def __init__(self,
emb_dim,
num_feats,
num_fields,
fc_dims=None,
dropout=None,
batch_norm=None,
out_type='binary'):
super(NFM, self).__init__()
self.emb_dim = emb_dim
self.num_feats = num_feats
self.num_fields = num_fields
self.first_order_weights = nn.Embedding(num_embeddings=num_feats,
embedding_dim=1)
nn.init.xavier_uniform_(self.first_order_weights.weight)
self.first_order_bias = nn.Parameter(torch.randn(1))
self.emb_layer = nn.Embedding(num_embeddings=num_feats,
embedding_dim=emb_dim)
nn.init.xavier_uniform_(self.emb_layer.weight)
self.bi_intaraction_layer = BiInteractionLayer()
if not fc_dims:
fc_dims = [32, 32]
self.fc_dims = fc_dims
self.fc_layers = MLP(emb_dim, fc_dims, dropout, batch_norm)
self.h = nn.Parameter(torch.zeros(1, fc_dims[-1])) # 1 * fc_dims[-1]
nn.init.xavier_uniform_(self.h.data)
self.output_layer = OutputLayer(in_dim=1, out_type=out_type)
def __init__(self,
emb_dim,
feat_dim,
num_fields,
fc_dims=None,
dropout=None,
batch_norm=None,
out_type='binary',
train_fm=True):
super(FNN, self).__init__()
# set model object to training FNN or training FM embedding
self.fm_trained = not train_fm
# embedding layer is embedded in the FM sub-module
self.emb_dim = emb_dim
# fc layers
if not fc_dims:
fc_dims = [32, 32]
self.fc_dims = fc_dims
self.num_fields = num_fields
self.fc_layers = MLP(emb_dim * num_fields, fc_dims, dropout,
batch_norm)
# fm model as the pre-trained embedding layer
self.fm = FM(emb_dim, feat_dim, out_type)
# output
self.output_layer = OutputLayer(fc_dims[-1], out_type)
def __init__(self, emb_dim, num_feats, num_cate_fields, num_cont_fields, num_cross_feats, fc_dims=None,
dropout=None, batch_norm=None, out_type='binary'):
super(WideAndDeep, self).__init__()
self.emb_dim = emb_dim
self.num_feats = num_feats
self.num_cate_fields = num_cate_fields
self.num_cont_fields = num_cont_fields
self.num_cross_feats = num_cross_feats
# first order weight for category features
self.cate_weights = nn.Embedding(num_embeddings=num_feats - num_cont_fields, embedding_dim=1)
nn.init.xavier_uniform_(self.cate_weights.weight)
# first order weight for continuous features
self.cont_weights = nn.Linear(in_features=num_cont_fields, out_features=1)
nn.init.xavier_uniform_(self.cont_weights)
self.wide_bias = nn.Parameter(torch.randn(1))
if not fc_dims:
fc_dims = [32, 32]
fc_dims.append(1)
self.fc_dims = fc_dims
# embedding for deep network
self.emb_layer = nn.Embedding(num_embeddings=num_feats - num_cont_fields, embedding_dim=emb_dim)
nn.init.xavier_uniform_(self.emb_layer.weight)
self.deep = MLP(num_cont_fields + num_cate_fields * emb_dim, fc_dims, dropout, batch_norm)
self.out_layer = OutputLayer(in_dim=1, out_type=out_type)
def __init__(self, emb_dim, feat_dim, num_fields, fc_dims=None, dropout=None, batch_norm=None, out_type='binary'):
super(NFM, self).__init__()
self.emb_dim = emb_dim
self.feat_dim = feat_dim
self.num_fields = num_fields
self.emb_layer = nn.Embedding(num_embeddings=feat_dim, embedding_dim=emb_dim)
self.bi_intaraction_layer = BiInteractionLayer()
if not fc_dims:
fc_dims = [32, 32]
self.fc_dims = fc_dims
self.fc_layers = MLP(emb_dim, fc_dims, dropout, batch_norm)
self.output_layer = OutputLayer(in_dim=fc_dims[-1], out_type=out_type)
def __init__(self, emb_dim, feat_dim, num_cate_fields, num_cont_fields, num_cross_feats, fc_dims=None, dropout=None,
batch_norm=None, out_type='binary'):
super(WideAndDeep, self).__init__()
self.emb_dim = emb_dim
self.feat_dim = feat_dim
self.num_cate_fields = num_cate_fields
self.num_cont_fields = num_cont_fields
self.num_cross_feats = num_cross_feats
if not fc_dims:
fc_dims = [32, 32]
self.emb_layer = nn.Embedding(num_embeddings=feat_dim - num_cont_fields, embedding_dim=emb_dim)
self.deep = MLP(num_cont_fields + num_cate_fields * emb_dim, fc_dims, dropout, batch_norm)
self.wide = LR(num_cross_feats, out_type='regression')
self.out_layer = OutputLayer(in_dim=fc_dims[-1] + 1, out_type=out_type)
def __init__(self,
emb_dim,
num_feats,
num_categories,
field_ranges,
fc_dims=None,
dropout=None,
batch_norm=None,
out_type='binary'):
super(FLEN, self).__init__()
self.num_feats = num_feats
self.emb_dim = emb_dim
self.num_categories = num_categories
if not field_ranges:
field_ranges = torch.tensor(range(num_categories))
self.field_ranges = field_ranges
self.num_fields = len(field_ranges)
# embedding layer
self.emb_layer = nn.Embedding(num_embeddings=num_feats,
embedding_dim=emb_dim)
nn.init.xavier_uniform_(self.emb_layer.weight)
# S part
self.first_order_weights = nn.Embedding(num_embeddings=num_categories,
embedding_dim=1)
nn.init.xavier_uniform_(self.first_order_weights.weight)
self.first_order_bias = nn.Parameter(torch.randn(1))
# MF part
self.num_pairs = self.num_fields * (self.num_fields - 1) / 2
self.r_mf = nn.Parameter(torch.zeros(self.num_pairs,
1)) # num_pairs * 1
nn.init.xavier_uniform_(self.r_mf.data)
# FM part
self.r_fm = nn.Parameter(torch.zeros(self.num_fields,
1)) # num_fields * 1
nn.init.xavier_uniform_(self.r_fm.data)
# dnn
if not fc_dims:
fc_dims = [32, 32, 32]
self.fc_dims = fc_dims
self.fc_layers = MLP(fc_dims, dropout, batch_norm)
self.output_layer = OutputLayer(fc_dims[-1] + 1 + self.emb_dim,
out_type)
def __init__(self,
emb_dim,
feat_dim,
num_fields,
fc_dims=None,
dropout=None,
batch_norm=None,
product_type='inner',
out_type='binary'):
super(PNN, self).__init__()
# embedding layer
self.emb_dim = emb_dim
self.feat_dim = feat_dim
self.num_fields = num_fields
self.emb_layer = nn.Embedding(num_embeddings=self.feat_dim,
embedding_dim=self.emb_dim)
nn.init.xavier_uniform_(self.emb_layer.weight)
# linear signal layer, named l_z
if not fc_dims:
fc_dims = [32, 32]
self.d1 = d1 = fc_dims[0]
self.product_type = product_type
if product_type == '*':
d1 *= 2
self.linear_signal_weights = nn.Linear(in_features=num_fields *
emb_dim,
out_features=d1)
nn.init.xavier_uniform_(self.linear_signal_weights.weight)
# product layer, named l_p
if product_type == 'inner':
self.product_layer = InnerProductLayer(num_fields, d1)
elif product_type == 'outer':
self.product_layer = OuterProductLayer(emb_dim, num_fields, d1)
else:
self.product_layer = HybridProductLayer(emb_dim, num_fields, d1)
# fc layers
# l_1=relu(l_z+l_p_b_1)
self.l1_layer = nn.ReLU()
self.l1_bias = nn.Parameter(torch.randn(d1))
# l_2 to l_n
self.fc_dims = fc_dims
self.fc_layers = MLP(d1, self.fc_dims, dropout, batch_norm)
# output layer
self.output_layer = OutputLayer(fc_dims[-1], out_type)
def __init__(self,
emb_dim,
num_feats,
num_cate_fields,
num_cont_fields,
cross_depth,
fc_dims=None,
dropout=None,
batch_norm=None,
out_type='binary'):
super(DCN, self).__init__()
self.emb_dim = emb_dim
self.num_feats = num_feats
self.num_cate_fields = num_cate_fields
self.num_cont_fields = num_cont_fields
self.cross_depth = cross_depth
# embedding for category features
self.emb_layer = nn.Embedding(num_embeddings=num_feats -
num_cont_fields,
embedding_dim=emb_dim)
nn.init.xavier_uniform_(self.emb_layer.weight)
# deep network
if not fc_dims:
fc_dims = [32, 32]
self.fc_dims = fc_dims
x0_dim = num_cont_fields + num_cate_fields * emb_dim
self.deep = MLP(x0_dim, fc_dims, dropout, batch_norm)
# cross network
cross_layers = []
for _ in range(cross_depth):
cross_layers.append(CrossLayer(x0_dim))
self.cross = nn.ModuleList(cross_layers)
self.out_layer = OutputLayer(in_dim=fc_dims[-1] + x0_dim,
out_type=out_type)
def __init__(self,
emb_dim,
feat_dim,
num_fields,
fc_dims=None,
dropout=None,
batch_norm=None,
out_type='binary'):
super(DeepFM, self).__init__()
# embedding layer is embedded in the FM sub-module
self.emb_dim = emb_dim
# fm
self.fm = FM(emb_dim, feat_dim, out_type='regression')
# dnn
if not fc_dims:
fc_dims = [32, 32, 32]
self.fc_dims = fc_dims
self.num_fields = num_fields
self.dnn = MLP(emb_dim * num_fields, fc_dims, dropout, batch_norm)
# output
self.output_layer = OutputLayer(fc_dims[-1] + 1, out_type)