Method _create_model

modelzoo/dien/train.py:472–554

(self)

    # create model
    def _create_model(self):
        # input layer to get embedding of features
        with tf.variable_scope('input_layer',
                               partitioner=self._input_layer_partitioner,
                               reuse=tf.AUTO_REUSE):
            uid_emb, item_emb, his_item_emb, noclk_his_item_emb, sequence_length = self._embedding_input_layer(
            )

            if self.bf16:
                uid_emb = tf.cast(uid_emb, tf.bfloat16)
                item_emb = tf.cast(item_emb, tf.bfloat16)
                his_item_emb = tf.cast(his_item_emb, tf.bfloat16)
                noclk_his_item_emb = tf.cast(noclk_his_item_emb, tf.bfloat16)

            item_his_eb_sum = tf.reduce_sum(his_item_emb, 1)
            mask = tf.sequence_mask(sequence_length)

        # RNN layer_1
        with tf.variable_scope('rnn_1'):
            run_output_1, _ = tf.nn.dynamic_rnn(
                tf.nn.rnn_cell.GRUCell(self._hidden_size),
                inputs=his_item_emb,
                sequence_length=sequence_length,
                dtype=self._data_type,
                scope='gru1')
            tf.summary.histogram('GRU_outputs', run_output_1)

        # Aux loss
        aux_loss_scope = tf.variable_scope(
            'aux_loss', partitioner=self._dense_layer_partitioner)
        with aux_loss_scope.keep_weights(dtype=tf.float32) if self.bf16 \
            else aux_loss_scope:
            self._aux_loss = self._auxiliary_loss(run_output_1[:, :-1, :],
                                                  his_item_emb[:, 1:, :],
                                                  noclk_his_item_emb[:, 1:, :],
                                                  mask[:, 1:],
                                                  dtype=self._data_type,
                                                  stag='gru')
            if self.bf16:
                self._aux_loss = tf.cast(self._aux_loss, tf.float32)

        # Attention layer
        attention_scope = tf.variable_scope('attention_layer')
        with attention_scope.keep_weights(dtype=tf.float32) if self.bf16 \
            else attention_scope:
            _, alphas = self._attention(item_emb,
                                        run_output_1,
                                        self._attention_size,
                                        mask,
                                        softmax_stag=1,
                                        stag='1_1',
                                        mode='LIST',
                                        return_alphas=True)
            tf.summary.histogram('alpha_outputs', alphas)

        # RNN layer_2
        with tf.variable_scope('rnn_2'):
            _, final_state2 = tf.nn.dynamic_rnn(