def create_input_processors(
observation_shapes: List[Tuple[int, ...]],
h_size: int,
vis_encode_type: EncoderType,
normalize: bool = False,
) -> Tuple[nn.ModuleList, nn.ModuleList, int]:
"""
Creates visual and vector encoders, along with their normalizers.
:param observation_shapes: List of Tuples that represent the action dimensions.
:param action_size: Number of additional un-normalized inputs to each vector encoder. Used for
conditioining network on other values (e.g. actions for a Q function)
:param h_size: Number of hidden units per layer.
:param vis_encode_type: Type of visual encoder to use.
:param unnormalized_inputs: Vector inputs that should not be normalized, and added to the vector
obs.
:param normalize: Normalize all vector inputs.
:return: Tuple of visual encoders and vector encoders each as a list.
"""
visual_encoders: List[nn.Module] = []
vector_encoders: List[nn.Module] = []
visual_encoder_class = ModelUtils.get_encoder_for_type(vis_encode_type)
vector_size = 0
visual_output_size = 0
for i, dimension in enumerate(observation_shapes):
if len(dimension) == 3:
ModelUtils._check_resolution_for_encoder(
dimension[0], dimension[1], vis_encode_type
)
visual_encoders.append(
visual_encoder_class(
dimension[0], dimension[1], dimension[2], h_size
)
)
visual_output_size += h_size
elif len(dimension) == 1:
vector_size += dimension[0]
else:
raise UnityTrainerException(
f"Unsupported shape of {dimension} for observation {i}"
)
if vector_size > 0:
vector_encoders.append(VectorInput(vector_size, normalize))
# Total output size for all inputs + CNNs
total_processed_size = vector_size + visual_output_size
return (
nn.ModuleList(visual_encoders),
nn.ModuleList(vector_encoders),
total_processed_size,
)
def create_input_processors(
observation_specs: List[ObservationSpec],
h_size: int,
vis_encode_type: EncoderType,
normalize: bool = False,
) -> Tuple[nn.ModuleList, List[int]]:
"""
Creates visual and vector encoders, along with their normalizers.
:param observation_specs: List of ObservationSpec that represent the observation dimensions.
:param action_size: Number of additional un-normalized inputs to each vector encoder. Used for
conditioning network on other values (e.g. actions for a Q function)
:param h_size: Number of hidden units per layer.
:param vis_encode_type: Type of visual encoder to use.
:param unnormalized_inputs: Vector inputs that should not be normalized, and added to the vector
obs.
:param normalize: Normalize all vector inputs.
:return: Tuple of visual encoders and vector encoders each as a list.
"""
encoders: List[nn.Module] = []
embedding_sizes: List[int] = []
for obs_spec in observation_specs:
encoder, embedding_size = ModelUtils.get_encoder_for_obs(
obs_spec.shape, normalize, h_size, vis_encode_type)
encoders.append(encoder)
embedding_sizes.append(embedding_size)
return (nn.ModuleList(encoders), embedding_sizes)
def create_input_processors(
observation_specs: List[ObservationSpec],
h_size: int,
vis_encode_type: EncoderType,
normalize: bool = False,
) -> Tuple[nn.ModuleList, List[int]]:
"""
Creates visual and vector encoders, along with their normalizers.
:param observation_specs: List of ObservationSpec that represent the observation dimensions.
:param action_size: Number of additional un-normalized inputs to each vector encoder. Used for
conditioning network on other values (e.g. actions for a Q function)
:param h_size: Number of hidden units per layer.
:param vis_encode_type: Type of visual encoder to use.
:param unnormalized_inputs: Vector inputs that should not be normalized, and added to the vector
obs.
:param normalize: Normalize all vector inputs.
:return: Tuple of :
- ModuleList of the encoders
- A list of embedding sizes (0 if the input requires to be processed with a variable length
observation encoder)
"""
encoders: List[nn.Module] = []
embedding_sizes: List[int] = []
for obs_spec in observation_specs:
encoder, embedding_size = ModelUtils.get_encoder_for_obs(
obs_spec, normalize, h_size, vis_encode_type)
encoders.append(encoder)
embedding_sizes.append(embedding_size)
x_self_size = sum(embedding_sizes) # The size of the "self" embedding
if x_self_size > 0:
for enc in encoders:
if isinstance(enc, EntityEmbedding):
enc.add_self_embedding(h_size)
return (nn.ModuleList(encoders), embedding_sizes)
def _create_policy_branches(self, hidden_size: int) -> nn.ModuleList:
branches = []
for size in self.act_sizes:
branch_output_layer = linear_layer(
hidden_size,
size,
kernel_init=Initialization.KaimingHeNormal,
kernel_gain=0.1,
bias_init=Initialization.Zero,
)
branches.append(branch_output_layer)
return nn.ModuleList(branches)