[torch.Tensor] = None,
) -> DqnPolicyActionSet:
assert q_scores.shape[0] == 1 and len(q_scores.shape) == 2
if possible_actions_presence is None:
possible_actions_presence = torch.ones_like(q_scores)
possible_actions_presence = possible_actions_presence.reshape(1, -1)
assert possible_actions_presence.shape == q_scores.shape
# set impossible actions so low that they can't be picked
q_scores -= (1.0 - possible_actions_presence) * 1e10 # type: ignore
q_scores_softmax = (
masked_softmax(q_scores, possible_actions_presence, softmax_temperature)
.detach()
.numpy()[0]
)
if np.isnan(q_scores_softmax).any() or np.max(q_scores_softmax) < 1e-3:
q_scores_softmax[:] = 1.0 / q_scores_softmax.shape[0]
greedy_act_idx = int(torch.argmax(q_scores))
softmax_act_idx = int(np.random.choice(q_scores.size()[1], p=q_scores_softmax))
return DqnPolicyActionSet(
greedy=greedy_act_idx,
softmax=softmax_act_idx,
greedy_act_name=action_names[greedy_act_idx],
softmax_act_name=action_names[softmax_act_idx],
softmax_act_prob=q_scores_softmax[softmax_act_idx],
)
def policy_given_q_values(
q_scores: torch.Tensor,
softmax_temperature: float,
possible_actions_presence: torch.Tensor,
) -> DqnPolicyActionSet:
assert q_scores.shape[0] == 1 and len(q_scores.shape) == 2
possible_actions_presence = possible_actions_presence.reshape(1, -1)
assert possible_actions_presence.shape == q_scores.shape
# set impossible actions so low that they can't be picked
q_scores -= (1.0 - possible_actions_presence) * 1e10
q_scores_softmax_numpy = (
masked_softmax(
q_scores.reshape(1, -1), possible_actions_presence, softmax_temperature
)
.detach()
.numpy()[0]
)
if (
np.isnan(q_scores_softmax_numpy).any()
or np.max(q_scores_softmax_numpy) < 1e-3
):
q_scores_softmax_numpy[:] = 1.0 / q_scores_softmax_numpy.shape[0]
greedy_act_idx = int(torch.argmax(q_scores))
softmax_act_idx = int(
np.random.choice(q_scores.size()[1], p=q_scores_softmax_numpy)
)
return DqnPolicyActionSet(
greedy=greedy_act_idx,
softmax=softmax_act_idx,
softmax_act_prob=float(q_scores_softmax_numpy[softmax_act_idx]),
)
def policy(
self, state: torch.Tensor, possible_actions_presence: torch.Tensor
) -> DqnPolicyActionSet:
assert state.size()[0] == 1, "Only pass in one state when getting a policy"
if self.use_gpu:
state = state.cuda()
q_scores = self.predict(state)
assert q_scores.shape[0] == 1
# set impossible actions so low that they can't be picked
q_scores -= (1.0 - possible_actions_presence) * 1e10 # type: ignore
q_scores_softmax = masked_softmax(
q_scores, possible_actions_presence, self.trainer.rl_temperature
).numpy()[0]
if np.isnan(q_scores_softmax).any() or np.max(q_scores_softmax) < 1e-3:
q_scores_softmax[:] = 1.0 / q_scores_softmax.shape[0]
return DqnPolicyActionSet(
greedy=int(torch.argmax(q_scores)),
softmax=int(np.random.choice(q_scores.size()[1], p=q_scores_softmax)),
)
def policy(
self,
states_tiled: torch.Tensor,
possible_actions_with_presence: Tuple[torch.Tensor, torch.Tensor],
):
possible_actions, possible_actions_presence = possible_actions_with_presence
assert states_tiled.size()[0] == possible_actions.size()[0]
assert possible_actions.size()[1] == self.action_dim
assert possible_actions.size()[0] == possible_actions_presence.size()[0]
if self.use_gpu:
states_tiled = states_tiled.cuda()
possible_actions = possible_actions.cuda()
q_scores = self.predict(states_tiled, possible_actions).reshape(
[1, self.action_dim]
)
possible_actions_presence = (possible_actions_presence.sum(dim=1) > 0).float()
# set impossible actions so low that they can't be picked
q_scores -= (
1.0 - possible_actions_presence.reshape(1, self.action_dim) # type: ignore
) * 1e10
q_scores_softmax_numpy = masked_softmax(
q_scores.reshape(1, -1),
possible_actions_presence.reshape(1, -1),
self.trainer.rl_temperature,
).numpy()[0]
if (
np.isnan(q_scores_softmax_numpy).any()
or np.max(q_scores_softmax_numpy) < 1e-3
):
q_scores_softmax_numpy[:] = 1.0 / q_scores_softmax_numpy.shape[0]
return DqnPolicyActionSet(
greedy=int(torch.argmax(q_scores)),
softmax=int(np.random.choice(q_scores.size()[1], p=q_scores_softmax_numpy)),
)
def create_from_tensors_parametric_dqn(
cls,
trainer: ParametricDQNTrainer,
mdp_ids: np.ndarray,
sequence_numbers: torch.Tensor,
states: rlt.PreprocessedFeatureVector,
actions: rlt.PreprocessedFeatureVector,
propensities: torch.Tensor,
rewards: torch.Tensor,
possible_actions_mask: torch.Tensor,
possible_actions: rlt.PreprocessedFeatureVector,
max_num_actions: int,
metrics: Optional[torch.Tensor] = None,
):
old_q_train_state = trainer.q_network.training
old_reward_train_state = trainer.reward_network.training
trainer.q_network.train(False)
trainer.reward_network.train(False)
state_action_pairs = rlt.PreprocessedStateAction(state=states, action=actions)
tiled_state = states.float_features.repeat(1, max_num_actions).reshape(
-1, states.float_features.shape[1]
)
assert possible_actions is not None
# Get Q-value of action taken
possible_actions_state_concat = rlt.PreprocessedStateAction(
state=rlt.PreprocessedFeatureVector(float_features=tiled_state),
action=possible_actions,
)
# FIXME: model_values, model_values_for_logged_action, and model_metrics_values
# should be calculated using q_network_cpe (as in discrete dqn).
# q_network_cpe has not been added in parametric dqn yet.
model_values = trainer.q_network(
possible_actions_state_concat
).q_value # type: ignore
optimal_q_values, _ = trainer.get_detached_q_values(
possible_actions_state_concat.state, possible_actions_state_concat.action
)
eval_action_idxs = None
assert (
model_values.shape[1] == 1
and model_values.shape[0]
== possible_actions_mask.shape[0] * possible_actions_mask.shape[1]
), (
"Invalid shapes: "
+ str(model_values.shape)
+ " != "
+ str(possible_actions_mask.shape)
)
model_values = model_values.reshape(possible_actions_mask.shape)
optimal_q_values = optimal_q_values.reshape(possible_actions_mask.shape)
model_propensities = masked_softmax(
optimal_q_values, possible_actions_mask, trainer.rl_temperature
)
rewards_and_metric_rewards = trainer.reward_network(
possible_actions_state_concat
).q_value # type: ignore
model_rewards = rewards_and_metric_rewards[:, :1]
assert (
model_rewards.shape[0] * model_rewards.shape[1]
== possible_actions_mask.shape[0] * possible_actions_mask.shape[1]
), (
"Invalid shapes: "
+ str(model_rewards.shape)
+ " != "
+ str(possible_actions_mask.shape)
)
model_rewards = model_rewards.reshape(possible_actions_mask.shape)
model_metrics = rewards_and_metric_rewards[:, 1:]
model_metrics = model_metrics.reshape(possible_actions_mask.shape[0], -1)
model_values_for_logged_action = trainer.q_network(state_action_pairs).q_value
model_rewards_and_metrics_for_logged_action = trainer.reward_network(
state_action_pairs
).q_value
model_rewards_for_logged_action = model_rewards_and_metrics_for_logged_action[
:, :1
]
action_dim = possible_actions.float_features.shape[1]
action_mask = torch.all(
possible_actions.float_features.view(-1, max_num_actions, action_dim)
== actions.float_features.unsqueeze(dim=1),
dim=2,
).float()
assert torch.all(action_mask.sum(dim=1) == 1)
num_metrics = model_metrics.shape[1] // max_num_actions
model_metrics_values = None
model_metrics_for_logged_action = None
model_metrics_values_for_logged_action = None
if num_metrics > 0:
# FIXME: calculate model_metrics_values when q_network_cpe is added
# to parametric dqn
model_metrics_values = model_values.repeat(1, num_metrics)
trainer.q_network.train(old_q_train_state) # type: ignore
trainer.reward_network.train(old_reward_train_state) # type: ignore
return cls(
mdp_id=mdp_ids,
sequence_number=sequence_numbers,
logged_propensities=propensities,
logged_rewards=rewards,
action_mask=action_mask,
model_rewards=model_rewards,
model_rewards_for_logged_action=model_rewards_for_logged_action,
model_values=model_values,
model_values_for_logged_action=model_values_for_logged_action,
model_metrics_values=model_metrics_values,
model_metrics_values_for_logged_action=model_metrics_values_for_logged_action,
model_propensities=model_propensities,
logged_metrics=metrics,
model_metrics=model_metrics,
model_metrics_for_logged_action=model_metrics_for_logged_action,
# Will compute later
logged_values=None,
logged_metrics_values=None,
possible_actions_mask=possible_actions_mask,
optimal_q_values=optimal_q_values,
eval_action_idxs=eval_action_idxs,
)
def create_from_tensors_dqn(
cls,
trainer: DQNTrainer,
mdp_ids: np.ndarray,
sequence_numbers: torch.Tensor,
states: rlt.PreprocessedFeatureVector,
actions: rlt.PreprocessedFeatureVector,
propensities: torch.Tensor,
rewards: torch.Tensor,
possible_actions_mask: torch.Tensor,
metrics: Optional[torch.Tensor] = None,
):
old_q_train_state = trainer.q_network.training
old_reward_train_state = trainer.reward_network.training
old_q_cpe_train_state = trainer.q_network_cpe.training
trainer.q_network.train(False)
trainer.reward_network.train(False)
trainer.q_network_cpe.train(False)
num_actions = trainer.num_actions
action_mask = actions.float() # type: ignore
rewards = trainer.boost_rewards(rewards, actions) # type: ignore
model_values = trainer.q_network_cpe(
rlt.PreprocessedState(state=states)
).q_values[:, 0:num_actions]
optimal_q_values, _ = trainer.get_detached_q_values(
states # type: ignore
)
eval_action_idxs = trainer.get_max_q_values( # type: ignore
optimal_q_values, possible_actions_mask
)[1]
model_propensities = masked_softmax(
optimal_q_values, possible_actions_mask, trainer.rl_temperature
)
assert model_values.shape == actions.shape, ( # type: ignore
"Invalid shape: "
+ str(model_values.shape) # type: ignore
+ " != "
+ str(actions.shape) # type: ignore
)
assert model_values.shape == possible_actions_mask.shape, ( # type: ignore
"Invalid shape: "
+ str(model_values.shape) # type: ignore
+ " != "
+ str(possible_actions_mask.shape) # type: ignore
)
model_values_for_logged_action = torch.sum(
model_values * action_mask, dim=1, keepdim=True
)
rewards_and_metric_rewards = trainer.reward_network(
rlt.PreprocessedState(state=states)
)
# In case we reuse the modular for Q-network
if hasattr(rewards_and_metric_rewards, "q_values"):
rewards_and_metric_rewards = rewards_and_metric_rewards.q_values
model_rewards = rewards_and_metric_rewards[:, 0:num_actions]
assert model_rewards.shape == actions.shape, ( # type: ignore
"Invalid shape: "
+ str(model_rewards.shape) # type: ignore
+ " != "
+ str(actions.shape) # type: ignore
)
model_rewards_for_logged_action = torch.sum(
model_rewards * action_mask, dim=1, keepdim=True
)
model_metrics = rewards_and_metric_rewards[:, num_actions:]
assert model_metrics.shape[1] % num_actions == 0, (
"Invalid metrics shape: "
+ str(model_metrics.shape)
+ " "
+ str(num_actions)
)
num_metrics = model_metrics.shape[1] // num_actions
if num_metrics == 0:
model_metrics_values = None
model_metrics_for_logged_action = None
model_metrics_values_for_logged_action = None
else:
model_metrics_values = trainer.q_network_cpe(
rlt.PreprocessedState(state=states)
)
# Backward compatility
if hasattr(model_metrics_values, "q_values"):
model_metrics_values = model_metrics_values.q_values
model_metrics_values = model_metrics_values[:, num_actions:]
assert (
model_metrics_values.shape[1] == num_actions * num_metrics
), ( # type: ignore
"Invalid shape: "
+ str(model_metrics_values.shape[1]) # type: ignore
+ " != "
+ str(actions.shape[1] * num_metrics) # type: ignore
)
model_metrics_for_logged_action_list = []
model_metrics_values_for_logged_action_list = []
for metric_index in range(num_metrics):
metric_start = metric_index * num_actions
metric_end = (metric_index + 1) * num_actions
model_metrics_for_logged_action_list.append(
torch.sum(
model_metrics[:, metric_start:metric_end] * action_mask,
dim=1,
keepdim=True,
)
)
model_metrics_values_for_logged_action_list.append(
torch.sum(
model_metrics_values[:, metric_start:metric_end] * action_mask,
dim=1,
keepdim=True,
)
)
model_metrics_for_logged_action = torch.cat(
model_metrics_for_logged_action_list, dim=1
)
model_metrics_values_for_logged_action = torch.cat(
model_metrics_values_for_logged_action_list, dim=1
)
trainer.q_network_cpe.train(old_q_cpe_train_state) # type: ignore
trainer.q_network.train(old_q_train_state) # type: ignore
trainer.reward_network.train(old_reward_train_state) # type: ignore
return cls(
mdp_id=mdp_ids,
sequence_number=sequence_numbers,
logged_propensities=propensities,
logged_rewards=rewards,
action_mask=action_mask,
model_rewards=model_rewards,
model_rewards_for_logged_action=model_rewards_for_logged_action,
model_values=model_values,
model_values_for_logged_action=model_values_for_logged_action,
model_metrics_values=model_metrics_values,
model_metrics_values_for_logged_action=model_metrics_values_for_logged_action,
model_propensities=model_propensities,
logged_metrics=metrics,
model_metrics=model_metrics,
model_metrics_for_logged_action=model_metrics_for_logged_action,
# Will compute later
logged_values=None,
logged_metrics_values=None,
possible_actions_mask=possible_actions_mask,
optimal_q_values=optimal_q_values,
eval_action_idxs=eval_action_idxs,
)
def _calculate_cpes(
self,
training_batch,
states,
next_states,
all_action_scores,
all_next_action_scores,
logged_action_idxs,
discount_tensor,
not_done_mask,
):
if not self.calc_cpe_in_training:
return None, None, None
if training_batch.extras.metrics is None:
metrics_reward_concat_real_vals = training_batch.training_input.reward
else:
metrics_reward_concat_real_vals = torch.cat(
(training_batch.training_input.reward,
training_batch.extras.metrics),
dim=1,
)
model_propensities_next_states = masked_softmax(
all_next_action_scores,
training_batch.training_input.possible_next_actions_mask if
self.maxq_learning else training_batch.training_input.next_action,
self.rl_temperature,
)
with torch.enable_grad():
######### Train separate reward network for CPE evaluation #############
# FIXME: the reward network should be outputting a tensor,
# not a q-value object
reward_estimates = self.reward_network(states).q_values
reward_estimates_for_logged_actions = reward_estimates.gather(
1, self.reward_idx_offsets + logged_action_idxs)
reward_loss = F.mse_loss(reward_estimates_for_logged_actions,
metrics_reward_concat_real_vals)
reward_loss.backward()
self._maybe_run_optimizer(self.reward_network_optimizer,
self.minibatches_per_step)
######### Train separate q-network for CPE evaluation #############
metric_q_values = self.q_network_cpe(states).q_values.gather(
1, self.reward_idx_offsets + logged_action_idxs)
all_metrics_target_q_values = torch.chunk(
self.q_network_cpe_target(next_states).q_values.detach(),
len(self.metrics_to_score),
dim=1,
)
target_metric_q_values = []
for i, per_metric_target_q_values in enumerate(
all_metrics_target_q_values):
per_metric_next_q_values = torch.sum(
per_metric_target_q_values *
model_propensities_next_states,
1,
keepdim=True,
)
per_metric_next_q_values = per_metric_next_q_values * not_done_mask
per_metric_target_q_values = metrics_reward_concat_real_vals[:, i:i + 1] + (
discount_tensor * per_metric_next_q_values)
target_metric_q_values.append(per_metric_target_q_values)
target_metric_q_values = torch.cat(target_metric_q_values, dim=1)
metric_q_value_loss = self.q_network_loss(metric_q_values,
target_metric_q_values)
metric_q_value_loss.backward()
self._maybe_run_optimizer(self.q_network_cpe_optimizer,
self.minibatches_per_step)
# Use the soft update rule to update target network
self._maybe_soft_update(
self.q_network_cpe,
self.q_network_cpe_target,
self.tau,
self.minibatches_per_step,
)
model_propensities = masked_softmax(
all_action_scores,
training_batch.training_input.possible_actions_mask
if self.maxq_learning else training_batch.training_input.action,
self.rl_temperature,
)
model_rewards = reward_estimates[:,
torch.arange(
self.reward_idx_offsets[0],
self.reward_idx_offsets[0] +
self.num_actions,
), ]
return reward_loss, model_rewards, model_propensities