class TabNetCombinerConfig:
size: int = schema.PositiveInteger(default=32) # N_a in the paper
output_size: int = schema.PositiveInteger(default=32) # N_d in the paper
num_steps: int = schema.NonNegativeInteger(default=1) # N_steps in the paper
num_total_blocks: int = schema.NonNegativeInteger(default=4)
num_shared_blocks: int = schema.NonNegativeInteger(default=2)
relaxation_factor: float = 1.5 # gamma in the paper
bn_epsilon: float = 1e-3
bn_momentum: float = 0.7 # m_B in the paper
# B_v from the paper
bn_virtual_bs: Optional[int] = schema.PositiveInteger()
sparsity: float = 1e-5 # lambda_sparse in the paper
dropout: float = schema.FloatRange(default=0.0, min=0, max=1)
class Meta:
unknown = INCLUDE
class TransformerCombinerConfig:
num_layers: int = schema.PositiveInteger(default=1)
hidden_size: int = schema.NonNegativeInteger(default=256)
num_heads: int = schema.NonNegativeInteger(default=8)
transformer_fc_size: int = schema.NonNegativeInteger(default=256)
dropout: float = schema.FloatRange(default=0.1, min=0, max=1)
fc_layers: Optional[List[Dict[str, Any]]] = schema.DictList()
num_fc_layers: int = schema.NonNegativeInteger(default=0)
fc_size: int = schema.PositiveInteger(default=256)
use_bias: bool = True
weights_initializer: Union[str, Dict] = schema.InitializerOrDict(default="xavier_uniform")
bias_initializer: Union[str, Dict] = schema.InitializerOrDict(default="zeros")
norm: Optional[str] = schema.StringOptions(["batch", "layer"])
norm_params: Optional[dict] = schema.Dict()
fc_activation: str = "relu"
fc_dropout: float = schema.FloatRange(default=0.0, min=0, max=1)
fc_residual: bool = False
reduce_output: Optional[str] = schema.ReductionOptions(default="mean")
class Meta:
unknown = INCLUDE
class TabNetCombinerConfig:
size: int = schema.PositiveInteger(default=32) # N_a in the paper
output_size: int = schema.PositiveInteger(default=32) # N_d in the paper
num_steps: int = schema.NonNegativeInteger(
default=1) # N_steps in the paper
num_total_blocks: int = schema.NonNegativeInteger(default=4)
num_shared_blocks: int = schema.NonNegativeInteger(default=2)
relaxation_factor: float = 1.5 # gamma in the paper
bn_epsilon: float = 1e-3
bn_momentum: float = 0.7 # m_B in the paper
# B_v from the paper
bn_virtual_bs: Optional[int] = schema.PositiveInteger()
sparsity: float = 1e-5 # lambda_sparse in the paper
entmax_mode: str = schema.StringOptions(
["entmax15", "sparsemax", "constant", "adaptive"], default="sparsemax")
entmax_alpha: float = schema.FloatRange(
default=1.5, min=1, max=2) # 1 corresponds to softmax, 2 is sparsemax.
dropout: float = schema.FloatRange(default=0.0, min=0, max=1)
class Meta:
unknown = INCLUDE
class ConcatCombinerConfig:
fc_layers: Optional[List[Dict[str, Any]]] = schema.DictList()
num_fc_layers: int = schema.NonNegativeInteger(default=0)
fc_size: int = schema.PositiveInteger(default=256)
use_bias: bool = True
weights_initializer: Union[str, Dict] = schema.InitializerOrDict(default="xavier_uniform")
bias_initializer: Union[str, Dict] = schema.InitializerOrDict(default="zeros")
norm: Optional[str] = schema.StringOptions(["batch", "layer"])
norm_params: Optional[dict] = schema.Dict()
activation: str = "relu"
dropout: float = schema.FloatRange(default=0.0, min=0, max=1)
flatten_inputs: bool = False
residual: bool = False
class Meta:
unknown = INCLUDE