Method _create_optimizer

modelzoo/ple/train.py:322–352

(self)

    # define optimizer and generate train_op
    def _create_optimizer(self):
        self.global_step = tf.train.get_or_create_global_step()
        print('self.tf = ', self.tf, ' self._optimizer_type = ', self._optimizer_type)
        if self.tf or self._optimizer_type == 'adam':
            optimizer = tf.train.AdamOptimizer(
                learning_rate=self._learning_rate,
                beta1=0.9,
                beta2=0.999,
                epsilon=1e-8)
        elif self._optimizer_type == 'adagrad':
            optimizer = tf.train.AdagradOptimizer(
                learning_rate=self._learning_rate,
                initial_accumulator_value=0.1,
                use_locking=False)
        elif self._optimizer_type == 'adamasync':
            optimizer = tf.train.AdamAsyncOptimizer(
                learning_rate=self._learning_rate,
                beta1=0.9,
                beta2=0.999,
                epsilon=1e-8)
        elif self._optimizer_type == 'adagraddecay':
            optimizer = tf.train.AdagradDecayOptimizer(
                learning_rate=self._learning_rate,
                global_step=self.global_step)
        else:
            raise ValueError("Optimizer type error.")

        update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
        with tf.control_dependencies(update_ops):
            self.train_op = optimizer.minimize(self.loss,
                                               global_step=self.global_step)

    # compute acc & auc
    def _create_metrics(self):