Method _create_model

modelzoo/bst/train.py:237–328

(self)

    # create model
    def _create_model(self):
        # input embeddings of user & item features
        for key in TAG_COLUMN:
            self._feature[key] = tf.strings.split(self._feature[key], '|')
            self._feature[key] = tf.sparse.slice(
                self._feature[key], [0, 0],
                [self._batch_size, self._max_seqence_length - 1])
        # input layer
        with tf.variable_scope('input_layer',
                               partitioner=self._input_layer_partitioner):
            # unsequence input
            key_dict = {}
            with tf.variable_scope('unseq_input_layer', reuse=tf.AUTO_REUSE):
                if self._adaptive_emb and not self.tf:
                    '''Adaptive Embedding Feature Part 1 of 2'''
                    adaptive_mask_tensors = {}
                    for col in INPUT_FEATURES:
                        adaptive_mask_tensors[col] = tf.ones([args.batch_size],
                                                             tf.int32)
                    unseq_emb = tf.feature_column.input_layer(
                        self._feature,
                        self._unseq_column,
                        adaptive_mask_tensors=adaptive_mask_tensors,
                        cols_to_output_tensors=key_dict)
                else:
                    unseq_emb = tf.feature_column.input_layer(
                        self._feature,
                        self._unseq_column,
                        cols_to_output_tensors=key_dict)

            # bst input
            with tf.variable_scope('bst_input_layer', reuse=tf.AUTO_REUSE):
                # tag input
                with tf.variable_scope('tag_input_layer', reuse=tf.AUTO_REUSE):
                    tag_emb, tag_len = tf.contrib.feature_column.sequence_input_layer(
                        self._feature, self._tag_column)

                # key input
                with tf.variable_scope('key_input_layer', reuse=tf.AUTO_REUSE):
                    key_emb_list = []
                    for key in self._key_column:
                        key_emb_list.append(key_dict[key])
                    key_emb = tf.concat(key_emb_list, axis=-1)

            if self.bf16:
                unseq_emb = tf.cast(unseq_emb, dtype=tf.bfloat16)
                tag_emb = tf.cast(tag_emb, dtype=tf.bfloat16)
                key_emb = tf.cast(key_emb, dtype=tf.bfloat16)

            bst_tower_fea = {
                'key': key_emb,
                'hist_seq_emb': tag_emb,
                'hist_seq_len': tag_len
            }

        # BST
        bst_scope = tf.variable_scope(
            'bst_tower',