def __init__(self, field_dims, embed_dim, mlp_layers=(400, 400, 400), dropouts=(0.5, 0.5)):
super(NeuralFactorizationMachineModel, self).__init__()
self.linear = FeaturesLinear(field_dims) # 线性部分
self.embedding = FeaturesEmbedding(field_dims, embed_dim) # 交互部分
self.fm = FactorizationMachine(embed_dim, reduce_sum=False)
self.mlp = MultiLayerPerceptron(embed_dim, mlp_layers, dropouts[1])
def __init__(self,
field_dims,
embed_dim,
mlp_layers=(128, 64),
dropout=0.5,
field_len=10):
super(DeepFactorizationMachineModel, self).__init__()
self.linear = FeaturesLinear(field_dims)
self.fm = FactorizationMachine(embed_dim=embed_dim, reduce_sum=True)
self.embedding = FeaturesEmbedding(field_dims, embed_dim)
self.embed_output_size = field_len * embed_dim
self.mlp = MultiLayerPerceptron(self.embed_output_size, mlp_layers,
dropout)
class NeuralFactorizationMachineModel(torch.nn.Module):
"""
A pytorch implementation of Neural Factorization Machine.
将FM的第二部分，
Reference:
X He and TS Chua, Neural Factorization Machines for Sparse Predictive Analytics, 2017.
"""
def __init__(self, field_dims, embed_dim, mlp_layers=(400, 400, 400), dropouts=(0.5, 0.5)):
super(NeuralFactorizationMachineModel, self).__init__()
self.linear = FeaturesLinear(field_dims) # 线性部分
self.embedding = FeaturesEmbedding(field_dims, embed_dim) # 交互部分
self.fm = FactorizationMachine(embed_dim, reduce_sum=False)
self.mlp = MultiLayerPerceptron(embed_dim, mlp_layers, dropouts[1])
def forward(self, x):
"""
:param x: [B,n_f]
:return: [B,1]
"""
x_fm = self.fm(self.embedding(x)) # [B,d]
x = self.linear(x) + self.mlp(x_fm)
return x
def get_l2_loss(self, lambdas=1e-5):
regularization_loss = 0
for parameter in self.mlp.parameters():
regularization_loss += torch.sum(parameter.pow(2))
return lambdas*regularization_loss
def __init__(self, field_dims, embed_dim, mlp_dims, dropout):
super().__init__()
self.linear = FeaturesLinear(field_dims)
self.fm = FactorizationMachine(reduce_sum=True)
self.embedding = FeaturesEmbedding(field_dims, embed_dim)
self.embed_output_dim = len(field_dims) * embed_dim
self.mlp = MultiLayerPerceptron(self.embed_output_dim, mlp_dims, dropout)
def __init__(self,
field_dims,
embed_dim,
mlp_dims,
dropout=0.5,
field_len=10):
super(WideAndDeepModel, self).__init__()
# wide 部分： linear + FM
self.linear = FeaturesLinear(field_dims)
self.embedding = FeaturesEmbedding(field_dims, embed_dim)
self.embed_output_dim = field_len * embed_dim
self.mlp = MultiLayerPerceptron(self.embed_output_dim,
mlp_dims,
dropout=dropout)
def __init__(self, field_dims, embed_dims, mlp_dims, dropout):
super().__init__()
self.linear = FeaturesLinear(field_dims)
self.embedding = FeaturesEmbedding(field_dims, embed_dims)
self.embed_output_dim = len(field_dims) * embed_dims
self.mlp = MultiLayerPerceptron(self.embed_output_dim, mlp_dims,
dropout)
def __init__(self, field_dims, embed_dim, mlp_dims, dropouts):
super().__init__()
self.embedding = FeaturesEmbedding(field_dims, embed_dim)
self.linear = FeaturesLinear(field_dims)
self.fm = torch.nn.Sequential(FactorizationMachine(reduce_sum=False),
torch.nn.BatchNorm1d(embed_dim),
torch.nn.Dropout(dropouts[0]))
self.mlp = MultiLayerPerceptron(embed_dim, mlp_dims, dropouts[1])
def __init__(self, field_dims, user_field_idx, item_field_idx, embed_dim, mlp_dims, dropout):
super().__init__()
self.user_field_idx = user_field_idx
self.item_field_idx = item_field_idx
self.embedding = FeaturesEmbedding(field_dims, embed_dim)
self.embed_output_dim = len(field_dims) * embed_dim
self.mlp = MultiLayerPerceptron(self.embed_output_dim, mlp_dims, dropout, output_layer=False)
self.fc = torch.nn.Linear(mlp_dims[-1] + embed_dim, 1)
def __init__(self, field_dims, embed_dim, mlp_layers=(400, 400, 400), dropouts=(0.5, 0.5)):
super(NON2, self).__init__()
self.linear = FeaturesLinear(field_dims) # 线性部分
# self.embedding = FeaturesEmbedding(field_dims, embed_dim) # embedding
self.embedding = FeaturesEmbeddingWithGlobalIn(field_dims, embed_dim) # embedding
self.dnn = MultiLayerPerceptron(embed_dim*len(field_dims), mlp_layers,dropout=dropouts[0], output_layer=False)
self.atten_embedding = torch.nn.Linear(embed_dim, 32)
self.atten_output_dim = len(field_dims) * 32
self.self_attns = torch.nn.ModuleList([
torch.nn.MultiheadAttention(32, 4, dropout=dropouts[0]) for _ in range(3)
])
self.input_dim = 400 + self.atten_output_dim + 1
self.mlp = MultiLayerPerceptron(self.input_dim, embed_dims=(64,32), dropout=dropouts[1])
def __init__(self, field_dims, embed_dim, mlp_layers=(400,400,400),dropouts=(0.5, 0.5),type="glu"):
super(GFRLNFM, self).__init__()
self.linear = FeaturesLinear(field_dims) # 线性部分
self.embedding_gate = FeaturesEmbeddingWithGlobalIn(field_dims,embed_dim,type=type)
self.fm = FactorizationMachine(reduce_sum=False)
self.mlp = MultiLayerPerceptron(embed_dim, mlp_layers, dropouts[1])
def __init__(self,
field_dims,
embed_dim,
mlp_layers=(256, 128, 64),
dropout=0.5):
super(FactorizationSupportedNeuralNetworkModel, self).__init__()
self.embedding = FeaturesEmbedding(field_dims, embed_dim)
self.mlp_input_dim = len(field_dims) * embed_dim
self.mlp = MultiLayerPerceptron(self.mlp_input_dim, mlp_layers)
def __init__(self, field_dims, embed_dim, mlp_dims, dropout=0.5):
super(GELWDL, self).__init__()
# wide 部分： linear + FM
self.linear = FeaturesLinear(field_dims)
self.embedding = FeaturesEmbeddingWithGlobalIn(field_dims, embed_dim)
self.embed_output_dim = len(field_dims) * embed_dim
self.mlp = MultiLayerPerceptron(self.embed_output_dim,
mlp_dims,
dropout=dropout)
def __init__(self,
field_dims,
embed_dim,
mlp_dims,
dropout,
cross_layer_sizes,
split_half=True):
super().__init__()
self.embedding = FeaturesEmbedding(field_dims, embed_dim)
self.embed_output_dim = len(field_dims) * embed_dim
self.cin = CompressedInteractionNetwork(len(field_dims),
cross_layer_sizes, split_half)
self.mlp = MultiLayerPerceptron(self.embed_output_dim, mlp_dims,
dropout)
self.linear = FeaturesLinear(field_dims)
def __init__(self, field_dims, embed_dim, mlp_layers=(400,400,400), glunum=2, dropouts=(0.1, 0.1)):
super().__init__()
self.linear = FeaturesLinear(field_dims) # 线性部分
# self.embedding_gate = FeaturesEmbeddingWithGate(field_dims,embed_dim,glu_num=glunum)
self.embedding = FeaturesEmbedding(field_dims, embed_dim) # 交互部分，先
# 第二层： 门机制，提取信息
#
self.gene_inter = GenerateConv()
glu_list = list()
for _ in range(glunum):
glu_list.append(GLUActivation1D(embed_dim, int(embed_dim*2)))
self.glus = torch.nn.Sequential(*glu_list)
# self.mlp_input = len(field_dims) * (len(field_dims)-1)/2
self.mlp = MultiLayerPerceptron(embed_dim, mlp_layers, dropouts[1])
def __init__(self, field_dims, embed_dim, mlp_dims, dropout):
"""
:param field_dims: Number of input dimensions
:param embed_dim: Number of dense embedding dimensions
:param mlp_dims: Number of hidden layers
:param dropout: dropout rate
"""
super().__init__()
# Wide Learning Component
self.linear = FeaturesLinear(field_dims)
# Deep Learning Component
self.embedding = FeaturesEmbedding(field_dims, embed_dim)
self.embed_output_dim = len(field_dims) * embed_dim
self.mlp = MultiLayerPerceptron(self.embed_output_dim, mlp_dims,
dropout)
def __init__(self,
field_dims,
embed_dim,
mlp_layers=(128, 64),
dropout=0.5,
type="glu"):
super(GELDFM, self).__init__()
self.linear = FeaturesLinear(field_dims)
self.fm = FactorizationMachine(reduce_sum=True)
self.embedding = FeaturesEmbeddingWithGlobalIn(field_dims,
embed_dim,
type=type)
self.embed_output_size = len(field_dims) * embed_dim
self.mlp = MultiLayerPerceptron(self.embed_output_size, mlp_layers,
dropout)
class WideAndDeepModel(torch.nn.Module):
"""
A pytorch implementation of wide and deep learning.
WD 的 工业功能比较强。 效果不错，因为参数少，运行快速。
Reference:
HT Cheng, et al. Wide & Deep Learning for Recommender Systems, 2016.
"""
def __init__(self,
field_dims,
embed_dim,
mlp_dims,
dropout=0.5,
field_len=10):
super(WideAndDeepModel, self).__init__()
# wide 部分： linear + FM
self.linear = FeaturesLinear(field_dims)
self.embedding = FeaturesEmbedding(field_dims, embed_dim)
self.embed_output_dim = field_len * embed_dim
self.mlp = MultiLayerPerceptron(self.embed_output_dim,
mlp_dims,
dropout=dropout)
# self.last_linear = nn.Linear(2,1)
def forward(self, x):
"""
:param x: [B,n_f]
:return: [B,1]
"""
embed_x = self.embedding(x)
# wide + DEEP, 和DeepFM比，没有FM部分，否则就是相同的
# TODO 改进，通过一个Linear
x = self.linear(x) + self.mlp(embed_x.view(x.size(0), -1))
# x = self.last_linear(torch.cat([self.linear(x),self.mlp(embed_x.view(x.size(0),-1))],dim=-1))
return x
def get_l2_loss(self, lambdas=1e-5):
regularization_loss = 0
for parameter in self.embedding.parameters():
regularization_loss += torch.sum(parameter.pow(2))
for parameter in self.mlp.parameters():
regularization_loss += torch.sum(parameter.pow(2))
return lambdas * regularization_loss
def __init__(self,
field_dims,
embed_dim,
mlp_dims=(400, 400, 400),
glunum=1,
dropout=0.1):
super(WDLGate, self).__init__()
# wide 部分： linear + FM
self.linear = FeaturesLinear(field_dims)
self.embedding_gate = FeaturesEmbeddingWithGate(field_dims,
embed_dim,
glu_num=glunum)
self.embed_output_dim = len(field_dims) * embed_dim
self.mlp = MultiLayerPerceptron(self.embed_output_dim,
mlp_dims,
dropout=dropout)
class DeepFactorizationMachineModel(nn.Module):
"""
DeepFM,主要由三部分组成
1、线性部分
2、FM 部分
3、Deep部分
"""
def __init__(self,
field_dims,
embed_dim,
mlp_layers=(128, 64),
dropout=0.5,
field_len=10):
super(DeepFactorizationMachineModel, self).__init__()
self.linear = FeaturesLinear(field_dims)
self.fm = FactorizationMachine(embed_dim=embed_dim, reduce_sum=True)
self.embedding = FeaturesEmbedding(field_dims, embed_dim)
self.embed_output_size = field_len * embed_dim
self.mlp = MultiLayerPerceptron(self.embed_output_size, mlp_layers,
dropout)
# self.last_fc = nn.Linear(3,1)
def forward(self, x):
"""
:param x:
:return:
"""
x_embed = self.embedding(x) # [B,n_f,e]
x_out = self.linear(x) + self.fm(x_embed) + self.mlp(
x_embed.view(x.size(0), -1))
# x_con = torch.cat([self.linear(x),self.fm(x_embed),self.mlp(x_embed.view(x.size(0),-1))],dim=1)
# x_out = self.last_fc(x_con)
return x_out
def get_l2_loss(self, lambdas=1e-5):
regularization_loss = 0
for parameter in self.embedding.parameters():
regularization_loss += torch.sum(parameter.pow(2))
for parameter in self.mlp.parameters():
regularization_loss += torch.sum(parameter.pow(2))
return lambdas * regularization_loss
def __init__(self,
field_dims,
embed_dim,
mlp_dims=(400, 400, 400),
dropout=0.5,
cross_layer_sizes=(200, 200, 200),
split_half=True):
super().__init__()
self.embedding = FeaturesEmbeddingWithGlobalIn(field_dims, embed_dim)
self.embed_output_dim = len(field_dims) * embed_dim
self.mlp = MultiLayerPerceptron(self.embed_output_dim, mlp_dims,
dropout)
# 修改cross CIN模块
self.cin = CompressedInteractionNetwork(len(field_dims),
cross_layer_sizes, split_half)
self.linear = FeaturesLinear(field_dims)