tf.saved_model.save

TensorFlow 2 version

View source on GitHub

Exports the Trackable object obj to SavedModel format.

tf.saved_model.save(
    obj, export_dir, signatures=None
)

Example usage:

class Adder(tf.Module):

  @tf.function(input_signature=[tf.TensorSpec(shape=None, dtype=tf.float32)])
  def add(self, x):
    return x + x + 1.

to_export = Adder()
tf.saved_model.save(to_export, '/tmp/adder')

The resulting SavedModel is then servable with an input named "x", its value having any shape and dtype float32.

The optional signatures argument controls which methods in obj will be available to programs which consume SavedModels, for example serving APIs. Python functions may be decorated with @tf.function(input_signature=...) and passed as signatures directly, or lazily with a call to get_concrete_function on the method decorated with @tf.function.

If the signatures argument is omitted, obj will be searched for @tf.function-decorated methods. If exactly one @tf.function is found, that method will be used as the default signature for the SavedModel. This behavior is expected to change in the future, when a corresponding tf.saved_model.load symbol is added. At that point signatures will be completely optional, and any @tf.function attached to obj or its dependencies will be exported for use with load.

When invoking a signature in an exported SavedModel, Tensor arguments are identified by name. These names will come from the Python function's argument names by default. They may be overridden by specifying a name=... argument in the corresponding tf.TensorSpec object. Explicit naming is required if multiple Tensors are passed through a single argument to the Python function.

The outputs of functions used as signatures must either be flat lists, in which case outputs will be numbered, or a dictionary mapping string keys to Tensor, in which case the keys will be used to name outputs.

Signatures are available in objects returned by tf.saved_model.load as a .signatures attribute. This is a reserved attribute: tf.saved_model.save on an object with a custom .signatures attribute will raise an exception.

Since tf.keras.Model objects are also Trackable, this function can be used to export Keras models. For example, exporting with a signature specified:

class Model(tf.keras.Model):

  @tf.function(input_signature=[tf.TensorSpec(shape=[None], dtype=tf.string)])
  def serve(self, serialized):
    ...

m = Model()
tf.saved_model.save(m, '/tmp/saved_model/')

Exporting from a function without a fixed signature:

class Model(tf.keras.Model):

  @tf.function
  def call(self, x):
    ...

m = Model()
tf.saved_model.save(
    m, '/tmp/saved_model/',
    signatures=m.call.get_concrete_function(
        tf.TensorSpec(shape=[None, 3], dtype=tf.float32, name="inp")))

tf.keras.Model instances constructed from inputs and outputs already have a signature and so do not require a @tf.function decorator or a signatures argument. If neither are specified, the model's forward pass is exported.

x = input_layer.Input((4,), name="x")
y = core.Dense(5, name="out")(x)
model = training.Model(x, y)
tf.saved_model.save(model, '/tmp/saved_model/')
# The exported SavedModel takes "x" with shape [None, 4] and returns "out"
# with shape [None, 5]

Variables must be tracked by assigning them to an attribute of a tracked object or to an attribute of obj directly. TensorFlow objects (e.g. layers from tf.keras.layers, optimizers from tf.train) track their variables automatically. This is the same tracking scheme that tf.train.Checkpoint uses, and an exported Checkpoint object may be restored as a training checkpoint by pointing tf.train.Checkpoint.restore to the SavedModel's "variables/" subdirectory. Currently variables are the only stateful objects supported by tf.saved_model.save, but others (e.g. tables) will be supported in the future.

tf.function does not hard-code device annotations from outside the function body, instead using the calling context's device. This means for example that exporting a model which runs on a GPU and serving it on a CPU will generally work, with some exceptions. tf.device annotations inside the body of the function will be hard-coded in the exported model; this type of annotation is discouraged. Device-specific operations, e.g. with "cuDNN" in the name or with device-specific layouts, may cause issues. Currently a DistributionStrategy is another exception: active distribution strategies will cause device placements to be hard-coded in a function. Exporting a single-device computation and importing under a DistributionStrategy is not currently supported, but may be in the future.

SavedModels exported with tf.saved_model.save strip default-valued attributes automatically, which removes one source of incompatibilities when the consumer of a SavedModel is running an older TensorFlow version than the producer. There are however other sources of incompatibilities which are not handled automatically, such as when the exported model contains operations which the consumer does not have definitions for.

Eager Compatibility

Not well supported when graph building. From TensorFlow 1.x, tf.compat.v1.enable_eager_execution() should run first. Calling tf.saved_model.save in a loop when graph building from TensorFlow 1.x will add new save operations to the default graph each iteration.

May not be called from within a function body.