Created by: zyz282994112

paddle v2环境，提前预训练后，使用is_static固定参数，但部分layer没有生效（例如time_layer），代码如下：

def network(self, is_predict=False, inputdata=None):
        word_dict_size = self._word_dict_size
        vocab_dict_size = self._vocab_dict_size
        hidden_size = self._hidden_size
        ########################
        #reset_parser()
        if not inputdata:
            inputdata = self.input_layer()
        query, query_time, month, year = inputdata
        query_emb = paddle.layer.embedding(input=query, size=hidden_size, name=self.model+'_wordemb', param_attr=paddle.attr.ParameterAttribute(is_static=True))
        query_time_emb = paddle.layer.embedding(input=query_time, size=hidden_size, name=self.model+'_wordtimeemb', param_attr=paddle.attr.ParameterAttribute(is_static=True))

        lstm_max_query = self.query_emb_func(query_emb, hidden_size)
        lstm_max_query_time = self.query_emb_func(query_time_emb, hidden_size)

        time_layer = paddle.layer.fc(size=32, input=[year, month], act=paddle.activation.Relu(), param_attr=paddle.attr.ParameterAttribute(is_static=True))

        feature_concat_layer = paddle.layer.concat(input=[lstm_max_query, lstm_max_query_time, time_layer])
        feature_concat_layer = paddle.layer.batch_norm(input=feature_concat_layer, act=paddle.activation.Relu(), param_attr=paddle.attr.ParameterAttribute(is_static=True))

        if self.model == 'ctr':
            output = paddle.layer.fc(size=2, input=feature_concat_layer, act=paddle.activation.Softmax(), param_attr=paddle.attr.ParameterAttribute(is_static=True))
        elif self.model == 'acp':
            output = paddle.layer.fc(size=1, input=feature_concat_layer, act=paddle.activation.Linear(), param_attr=paddle.attr.ParameterAttribute(is_static=True))
        elif self.model == 'cpm':
            std_acp = self.acp_model.network(is_predict=True, inputdata=inputdata)
            #acp = std_acp * 4.0 + 3.0
            acp = paddle.layer.scale_shift(input=std_acp, bias_attr=True)
            ctr = self.ctr_model.network(is_predict=True, inputdata=inputdata)
            with paddle.layer.mixed(size=1) as truectr:
                truectr += paddle.layer.identity_projection(input=ctr, offset=1)
            output = truectr * acp
        else:
            exit(1)
        if is_predict:
            if self.model == 'cpm':
                output = output * 1000.0
            return output
        else:
            if self.model == 'ctr':
                weight = paddle.layer.data(name="weight", type=paddle.data_type.dense_vector(1))
                label = paddle.layer.data(name="label", type=paddle.data_type.integer_value(2))
                cost = paddle.layer.cross_entropy_cost(input=output, label=label, weight=weight)
            elif self.model == 'acp':
                weight = paddle.layer.data(name="weight", type=paddle.data_type.dense_vector(1))
                acp = paddle.layer.data(name="std_acp", type=paddle.data_type.dense_vector(1))
                cost = paddle.layer.square_error_cost(input=output, label=acp, weight=weight)
            elif self.model == 'cpm':
                ctr_weight = paddle.layer.data(name="ctr_weight", type=paddle.data_type.dense_vector(1))
                ctr_label = paddle.layer.data(name="label", type=paddle.data_type.integer_value(2))
                ctr_cost = paddle.layer.cross_entropy_cost(input=ctr, label=ctr_label, weight=ctr_weight)

                acp_weight = paddle.layer.data(name="acp_weight", type=paddle.data_type.dense_vector(1))
                acp_label = paddle.layer.data(name="std_acp", type=paddle.data_type.dense_vector(1))
                acp_cost = paddle.layer.square_error_cost(input=std_acp, label=acp_label, weight=acp_weight)

                cpm_weight = paddle.layer.data(name="cpm_weight", type=paddle.data_type.dense_vector(1))
                cpm_label = paddle.layer.data(name="cpm", type=paddle.data_type.dense_vector(1))
                cpm_label = cpm_label * 0.001
                cpm_cost = paddle.layer.square_error_cost(input=output, label=cpm_label, weight=cpm_weight)
                cost = [ctr_cost, acp_cost, cpm_cost]
            else:
                exit(1)
            return cost