def __init__(self, obs_shapes: Dict[str, Sequence[int]], non_lin: Union[str, type(nn.Module)], hidden_units: List[int]): self.obs_shapes = obs_shapes # Maze relies on dictionaries to represent the inference graph self.perception_dict = OrderedDict() # build latent feature embedding block self.perception_dict['latent_inventory'] = DenseBlock( in_keys='observation_inventory', out_keys='latent_inventory', in_shapes=obs_shapes['observation_inventory'], hidden_units=[128], non_lin=non_lin) # build latent pixel embedding block self.perception_dict['latent_screen'] = VGGConvolutionDenseBlock( in_keys='observation_screen', out_keys='latent_screen', in_shapes=obs_shapes['observation_screen'], non_lin=non_lin, hidden_channels=[8, 16, 32], hidden_units=[32]) # Concatenate latent features self.perception_dict['latent_concat'] = ConcatenationBlock( in_keys=['latent_inventory', 'latent_screen'], out_keys='latent_concat', in_shapes=self.perception_dict['latent_inventory'].out_shapes() + self.perception_dict['latent_screen'].out_shapes(), concat_dim=-1) # Add latent dense block self.perception_dict['latent_dense'] = DenseBlock( in_keys='latent_concat', out_keys='latent_dense', hidden_units=hidden_units, non_lin=non_lin, in_shapes=self.perception_dict['latent_concat'].out_shapes() ) # Add recurrent block self.perception_dict['latent'] = LSTMLastStepBlock( in_keys='latent_dense', out_keys='latent', in_shapes=self.perception_dict['latent_dense'].out_shapes(), hidden_size=32, num_layers=1, bidirectional=False, non_lin=non_lin )
def __init__(self, obs_shapes: Dict[str, Sequence[int]], non_lin: Union[str, type(nn.Module)], hidden_units: List[int]): nn.Module.__init__(self) # Maze relies on dictionaries to represent the inference graph self.perception_dict = OrderedDict() # build latent feature embedding block self.perception_dict['latent_inventory'] = DenseBlock( in_keys='observation_inventory', out_keys='latent_inventory', in_shapes=obs_shapes['observation_inventory'], hidden_units=[128], non_lin=non_lin) # Concatenate latent features self.perception_dict['latent_concat'] = ConcatenationBlock( in_keys=['latent_inventory', 'latent_screen'], out_keys='latent_concat', in_shapes=self.perception_dict['latent_inventory'].out_shapes() + [obs_shapes['latent_screen']], concat_dim=-1) # Add latent dense block self.perception_dict['latent_dense'] = DenseBlock( in_keys='latent_concat', out_keys='latent_dense', hidden_units=hidden_units, non_lin=non_lin, in_shapes=self.perception_dict['latent_concat'].out_shapes()) # Add recurrent block self.perception_dict['latent'] = LSTMLastStepBlock( in_keys='latent_dense', out_keys='latent', in_shapes=self.perception_dict['latent_dense'].out_shapes(), hidden_size=32, num_layers=1, bidirectional=False, non_lin=non_lin) # build action heads self.perception_dict['value'] = LinearOutputBlock( in_keys='latent', out_keys='value', in_shapes=self.perception_dict['latent'].out_shapes(), output_units=1) # build inference block in_keys = list(obs_shapes.keys()) self.perception_net = InferenceBlock( in_keys=in_keys, out_keys='value', in_shapes=[obs_shapes[key] for key in in_keys], perception_blocks=self.perception_dict) # apply weight init self.perception_net.apply(make_module_init_normc(1.0)) self.perception_dict['value'].apply(make_module_init_normc(0.01))
def __init__(self, obs_shapes: Dict[str, Sequence[int]], non_lin: Union[str, type(nn.Module)]): nn.Module.__init__(self) self.obs_shapes = obs_shapes hidden_units = 32 self.perception_dict = OrderedDict() self.perception_dict['order_feat'] = DenseBlock( in_keys='ordered_piece', out_keys='order_feat', in_shapes=self.obs_shapes['ordered_piece'], hidden_units=[hidden_units], non_lin=non_lin) self.perception_dict['selected_feat'] = DenseBlock( in_keys='selected_piece', out_keys='selected_feat', in_shapes=self.obs_shapes['selected_piece'], hidden_units=[hidden_units], non_lin=non_lin) self.perception_dict['latent'] = ConcatenationBlock( in_keys=['order_feat', 'selected_feat'], out_keys='latent', in_shapes=[[hidden_units], [hidden_units], [hidden_units]], concat_dim=-1) self.perception_dict['value'] = LinearOutputBlock( in_keys='latent', out_keys='value', in_shapes=self.perception_dict['latent'].out_shapes(), output_units=1) in_keys = ['ordered_piece', 'selected_piece'] self.perception_net = InferenceBlock( in_keys=in_keys, out_keys='value', in_shapes=[self.obs_shapes[key] for key in in_keys], perception_blocks=self.perception_dict) # initialize model weights self.perception_net.apply(make_module_init_normc(1.0)) self.perception_dict['value'].apply(make_module_init_normc(0.01))
class ValueNet(nn.Module): """Simple feed forward value network. """ def __init__(self, obs_shapes, non_lin=nn.Tanh): super().__init__() # build perception part self.perception_network = DenseBlock(in_keys="observation", out_keys="latent", in_shapes=obs_shapes['observation'], hidden_units=[32, 32], non_lin=non_lin) module_init = make_module_init_normc(std=1.0) self.perception_network.apply(module_init) # build action head self.value_head = LinearOutputBlock(in_keys="latent", out_keys="value", in_shapes=self.perception_network.out_shapes(), output_units=1) module_init = make_module_init_normc(std=0.01) self.value_head.apply(module_init) # compile inference model self.net = InferenceBlock(in_keys="observation", out_keys="value", in_shapes=list(obs_shapes.values()), perception_blocks={"latent": self.perception_network, "value": self.value_head}) def forward(self, x): """ forward pass. """ return self.net(x)
def __init__(self, obs_shapes: Dict[str, Sequence[int]], action_spaces_dict: Dict[Union[str, int], spaces.Space], non_lin: Union[str, type(nn.Module)]): super().__init__() self.obs_shapes = obs_shapes # build perception part self.perception_dict = OrderedDict() self.perception_dict['latent-obs'] = DenseBlock( in_keys="observation", out_keys="latent-obs", in_shapes=obs_shapes['observation'], hidden_units=[256], non_lin=non_lin) self.perception_dict['latent-act'] = DenseBlock( in_keys="action", out_keys="latent-act", in_shapes=obs_shapes['action'], hidden_units=[256], non_lin=non_lin) self.perception_dict['concat'] = ConcatenationBlock( in_keys=['latent-obs', 'latent-act'], in_shapes=self.perception_dict['latent-obs'].out_shapes() + self.perception_dict['latent-act'].out_shapes(), concat_dim=-1, out_keys='concat') self.perception_dict['latent'] = DenseBlock( in_keys="concat", out_keys="latent", in_shapes=self.perception_dict['concat'].out_shapes(), hidden_units=[256], non_lin=non_lin) # build action head self.perception_dict['q_value'] = LinearOutputBlock( in_keys="latent", out_keys="q_value", in_shapes=self.perception_dict['latent'].out_shapes(), output_units=1) self.perception_net = InferenceBlock( in_keys=['observation', 'action'], out_keys='q_value', in_shapes=[ self.obs_shapes['observation'], self.obs_shapes['action'] ], perception_blocks=self.perception_dict) # initialize model weights self.perception_net.apply(make_module_init_normc(1.0)) self.perception_dict['q_value'].apply(make_module_init_normc(0.01))
def __init__(self, obs_shapes: Dict[str, Sequence[int]], action_logits_shapes: Dict[str, Sequence[int]], non_lin: Union[str, type(nn.Module)]): super().__init__() self.obs_shapes = obs_shapes action_key = list(action_logits_shapes.keys())[0] # build perception part self.perception_dict = OrderedDict() self.perception_dict['embedding'] = DenseBlock( in_keys="observation", out_keys="embedding", in_shapes=obs_shapes['observation'], hidden_units=[256, 256], non_lin=non_lin) # build action head self.perception_dict[action_key] = LinearOutputBlock( in_keys="embedding", out_keys=action_key, in_shapes=self.perception_dict['embedding'].out_shapes(), output_units=action_logits_shapes[action_key][0]) self.perception_net = InferenceBlock( in_keys='observation', out_keys=action_key, in_shapes=[self.obs_shapes['observation']], perception_blocks=self.perception_dict) # initialize model weights self.perception_net.apply(make_module_init_normc(1.0)) self.perception_dict[action_key].apply(make_module_init_normc(0.01))
def __init__(self, in_keys: Union[str, List[str]], out_keys: Union[str, List[str]], in_shapes: Union[Sequence[int], List[Sequence[int]]], hidden_channels: List[int], hidden_units: List[int], non_lin: Union[str, type(nn.Module)]): super().__init__(in_keys=in_keys, out_keys=out_keys, in_shapes=in_shapes) out_keys_conv = [f"{k}_conv" for k in self.out_keys] self.conv_block = VGGConvolutionBlock(in_keys=in_keys, out_keys=out_keys_conv, in_shapes=in_shapes, hidden_channels=hidden_channels, non_lin=non_lin) out_keys_flatten = [f"{k}_flat" for k in out_keys_conv ] if len(hidden_units) > 0 else out_keys self.flatten_block = FlattenBlock( in_keys=out_keys_conv, out_keys=out_keys_flatten, in_shapes=self.conv_block.out_shapes(), num_flatten_dims=3) self.dense_block = None if len(hidden_units) > 0: self.dense_block = DenseBlock( in_keys=out_keys_flatten, out_keys=out_keys, in_shapes=self.flatten_block.out_shapes(), hidden_units=hidden_units, non_lin=non_lin)
def __init__(self, obs_shapes: Dict[str, Sequence[int]], non_lin: type(nn.Module)): super().__init__(obs_shapes, non_lin) self.perception_dict['value_head_net'] = DenseBlock( in_keys='hidden_out', in_shapes=self.perception_dict['hidden_out'].out_shapes(), out_keys='value_head_net', hidden_units=[5, 2], non_lin=non_lin) self.perception_dict['value'] = LinearOutputBlock( in_keys='value_head_net', in_shapes=self.perception_dict['value_head_net'].out_shapes(), out_keys='value', output_units=1) # Set up inference block self.perception_net = InferenceBlock( in_keys=list(self.obs_shapes.keys()), out_keys='value', in_shapes=[self.obs_shapes[key] for key in self.obs_shapes.keys()], perception_blocks=self.perception_dict) # initialize model weights self.perception_net.apply(make_module_init_normc(1.0)) self.perception_dict['value'].apply(make_module_init_normc(0.01))
def __init__(self, obs_shapes: Dict[str, Sequence[int]], head_units: List[int], non_lin: nn.Module): super().__init__() self.perception_dict: Dict[str, PerceptionBlock] = dict() # build action head # build perception part self.perception_dict["head"] = DenseBlock( in_keys="latent", out_keys="head", in_shapes=obs_shapes["latent"], hidden_units=head_units, non_lin=non_lin) self.perception_dict["value"] = LinearOutputBlock( in_keys="head", out_keys="value", in_shapes=self.perception_dict["head"].out_shapes(), output_units=1) self.perception_dict['head'].apply(make_module_init_normc(std=1.0)) self.perception_dict["value"].apply(make_module_init_normc(std=0.01)) # compile inference model self.net = InferenceBlock(in_keys=list(obs_shapes.keys()), out_keys="value", in_shapes=list(obs_shapes.values()), perception_blocks=self.perception_dict)
def __init__(self, obs_shapes: Dict[str, Sequence[int]], action_logits_shapes: Dict[str, Sequence[int]], hidden_units: List[int], head_units: List[int], non_lin=nn.Module): super().__init__(obs_shapes, hidden_units, non_lin) # build perception part self.perception_dict["head"] = DenseBlock( in_keys="latent", out_keys="head", in_shapes=self.perception_dict["latent"].out_shapes(), hidden_units=head_units, non_lin=self.non_lin) self.perception_dict['head'].apply(make_module_init_normc(std=1.0)) # build action head for action, shape in action_logits_shapes.items(): self.perception_dict[action] = LinearOutputBlock( in_keys="head", out_keys=action, in_shapes=self.perception_dict["head"].out_shapes(), output_units=action_logits_shapes[action][-1]) module_init = make_module_init_normc(std=0.01) self.perception_dict[action].apply(module_init) # compile inference model self.net = InferenceBlock(in_keys=list(obs_shapes.keys()), out_keys=list(action_logits_shapes.keys()) + ['latent'], in_shapes=list(obs_shapes.values()), perception_blocks=self.perception_dict)
def __init__(self, obs_shapes: Dict[str, Sequence[int]], action_logits_shapes: Dict[str, Sequence[int]], non_lin: Union[str, type(nn.Module)], hidden_units: List[int]): super().__init__() # Maze relies on dictionaries to represent the inference graph self.perception_dict = OrderedDict() # build latent embedding block self.perception_dict['latent'] = DenseBlock( in_keys='observation', out_keys='latent', in_shapes=obs_shapes['observation'], hidden_units=hidden_units,non_lin=non_lin) # build action head self.perception_dict['action'] = LinearOutputBlock( in_keys='latent', out_keys='action', in_shapes=self.perception_dict['latent'].out_shapes(), output_units=int(np.prod(action_logits_shapes["action"]))) # build inference block self.perception_net = InferenceBlock( in_keys='observation', out_keys='action', in_shapes=obs_shapes['observation'], perception_blocks=self.perception_dict) # apply weight init self.perception_net.apply(make_module_init_normc(1.0)) self.perception_dict['action'].apply(make_module_init_normc(0.01))
def __init__(self, obs_shapes: Dict[str, Sequence[int]], action_logits_shapes: Dict[str, Sequence[int]], non_lin: type(nn.Module)): super().__init__(obs_shapes, non_lin) for action_head_name in action_logits_shapes.keys(): head_hidden_units = [lambda out_shape: out_shape[0] * 5, lambda out_shape: out_shape[0] * 2, lambda out_shape: out_shape[0]] head_hidden_units = [func(action_logits_shapes[action_head_name]) for func in head_hidden_units] self.perception_dict[f'{action_head_name}_net'] = DenseBlock( in_keys='hidden_out', in_shapes=self.perception_dict['hidden_out'].out_shapes(), out_keys=f'{action_head_name}_net', hidden_units=head_hidden_units[:-1], non_lin=non_lin) self.perception_dict[f'{action_head_name}'] = LinearOutputBlock( in_keys=f'{action_head_name}_net', in_shapes=self.perception_dict[f'{action_head_name}_net'].out_shapes(), out_keys=action_head_name, output_units=head_hidden_units[-1] ) # Set up inference block self.perception_net = InferenceBlock( in_keys=list(self.obs_shapes.keys()), out_keys=list(action_logits_shapes.keys()), in_shapes=[self.obs_shapes[key] for key in self.obs_shapes.keys()], perception_blocks=self.perception_dict) self.perception_net.apply(make_module_init_normc(1.0)) for action_head_name in action_logits_shapes.keys(): self.perception_dict[f'{action_head_name}'].apply(make_module_init_normc(0.01))
def __init__(self, obs_shapes: Dict[str, Sequence[int]], non_lin: Union[str, type(nn.Module)]): nn.Module.__init__(self) # initialize the perception dictionary self.perception_dict = OrderedDict() # concatenate all observations in dictionary self.perception_dict['concat'] = ConcatenationBlock( in_keys=[ 'cart_position', 'cart_velocity', 'pole_angle', 'pole_angular_velocity' ], out_keys='concat', in_shapes=[ obs_shapes['cart_position'], obs_shapes['cart_velocity'], obs_shapes['pole_angle'], obs_shapes['pole_angular_velocity'] ], concat_dim=-1) # process concatenated representation with two dense layers self.perception_dict['embedding'] = DenseBlock( in_keys='concat', in_shapes=self.perception_dict['concat'].out_shapes(), hidden_units=[128, 128], non_lin=non_lin, out_keys='embedding') # add a linear output block self.perception_dict['value'] = LinearOutputBlock( in_keys='embedding', out_keys='value', in_shapes=self.perception_dict['embedding'].out_shapes(), output_units=1) # compile an inference block self.perception_net = InferenceBlock( in_keys=[ 'cart_position', 'cart_velocity', 'pole_angle', 'pole_angular_velocity' ], out_keys='value', in_shapes=[ obs_shapes[key] for key in [ 'cart_position', 'cart_velocity', 'pole_angle', 'pole_angular_velocity' ] ], perception_blocks=self.perception_dict) # initialize model weights self.perception_net.apply(make_module_init_normc(1.0)) self.perception_dict['value'].apply(make_module_init_normc(0.01))
def __init__(self, obs_shapes, action_logits_shapes, non_lin=nn.Tanh): super().__init__() # build perception part self.perception_network = DenseBlock(in_keys="observation", out_keys="latent", in_shapes=obs_shapes['observation'], hidden_units=[32, 32], non_lin=non_lin) module_init = make_module_init_normc(std=1.0) self.perception_network.apply(module_init) # build action head self.action_head = LinearOutputBlock(in_keys="latent", out_keys="action", in_shapes=self.perception_network.out_shapes(), output_units=action_logits_shapes['action'][-1]) module_init = make_module_init_normc(std=0.01) self.action_head.apply(module_init) # compile inference model self.net = InferenceBlock(in_keys="observation", out_keys="action", in_shapes=list(obs_shapes.values()), perception_blocks={"latent": self.perception_network, "action": self.action_head})
def __init__(self, in_keys: Union[str, List[str]], out_keys: Union[str, List[str]], in_shapes: Union[Sequence[int], List[Sequence[int]]], hidden_channels: List[int], hidden_kernels: List[Union[int, Tuple[int, ...]]], convolution_dimension: int, hidden_strides: Optional[List[Union[int, Tuple[int, ...]]]], hidden_dilations: Optional[List[Union[int, Tuple[int, ...]]]], hidden_padding: Optional[List[Union[int, Tuple[int, ...]]]], padding_mode: Optional[str], hidden_units: List[int], non_lin: Union[str, type(nn.Module)]): super().__init__(in_keys=in_keys, out_keys=out_keys, in_shapes=in_shapes) out_keys_conv = [f"{k}_conv" for k in self.out_keys] self.conv_block = StridedConvolutionBlock( in_keys=in_keys, out_keys=out_keys_conv, in_shapes=in_shapes, hidden_channels=hidden_channels, hidden_kernels=hidden_kernels, convolution_dimension=convolution_dimension, hidden_strides=hidden_strides, hidden_dilations=hidden_dilations, hidden_padding=hidden_padding, padding_mode=padding_mode, non_lin=non_lin) out_keys_flatten = [f"{k}_flat" for k in out_keys_conv ] if len(hidden_units) > 0 else out_keys self.flatten_block = FlattenBlock( in_keys=out_keys_conv, out_keys=out_keys_flatten, in_shapes=self.conv_block.out_shapes(), num_flatten_dims=3) if len(hidden_units) > 0: self.dense_block = DenseBlock( in_keys=out_keys_flatten, out_keys=out_keys, in_shapes=self.flatten_block.out_shapes(), hidden_units=hidden_units, non_lin=non_lin) else: self.dense_block = None
def __init__(self, obs_shapes: Dict[str, Sequence[int]], non_lin: type(nn.Module)): nn.Module.__init__(self) self.obs_shapes = obs_shapes perception_dict: Dict[str, PerceptionBlock] = dict() for in_key, in_shape in self.obs_shapes.items(): if len(in_shape) > 1: next_in_key = f'{in_key}_flat' perception_dict[next_in_key] = FlattenBlock( in_keys=in_key, in_shapes=in_shape, out_keys=next_in_key, num_flatten_dims=len(in_shape)) next_in_shape = perception_dict[next_in_key].out_shapes() else: next_in_key = in_key next_in_shape = in_shape perception_dict[f'{in_key}_encoded_feat'] = DenseBlock( in_keys=next_in_key, in_shapes=next_in_shape, out_keys=f'{in_key}_encoded_feat', non_lin=non_lin, hidden_units=[16]) perception_dict[f'{in_key}_encoded_layer'] = LinearOutputBlock( in_keys=f'{in_key}_encoded_feat', in_shapes=perception_dict[f'{in_key}_encoded_feat'].out_shapes( ), out_keys=f'{in_key}_encoded_layer', output_units=8) concat_in_keys = [ key for key in perception_dict.keys() if '_encoded_layer' in key ] perception_dict['hidden_out'] = ConcatenationBlock( in_keys=concat_in_keys, in_shapes=sum( [perception_dict[key].out_shapes() for key in concat_in_keys], []), out_keys='hidden_out', concat_dim=-1) self.perception_dict = perception_dict
def __init__(self, obs_shapes: Dict[str, Sequence[int]], hidden_units: List[int], non_lin: nn.Module): super().__init__() self.hidden_units = hidden_units self.non_lin = non_lin self.perception_dict: Dict[str, PerceptionBlock] = dict() # first, flatten all observations flat_keys = [] for obs, shape in obs_shapes.items(): out_key = f'{obs}_flat' flat_keys.append(out_key) self.perception_dict[out_key] = FlattenBlock( in_keys=obs, out_keys=out_key, in_shapes=shape, num_flatten_dims=len(shape)) # next, concatenate flat observations in_shapes = [ self.perception_dict[k].out_shapes()[0] for k in flat_keys ] self.perception_dict["concat"] = ConcatenationBlock( in_keys=flat_keys, out_keys='concat', in_shapes=in_shapes, concat_dim=-1) # build perception part self.perception_dict["latent"] = DenseBlock( in_keys="concat", out_keys="latent", in_shapes=self.perception_dict["concat"].out_shapes(), hidden_units=self.hidden_units, non_lin=self.non_lin) # initialize model weights module_init = make_module_init_normc(std=1.0) for key in self.perception_dict.keys(): self.perception_dict[key].apply(module_init)
def build_perception_dict(): """ helper function """ in_dict = build_multi_input_dict(dims=[[100, 1, 16], [100, 1, 8]]) perception_dict = dict() for in_key, in_tensor in in_dict.items(): # compile network block net = DenseBlock(in_keys=in_key, out_keys=f"{in_key}_feat", in_shapes=[in_tensor.shape[-1:]], hidden_units=[32, 32], non_lin=nn.ReLU) perception_dict[f"{in_key}_feat"] = net net = ConcatenationBlock(in_keys=list(perception_dict.keys()), out_keys="concat", in_shapes=[(32, ), (32, )], concat_dim=-1) perception_dict["concat"] = net return in_dict, perception_dict
def build_perception_dict(): """ helper function """ obs_keys = ["obs_inventory", "obs_screen"] in_dict = build_multi_input_dict(dims=[[1, 16], [1, 3, 64, 64]], names=obs_keys) perception_dict = dict() # --- block --- net = DenseBlock(in_keys="obs_inventory", out_keys="obs_inventory_latent", in_shapes=[in_dict["obs_inventory"].shape[-1:]], hidden_units=[32, 32], non_lin=nn.ReLU) perception_dict["obs_inventory_latent"] = net # --- block --- net = VGGConvolutionDenseBlock(in_keys="obs_screen", out_keys="obs_screen_latent", in_shapes=[in_dict["obs_screen"].shape[-3:]], hidden_channels=[8, 16, 32], hidden_units=[32], non_lin=nn.ReLU) perception_dict["obs_screen_latent"] = net # --- block --- net = ConcatenationBlock(in_keys=list(perception_dict.keys()), out_keys="concat", in_shapes=[(32,), (32,)], concat_dim=-1) perception_dict["concat"] = net # --- block --- net = LinearOutputBlock(in_keys=["concat"], out_keys="action_move", in_shapes=[(64,)], output_units=4) perception_dict["action_move"] = net # --- block --- net = LinearOutputBlock(in_keys=["concat"], out_keys="action_use", in_shapes=[(64,)], output_units=16) perception_dict["action_use"] = net return in_dict, perception_dict
def __init__(self, in_keys: Union[str, List[str]], out_keys: Union[str, List[str]], in_shapes: Union[Sequence[int], List[Sequence[int]]], num_flatten_dims: int, hidden_units: List[int], non_lin: Union[str, type(nn.Module)]): super().__init__(in_keys=in_keys, out_keys=out_keys, in_shapes=in_shapes) out_keys_flatten = [f"{k}_flat" for k in self.out_keys] self.flatten_block = FlattenBlock(in_keys=self.in_keys, out_keys=out_keys_flatten, in_shapes=self.in_shapes, num_flatten_dims=num_flatten_dims) self.dense_block = DenseBlock( in_keys=out_keys_flatten, out_keys=out_keys, in_shapes=self.flatten_block.out_shapes(), hidden_units=hidden_units, non_lin=non_lin)
def test_mlp_and_concat(): """ perception test """ in_dict = build_multi_input_dict(dims=[[100, 1, 16], [100, 1, 8]]) feat_dict = dict() for in_key, in_tensor in in_dict.items(): # compile network block net = DenseBlock(in_keys=in_key, out_keys=f"{in_key}_feat", in_shapes=(in_tensor.shape[-1], ), hidden_units=[32, 32], non_lin=nn.ReLU) # update output dictionary feat_dict.update(net(in_dict)) net = ConcatenationBlock(in_keys=list(feat_dict.keys()), out_keys="concat", in_shapes=[(32, ), (32, )], concat_dim=-1) out_dict = net(feat_dict) assert out_dict["concat"].ndim == 3 assert out_dict["concat"].shape[-1] == 64
def __init__(self, obs_shapes: Dict[str, Sequence[int]], action_logits_shapes: Dict[str, Sequence[int]], non_lin: Union[str, type(nn.Module)], with_mask: bool): nn.Module.__init__(self) self.obs_shapes = obs_shapes hidden_units = 32 self.perception_dict = OrderedDict() self.perception_dict['selected_feat'] = DenseBlock( in_keys='selected_piece', out_keys='selected_feat', in_shapes=self.obs_shapes['selected_piece'], hidden_units=[hidden_units], non_lin=non_lin) self.perception_dict['order_feat'] = DenseBlock( in_keys='ordered_piece', out_keys='order_feat', in_shapes=self.obs_shapes['ordered_piece'], hidden_units=[hidden_units], non_lin=non_lin) self.perception_dict['latent'] = ConcatenationBlock( in_keys=['selected_feat', 'order_feat'], out_keys='latent', in_shapes=[[hidden_units], [hidden_units]], concat_dim=-1) rotation_out_key = 'cut_rotation_logits' if with_mask else 'cut_rotation' self.perception_dict[rotation_out_key] = LinearOutputBlock( in_keys='latent', out_keys=rotation_out_key, in_shapes=self.perception_dict['latent'].out_shapes(), output_units=action_logits_shapes['cut_rotation'][0]) if with_mask: self.perception_dict['cut_rotation'] = ActionMaskingBlock( in_keys=['cut_rotation_logits', 'cutting_mask'], out_keys='cut_rotation', in_shapes=self.perception_dict['cut_rotation_logits']. out_shapes() + [self.obs_shapes['cutting_mask']], num_actors=1, num_of_actor_actions=None) self.perception_dict['cut_order'] = LinearOutputBlock( in_keys='latent', out_keys='cut_order', in_shapes=self.perception_dict['latent'].out_shapes(), output_units=action_logits_shapes['cut_order'][0]) in_keys = ['selected_piece', 'ordered_piece'] if with_mask: in_keys.append('cutting_mask') self.perception_net = InferenceBlock( in_keys=in_keys, out_keys=['cut_rotation', 'cut_order'], in_shapes=[self.obs_shapes[key] for key in in_keys], perception_blocks=self.perception_dict) # initialize model weights self.perception_net.apply(make_module_init_normc(1.0)) self.perception_dict[rotation_out_key].apply( make_module_init_normc(0.01)) self.perception_dict['cut_order'].apply(make_module_init_normc(0.01))
def __init__(self, obs_shapes: Dict[str, Sequence[int]], action_logits_shapes: Dict[str, Sequence[int]], non_lin: Union[str, type(nn.Module)], with_mask: bool): nn.Module.__init__(self) self.obs_shapes = obs_shapes hidden_units, embedding_dim = 32, 7 self.perception_dict = OrderedDict() # embed inventory # --------------- self.perception_dict['inventory_feat'] = DenseBlock( in_keys='inventory', out_keys='inventory_feat', in_shapes=self.obs_shapes['inventory'], hidden_units=[hidden_units], non_lin=non_lin) self.perception_dict['inventory_embed'] = LinearOutputBlock( in_keys='inventory_feat', out_keys='inventory_embed', in_shapes=self.perception_dict['inventory_feat'].out_shapes(), output_units=embedding_dim) # embed ordered_piece # ------------------_ self.perception_dict['order_unsqueezed'] = FunctionalBlock( in_keys='ordered_piece', out_keys='order_unsqueezed', in_shapes=self.obs_shapes['ordered_piece'], func=lambda x: torch.unsqueeze(x, dim=-2)) self.perception_dict['order_feat'] = DenseBlock( in_keys='order_unsqueezed', out_keys='order_feat', in_shapes=self.perception_dict['order_unsqueezed'].out_shapes(), hidden_units=[hidden_units], non_lin=non_lin) self.perception_dict['order_embed'] = LinearOutputBlock( in_keys='order_feat', out_keys='order_embed', in_shapes=self.perception_dict['order_feat'].out_shapes(), output_units=embedding_dim) # compute dot product score # ------------------------- in_shapes = self.perception_dict['inventory_embed'].out_shapes() in_shapes += self.perception_dict['order_embed'].out_shapes() out_key = 'corr_score' if with_mask else 'piece_idx' self.perception_dict[out_key] = CorrelationBlock( in_keys=['inventory_embed', 'order_embed'], out_keys=out_key, in_shapes=in_shapes, reduce=True) # apply action masking if with_mask: self.perception_dict['piece_idx'] = ActionMaskingBlock( in_keys=['corr_score', 'inventory_mask'], out_keys='piece_idx', in_shapes=self.perception_dict['corr_score'].out_shapes() + [self.obs_shapes['inventory_mask']], num_actors=1, num_of_actor_actions=None) assert self.perception_dict['piece_idx'].out_shapes( )[0][0] == action_logits_shapes['piece_idx'][0] in_keys = ['ordered_piece', 'inventory'] if with_mask: in_keys.append('inventory_mask') self.perception_net = InferenceBlock( in_keys=in_keys, out_keys='piece_idx', in_shapes=[self.obs_shapes[key] for key in in_keys], perception_blocks=self.perception_dict) # initialize model weights self.perception_net.apply(make_module_init_normc(1.0)) self.perception_dict['inventory_embed'].apply( make_module_init_normc(0.01)) self.perception_dict['order_embed'].apply(make_module_init_normc(0.01))