def __init__(self,
state_rep: state_representation.StateRepresentation,
embedding_size: int,
zero_out: bool = True):
super(DynamicEnvironmentEmbedder, self).__init__()
self._card_count_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(embedding_size,
state_rep.get_card_counts(),
add_unk=False,
zero_out=zero_out)
self._card_color_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(embedding_size,
state_rep.get_card_colors(),
add_unk=False,
zero_out=zero_out)
self._card_shape_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(embedding_size,
state_rep.get_card_shapes(),
add_unk=False,
zero_out=zero_out)
self._card_selection_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size, state_rep.get_card_selection(), add_unk=False, zero_out=zero_out)
self._leader_rotation_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size, state_rep.get_leader_rotation(), add_unk=False, zero_out=zero_out)
self._follower_rotation_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size, state_rep.get_follower_rotation(), add_unk=False, zero_out=zero_out)
self._embedders: List[word_embedder.WordEmbedder] = [self._card_count_embedder,
self._card_color_embedder,
self._card_shape_embedder,
self._card_selection_embedder,
self._leader_rotation_embedder,
self._follower_rotation_embedder]
def __init__(self, args: text_encoder_args.TextEncoderArgs,
vocabulary: List[str],
dropout: float = 0.):
super(TextEncoder, self).__init__()
self._args = args
self._embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(self._args.get_word_embedding_size(),
vocabulary)
self._each_dir_hidden_size: int = int(self._args.get_hidden_size() / 2)
self._dropout = nn.Dropout(dropout)
self._rnn: nn.LSTM = nn.LSTM(self._args.get_word_embedding_size(),
self._each_dir_hidden_size,
self._args.get_number_of_layers(),
batch_first=True,
bidirectional=True,
dropout=dropout)
initialization.initialize_rnn(self._rnn)
def __init__(self,
state_rep: state_representation.StateRepresentation,
embedding_size: int,
zero_out: bool = False):
super(StaticEnvironmentEmbedder, self).__init__()
self._state_rep: state_representation.StateRepresentation = state_rep
if zero_out:
raise ValueError(
'Zeroing out zero-value properties is no longer supported due to optimizations.'
)
all_embeddings = state_rep.get_terrains() + state_rep.get_hut_colors() + state_rep.get_hut_rotations() + \
state_rep.get_windmill_rotations() + state_rep.get_tower_rotations() + \
state_rep.get_tent_rotations() + state_rep.get_tree_types() + state_rep.get_plant_types() + \
state_rep.get_prop_types()
self._embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size,
all_embeddings,
add_unk=False,
must_be_unique=False)
def __init__(self,
state_rep: state_representation.StateRepresentation,
embedding_size: int,
zero_out: bool = False):
super(DynamicEnvironmentEmbedder, self).__init__()
self._state_rep: state_representation.StateRepresentation = state_rep
if zero_out:
raise ValueError(
'Zeroing out zero-value properties is no longer supported due to optimizations.'
)
# A single embedding lookup is used for all properties.
all_embeddings = state_rep.get_card_counts() + state_rep.get_card_colors() + state_rep.get_card_shapes() + \
state_rep.get_card_selection() + state_rep.get_leader_rotation() + \
state_rep.get_follower_rotation()
self._embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size,
all_embeddings,
add_unk=False,
must_be_unique=False)
def __init__(self,
state_rep: state_representation.StateRepresentation,
embedding_size: int,
zero_out: bool = True):
super(StaticEnvironmentEmbedder, self).__init__()
self._terrain_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size,
state_rep.get_terrains(),
add_unk=False,
zero_out=False)
self._hut_colors_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size,
state_rep.get_hut_colors(),
add_unk=False,
zero_out=zero_out)
self._hut_rotations_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size,
state_rep.get_hut_rotations(),
add_unk=False,
zero_out=zero_out)
self._windmill_rotations_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size,
state_rep.get_windmill_rotations(),
add_unk=False,
zero_out=zero_out)
self._tower_rotations_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size,
state_rep.get_tower_rotations(),
add_unk=False,
zero_out=zero_out)
self._tent_rotations_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size,
state_rep.get_tent_rotations(),
add_unk=False,
zero_out=zero_out)
self._tree_types_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size,
state_rep.get_tree_types(),
add_unk=False,
zero_out=zero_out)
self._plant_types_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size,
state_rep.get_plant_types(),
add_unk=False,
zero_out=zero_out)
self._prop_types_embedder: word_embedder.WordEmbedder = word_embedder.WordEmbedder(
embedding_size,
state_rep.get_prop_types(),
add_unk=False,
zero_out=zero_out)
self._embedders: List[word_embedder.WordEmbedder] = [
self._prop_types_embedder, self._hut_colors_embedder,
self._hut_rotations_embedder, self._tree_types_embedder,
self._plant_types_embedder, self._windmill_rotations_embedder,
self._tower_rotations_embedder, self._tent_rotations_embedder,
self._terrain_embedder
]
def __init__(self,
args: model_args.ModelArgs,
input_vocabulary: List[str],
auxiliaries: List[auxiliary.Auxiliary],
load_pretrained: bool = True,
end_to_end: bool = False):
super(ActionGeneratorModel, self).__init__()
self._args: model_args.ModelArgs = args
self._end_to_end = end_to_end
# If end-to-end, also add in the hex predictor model.
self._plan_predictor: Optional[
plan_predictor_model.PlanPredictorModel] = None
if self._end_to_end or load_pretrained:
self._plan_predictor: plan_predictor_model.PlanPredictorModel = plan_predictor_model.PlanPredictorModel(
args, input_vocabulary, auxiliaries)
self._output_layer = None
self._rnn = None
self._action_embedder = None
if self._args.get_decoder_args().use_recurrence():
self._action_embedder: word_embedder.WordEmbedder = \
word_embedder.WordEmbedder(
self._args.get_decoder_args().get_action_embedding_size(),
[str(action) for action in agent_actions.AGENT_ACTIONS],
add_unk=False)
self._rnn: nn.Module = nn.LSTM(
self._args.get_decoder_args().get_state_internal_size() +
self._args.get_decoder_args().get_action_embedding_size(),
self._args.get_decoder_args().get_hidden_size(),
self._args.get_decoder_args().get_num_layers(),
batch_first=True)
# Add a different output layer
self._output_layer: nn.Module = nn.Linear(
self._args.get_decoder_args().get_hidden_size() +
self._args.get_decoder_args().get_state_internal_size(),
len(agent_actions.AGENT_ACTIONS))
torch.nn.init.orthogonal_(
self._output_layer.weight,
torch.nn.init.calculate_gain("leaky_relu"))
self._output_layer.bias.data.fill_(0)
distribution_num_channels: int = 0
if self._args.get_decoder_args().use_trajectory_distribution():
distribution_num_channels += 1
if self._args.get_decoder_args().use_goal_probabilities():
distribution_num_channels += 1
if self._args.get_decoder_args().use_obstacle_probabilities():
distribution_num_channels += 1
if self._args.get_decoder_args().use_avoid_probabilities():
distribution_num_channels += 1
self._map_distribution_embedder: map_distribution_embedder.MapDistributionEmbedder = \
map_distribution_embedder.MapDistributionEmbedder(
distribution_num_channels,
self._args.get_decoder_args().get_state_internal_size(),
self._args.get_decoder_args().get_state_internal_size()
if self._args.get_decoder_args().use_recurrence() else len(
agent_actions.AGENT_ACTIONS),
self._args.get_decoder_args().get_crop_size(),
self._args.get_decoder_args().convolution_encode_map_distributions(),
self._args.get_decoder_args().use_recurrence())
if load_pretrained:
if self._args.get_decoder_args().pretrained_generator():
initialization.load_pretrained_parameters(
self._args.get_decoder_args(
).pretrained_action_generator_filepath(),
module=self)
if self._args.get_decoder_args().pretrained_plan_predictor():
initialization.load_pretrained_parameters(
self._args.get_decoder_args(
).pretrained_plan_predictor_filepath(),
module=self._plan_predictor)