我基于bert的发布脚本使用了tf.estimator.train_and_evaluate。

if FLAGS.do_train and FLAGS.do_eval:
        tf.logging.info("***** Running training and evaluation *****")
        tf.logging.info("  Num examples of training data = %d", train_examples_len)
        tf.logging.info("  Batch size of training data = %d", FLAGS.train_batch_size)
        tf.logging.info("  Num steps of training = %d", num_train_steps)
        tf.logging.info("  Num examples of eval data = %d", eval_examples_len)
        tf.logging.info("  Batch size of eval data = %d", FLAGS.eval_batch_size)

        train_input_fn = file_based_input_fn_builder(
            input_file=train_file,
            seq_length=FLAGS.max_seq_length,
            is_training=True,
            drop_remainder=True)
        train_spec = tf.estimator.TrainSpec(input_fn=train_input_fn, max_steps=num_train_steps)

        eval_drop_remainder = True if FLAGS.use_tpu else False
        eval_input_fn = file_based_input_fn_builder(
            input_file=eval_file,
            seq_length=FLAGS.max_seq_length,
            is_training=False,
            drop_remainder=eval_drop_remainder)
        eval_spec = tf.estimator.EvalSpec(input_fn=eval_input_fn, steps=500, start_delay_secs=0,
                                          throttle_secs=200)
        tf.estimator.train_and_evaluate(estimator, train_spec=train_spec, eval_spec=eval_spec)

我发现训练在GPU上进行，而评估在CPU上进行。如何让评估也在GPU上进行？