def __init__(self,
embed_size: int,
num_fields: int,
deep_output_size: int,
deep_layer_sizes: List[int],
ffm_dropout_p: Optional[float] = None,
deep_dropout_p: Optional[List[float]] = None,
deep_activation: Optional[nn.Module] = nn.ReLU()):
"""
Initialize DeepFieldAwareFactorizationMachineModel
Args:
embed_size (int): size of embedding tensor
num_fields (int): number of inputs' fields
deep_output_size (int): output size of dense network
deep_layer_sizes (List[int]): layer sizes of dense network
ffm_dropout_p (float, optional): probability of Dropout in FFM. Defaults to 0.0
deep_dropout_p (List[float], optional): probability of Dropout in dense network. Defaults to None
deep_activation (torch.nn.Module, optional): activation function of dense network. Defaults to nn.ReLU()
"""
super().__init__()
self.ffm = FFMLayer(num_fields=num_fields, dropout_p=ffm_dropout_p)
inputs_size = combination(num_fields, 2)
inputs_size *= embed_size
self.deep = DNNLayer(inputs_size=inputs_size,
output_size=deep_output_size,
layer_sizes=deep_layer_sizes,
dropout_p=deep_dropout_p,
activation=deep_activation)
def __init__(self,
embed_size: int,
num_fields: int,
bilinear_type: str = "all",
bias: bool = True):
"""Initialize BilinearInteractionLayer
Args:
embed_size (int): Size of embedding tensor
num_fields (int): Number of inputs' fields
bilinear_type (str, optional): Type of bilinear to calculate interactions.
Defaults to "all".
bias (bool, optional): Flag to control using bias.
Defaults to True.
Attribute:
row_idx (T), dtype = torch.long: 1st indices to index inputs in 2nd dimension for inner product.
col_idx (T), dtype = torch.long: 2nd indices to index inputs in 2nd dimension for inner product.
bilinear (nn.Module): Module of bilinear-interaction.
bilinear_type (str): Type of bilinear to calculate interactions.
Raises:
NotImplementedError: /
ValueError: when bilinear_type is not in ["all", "each", "interaction"]
"""
# Refer to parent class
super(BilinearInteractionLayer, self).__init__()
# Create row_idx and col_idx to index inputs for outer product
self.row_idx = list()
self.col_idx = list()
for i in range(num_fields - 1):
for j in range(i + 1, num_fields):
self.row_idx.append(i)
self.col_idx.append(j)
self.row_idx = torch.LongTensor(self.row_idx)
self.col_idx = torch.LongTensor(self.col_idx)
# calculate number of interactive fields
num_interaction = combination(num_fields, 2)
if bilinear_type == "all":
self.bilinear = FieldAllTypeBilinear(embed_size,
embed_size,
bias=bias)
elif bilinear_type == "each":
self.bilinear = FieldEachTypeBilinear(num_interaction,
embed_size,
embed_size,
bias=bias)
elif bilinear_type == "interaction":
raise NotImplementedError()
# self.bilinear = FieldInteractionTypeBilinear(num_interaction, embed_size, embed_size, bias=bias)
else:
raise ValueError(
'bilinear_type only allows: ["all", "each", "interaction"].')
# Bind kernel_type to kernel_type
self.bilinear_type = bilinear_type
def __init__(self,
embed_size: int,
num_fields: int,
deep_layer_sizes: List[int],
output_size: int = 1,
prod_method: str = 'inner',
use_bias: Optional[bool] = True,
deep_dropout_p: Optional[List[float]] = None,
deep_activation: Optional[nn.Module] = nn.ReLU(),
**kwargs):
"""
Initialize ProductNeuralNetworkModel
Args:
embed_size (int): size of embedding tensor
num_fields (int): number of inputs' fields
deep_layer_sizes (List[int]): layer sizes of dense network
output_size (int): output size of model. i.e. output size of dense network. Defaults to 1
prod_method (str): method of product neural network. Allow: ["inner", "outer"]. Defaults to inner
use_bias (bool, optional): whether the bias constant is concatenated to the input. Defaults to True
deep_dropout_p (List[float], optional): probability of Dropout in dense network. Defaults to None
deep_activation (torch.nn.Module, optional): activation function of dense network. Defaults to nn.ReLU()
Arguments:
kernel_type (str): type of kernel to compress outer-product.
"""
super().__init__()
if prod_method == 'inner':
self.pnn = InnerProductNetworkLayer(num_fields=num_fields)
elif prod_method == 'outer':
self.pnn = OuterProductNetworkLayer(embed_size=embed_size,
num_fields=num_fields,
kernel_type=kwargs.get(
'kernel_type', 'mat'))
else:
raise ValueError(
f'{prod_method} is not allowed in prod_method. Required: ["inner", "outer"].'
)
self.use_bias = use_bias
cat_size = combination(num_fields, 2) + num_fields
if self.use_bias:
cat_size += 1
self.deep = DNNLayer(output_size=output_size,
layer_sizes=deep_layer_sizes,
inputs_size=cat_size,
dropout_p=deep_dropout_p,
activation=deep_activation)
if self.use_bias:
self.bias = nn.Parameter(torch.zeros((1, 1), names=(
'B',
'O',
)))
nn.init.uniform_(self.bias.data)
def __init__(self,
embed_size: int,
num_fields: int,
deep_output_size: int,
deep_layer_sizes: List[int],
reduction: int,
ffm_dropout_p: float = 0.0,
deep_dropout_p: List[float] = None,
deep_activation: Callable[[torch.Tensor],
torch.Tensor] = nn.ReLU()):
"""Initialize FieldAttentiveDeepFieldAwareFactorizationMachineModel
Args:
embed_size (int): Size of embedding tensor
num_fields (int): Number of inputs' fields
deep_output_size (int): Output size of dense network
deep_layer_sizes (List[int]): Layer sizes of dense network
reduction (int): Reduction of CIN layer
ffm_dropout_p (float, optional): Probability of Dropout in FFM.
Defaults to 0.0.
deep_dropout_p (List[float], optional): Probability of Dropout in dense network.
Defaults to None.
deep_activation (Callable[[T], T], optional): Activation function of dense network.
Defaults to nn.ReLU().
Attributes:
cen (nn.Module): Module of compose excitation network layer.
ffm (nn.Module): Module of field-aware factorization machine layer.
deep (nn.Module): Module of dense layer.
"""
# refer to parent class
super(FieldAttentiveDeepFieldAwareFactorizationMachineModel,
self).__init__()
# initialize compose excitation network
self.cen = CENLayer(num_fields, reduction)
# initialize ffm layer
self.ffm = FFMLayer(num_fields=num_fields, dropout_p=ffm_dropout_p)
# calculate the output's size of ffm, i.e. inputs' size of DNNLayer
inputs_size = combination(num_fields, 2)
inputs_size *= embed_size
# initialize dense layer
self.deep = DNNLayer(inputs_size=inputs_size,
output_size=deep_output_size,
layer_sizes=deep_layer_sizes,
dropout_p=deep_dropout_p,
activation=deep_activation)
def __init__(self,
embed_size: int,
num_fields: int,
bilinear_type: str = 'all',
bias: bool = True):
"""
Initialize BilinearInteractionLayer
Args:
embed_size (int): size of embedding tensor
num_fields (int): number of inputs' fields
bilinear_type (str, optional): type of bilinear to calculate interactions. Defaults to "all"
bias (bool, optional): flag to control using bias. Defaults to True
Raises:
NotImplementedError: /
ValueError: when bilinear_type is not in ["all", "each", "interaction"]
"""
super().__init__()
self.row_idx = []
self.col_idx = []
for i in range(num_fields - 1):
for j in range(i + 1, num_fields):
self.row_idx.append(i)
self.col_idx.append(j)
self.row_idx = torch.LongTensor(self.row_idx)
self.col_idx = torch.LongTensor(self.col_idx)
num_interaction = combination(num_fields, 2)
self.bilinear_type = bilinear_type
if bilinear_type == 'all':
self.bilinear = FieldAllTypeBilinear(embed_size,
embed_size,
bias=bias)
elif bilinear_type == 'each':
self.bilinear = FieldEachTypeBilinear(num_interaction,
embed_size,
embed_size,
bias=bias)
elif bilinear_type == 'interaction':
# self.bilinear = FieldInteractionTypeBilinear(num_interaction, embed_size, embed_size, bias=bias)
raise NotImplementedError()
else:
raise ValueError(
'bilinear_type only allows: ["all", "each", "interaction"].')
def __init__(self,
embed_size: int,
num_fields: int,
senet_reduction: int,
deep_output_size: int,
deep_layer_sizes: List[int],
bilinear_type: Optional[str] = 'all',
bilinear_bias: Optional[bool] = True,
deep_dropout_p: Optional[List[float]] = None,
deep_activation: Optional[nn.Module] = nn.ReLU()):
"""
Initialize FeatureImportanceAndBilinearFeatureInteractionNetwork
Args:
embed_size (int): size of embedding tensor
num_fields (int): number of inputs' fields
senet_reduction (int): size of reduction in dense layer of senet.
deep_output_size (int): output size of dense network
deep_layer_sizes (List[int]): layer sizes of dense network
bilinear_type (str, optional): type of bilinear to calculate interactions. Defaults to "all"
bilinear_bias (bool, optional): flag to control using bias in bilinear-interactions. Defaults to True
deep_dropout_p (List[float], optional): probability of Dropout in dense network. Defaults to None
deep_activation (torch.nn.Module, optional): activation function of dense network. Defaults to nn.ReLU()
"""
super().__init__()
inputs_size = combination(num_fields, 2)
inputs_size = inputs_size * embed_size * 2
self.senet = SENETLayer(num_fields, senet_reduction, squared=False)
self.emb_bilinear = BilinearInteractionLayer(embed_size, num_fields,
bilinear_type,
bilinear_bias)
self.senet_bilinear = BilinearInteractionLayer(embed_size, num_fields,
bilinear_type,
bilinear_bias)
self.deep = DNNLayer(inputs_size=inputs_size,
output_size=deep_output_size,
layer_sizes=deep_layer_sizes,
dropout_p=deep_dropout_p,
activation=deep_activation)
def __init__(self,
embed_size: int,
num_fields: int,
deep_layer_sizes: List[int],
output_size: int = 1,
prod_method: str = "inner",
deep_dropout_p: List[float] = None,
deep_activation: Callable[[torch.Tensor],
torch.Tensor] = nn.ReLU(),
**kwargs):
r"""Initialize ProductNeuralNetworkModel
Args:
embed_size (int): Size of embedding tensor
num_fields (int): Number of inputs' fields
deep_layer_sizes (List[int]): Layer sizes of dense network
output_size (int): Output size of model
i.e. output size of dense network.
Defaults to 1.
prod_method (str): Method of product neural network.
Allow: [inner, outer].
Defaults to inner.
deep_dropout_p (List[float], optional): Probability of Dropout in dense network.
Defaults to None.
deep_activation (Callable[[T], T], optional): Activation function of dense network.
Defaults to nn.ReLU().
Arguments:
kernel_type (str): Type of kernel to compress outer-product.
Attributes:
pnn (nn.Module): Module of product neural network.
deep (nn.Module): Module of dense layer.
bias (nn.Parameter): Parameter of bias of field-aware factorization machine.
Raises:
ValueError: when prod_method is not in [inner, outer].
"""
# Refer to parent class
super(ProductNeuralNetworkModel, self).__init__()
# Initialize product network
if prod_method == "inner":
self.pnn = InnerProductNetworkLayer(num_fields=num_fields)
elif prod_method == "outer":
self.pnn = OuterProductNetworkLayer(embed_size=embed_size,
num_fields=num_fields,
kernel_type=kwargs.get(
"kernel_type", "mat"))
else:
raise ValueError(
"'%s' is not allowed in prod_method. Please use ['inner', 'outer']."
)
# Calculate size of inputs of dense layer
cat_size = combination(num_fields, 2) + num_fields + 1
# Initialize dense layer
self.deep = DNNLayer(output_size=output_size,
layer_sizes=deep_layer_sizes,
inputs_size=cat_size,
dropout_p=deep_dropout_p,
activation=deep_activation)
# Initialize bias parameter
self.bias = nn.Parameter(torch.zeros((1, 1), names=("B", "O")))
nn.init.uniform_(self.bias.data)
