Using the SavedModel format

View on TensorFlow.org

Run in Google Colab

View source on GitHub

Download notebook

A SavedModel contains a complete TensorFlow program, including trained parameters (i.e, tf.Variables) and computation. It does not require the original model building code to run, which makes it useful for sharing or deploying with TFLite, TensorFlow.js, TensorFlow Serving, or TensorFlow Hub.

You can save and load a model in the SavedModel format using the following APIs:

• Low-level tf.saved_model API. This document describes how to use this API in detail.
  • Save: tf.saved_model.save(model, path_to_dir)
  • Load: model = tf.saved_model.load(path_to_dir)
• High-level tf.keras.Model API. Refer to the keras save and serialize guide.
• If you just want to save/load weights during training, refer to the checkpoints guide.

Creating a SavedModel from Keras

For a quick introduction, this section exports a pre-trained Keras model and serves image classification requests with it. The rest of the guide will fill in details and discuss other ways to create SavedModels.

import os
import tempfile

from matplotlib import pyplot as plt
import numpy as np
import tensorflow as tf

tmpdir = tempfile.mkdtemp()

physical_devices = tf.config.list_physical_devices('GPU')
for device in physical_devices:
  tf.config.experimental.set_memory_growth(device, True)

file = tf.keras.utils.get_file(
    "grace_hopper.jpg",
    "https://storage.googleapis.com/download.tensorflow.org/example_images/grace_hopper.jpg")
img = tf.keras.utils.load_img(file, target_size=[224, 224])
plt.imshow(img)
plt.axis('off')
x = tf.keras.utils.img_to_array(img)
x = tf.keras.applications.mobilenet.preprocess_input(
    x[tf.newaxis,...])

You'll use an image of Grace Hopper as a running example, and a Keras pre-trained image classification model since it's easy to use. Custom models work too, and are covered in detail later.

labels_path = tf.keras.utils.get_file(
    'ImageNetLabels.txt',
    'https://storage.googleapis.com/download.tensorflow.org/data/ImageNetLabels.txt')
imagenet_labels = np.array(open(labels_path).read().splitlines())

pretrained_model = tf.keras.applications.MobileNet()
result_before_save = pretrained_model(x)

decoded = imagenet_labels[np.argsort(result_before_save)[0,::-1][:5]+1]

print("Result before saving:\n", decoded)

The top prediction for this image is "military uniform".

mobilenet_save_path = os.path.join(tmpdir, "mobilenet/1/")
tf.saved_model.save(pretrained_model, mobilenet_save_path)

The save-path follows a convention used by TensorFlow Serving where the last path component (1/ here) is a version number for your model - it allows tools like Tensorflow Serving to reason about the relative freshness.

You can load the SavedModel back into Python with tf.saved_model.load and see how Admiral Hopper's image is classified.

loaded = tf.saved_model.load(mobilenet_save_path)
print(list(loaded.signatures.keys()))  # ["serving_default"]

Imported signatures always return dictionaries. To customize signature names and output dictionary keys, see Specifying signatures during export.

infer = loaded.signatures["serving_default"]
print(infer.structured_outputs)

Running inference from the SavedModel gives the same result as the original model.

labeling = infer(tf.constant(x))[pretrained_model.output_names[0]]

decoded = imagenet_labels[np.argsort(labeling)[0,::-1][:5]+1]

print("Result after saving and loading:\n", decoded)

Running a SavedModel in TensorFlow Serving

SavedModels are usable from Python (more on that below), but production environments typically use a dedicated service for inference without running Python code. This is easy to set up from a SavedModel using TensorFlow Serving.

See the TensorFlow Serving REST tutorial for an end-to-end tensorflow-serving example.

The SavedModel format on disk

A SavedModel is a directory containing serialized signatures and the state needed to run them, including variable values and vocabularies.

ls {mobilenet_save_path}

The saved_model.pb file stores the actual TensorFlow program, or model, and a set of named signatures, each identifying a function that accepts tensor inputs and produces tensor outputs.

SavedModels may contain multiple variants of the model (multiple v1.MetaGraphDefs, identified with the --tag_set flag to saved_model_cli), but this is rare. APIs which create multiple variants of a model include tf.Estimator.experimental_export_all_saved_models and in TensorFlow 1.x tf.saved_model.Builder.

saved_model_cli show --dir {mobilenet_save_path} --tag_set serve

The variables directory contains a standard training checkpoint (see the guide to training checkpoints).

ls {mobilenet_save_path}/variables

The assets directory contains files used by the TensorFlow graph, for example text files used to initialize vocabulary tables. It is unused in this example.

SavedModels may have an assets.extra directory for any files not used by the TensorFlow graph, for example information for consumers about what to do with the SavedModel. TensorFlow itself does not use this directory.

The fingerprint.pb file contains the fingerprint of the SavedModel, which is composed of several 64-bit hashes that uniquely identify the contents of the SavedModel. The fingerprinting API is currently experimental, but tf.saved_model.experimental.read_fingerprint can be used to read the SavedModel fingerprint into a tf.saved_model.experimental.Fingerprint object.

Saving a custom model

tf.saved_model.save supports saving tf.Module objects and its subclasses, like tf.keras.Layer and tf.keras.Model.

Let's look at an example of saving and restoring a tf.Module.

class CustomModule(tf.Module):

  def __init__(self):
    super(CustomModule, self).__init__()
    self.v = tf.Variable(1.)

  @tf.function
  def __call__(self, x):
    print('Tracing with', x)
    return x * self.v

  @tf.function(input_signature=[tf.TensorSpec([], tf.float32)])
  def mutate(self, new_v):
    self.v.assign(new_v)

module = CustomModule()

When you save a tf.Module, any tf.Variable attributes, tf.function-decorated methods, and tf.Modules found via recursive traversal are saved. (See the Checkpoint tutorial for more about this recursive traversal.) However, any Python attributes, functions, and data are lost. This means that when a tf.function is saved, no Python code is saved.

If no Python code is saved, how does SavedModel know how to restore the function?

Briefly, tf.function works by tracing the Python code to generate a ConcreteFunction (a callable wrapper around tf.Graph). When saving a tf.function, you're really saving the tf.function's cache of ConcreteFunctions.

To learn more about the relationship between tf.function and ConcreteFunctions, refer to the tf.function guide.

module_no_signatures_path = os.path.join(tmpdir, 'module_no_signatures')
module(tf.constant(0.))
print('Saving model...')
tf.saved_model.save(module, module_no_signatures_path)

Loading and using a custom model

When you load a SavedModel in Python, all tf.Variable attributes, tf.function-decorated methods, and tf.Modules are restored in the same object structure as the original saved tf.Module.

imported = tf.saved_model.load(module_no_signatures_path)
assert imported(tf.constant(3.)).numpy() == 3
imported.mutate(tf.constant(2.))
assert imported(tf.constant(3.)).numpy() == 6

Because no Python code is saved, calling a tf.function with a new input signature will fail:

imported(tf.constant([3.]))

ValueError: Could not find matching function to call for canonicalized inputs ((,), {}). Only existing signatures are [((TensorSpec(shape=(), dtype=tf.float32, name=u'x'),), {})].

Basic fine-tuning

Variable objects are available, and you can backprop through imported functions. That is enough to fine-tune (i.e. retrain) a SavedModel in simple cases.

optimizer = tf.keras.optimizers.SGD(0.05)

def train_step():
  with tf.GradientTape() as tape:
    loss = (10. - imported(tf.constant(2.))) ** 2
  variables = tape.watched_variables()
  grads = tape.gradient(loss, variables)
  optimizer.apply_gradients(zip(grads, variables))
  return loss

for _ in range(10):
  # "v" approaches 5, "loss" approaches 0
  print("loss={:.2f} v={:.2f}".format(train_step(), imported.v.numpy()))

General fine-tuning

A SavedModel from Keras provides more details than a plain __call__ to address more advanced cases of fine-tuning. TensorFlow Hub recommends to provide the following of those, if applicable, in SavedModels shared for the purpose of fine-tuning:

• If the model uses dropout or another technique in which the forward pass differs between training and inference (like batch normalization), the __call__ method takes an optional, Python-valued training= argument that defaults to False but can be set to True.
• Next to the __call__ attribute, there are .variable and .trainable_variable attributes with the corresponding lists of variables. A variable that was originally trainable but is meant to be frozen during fine-tuning is omitted from .trainable_variables.
• For the sake of frameworks like Keras that represent weight regularizers as attributes of layers or sub-models, there can also be a .regularization_losses attribute. It holds a list of zero-argument functions whose values are meant for addition to the total loss.

Going back to the initial MobileNet example, you can see some of those in action:

loaded = tf.saved_model.load(mobilenet_save_path)
print("MobileNet has {} trainable variables: {}, ...".format(
          len(loaded.trainable_variables),
          ", ".join([v.name for v in loaded.trainable_variables[:5]])))

trainable_variable_ids = {id(v) for v in loaded.trainable_variables}
non_trainable_variables = [v for v in loaded.variables
                           if id(v) not in trainable_variable_ids]
print("MobileNet also has {} non-trainable variables: {}, ...".format(
          len(non_trainable_variables),
          ", ".join([v.name for v in non_trainable_variables[:3]])))

Specifying signatures during export

Tools like TensorFlow Serving and saved_model_cli can interact with SavedModels. To help these tools determine which ConcreteFunctions to use, you need to specify serving signatures. tf.keras.Models automatically specify serving signatures, but you'll have to explicitly declare a serving signature for our custom modules.

By default, no signatures are declared in a custom tf.Module.

assert len(imported.signatures) == 0

To declare a serving signature, specify a ConcreteFunction using the signatures kwarg. When specifying a single signature, its signature key will be 'serving_default', which is saved as the constant tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY.

module_with_signature_path = os.path.join(tmpdir, 'module_with_signature')
call = module.__call__.get_concrete_function(tf.TensorSpec(None, tf.float32))
tf.saved_model.save(module, module_with_signature_path, signatures=call)

imported_with_signatures = tf.saved_model.load(module_with_signature_path)
list(imported_with_signatures.signatures.keys())

To export multiple signatures, pass a dictionary of signature keys to ConcreteFunctions. Each signature key corresponds to one ConcreteFunction.

module_multiple_signatures_path = os.path.join(tmpdir, 'module_with_multiple_signatures')
signatures = {"serving_default": call,
              "array_input": module.__call__.get_concrete_function(tf.TensorSpec([None], tf.float32))}

tf.saved_model.save(module, module_multiple_signatures_path, signatures=signatures)

imported_with_multiple_signatures = tf.saved_model.load(module_multiple_signatures_path)
list(imported_with_multiple_signatures.signatures.keys())

By default, the output tensor names are fairly generic, like output_0. To control the names of outputs, modify your tf.function to return a dictionary that maps output names to outputs. The names of inputs are derived from the Python function arg names.

class CustomModuleWithOutputName(tf.Module):
  def __init__(self):
    super(CustomModuleWithOutputName, self).__init__()
    self.v = tf.Variable(1.)

  @tf.function(input_signature=[tf.TensorSpec(None, tf.float32)])
  def __call__(self, x):
    return {'custom_output_name': x * self.v}

module_output = CustomModuleWithOutputName()
call_output = module_output.__call__.get_concrete_function(tf.TensorSpec(None, tf.float32))
module_output_path = os.path.join(tmpdir, 'module_with_output_name')
tf.saved_model.save(module_output, module_output_path,
                    signatures={'serving_default': call_output})

imported_with_output_name = tf.saved_model.load(module_output_path)
imported_with_output_name.signatures['serving_default'].structured_outputs

Proto-splitting

Due to limits of the protobuf implementation, proto sizes cannot exceed 2GB. This can lead to the following errors when attempting to save very large models:

ValueError: Message tensorflow.SavedModel exceeds maximum protobuf size of 2GB: ...

google.protobuf.message.DecodeError: Error parsing message as the message exceeded the protobuf limit with type 'tensorflow.GraphDef'

If you wish to save models that exceed the 2GB limit, then you'll need to save using the new proto-splitting option:

tf.saved_model.save(
  ...,
  options=tf.saved_model.SaveOptions(experimental_image_format=True)
)

More information can be found in the Proto Splitter / Merger Library guide.

Load a SavedModel in C++

The C++ version of the SavedModel loader provides an API to load a SavedModel from a path, while allowing SessionOptions and RunOptions. You have to specify the tags associated with the graph to be loaded. The loaded version of SavedModel is referred to as SavedModelBundle and contains the MetaGraphDef and the session within which it is loaded.

const string export_dir = ...
SavedModelBundle bundle;
...
LoadSavedModel(session_options, run_options, export_dir, {kSavedModelTagTrain},
               &bundle);

$ bazel build //tensorflow/python/tools:saved_model_cli

A SavedModel contains one or more model variants (technically, v1.MetaGraphDefs), identified by their tag-sets. To serve a model, you might wonder what kind of SignatureDefs are in each model variant, and what are their inputs and outputs. The show command let you examine the contents of the SavedModel in hierarchical order. Here's the syntax:

usage: saved_model_cli show [-h] --dir DIR [--all]
[--tag_set TAG_SET] [--signature_def SIGNATURE_DEF_KEY]

For example, the following command shows all available tag-sets in the SavedModel:

$ saved_model_cli show --dir /tmp/saved_model_dir
The given SavedModel contains the following tag-sets:
serve
serve, gpu

The following command shows all available SignatureDef keys for a tag set:

$ saved_model_cli show --dir /tmp/saved_model_dir --tag_set serve
The given SavedModel `MetaGraphDef` contains `SignatureDefs` with the
following keys:
SignatureDef key: "classify_x2_to_y3"
SignatureDef key: "classify_x_to_y"
SignatureDef key: "regress_x2_to_y3"
SignatureDef key: "regress_x_to_y"
SignatureDef key: "regress_x_to_y2"
SignatureDef key: "serving_default"

If there are multiple tags in the tag-set, you must specify all tags, each tag separated by a comma. For example:

$ saved_model_cli show --dir /tmp/saved_model_dir --tag_set serve,gpu

To show all inputs and outputs TensorInfo for a specific SignatureDef, pass in the SignatureDef key to signature_def option. This is very useful when you want to know the tensor key value, dtype and shape of the input tensors for executing the computation graph later. For example:

$ saved_model_cli show --dir \
/tmp/saved_model_dir --tag_set serve --signature_def serving_default
The given SavedModel SignatureDef contains the following input(s):
  inputs['x'] tensor_info:
      dtype: DT_FLOAT
      shape: (-1, 1)
      name: x:0
The given SavedModel SignatureDef contains the following output(s):
  outputs['y'] tensor_info:
      dtype: DT_FLOAT
      shape: (-1, 1)
      name: y:0
Method name is: tensorflow/serving/predict

To show all available information in the SavedModel, use the --all option. For example:

$ saved_model_cli show --dir /tmp/saved_model_dir --all
MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:

signature_def['classify_x2_to_y3']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['inputs'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 1)
        name: x2:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['scores'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 1)
        name: y3:0
  Method name is: tensorflow/serving/classify

...

signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['x'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 1)
        name: x:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['y'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 1)
        name: y:0
  Method name is: tensorflow/serving/predict

run command

Invoke the run command to run a graph computation, passing inputs and then displaying (and optionally saving) the outputs. Here's the syntax:

usage: saved_model_cli run [-h] --dir DIR --tag_set TAG_SET --signature_def
                           SIGNATURE_DEF_KEY [--inputs INPUTS]
                           [--input_exprs INPUT_EXPRS]
                           [--input_examples INPUT_EXAMPLES] [--outdir OUTDIR]
                           [--overwrite] [--tf_debug]

The run command provides the following three ways to pass inputs to the model:

• --inputs option enables you to pass numpy ndarray in files.
• --input_exprs option enables you to pass Python expressions.
• --input_examples option enables you to pass tf.train.Example.

--inputs

To pass input data in files, specify the --inputs option, which takes the following general format:

--inputs <INPUTS>

where INPUTS is either of the following formats:

• <input_key>=<filename>
• <input_key>=<filename>[<variable_name>]

You may pass multiple INPUTS. If you do pass multiple inputs, use a semicolon to separate each of the INPUTS.

saved_model_cli uses numpy.load to load the filename. The filename may be in any of the following formats:

• .npy
• .npz
• pickle format

A .npy file always contains a numpy ndarray. Therefore, when loading from a .npy file, the content will be directly assigned to the specified input tensor. If you specify a variable_name with that .npy file, the variable_name will be ignored and a warning will be issued.

When loading from a .npz (zip) file, you may optionally specify a variable_name to identify the variable within the zip file to load for the input tensor key. If you don't specify a variable_name, the SavedModel CLI will check that only one file is included in the zip file and load it for the specified input tensor key.

When loading from a pickle file, if no variable_name is specified in the square brackets, whatever that is inside the pickle file will be passed to the specified input tensor key. Otherwise, the SavedModel CLI will assume a dictionary is stored in the pickle file and the value corresponding to the variable_name will be used.

--input_exprs

To pass inputs through Python expressions, specify the --input_exprs option. This can be useful for when you don't have data files lying around, but still want to sanity check the model with some simple inputs that match the dtype and shape of the model's SignatureDefs. For example:

`<input_key>=[[1],[2],[3]]`

In addition to Python expressions, you may also pass numpy functions. For example:

`<input_key>=np.ones((32,32,3))`

(Note that the numpy module is already available to you as np.)

--input_examples

To pass tf.train.Example as inputs, specify the --input_examples option. For each input key, it takes a list of dictionary, where each dictionary is an instance of tf.train.Example. The dictionary keys are the features and the values are the value lists for each feature. For example:

`<input_key>=[{"age":[22,24],"education":["BS","MS"]}]`

Save output

By default, the SavedModel CLI writes output to stdout. If a directory is passed to --outdir option, the outputs will be saved as .npy files named after output tensor keys under the given directory.

Use --overwrite to overwrite existing output files.