Exporting LTR Model in SavedModel Format

Could we please get an example of how to save the produced model in https://github.com/tensorflow/ranking/blob/master/tensorflow_ranking/examples/tf_ranking_libsvm.ipynb in SavedModel format (ref: https://www.tensorflow.org/guide/saved_model). I’ve tried a bunch of variations of exporting from the TensorFlow API docs and GitHub examples with no luck.

feature_columns = example_feature_columns()
feature_spec = tf.feature_column.make_parse_example_spec(feature_columns)
export_fn = tf.estimator.export.build_parsing_serving_input_receiver_fn(feature_spec)
ranker.export_savedmodel("savedmodel", export_fn)

With the example above I get KeyError: NumericColumn(key='1', shape=(1,), default_value=(0.0,), dtype=tf.float32, normalizer_fn=None)

Some guidance on this problem would be greatly appreciated.

[kunalkishore]

@hodeld This might help you:

def make_serving_input_fn():
    """Returns serving input fn."""
    context_feature_spec = tf.feature_column.make_parse_example_spec(
      context_feature_columns().values())
    example_feature_spec = tf.feature_column.make_parse_example_spec(
      example_feature_columns().values())
    return tfr.data.build_ranking_serving_input_receiver_fn(
        data_format="example_list_with_context",
        context_feature_spec=context_feature_spec,
        example_feature_spec=example_feature_spec)

estimator_path = ranker.export_saved_model("path/to/export", make_serving_input_fn())

[working-estimate]

can you share an example of how this works with build_raw_serving_input_receiver_fn? The purpose here is to interface with requests that feed features directly instead of through a tf.Example.

def serving_fn():
    features={str(k):tf.placeholder(shape=(1,),dtype=tf.float32) for k in range(_NUM_FEATURES+1)}
    return tf.estimator.export.build_raw_serving_input_receiver_fn(features)()
ranker.export_saved_model(export_dir, serving_fn)

gives the error: ValueError: slice index 1 of dimension 0 out of bounds. for 'groupwise_dnn_v2/infer_sizes/strided_slice_1' (op: 'StridedSlice') with input shapes: [1], [1], [1], [1] and with computed input tensors: input[1] = <1>, input[2] = <2>, input[3] = <1>.