TensorFlow: “Cannot capture a stateful node by value” in tf.contrib.data API

Question

For transfer learning, one often uses a network as a feature extractor to create a dataset of features, on which another classifier is trained (e.g. a SVM).

I want to implement this using the Dataset API (tf.contrib.data) and dataset.map():

# feature_extractor will create a CNN on top of the given tensor
def features(feature_extractor, ...):
    dataset = inputs(...)  # This creates a dataset of (image, label) pairs

    def map_example(image, label):
        features = feature_extractor(image, trainable=False)
        #  Leaving out initialization from a checkpoint here... 
        return features, label

    dataset = dataset.map(map_example)

    return dataset

Doing this fails when creating an iterator for the dataset.

ValueError: Cannot capture a stateful node by value.

This is true, the kernels and biases of the network are variables and thus stateful. For this particular example they don't have to be though.

Is there a way to make Ops and specifically tf.Variable objects stateless?

Since I'm using tf.layers I cannot simply create them as constants, and setting trainable=False won't create constants neither but just won't add the variables to the GraphKeys.TRAINABLE_VARIABLES collection.

Answer 1

Unfortunately, tf.Variable is inherently stateful. However, this error only arises if you use Dataset.make_one_shot_iterator() to create the iterator.* To avoid the problem, you can instead use Dataset.make_initializable_iterator(), with the caveat that you must also run iterator.initializer on the returned iterator after running the initializer for the tf.Variable objects used in the input pipeline.

---

* The reason for this limitation is an implementation detail of Dataset.make_one_shot_iterator() and the work-in-progress TensorFlow function (Defun) support that it uses to encapsulate the dataset definition. Since using stateful resources like lookup tables and variables has been more popular than we initially imagined, we're looking into ways to relax this restriction.