我在尝试使用Tensorflow的tf.contrib.layers包进行分类时遇到了一个问题，实在是搞不明白。从一些例子来看（例如这个及其教程），图的处理都是由API完成的。我可以在我的环境中下载并运行同样的代码，没有任何问题。

然而，当我运行我的代码时，我得到了一个错误，提示我的全局步骤与我的损失不是来自同一个图，这看起来很奇怪：ValueError: Tensor("global_step:0", shape=(), dtype=int64_ref) must be from the same graph as Tensor("softmax_cross_entropy_loss/value:0", shape=(), dtype=float32).这个错误是在构建train_op时发生的。

这是我的Tensorflow代码（我确实有一些其他代码来处理数据的加载，但它们没有使用Tensorflow的任何东西）。抱歉代码现在有点乱：我一直在尝试拆解它来找出这个错误。

import numpy as np
import tensorflow as tf
from tensorflow.contrib.learn.python.learn.estimators import model_fn as model_fn_lib
import data # 我的数据加载模块
def train(training_file, vocab_path, hidden_units=[10, 20, 10], estimator=tf.contrib.learn.DNNClassifier):
    """
    给定一个训练CSV文件，训练一个Tensorflow神经网络
    """
    training_set = data.load(training_file)
    vocab = tf.contrib.learn.preprocessing.VocabularyProcessor(data.DOC_LENGTH)
    vocab = vocab.restore(vocab_path)
    training_data = tf.one_hot(training_set.data, len(vocab.vocabulary_._mapping), dtype=tf.float32)
    training_targets = tf.constant(np.array(training_set.targets, dtype=np.int32))
    classifier = tf.contrib.learn.Estimator(model_fn=lambda features, targets, mode, params: model_fn(features, targets, mode, params, hidden_units))
    classifier.fit(input_fn=lambda: (training_data, training_targets), steps=2000)
    return classifier
def model_fn(features, targets, mode, params, hidden_units):
    if len(hidden_units) <= 0:
        raise ValueError("隐藏单元必须是一个长度大于等于1的整数可迭代对象")
    # 定义网络
    network = tf.contrib.layers.relu(features, hidden_units[0])
    for i in range(1, len(hidden_units)):
        network = tf.contrib.layers.relu(network, hidden_units[i])
    # 展平网络
    network = tf.reshape(network, [-1, hidden_units[-1] * data.DOC_LENGTH])
    # 添加dropout以增强特征使用
    network = tf.layers.dropout(inputs=network, rate=0.5, training=mode == tf.contrib.learn.ModeKeys.TRAIN)
    # 计算logits
    logits = tf.contrib.layers.fully_connected(network, 15)
    loss = None
    train_op = None
    if mode != tf.contrib.learn.ModeKeys.INFER:
        targets = tf.cast(tf.one_hot(targets, 15, 1, 0), dtype=tf.float32)
        loss = tf.losses.softmax_cross_entropy(logits=logits, onehot_labels=targets)
    if mode == tf.contrib.learn.ModeKeys.TRAIN:
        # 这个train_op会导致错误
        train_op = tf.contrib.layers.optimize_loss(
          loss=loss,
          global_step=tf.train.get_global_step(),
          optimizer='Adam',
          learning_rate=0.01)
    predictions = {
          "classes": tf.argmax(input=logits, axis=1),
          "probabilities": tf.nn.softmax(logits, name="softmax_tensor")
    }
    return model_fn_lib.ModelFnOps(mode=mode, predictions=predictions, loss=loss, train_op=train_op)
def main(unusedargv):
    # ... 解析参数
    classifier = train(args.train_data, args.vocab)
    print(evaluate(classifier, args.train_data))
    print(evaluate(classifier, args.test_data))
if __name__ == "__main__":
    tf.app.run()

这是完整的堆栈跟踪：

File "categorize.py", line 126, in main
  classifier = train(args.train_data, args.vocab)
File "categorize.py", line 39, in train
  classifier.fit(input_fn=lambda: (training_data, training_targets), steps=2000)
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/util/deprecation.py", line 280, in new_func
  return func(*args, **kwargs)
File "/usr/local/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 426, in fit
  loss = self._train_model(input_fn=input_fn, hooks=hooks)
File "/usr/local/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 934, in _train_model
  model_fn_ops = self._call_legacy_get_train_ops(features, labels)
File "/usr/local/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 1003, in _call_legacy_get_train_ops
  train_ops = self._get_train_ops(features, labels)
File "/usr/local/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 1162, in _get_train_ops
  return self._call_model_fn(features, labels, model_fn_lib.ModeKeys.TRAIN)
File "/usr/local/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 1133, in _call_model_fn
  model_fn_results = self._model_fn(features, labels, **kwargs)
File "categorize.py", line 37, in <lambda>
  classifier = tf.contrib.learn.Estimator(model_fn=lambda features, targets, mode, params: model_fn(features, targets, mode, params, hidden_units))
File "categorize.py", line 73, in model_fn
  learning_rate=0.01)
File "/usr/local/lib/python3.6/site-packages/tensorflow/contrib/layers/python/layers/optimizers.py", line 152, in optimize_loss
  with vs.variable_scope(name, "OptimizeLoss", [loss, global_step]):
File "/usr/local/Cellar/python3/3.6.0_1/Frameworks/Python.framework/Versions/3.6/lib/python3.6/contextlib.py", line 82, in __enter__
  return next(self.gen)
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/ops/variable_scope.py", line 1410, in variable_scope
  g = ops._get_graph_from_inputs(values)  # pylint: disable=protected-access
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 3968, in _get_graph_from_inputs
  _assert_same_graph(original_graph_element, graph_element)
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 3907, in _assert_same_graph
  "%s must be from the same graph as %s." % (item, original_item))
ValueError: Tensor("global_step:0", shape=(), dtype=int64_ref) must be from the same graph as Tensor("softmax_cross_entropy_loss/value:0", shape=(), dtype=float32).

这是我的代码：

回答：

我已经找出了问题所在！这可能与Estimator接口有关，但基本上我需要将我的Tensorflow变量定义移到Estimator中。我最终创建了一个方法来做这件事，但当我在lambda中定义变量时也同样有效：

def train(training_file, vocab_path, hidden_units=[10, 20, 10]):
    """
    给定一个训练CSV文件，训练一个Tensorflow神经网络
    """
    training_set = data.load(training_file)
    vocab = tf.contrib.learn.preprocessing.VocabularyProcessor(data.DOC_LENGTH)
    vocab = vocab.restore(vocab_path)
    # 注意这里没有定义变量
    training_data = training_set.data
    training_targets = np.array(training_set.targets, dtype=np.int32)
    classifier = tf.contrib.learn.Estimator(model_fn=lambda features, targets, mode, params: model_fn(features, targets, mode, params, hidden_units))
    # 注意这里定义了变量
    classifier.fit(
        input_fn=lambda:
            (tf.one_hot(training_data, len(vocab.vocabulary_._mapping), dtype=tf.float32),
             tf.constant(training_targets)),
        steps=2000)
    return classifier