How correctly export_savedmodel() for TensorFlow WALS Matrix Factorization estimator?

Question

I am using this WALS Matrix Factorization module on TensorFlow. After fitting the estimator, I'm trying to save the model using the export_savedmodel() method but I am unable to provide the correct serving_input_fn argument. The code is here:

from tensorflow.contrib.factorization.python.ops import wals as wals_lib

# dense input array that shows the user X item interactions
# here set as dummy array
dense_array = np.ones((10,10))
num_rows, num_cols = dense_array.shape
emebedding_dim = 5 # manually setting hidden factor

factorizer = wals_lib.WALSMatrixFactorization(num_rows, num_cols, embedding_dim, max_sweeps=10)

# this generate_input_fn() is not shown here but it's a copy of
# https://github.com/tensorflow/tensorflow/blob/r1.13/tensorflow/contrib/factorization/python/ops/wals_test.py#L82
input_fn, _, _ = generate_input_fn(np_matrix=dense_array, batch_size=32, mode=model_fn.ModeKeys.TRAIN)
factorizer.fit(input_fn, steps=10)

# MY PROBLEM IS HERE
# How to define the correct serving input function?
factorizer.export_savedmodel('path/to/save/model', serving_input_fn=???)

The tricky part here is that the WALS module is, I believe, using an older paradigm of TensorFlow, where the serving_input_fn argument is expecting a callable function that returns an InputFnOps . However, the more updated Estimators, such as this one, expects a function that returns a tf.estimator.export.ServingInputReceiver or tf.estimator.export.TensorServingInputReceiver . I admit I'm not completely fluent in input functions of TensorFlow yet, but any help for my specific use case of saving my WALS estimator will be greatly appreciated. Thanks!

Answer 1

You're right about the deprecated parts of Tensorflow 1.x. For the WALSMatrixFactorization, your serving_input_fn needs to return an InputFnOps object. A correct input function would be therefore:

def serving_input_receiver_fn():
    # some example input
    receiver_tensors = {'my_input': tf.placeholder(dtype=tf.string, shape=[None, 1], name='foo')}
    # some example feature that completely ignores the input
    features = {
        WALSMatrixFactorization.INPUT_ROWS: tf.SparseTensor(indices=[[0, 0], [1, 2]], values=[1., 2.], dense_shape=[3, 4]),
        WALSMatrixFactorization.INPUT_COLS: tf.SparseTensor(indices=[[0, 0], [1, 2]], values=[3., 4.], dense_shape=[3, 4]),
        WALSMatrixFactorization.PROJECT_ROW: tf.constant(True),
    }
    return tf.contrib.learn.utils.InputFnOps(features, None, receiver_tensors)