Tensorflow: ValueError: The last dimension of the inputs to `Dense` should be defined. Found `None`

Question

I've created a keras model with a tensorflow backend but am having difficulty exporting my model for use on ML Engine (as a saved_model.pb). Here's what I'm doing:

dataset = tf.data.Dataset.from_tensor_slices((data_train, labels_train))
dataset = dataset.map(lambda x, y: ({'reviews': x}, y))
val_dataset = tf.data.Dataset.from_tensor_slices((data_test, labels_test))
val_dataset = val_dataset.map(lambda x, y: ({'reviews': x}, y))
dataset = dataset.batch(self.batch_size).repeat()  # repeat infinitely
val_dataset = val_dataset.batch(self.batch_size).repeat()

Then I perform some preprocessing on my Dataset objects:

dataset = dataset.map(lambda x, y: preprocess_text_and_y(x,y))
val_dataset = val_dataset.map(lambda x, y: preprocess_text_and_y(x,y))

I build my keras model and call .fit(...). It all works.

Then I try to export my model, with something like this:

def export(data_vocab):

    estimator = tf.keras.estimator.model_to_estimator(model)

    def serving():
        data_table = tf.contrib.lookup.index_table_from_tensor(tf.constant(self.data_vocab),
                                                                    default_value=0)
        inputs = {
            'reviews': tf.placeholder(shape=[1], dtype=tf.string)
        }
        preproc = inputs.copy()
        preproc = preprocess_text(preproc, data_table)
        return tf.estimator.export.ServingInputReceiver(preproc, inputs)

    estimator.export_savedmodel('./test_export', serving)

And unfortunately, I get back:

ValueError: The last dimension of the inputs to `Dense` should be defined. Found `None`.

I googled around and found this:

How to use TensorFlow Dataset API in combination with dense layers

which says I need to call tf.set_shape(...). I'm preprocessing strings into an array of integers with length 100. I've tried adding x['reviews'].set_shape([100]) in my preprocess_text function

But then that breaks training with:

ValueError: Shapes must be equal rank, but are 2 and 1

Any thoughts on how to fix?

Thanks!

Solution

If you set the shape after batching, you will need to set it to [None, 100] to include the batch axis:

x['reviews'].set_shape([None, 100])