why to expand the 1-dim of response_vecs?

[llllly26]

why you do this? responses_vec = responses_vec.view(1, batch_size, -1).expand(batch_size, batch_size, self.vec_dim)。in this way, the dimension of 'responses_vec' is (bs,bs,64).when calculate dot_attention that final_context_vec = dot_attention(responses_vec, context_vecs, context_vecs, None, self.dropout), attention_weights is (bs,bs,16).so when weighted,one query may gat the all information in batch,why? thanks!