Using classes with predict_signature_def

[viksit]

When using classes in the following,

# class_list = ["a", "B", ...]
class_tensor = tf.constant(class_list)
indices = tf.constant(list(range(0,len(class_list))))
table = tf.contrib.lookup.index_to_string_table_from_tensor(class_tensor)
classes = table.lookup(tf.to_int64(indices))

signature = predict_signature_def(inputs={"input": inputs},
                                  outputs={
                                  "scores": outputs,
                                  "classes": classes})

I get the following error,

{
  "error": "Prediction failed during encoding instances"
}

Without the classes key though, I get the output back just fine. What could be causing this? All examples in the repo seem to use signature_def_utils.build_signature_def() vs just predict_signature_def, so I'm not sure what could be going wrong here.

[kirilg]

build_signature_def() takes inputs and outputs as maps of string_key->TensorInfo protos.
predict_signature_def() takes inputs and outputs as dicts of type Tensor. It builds TensorInfos out of Tensors, and if there is one input and one output, standardizes the keys to "inputs" and "outputs" (compare that to your keys of "input", "scores", and "classes"). I suspect there is a mismatch of keys in the output which changes to "outputs" when you remove scores because then you're down to just one tensor.

Where are you getting the error "Prediction failed during encoding instances"? Is this at training time, at request time when querying the model that's loaded in a model_server? If that latter, could you also include a sample request?

[sukritiramesh]

Closing given @kirilg's response; please reopen if required.