Running a loaded TF2 SavedModel gives TensorfFlowException TF_INVALID_ARGUMENT

[daanrs]

I'm trying to load a TF2 SavedModel. I'm not sure if I'm doing it right, but I get the same error no matter what I do:

*** Exception: TensorFlowException TF_INVALID_ARGUMENT "Failed to find function \"__inference_signature_wrapper_9381\" in function library: "

I get this using any model I try. Example of a model I tried

import tensorflow as tf

model = tf.keras.applications.MobileNet()

mobilenet_save_path = "mobilenet/1/"
tf.saved_model.save(model, mobilenet_save_path)

I use something like

import "bytestring" Data.ByteString as BS (readFile)
import "proto-lens" Data.ProtoLens (decodeMessageOrDie)
import "vector" Data.Vector.Storable as V
import "tensorflow-proto" Proto.Tensorflow.Core.Protobuf.MetaGraph_Fields (graphDef)
import "tensorflow-proto" Proto.Tensorflow.Core.Protobuf.SavedModel (SavedModel)
import "tensorflow-proto" Proto.Tensorflow.Core.Protobuf.SavedModel_Fields (metaGraphs)
import qualified "tensorflow" TensorFlow.Core as TF
import qualified "tensorflow" TensorFlow.Tensor as TF

test_main :: IO ()
test_main = do
  let path = "mobilenet/1/saved_model.pb"
  savedModel <- decodeMessageOrDie <$> BS.readFile path :: IO SavedModel
  let [metaGraph] = view metaGraphs savedModel
      graphDef1 = view graphDef metaGraph

  tensor <- TF.runSession $ do
    TF.build $ TF.addGraphDef graphDef1
    TF.run $ TF.tensorRefFromName "serving_default_input_1" :: TF.Session (V.Vector Double)
  print tensor

Help would be much appreciated!

[blackgnezdo]

Chances are the saved file format is too new for the TF library you use. You can see that our version is 2.3.0 and it is about 2 years old now. You can look #267 to figure out what it takes to update the library version.

[Minnozz]

I'm updating the bindings to TF 2.10.1 and am getting the same error initially. I will try out a different method of loading the protobuf (there is a new function in the C bindings) and see if that makes any difference.