How to use the dnn.pytorch.layer.self_attention_layer function in dnn

def init_weights(self):
        for m in self.modules():
            if isinstance(m, layer.RNN_layer):
                m.init_weights()
            if isinstance(m, layer.self_attention_layer):
                m.init_weights()
            if isinstance(m, layer.softmax_layer):
                m.init_weights()

base.base.__init__(self, args)

        self.n_time = n_time
        self.bi_direction_num = 2 if self.bi_direction else 1
        out_n_hidden = self.n_hidden * self.bi_direction_num
        self.drop_out = nn.Dropout(self.drop_prob)
        self.embedding_layer(emb_matrix)

        self.extractors = nn.ModuleList()
        self.attentions = nn.ModuleList()
        self.predictors = nn.ModuleList()
        for _ in range(n_time):
            self.extractors.append(
                nn.ModuleList([layer.CNN_layer(self.emb_dim, 1, self.n_hidden, kw) for kw in range(1, 3)])
            )  # index 0 -> (nt-1)
            self.attentions.append(layer.self_attention_layer(out_n_hidden))
            self.predictors.append(layer.softmax_layer(out_n_hidden, self.n_class))  # index 0 -> (nt-1)
        self.connections = nn.ModuleList()
        self.connections.append(None)
        for _ in range(n_time - 1):
            self.connections.append(
                nn.Sequential(
                    nn.Linear(2 * out_n_hidden, out_n_hidden, bias=False),
                    nn.Sigmoid()
                )

* args [dict]: all model arguments
        * mode [str]: use 'classify'/'sequence' to get the result
        """
        nn.Module.__init__(self)
        base.base.__init__(self, args)

        self.n_hierarchy = n_hierarchy
        self.mode = mode
        self.bi_direction_num = 2 if self.bi_direction else 1

        self.emb_mat = layer.embedding_layer(emb_matrix, self.emb_type)
        self.drop_out = nn.Dropout(self.drop_prob)

        rnn_params = (self.n_hidden, self.n_layer, self.drop_prob, self.bi_direction, self.rnn_type)
        self.rnn = nn.ModuleList([layer.RNN_layer(self.emb_dim, *rnn_params)])
        self.att = nn.ModuleList([layer.self_attention_layer(self.bi_direction_num * self.n_hidden)])
        for _ in range(self.n_hierarchy - 1):
            self.rnn.append(layer.RNN_layer(self.bi_direction_num * self.n_hidden, *rnn_params))
            if self.use_attention:
                self.att.append(layer.self_attention_layer(self.bi_direction_num * self.n_hidden))
        self.predict = layer.softmax_layer(self.n_hidden * self.bi_direction_num, self.n_class)

base.base.__init__(self, args)

        self.n_hierarchy = n_hierarchy
        self.mode = mode
        self.bi_direction_num = 2 if self.bi_direction else 1

        self.emb_mat = layer.embedding_layer(emb_matrix, self.emb_type)
        self.drop_out = nn.Dropout(self.drop_prob)

        rnn_params = (self.n_hidden, self.n_layer, self.drop_prob, self.bi_direction, self.rnn_type)
        self.rnn = nn.ModuleList([layer.RNN_layer(self.emb_dim, *rnn_params)])
        self.att = nn.ModuleList([layer.self_attention_layer(self.bi_direction_num * self.n_hidden)])
        for _ in range(self.n_hierarchy - 1):
            self.rnn.append(layer.RNN_layer(self.bi_direction_num * self.n_hidden, *rnn_params))
            if self.use_attention:
                self.att.append(layer.self_attention_layer(self.bi_direction_num * self.n_hidden))
        self.predict = layer.softmax_layer(self.n_hidden * self.bi_direction_num, self.n_class)