_ff.py 文件源码

def build_training(self, global_steps, inputs, inferences):
    with tf.name_scope('target'):
      label_indices = self.classification.target_label_indices(inputs)

    with tf.name_scope('error'):
      cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=inferences,
                                                              labels=label_indices,
                                                              name='softmax_cross_entropy')
      loss = tf.reduce_mean(cross_entropy, name='loss')

      averager = tf.train.ExponentialMovingAverage(0.99, name='loss_averager')
      averaging = averager.apply([loss])

    with tf.name_scope(''):
      tf.summary.scalar('metrics/loss', loss)
      tf.summary.scalar('metrics/loss.average', averager.average(loss))

    with tf.control_dependencies([averaging]):
      with tf.name_scope(self.args.optimizer.get_name()):
        gradients = self.args.optimizer.compute_gradients(loss, var_list=tf.trainable_variables())
        train = self.args.optimizer.apply_gradients(gradients, global_steps, name='optimize')

      with tf.name_scope(''):
        for gradient, t in gradients:
          if gradient is not None:
            tf.summary.histogram(t.op.name + '.gradients', gradient)

    return loss, train