How to export a SavedModel in TensorFlow

Model training and model delivery are separate lifecycle stages. A SavedModel packages the TensorFlow graph, learned variables, assets, and callable endpoints into a directory that an inference runtime can load without the original model-building code.

For a current Keras model, model.export() writes an inference artifact rather than a full training archive. Use the native .keras format instead when another Keras process must restore optimizer state or resume training.

The exported artifact must remain a complete directory, including saved_model.pb and the variables subdirectory. Load it through tf.saved_model.load() and call the default serving signature before handing its input and output contract to TensorFlow Serving or another consumer.

1. Create export_savedmodel_demo.py with the TensorFlow import, deterministic seed, and versioned export path.

   export_savedmodel_demo.py

   import os
   from pathlib import Path
    
   os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
    
   import tensorflow as tf
    
   tf.get_logger().setLevel("ERROR")
   tf.keras.utils.set_random_seed(7)
    
   export_dir = Path("exported/number_classifier/1")

2. Append the training tensors to export_savedmodel_demo.py.

   export_savedmodel_demo.py

   features = tf.constant(
       [
           [0.0, 0.0, 0.0, 0.0],
           [0.0, 1.0, 0.0, 1.0],
           [1.0, 0.0, 1.0, 0.0],
           [1.0, 1.0, 1.0, 1.0],
           [0.2, 0.9, 0.1, 0.8],
           [0.8, 0.2, 0.9, 0.1],
       ],
       dtype=tf.float32,
   )
   labels = tf.constant([[0.0], [0.0], [1.0], [1.0], [0.0], [1.0]], dtype=tf.float32)

3. Append the model definition and training pass to export_savedmodel_demo.py.

   export_savedmodel_demo.py

   inputs = tf.keras.Input(shape=(4,), name="features")
   hidden = tf.keras.layers.Dense(8, activation="relu")(inputs)
   scores = tf.keras.layers.Dense(1, activation="sigmoid")(hidden)
   model = tf.keras.Model(inputs=inputs, outputs=scores, name="number_classifier")
   model.compile(optimizer="adam", loss="binary_crossentropy")
   model.fit(features, labels, epochs=12, verbose=0)

4. Append the export and reloaded-signature check to export_savedmodel_demo.py.

   export_savedmodel_demo.py

   model.export(export_dir)
    
   restored = tf.saved_model.load(str(export_dir))
   signature = restored.signatures["serving_default"]
   _, keyword_inputs = signature.structured_input_signature
   input_name, input_spec = next(iter(keyword_inputs.items()))
   sample_batch = tf.constant(
       [
           [0.1, 0.9, 0.2, 0.8],
           [0.9, 0.1, 0.8, 0.2],
       ],
       dtype=input_spec.dtype,
   )
   prediction = signature(**{input_name: sample_batch})
   output_name, output_tensor = next(iter(prediction.items()))
    
   print(f"TensorFlow: {tf.__version__}")
   print(f"SavedModel directory: {export_dir}")
   print(f"Contains SavedModel: {tf.saved_model.contains_saved_model(str(export_dir))}")
   print(f"Signature names: {sorted(restored.signatures.keys())}")
   print(f"Input: {input_name} shape={tuple(input_spec.shape)} dtype={input_spec.dtype.name}")
   print(f"Output: {output_name} shape={tuple(output_tensor.shape)}")

   The numeric 1 directory follows the version layout used by TensorFlow Serving, where each changed serving artifact needs its own version number.

5. Run the completed script to exercise the SavedModel export through a separately loaded serving signature.

   $ python3 export_savedmodel_demo.py
   Saved artifact at 'exported/number_classifier/1'. The following endpoints are available:
   
   * Endpoint 'serve'
     args_0 (POSITIONAL_ONLY): TensorSpec(shape=(None, 4), dtype=tf.float32, name='features')
   Output Type:
     TensorSpec(shape=(None, 1), dtype=tf.float32, name=None)
   ##### snipped #####
   TensorFlow: 2.21.0
   SavedModel directory: exported/number_classifier/1
   Contains SavedModel: True
   Signature names: ['serve', 'serving_default']
   Input: features shape=(None, 4) dtype=float32
   Output: output_0 shape=(2, 1)

   Contains SavedModel: True confirms the export format, while the final shape comes from inference through the separately loaded signature.

6. Inspect the exported directory for the SavedModel protobuf and variable checkpoint.

   $ ls -R exported/number_classifier/1
   exported/number_classifier/1:
   assets
   fingerprint.pb
   saved_model.pb
   variables
   
   exported/number_classifier/1/assets:
   
   exported/number_classifier/1/variables:
   variables.data-00000-of-00001
   variables.index

   The whole version directory is the deployable unit because model variables and assets can live outside saved_model.pb.

7. Inspect the default serving signature before handing the export to another runtime.

   $ saved_model_cli show --dir exported/number_classifier/1 --tag_set serve --signature_def serving_default
   The given SavedModel SignatureDef contains the following input(s):
     inputs['features'] tensor_info:
         dtype: DT_FLOAT
         shape: (-1, 4)
         name: serving_default_features:0
   The given SavedModel SignatureDef contains the following output(s):
     outputs['output_0'] tensor_info:
         dtype: DT_FLOAT
         shape: (-1, 1)
         name: StatefulPartitionedCall_1:0
   Method name is: tensorflow/serving/predict

   The features input and output_0 output are the field names a serving client must map.
   Related: How to inspect a TensorFlow SavedModel with saved_model_cli
   Related: How to deploy TensorFlow Serving with Docker

8. Run the exported serving_default signature with two four-value feature rows to prove its checkpoint variables produce predictions.

   $ saved_model_cli run \
     --dir exported/number_classifier/1 \
     --tag_set serve \
     --signature_def serving_default \
     --input_exprs 'features=[[0.1,0.9,0.2,0.8],[0.9,0.1,0.8,0.2]]'
   ##### snipped #####
   INFO:tensorflow:Restoring parameters from exported/number_classifier/1/variables/variables
   Result for output key output_0:
   [[0.56982714]
    [0.41488308]]

   The restore message confirms that saved_model_cli loaded the variable checkpoint before the two prediction rows were returned.