Model export

Could you please give a code example of how to export a model for TensorFlow Serving? No luck with estimator.export_saved_model or tf.estimator.BestExporter. I must be doing something wrong with feature_spec.

[eggie5]

I ran the exporter (above) on my ranker. (my model only has 1 input, a sparse id (used for an embedding lookup) and it has group size 1. )

I understand the input shape (-1) is a protobuf string. However, I thought the output shape was interesting. I would score my documents using the predict signature right? It has a (-1, -1) shape. What does this mean? I would expect [-1, 1] for a batch of scalars, where each scalar is a document score… @ramakumar1729 ?

$ saved_model_cli show --dir . --tag_set serve --signature_def predict
The given SavedModel SignatureDef contains the following input(s):
  inputs['examples'] tensor_info:
      dtype: DT_STRING
      shape: (-1)
      name: input_example_tensor:0
The given SavedModel SignatureDef contains the following output(s):
  outputs['output'] tensor_info:
      dtype: DT_FLOAT
      shape: (-1, -1)
      name: groupwise_dnn_v2/accumulate_scores/truediv:0
Method name is: tensorflow/serving/predict

The example below will score 1 document. How can I score 2,3 or a batch of docs?

saved_model_cli run \
    --dir . \
    --tag_set serve \
    --signature_def predict \
    --input_examples 'examples=[{"1":[534]}]'

Result for output key output:
[[0.8916328]]

[ramakumar1729]

Thanks for the quick reaction. I think I’ve figured it out (assuming this is executed in examples/tf_ranking_libsvm.py):

def serving_input_receiver_fn():
    feature_names = ["{}".format(i + 1) for i in range(FLAGS.num_features)]
    feature_columns = [tf.feature_column.numeric_column(
          name, shape=(1,), default_value=0.0) for name in feature_names]
    feature_spec = tf.feature_column.make_parse_example_spec(feature_columns)

    return  tf.estimator.export.build_parsing_serving_input_receiver_fn(feature_spec)()

I’m not sure if the last brackets are needed. Rest of the code looks fine.

estimator.export_savedmodel(FLAGS.output_dir + ‘/export’, serving_input_receiver_fn)

Does it make sense?