| 95 | |
| 96 | |
| 97 | class DCN(): |
| 98 | def __init__(self, |
| 99 | emb_stacking_columns=None, |
| 100 | dnn_hidden_units=[1024, 512, 256], |
| 101 | optimizer_type='adam', |
| 102 | cross_learning_rate=0.001, |
| 103 | deep_learning_rate=0.001, |
| 104 | inputs=None, |
| 105 | bf16=False, |
| 106 | stock_tf=None, |
| 107 | adaptive_emb=False, |
| 108 | input_layer_partitioner=None, |
| 109 | dense_layer_partitioner=None): |
| 110 | if not inputs: |
| 111 | raise ValueError("Dataset is not defined.") |
| 112 | self._feature = inputs[0] |
| 113 | self._label = inputs[1] |
| 114 | |
| 115 | self._emb_stacking_columns = emb_stacking_columns |
| 116 | if not emb_stacking_columns: |
| 117 | raise ValueError("Embedding and stacking column is not defined.") |
| 118 | |
| 119 | self.tf = stock_tf |
| 120 | self.bf16 = False if self.tf else bf16 |
| 121 | self.is_training = True |
| 122 | self._adaptive_emb = adaptive_emb |
| 123 | |
| 124 | self._dnn_hidden_units = dnn_hidden_units |
| 125 | self._deep_learning_rate = deep_learning_rate |
| 126 | self._cross_learning_rate = cross_learning_rate |
| 127 | self._optimizer_type = optimizer_type |
| 128 | self._input_layer_partitioner = input_layer_partitioner |
| 129 | self._dense_layer_partitioner = dense_layer_partitioner |
| 130 | |
| 131 | self._create_model() |
| 132 | with tf.name_scope('head'): |
| 133 | self._create_loss() |
| 134 | self._create_optimizer() |
| 135 | self._create_metrics() |
| 136 | |
| 137 | # used to add summary in tensorboard |
| 138 | def _add_layer_summary(self, value, tag): |
| 139 | tf.summary.scalar('%s/fraction_of_zero_values' % tag, |
| 140 | tf.nn.zero_fraction(value)) |
| 141 | tf.summary.histogram('%s/activation' % tag, value) |
| 142 | |
| 143 | def _dnn(self, dnn_input, dnn_hidden_units=None, layer_name=''): |
| 144 | for layer_id, num_hidden_units in enumerate(dnn_hidden_units): |
| 145 | with tf.variable_scope(layer_name + '_%d' % layer_id, |
| 146 | partitioner=self._dense_layer_partitioner, |
| 147 | reuse=tf.AUTO_REUSE) as dnn_layer_scope: |
| 148 | dnn_input = tf.layers.dense( |
| 149 | dnn_input, |
| 150 | units=num_hidden_units, |
| 151 | activation=tf.nn.relu, |
| 152 | kernel_initializer=tf.glorot_uniform_initializer(), |
| 153 | name=dnn_layer_scope) |
| 154 | |