tf predict 模型的预测模式

训练模型就是为了让其作预测. 笔记如下.

model_fn

预测的原理是, tf 进程构建了预测 mode下的计算图, 然后从 model_dir 中恢复变量, 就绪后作预测.

# _model_fn 定义
def _model_fn(features,  # This is batch_features from input_fn
              labels,  # This is batch_labels from input_fn
              mode,  # An instance of tf.estimator.ModeKeys
              params):
    # ...
    if mode == tf.estimator.ModeKeys.PREDICT:
        predictions = dict()
        score = tf.sigmoid(logits)
        predictions['sample_key'] = sample_key_tensor
        predictions['score'] = score
        # predictions['doubtful_tensor'] = score
        estimator_spec = tf.estimator.EstimatorSpec(
            mode=tf.estimator.ModeKeys.PREDICT,
            predictions=predictions
        )
        # ...

遍历预测结果

# 遍历预测结果
pred_result_generator = estimator.predict(input_fn=pred_input_fn)
for pred_no, pred_dict in enumerate(pred_result_generator):  # 不再是batch_size
    sample_key = pred_dict['sample_key']  # 单个int
    pred_ctr = pred_dict['score']  # 单个float
    print(sample_key, pred_ctr)

常见问题

• 计算图要严格对应
  若不能严格一致
  /lib/python2.7/site-packages/tensorflow/python/client/session.py", line 1417, in _do_call
  raise type(e)(node_def, op, message)
  tensorflow.python.framework.errors_impl.NotFoundError