Understanding DCN Network

[nishank974]

Describe the question(问题描述)

1. In DCN network why do we have a linear_model(X) in forward function? What is the use of linear_model?

2. Why do you have separate embedding layers for linear_model and for dnn and cross.

` def forward(self, X):

    logit = self.linear_model(X) # this line Q1
    sparse_embedding_list, dense_value_list = self.input_from_feature_columns(X, self.dnn_feature_columns,
                                                                              self.embedding_dict) # this line Q2

    dnn_input = combined_dnn_input(sparse_embedding_list, dense_value_list)

`

[zanshuxun]

1. This is an extended implementation, i.e. the linear part in the WDL model. If you don't need the linear part in actual use, just leave linear_feature_columns a empty list and use only dnn_feature_columns. Use it according to actual needs.
2. Unlike the dnn part, the embedding dimension of the linear part is always 1 (in function create_embedding_matrix, see below), which is an implementation of the linear part in WDL:
   

   DeepCTR-Torch/deepctr_torch/inputs.py

   Lines 167 to 171 in f685425

   	embedding_dict = nn.ModuleDict(
   	{feat.embedding_name: nn.Embedding(feat.vocabulary_size, feat.embedding_dim if not linear else 1, sparse=sparse)
   	for feat in
   	sparse_feature_columns + varlen_sparse_feature_columns}
   	)

Thus they need to be embedded separately.
dnn_feature_columns:

DeepCTR-Torch/deepctr_torch/models/basemodel.py

Line 115 in f685425

self.embedding_dict = create_embedding_matrix(dnn_feature_columns, init_std, sparse=False, device=device)

linear_feature_columns:

DeepCTR-Torch/deepctr_torch/models/basemodel.py

Lines 47 to 48 in f685425

	self.embedding_dict = create_embedding_matrix(feature_columns, init_std, linear=True, sparse=False,
	device=device)