For dcn model

[jiarenyf]

In

dnn_ctr/model/DCN.py

Line 232 in c750fec

x_l = torch.sum(x_0 * x_l, 1).view([-1,1]) * getattr(self,'cross_weight_'+str(i+1)).view([1,-1]) + getattr(self,'cross_bias_'+str(i+1)) + x_l

, x_0 * x_l should be replaced by torch.matmul(x_0, x_l.t()), right ?

[nzc]

The x_l's size should be batch_size * [field_size * embedding_size]，in DCN paper，x_l's size should be [filed_sizeembedding], so result of x_0x_l^T is rank-one. Thinking of expanding to batch_size [field_sizeembedding_size], the result's size should be batch_size1. If use torch.matmul(x_0, x_l.t()), the size is wrong.

[nzc]

@jiarenyf

[jiarenyf]

The shapes of x0, xl, wl, bl are all [fieldSzie*embeddingSize, 1] (ignoring the batch_axis), and the formulation of xl should be xl = matmul(matmul(x0, xl.T), wl) + bl + xl, as shown in the following image:
...

I am not familiar with pytorch, but in mxnet I use batch_dot to implement the calculation of xl, as in here ...

[nzc]

@jiarenyf The x_0 in my code is two-dimension . And I do the same thing as batch_dot in my code

[jiarenyf]

But the result of x0*xl^T is not rank-one, it should be [batchSize, fieldSize*embeddingSize, fieldSize*embeddingSize], and the shape of x0*xl^T*wl is [batchSize, fieldSize*embeddingSize] ... Here I use * to represent matmul ...