def generate_samples(self,
num_transitions,
epsilon,
with_possible=True) -> Samples:
samples = GridworldContinuous.generate_samples(self, num_transitions,
epsilon, with_possible)
enum_states = []
for state in samples.states:
enum_states.append({0: float(list(state.keys())[0])})
enum_next_states = []
for state in samples.next_states:
enum_next_states.append({0: float(list(state.keys())[0])})
return Samples(
mdp_ids=samples.mdp_ids,
sequence_numbers=samples.sequence_numbers,
states=enum_states,
actions=samples.actions,
propensities=samples.propensities,
rewards=samples.rewards,
next_states=enum_next_states,
next_actions=samples.next_actions,
terminals=samples.terminals,
possible_next_actions=samples.possible_next_actions,
reward_timelines=samples.reward_timelines,
)
def generate_samples(self,
num_transitions,
epsilon,
with_possible=True) -> Samples:
samples = self.generate_samples_discrete(num_transitions, epsilon,
with_possible)
continuous_actions = [
self.action_to_features(a) for a in samples.actions
]
continuous_next_actions = [
self.action_to_features(a) if a is not "" else {}
for a in samples.next_actions
]
continuous_possible_next_actions = []
for possible_next_action in samples.possible_next_actions:
continuous_possible_next_actions.append([
self.action_to_features(a) if a is not None else {}
for a in possible_next_action
])
return Samples(
states=samples.states,
actions=continuous_actions,
propensities=samples.propensities,
rewards=samples.rewards,
next_states=samples.next_states,
next_actions=continuous_next_actions,
is_terminal=samples.is_terminal,
possible_next_actions=continuous_possible_next_actions,
reward_timelines=samples.reward_timelines,
)
def test_predictor_export(self):
"""Verify that q-values before model export equal q-values after
model export. Meant to catch issues with export logic."""
environment = Gridworld()
trainer = trainer = self.get_sarsa_trainer(environment, False)
samples = Samples(
mdp_ids=["0"],
sequence_numbers=[0],
states=[{0: 1.0, 1: 1.0, 2: 1.0, 3: 1.0, 4: 1.0, 5: 1.0, 15: 1.0, 24: 1.0}],
actions=["D"],
action_probabilities=[0.5],
rewards=[0],
possible_actions=[["R", "D"]],
next_states=[{5: 1.0}],
next_actions=["U"],
terminals=[False],
possible_next_actions=[["R", "U", "D"]],
)
tdps = environment.preprocess_samples(samples, 1)
pre_export_q_values = trainer.q_network(tdps[0].states).detach().numpy()
predictor = trainer.predictor()
with tempfile.TemporaryDirectory() as tmpdirname:
tmp_path = os.path.join(tmpdirname, "model")
predictor.save(tmp_path, "minidb")
new_predictor = DQNPredictor.load(tmp_path, "minidb", False)
post_export_q_values = new_predictor.predict([samples.states[0]])
for i, action in enumerate(environment.ACTIONS):
self.assertAlmostEquals(
pre_export_q_values[0][i], post_export_q_values[0][action], places=4
)
def generate_samples(self, num_transitions, epsilon,
discount_factor) -> Samples:
samples = GridworldContinuous.generate_samples(self, num_transitions,
epsilon,
discount_factor)
enum_states = []
for state in samples.states:
enum_states.append({0: float(list(state.keys())[0])})
enum_next_states = []
for state in samples.next_states:
enum_next_states.append({0: float(list(state.keys())[0])})
return Samples(
mdp_ids=samples.mdp_ids,
sequence_numbers=samples.sequence_numbers,
states=enum_states,
actions=samples.actions,
action_probabilities=samples.action_probabilities,
rewards=samples.rewards,
possible_actions=samples.possible_actions,
next_states=enum_next_states,
next_actions=samples.next_actions,
terminals=samples.terminals,
possible_next_actions=samples.possible_next_actions,
episode_values=samples.episode_values,
)
def generate_samples(
self,
num_transitions,
epsilon,
discount_factor,
multi_steps: Optional[int] = None,
) -> Samples:
samples = self.generate_samples_discrete(num_transitions, epsilon,
discount_factor, multi_steps)
continuous_actions = [
self.action_to_features(a) for a in samples.actions
]
continuous_possible_actions = []
for possible_action in samples.possible_actions:
continuous_possible_actions.append([
self.action_to_features(a) if a is not None else {}
for a in possible_action
])
if multi_steps is not None:
continuous_next_actions = [[
self.action_to_features(a) if a is not "" else {}
for a in next_action
] for next_action in samples.next_actions]
continuous_possible_next_actions = []
for possible_next_actions in samples.possible_next_actions:
continuous_possible_next_actions.append([[
self.action_to_features(a) if a is not None else {}
for a in pna
] for pna in possible_next_actions])
else:
continuous_next_actions = [
self.action_to_features(a) if a is not "" else {}
for a in samples.next_actions
]
continuous_possible_next_actions = []
for possible_next_action in samples.possible_next_actions:
continuous_possible_next_actions.append([
self.action_to_features(a) if a is not None else {}
for a in possible_next_action
])
return Samples(
mdp_ids=samples.mdp_ids,
sequence_numbers=samples.sequence_numbers,
states=samples.states,
actions=continuous_actions,
action_probabilities=samples.action_probabilities,
rewards=samples.rewards,
possible_actions=continuous_possible_actions,
next_states=samples.next_states,
next_actions=continuous_next_actions,
terminals=samples.terminals,
possible_next_actions=continuous_possible_next_actions,
)
def test_predictor_torch_export(self):
"""Verify that q-values before model export equal q-values after
model export. Meant to catch issues with export logic."""
environment = Gridworld()
samples = Samples(
mdp_ids=["0"],
sequence_numbers=[0],
sequence_number_ordinals=[1],
states=[{0: 1.0, 1: 1.0, 2: 1.0, 3: 1.0, 4: 1.0, 5: 1.0, 15: 1.0, 24: 1.0}],
actions=["D"],
action_probabilities=[0.5],
rewards=[0],
possible_actions=[["R", "D"]],
next_states=[{5: 1.0}],
next_actions=["U"],
terminals=[False],
possible_next_actions=[["R", "U", "D"]],
)
tdps = environment.preprocess_samples(samples, 1)
assert len(tdps) == 1, "Invalid number of data pages"
trainer, exporter = self.get_modular_sarsa_trainer_exporter(
environment, {}, False
)
input = rlt.PreprocessedState.from_tensor(tdps[0].states)
pre_export_q_values = trainer.q_network(input).q_values.detach().numpy()
preprocessor = Preprocessor(environment.normalization, False)
cpu_q_network = trainer.q_network.cpu_model()
cpu_q_network.eval()
dqn_with_preprocessor = DiscreteDqnWithPreprocessor(cpu_q_network, preprocessor)
serving_module = DiscreteDqnPredictorWrapper(
dqn_with_preprocessor, action_names=environment.ACTIONS
)
with tempfile.TemporaryDirectory() as tmpdirname:
buf = export_module_to_buffer(serving_module)
tmp_path = os.path.join(tmpdirname, "model")
with open(tmp_path, "wb") as f:
f.write(buf.getvalue())
f.close()
predictor = DiscreteDqnTorchPredictor(torch.jit.load(tmp_path))
post_export_q_values = predictor.predict([samples.states[0]])
for i, action in enumerate(environment.ACTIONS):
self.assertAlmostEqual(
float(pre_export_q_values[0][i]),
float(post_export_q_values[0][action]),
places=4,
)
def generate_samples(self, num_transitions, epsilon, discount_factor) -> Samples:
samples = self.generate_samples_discrete(
num_transitions, epsilon, discount_factor
)
continuous_actions = [self.action_to_features(a) for a in samples.actions]
continuous_next_actions = [
self.action_to_features(a) if a is not "" else {}
for a in samples.next_actions
]
continuous_possible_actions = []
for possible_action in samples.possible_actions:
continuous_possible_actions.append(
[
self.action_to_features(a) if a is not None else {}
for a in possible_action
]
)
continuous_possible_next_actions = []
for possible_next_action in samples.possible_next_actions:
continuous_possible_next_actions.append(
[
self.action_to_features(a) if a is not None else {}
for a in possible_next_action
]
)
return Samples(
mdp_ids=samples.mdp_ids,
sequence_numbers=samples.sequence_numbers,
states=samples.states,
actions=continuous_actions,
action_probabilities=samples.action_probabilities,
rewards=samples.rewards,
possible_actions=continuous_possible_actions,
next_states=samples.next_states,
next_actions=continuous_next_actions,
terminals=samples.terminals,
possible_next_actions=continuous_possible_next_actions,
episode_values=samples.episode_values,
)
def preprocess_samples(
self,
samples: Samples,
minibatch_size: int,
use_gpu: bool = False,
one_hot_action: bool = True,
normalize_actions: bool = True,
) -> List[TrainingDataPage]:
logger.info("Shuffling...")
samples.shuffle()
logger.info("Sparse2Dense...")
net = core.Net("gridworld_preprocessing")
C2.set_net(net)
saa = StackedAssociativeArray.from_dict_list(samples.states, "states")
sorted_state_features, _ = sort_features_by_normalization(self.normalization)
state_matrix, _ = sparse_to_dense(
saa.lengths, saa.keys, saa.values, sorted_state_features
)
saa = StackedAssociativeArray.from_dict_list(samples.next_states, "next_states")
next_state_matrix, _ = sparse_to_dense(
saa.lengths, saa.keys, saa.values, sorted_state_features
)
sorted_action_features, _ = sort_features_by_normalization(
self.normalization_action
)
saa = StackedAssociativeArray.from_dict_list(samples.actions, "action")
action_matrix, _ = sparse_to_dense(
saa.lengths, saa.keys, saa.values, sorted_action_features
)
saa = StackedAssociativeArray.from_dict_list(
samples.next_actions, "next_action"
)
next_action_matrix, _ = sparse_to_dense(
saa.lengths, saa.keys, saa.values, sorted_action_features
)
action_probabilities = torch.tensor(
samples.action_probabilities, dtype=torch.float32
).reshape(-1, 1)
rewards = torch.tensor(samples.rewards, dtype=torch.float32).reshape(-1, 1)
pnas_lengths_list = []
pnas_flat: List[List[str]] = []
for pnas in samples.possible_next_actions:
pnas_lengths_list.append(len(pnas))
pnas_flat.extend(pnas)
saa = StackedAssociativeArray.from_dict_list(pnas_flat, "possible_next_actions")
pnas_lengths = torch.tensor(pnas_lengths_list, dtype=torch.int32)
pna_lens_blob = "pna_lens_blob"
workspace.FeedBlob(pna_lens_blob, pnas_lengths.numpy())
possible_next_actions_matrix, _ = sparse_to_dense(
saa.lengths, saa.keys, saa.values, sorted_action_features
)
state_pnas_tile_blob = C2.LengthsTile(next_state_matrix, pna_lens_blob)
workspace.RunNetOnce(net)
logger.info("Preprocessing...")
state_preprocessor = Preprocessor(self.normalization, False)
action_preprocessor = Preprocessor(self.normalization_action, False)
states_ndarray = workspace.FetchBlob(state_matrix)
states_ndarray = state_preprocessor.forward(states_ndarray)
actions_ndarray = torch.from_numpy(workspace.FetchBlob(action_matrix))
if normalize_actions:
actions_ndarray = action_preprocessor.forward(actions_ndarray)
next_states_ndarray = workspace.FetchBlob(next_state_matrix)
next_states_ndarray = state_preprocessor.forward(next_states_ndarray)
next_actions_ndarray = torch.from_numpy(workspace.FetchBlob(next_action_matrix))
if normalize_actions:
next_actions_ndarray = action_preprocessor.forward(next_actions_ndarray)
logged_possible_next_actions = action_preprocessor.forward(
workspace.FetchBlob(possible_next_actions_matrix)
)
state_pnas_tile = state_preprocessor.forward(
workspace.FetchBlob(state_pnas_tile_blob)
)
logged_possible_next_state_actions = torch.cat(
(state_pnas_tile, logged_possible_next_actions), dim=1
)
logger.info("Reward Timeline to Torch...")
possible_next_actions_ndarray = logged_possible_next_actions
possible_next_actions_state_concat = logged_possible_next_state_actions
time_diffs = torch.ones([len(samples.states), 1])
tdps = []
pnas_start = 0
logger.info("Batching...")
for start in range(0, states_ndarray.shape[0], minibatch_size):
end = start + minibatch_size
if end > states_ndarray.shape[0]:
break
pnas_end = pnas_start + torch.sum(pnas_lengths[start:end])
pnas = possible_next_actions_ndarray[pnas_start:pnas_end]
pnas_concat = possible_next_actions_state_concat[pnas_start:pnas_end]
pnas_start = pnas_end
tdp = TrainingDataPage(
states=states_ndarray[start:end],
actions=actions_ndarray[start:end],
propensities=action_probabilities[start:end],
rewards=rewards[start:end],
next_states=next_states_ndarray[start:end],
next_actions=next_actions_ndarray[start:end],
possible_next_actions=None,
not_terminals=(pnas_lengths[start:end] > 0).reshape(-1, 1),
time_diffs=time_diffs[start:end],
possible_next_actions_lengths=pnas_lengths[start:end],
possible_next_actions_state_concat=pnas_concat,
)
tdp.set_type(torch.cuda.FloatTensor if use_gpu else torch.FloatTensor)
tdps.append(tdp)
return tdps
def preprocess_samples(self, samples: Samples,
minibatch_size: int) -> List[TrainingDataPage]:
samples.shuffle()
net = core.Net("gridworld_preprocessing")
C2.set_net(net)
preprocessor = PreprocessorNet(True)
saa = StackedAssociativeArray.from_dict_list(samples.states, "states")
state_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization,
"state_norm",
False,
False,
)
saa = StackedAssociativeArray.from_dict_list(samples.next_states,
"next_states")
next_state_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization,
"next_state_norm",
False,
False,
)
saa = StackedAssociativeArray.from_dict_list(samples.actions, "action")
action_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization_action,
"action_norm",
False,
False,
)
saa = StackedAssociativeArray.from_dict_list(samples.next_actions,
"next_action")
next_action_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization_action,
"next_action_norm",
False,
False,
)
propensities = np.array(samples.propensities,
dtype=np.float32).reshape(-1, 1)
rewards = np.array(samples.rewards, dtype=np.float32).reshape(-1, 1)
pnas_lengths_list = []
pnas_flat: List[List[str]] = []
for pnas in samples.possible_next_actions:
pnas_lengths_list.append(len(pnas))
pnas_flat.extend(pnas)
saa = StackedAssociativeArray.from_dict_list(pnas_flat,
"possible_next_actions")
pnas_lengths = np.array(pnas_lengths_list, dtype=np.int32)
possible_next_actions_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization_action,
"possible_next_action_norm",
False,
False,
)
workspace.RunNetOnce(net)
states_ndarray = workspace.FetchBlob(state_matrix)
actions_ndarray = workspace.FetchBlob(action_matrix)
next_states_ndarray = workspace.FetchBlob(next_state_matrix)
next_actions_ndarray = workspace.FetchBlob(next_action_matrix)
possible_next_actions_ndarray = workspace.FetchBlob(
possible_next_actions_matrix)
tdps = []
pnas_start = 0
for start in range(0, states_ndarray.shape[0], minibatch_size):
end = start + minibatch_size
if end > states_ndarray.shape[0]:
break
pnas_end = pnas_start + np.sum(pnas_lengths[start:end])
pnas = possible_next_actions_ndarray[pnas_start:pnas_end]
pnas_start = pnas_end
tdps.append(
TrainingDataPage(
states=states_ndarray[start:end],
actions=actions_ndarray[start:end],
propensities=propensities[start:end],
rewards=rewards[start:end],
next_states=next_states_ndarray[start:end],
next_actions=next_actions_ndarray[start:end],
possible_next_actions=StackedArray(pnas_lengths[start:end],
pnas),
not_terminals=(pnas_lengths[start:end] > 0).reshape(-1, 1),
reward_timelines=samples.reward_timelines[start:end]
if samples.reward_timelines else None,
))
return tdps
def preprocess_samples(self, samples: Samples,
minibatch_size: int) -> List[TrainingDataPage]:
samples.shuffle()
net = core.Net("gridworld_preprocessing")
C2.set_net(net)
preprocessor = PreprocessorNet(True)
saa = StackedAssociativeArray.from_dict_list(samples.states, "states")
state_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization,
"state_norm",
False,
False,
False,
)
saa = StackedAssociativeArray.from_dict_list(samples.next_states,
"next_states")
next_state_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization,
"next_state_norm",
False,
False,
False,
)
saa = StackedAssociativeArray.from_dict_list(samples.actions, "action")
action_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization_action,
"action_norm",
False,
False,
False,
)
saa = StackedAssociativeArray.from_dict_list(samples.next_actions,
"next_action")
next_action_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization_action,
"next_action_norm",
False,
False,
False,
)
propensities = np.array(samples.propensities,
dtype=np.float32).reshape(-1, 1)
rewards = np.array(samples.rewards, dtype=np.float32).reshape(-1, 1)
pnas_lengths_list = []
pnas_flat: List[List[str]] = []
for pnas in samples.possible_next_actions:
pnas_lengths_list.append(len(pnas))
pnas_flat.extend(pnas)
saa = StackedAssociativeArray.from_dict_list(pnas_flat,
"possible_next_actions")
pnas_lengths = np.array(pnas_lengths_list, dtype=np.int32)
pna_lens_blob = "pna_lens_blob"
workspace.FeedBlob(pna_lens_blob, pnas_lengths)
possible_next_actions_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization_action,
"possible_next_action_norm",
False,
False,
False,
)
state_pnas_tile_blob = C2.LengthsTile(next_state_matrix, pna_lens_blob)
workspace.RunNetOnce(net)
state_preprocessor = Preprocessor(self.normalization, False)
action_preprocessor = Preprocessor(self.normalization_action, False)
states_ndarray = workspace.FetchBlob(state_matrix)
states_ndarray = state_preprocessor.forward(states_ndarray).numpy()
actions_ndarray = workspace.FetchBlob(action_matrix)
actions_ndarray = action_preprocessor.forward(actions_ndarray).numpy()
next_states_ndarray = workspace.FetchBlob(next_state_matrix)
next_states_ndarray = state_preprocessor.forward(
next_states_ndarray).numpy()
next_actions_ndarray = workspace.FetchBlob(next_action_matrix)
next_actions_ndarray = action_preprocessor.forward(
next_actions_ndarray).numpy()
logged_possible_next_actions = action_preprocessor.forward(
workspace.FetchBlob(possible_next_actions_matrix))
state_pnas_tile = state_preprocessor.forward(
workspace.FetchBlob(state_pnas_tile_blob))
logged_possible_next_state_actions = torch.cat(
(state_pnas_tile, logged_possible_next_actions), dim=1)
possible_next_actions_ndarray = logged_possible_next_actions.cpu(
).numpy()
next_state_pnas_concat = logged_possible_next_state_actions.cpu(
).numpy()
time_diffs = np.ones(len(states_ndarray))
episode_values = None
if samples.reward_timelines is not None:
episode_values = np.zeros(rewards.shape, dtype=np.float32)
for i, reward_timeline in enumerate(samples.reward_timelines):
for time_diff, reward in reward_timeline.items():
episode_values[i, 0] += reward * (DISCOUNT**time_diff)
tdps = []
pnas_start = 0
for start in range(0, states_ndarray.shape[0], minibatch_size):
end = start + minibatch_size
if end > states_ndarray.shape[0]:
break
pnas_end = pnas_start + np.sum(pnas_lengths[start:end])
pnas = possible_next_actions_ndarray[pnas_start:pnas_end]
pnas_concat = next_state_pnas_concat[pnas_start:pnas_end]
pnas_start = pnas_end
tdps.append(
TrainingDataPage(
states=states_ndarray[start:end],
actions=actions_ndarray[start:end],
propensities=propensities[start:end],
rewards=rewards[start:end],
next_states=next_states_ndarray[start:end],
next_actions=next_actions_ndarray[start:end],
possible_next_actions=StackedArray(pnas_lengths[start:end],
pnas),
not_terminals=(pnas_lengths[start:end] > 0).reshape(-1, 1),
episode_values=episode_values[start:end]
if episode_values is not None else None,
time_diffs=time_diffs[start:end],
possible_next_actions_lengths=pnas_lengths[start:end],
next_state_pnas_concat=pnas_concat,
))
return tdps
def preprocess_samples(
self,
samples: Samples,
minibatch_size: int,
use_gpu: bool = False,
one_hot_action: bool = True,
normalize_actions: bool = True,
) -> List[TrainingDataPage]:
logger.info("Shuffling...")
samples.shuffle()
logger.info("Sparse2Dense...")
net = core.Net("gridworld_preprocessing")
C2.set_net(net)
saa = StackedAssociativeArray.from_dict_list(samples.states, "states")
sorted_state_features, _ = sort_features_by_normalization(
self.normalization)
state_matrix, _ = sparse_to_dense(saa.lengths, saa.keys, saa.values,
sorted_state_features)
saa = StackedAssociativeArray.from_dict_list(samples.next_states,
"next_states")
next_state_matrix, _ = sparse_to_dense(saa.lengths, saa.keys,
saa.values,
sorted_state_features)
sorted_action_features, _ = sort_features_by_normalization(
self.normalization_action)
saa = StackedAssociativeArray.from_dict_list(samples.actions, "action")
action_matrix, _ = sparse_to_dense(saa.lengths, saa.keys, saa.values,
sorted_action_features)
saa = StackedAssociativeArray.from_dict_list(samples.next_actions,
"next_action")
next_action_matrix, _ = sparse_to_dense(saa.lengths, saa.keys,
saa.values,
sorted_action_features)
action_probabilities = torch.tensor(samples.action_probabilities,
dtype=torch.float32).reshape(
-1, 1)
rewards = torch.tensor(samples.rewards,
dtype=torch.float32).reshape(-1, 1)
max_action_size = 4
pnas_mask_list: List[List[int]] = []
pnas_flat: List[Dict[str, float]] = []
for pnas in samples.possible_next_actions:
pnas_mask_list.append([1] * len(pnas) + [0] *
(max_action_size - len(pnas)))
pnas_flat.extend(pnas)
for _ in range(max_action_size - len(pnas)):
pnas_flat.append({}) # Filler
saa = StackedAssociativeArray.from_dict_list(pnas_flat,
"possible_next_actions")
pnas_mask = torch.Tensor(pnas_mask_list)
possible_next_actions_matrix, _ = sparse_to_dense(
saa.lengths, saa.keys, saa.values, sorted_action_features)
workspace.RunNetOnce(net)
logger.info("Preprocessing...")
state_preprocessor = Preprocessor(self.normalization, False)
action_preprocessor = Preprocessor(self.normalization_action, False)
states_ndarray = workspace.FetchBlob(state_matrix)
states_ndarray = state_preprocessor.forward(states_ndarray)
actions_ndarray = torch.from_numpy(workspace.FetchBlob(action_matrix))
if normalize_actions:
actions_ndarray = action_preprocessor.forward(actions_ndarray)
next_states_ndarray = workspace.FetchBlob(next_state_matrix)
next_states_ndarray = state_preprocessor.forward(next_states_ndarray)
state_pnas_tile = next_states_ndarray.repeat(
1, max_action_size).reshape(-1, next_states_ndarray.shape[1])
next_actions_ndarray = torch.from_numpy(
workspace.FetchBlob(next_action_matrix))
if normalize_actions:
next_actions_ndarray = action_preprocessor.forward(
next_actions_ndarray)
logged_possible_next_actions = action_preprocessor.forward(
workspace.FetchBlob(possible_next_actions_matrix))
assert state_pnas_tile.shape[0] == logged_possible_next_actions.shape[
0], ("Invalid shapes: " + str(state_pnas_tile.shape) + " != " +
str(logged_possible_next_actions.shape))
logged_possible_next_state_actions = torch.cat(
(state_pnas_tile, logged_possible_next_actions), dim=1)
logger.info("Reward Timeline to Torch...")
time_diffs = torch.ones([len(samples.states), 1])
tdps = []
pnas_start = 0
logger.info("Batching...")
for start in range(0, states_ndarray.shape[0], minibatch_size):
end = start + minibatch_size
if end > states_ndarray.shape[0]:
break
pnas_end = pnas_start + (minibatch_size * max_action_size)
tdp = TrainingDataPage(
states=states_ndarray[start:end],
actions=actions_ndarray[start:end],
propensities=action_probabilities[start:end],
rewards=rewards[start:end],
next_states=next_states_ndarray[start:end],
next_actions=next_actions_ndarray[start:end],
possible_next_actions=None,
not_terminal=(pnas_mask[start:end, :].sum(dim=1, keepdim=True)
> 0),
time_diffs=time_diffs[start:end],
possible_next_actions_mask=pnas_mask[start:end, :],
possible_next_actions_state_concat=
logged_possible_next_state_actions[pnas_start:pnas_end, :],
)
pnas_start = pnas_end
tdp.set_type(
torch.cuda.FloatTensor if use_gpu else torch.FloatTensor)
tdps.append(tdp)
return tdps
