Method _create_model

modelzoo/esmm/train.py:194–286

(self)

    # create model
    def _create_model(self):
        with tf.variable_scope('user_input_layer',
                               partitioner=self._input_layer_partitioner,
                               reuse=tf.AUTO_REUSE):
            if not self._tf and self._adaptive_emb:
                '''Adaptive Embedding Feature Part 1 of 2'''
                adaptive_mask_tensors = {}
                for col in USER_COLUMN:
                    adaptive_mask_tensors[col] = tf.ones([self._batch_size],
                                                         tf.int32)
                user_emb = tf.feature_column.input_layer(
                    self.feature,
                    self._user_column,
                    adaptive_mask_tensors=adaptive_mask_tensors)
            else:
                user_emb = tf.feature_column.input_layer(self.feature,
                                                         self._user_column)
        with tf.variable_scope('item_input_layer',
                               partitioner=self._input_layer_partitioner,
                               reuse=tf.AUTO_REUSE):
            if not self._tf and self._adaptive_emb:
                '''Adaptive Embedding Feature Part 1 of 2'''
                adaptive_mask_tensors = {}
                for col in ITEM_COLUMN:
                    adaptive_mask_tensors[col] = tf.ones([self._batch_size],
                                                         tf.int32)
                item_emb = tf.feature_column.input_layer(
                    self.feature,
                    self._item_column,
                    adaptive_mask_tensors=adaptive_mask_tensors)
            else:
                item_emb = tf.feature_column.input_layer(self.feature,
                                                         self._item_column)
        with tf.variable_scope('combo_input_layer',
                               partitioner=self._input_layer_partitioner,
                               reuse=tf.AUTO_REUSE):
            if not self._tf and self._adaptive_emb:
                '''Adaptive Embedding Feature Part 1 of 2'''
                adaptive_mask_tensors = {}
                for col in COMBO_COLUMN:
                    adaptive_mask_tensors[col] = tf.ones([self._batch_size],
                                                         tf.int32)
                combo_emb = tf.feature_column.input_layer(
                    self.feature,
                    self._combo_column,
                    adaptive_mask_tensors=adaptive_mask_tensors)
            else:
                for key in TAG_COLUMN:
                    self.feature[key] = tf.strings.split(self.feature[key], '|')
                combo_emb = tf.feature_column.input_layer(self.feature,
                                                          self._combo_column)

        with self._make_scope('ESMM', self._bf16):
            if self._bf16:
                user_emb = tf.cast(user_emb, dtype=tf.bfloat16)
                item_emb = tf.cast(item_emb, dtype=tf.bfloat16)
                combo_emb = tf.cast(combo_emb, dtype=tf.bfloat16)