View source on GitHub |
ExportArchive is used to write SavedModel artifacts (e.g. for inference).
tf.keras.export.ExportArchive()
If you have a Keras model or layer that you want to export as SavedModel for
serving (e.g. via TensorFlow-Serving), you can use ExportArchive
to configure the different serving endpoints you need to make available,
as well as their signatures. Simply instantiate an ExportArchive
,
use track()
to register the layer(s) or model(s) to be used,
then use the add_endpoint()
method to register a new serving endpoint.
When done, use the write_out()
method to save the artifact.
The resulting artifact is a SavedModel and can be reloaded via
tf.saved_model.load
.
Examples:
Here's how to export a model for inference.
export_archive = ExportArchive()
export_archive.track(model)
export_archive.add_endpoint(
name="serve",
fn=model.call,
input_signature=[tf.TensorSpec(shape=(None, 3), dtype=tf.float32)],
)
export_archive.write_out("path/to/location")
# Elsewhere, we can reload the artifact and serve it.
# The endpoint we added is available as a method:
serving_model = tf.saved_model.load("path/to/location")
outputs = serving_model.serve(inputs)
Here's how to export a model with one endpoint for inference and one endpoint for a training-mode forward pass (e.g. with dropout on).
export_archive = ExportArchive()
export_archive.track(model)
export_archive.add_endpoint(
name="call_inference",
fn=lambda x: model.call(x, training=False),
input_signature=[tf.TensorSpec(shape=(None, 3), dtype=tf.float32)],
)
export_archive.add_endpoint(
name="call_training",
fn=lambda x: model.call(x, training=True),
input_signature=[tf.TensorSpec(shape=(None, 3), dtype=tf.float32)],
)
export_archive.write_out("path/to/location")
Note on resource tracking:
ExportArchive
is able to automatically track all tf.Variables
used
by its endpoints, so most of the time calling .track(model)
is not strictly required. However, if your model uses lookup layers such
as IntegerLookup
, StringLookup
, or TextVectorization
,
it will need to be tracked explicitly via .track(model)
.
Explicit tracking is also required if you need to be able to access
the properties variables
, trainable_variables
, or
non_trainable_variables
on the revived archive.
Methods
add_endpoint
add_endpoint(
name, fn, input_signature=None
)
Register a new serving endpoint.
Arguments | |
---|---|
name
|
Str, name of the endpoint. |
fn
|
A function. It should only leverage resources
(e.g. tf.Variable objects or tf.lookup.StaticHashTable
objects) that are available on the models/layers
tracked by the ExportArchive (you can call .track(model)
to track a new model).
The shape and dtype of the inputs to the function must be
known. For that purpose, you can either 1) make sure that
fn is a tf.function that has been called at least once, or
2) provide an |
input_signature
|
Used to specify the shape and dtype of the
inputs to fn . List of tf.TensorSpec objects (one
per positional input argument of fn ). Nested arguments are
allowed (see below for an example showing a Functional model
with 2 input arguments).
|
Example:
Adding an endpoint using the input_signature
argument when the
model has a single input argument:
export_archive = ExportArchive()
export_archive.track(model)
export_archive.add_endpoint(
name="serve",
fn=model.call,
input_signature=[tf.TensorSpec(shape=(None, 3), dtype=tf.float32)],
)
Adding an endpoint using the input_signature
argument when the
model has two positional input arguments:
export_archive = ExportArchive()
export_archive.track(model)
export_archive.add_endpoint(
name="serve",
fn=model.call,
input_signature=[
tf.TensorSpec(shape=(None, 3), dtype=tf.float32),
tf.TensorSpec(shape=(None, 4), dtype=tf.float32),
],
)
Adding an endpoint using the input_signature
argument when the
model has one input argument that is a list of 2 tensors (e.g.
a Functional model with 2 inputs):
model = keras.Model(inputs=[x1, x2], outputs=outputs)
export_archive = ExportArchive()
export_archive.track(model)
export_archive.add_endpoint(
name="serve",
fn=model.call,
input_signature=[
[
tf.TensorSpec(shape=(None, 3), dtype=tf.float32),
tf.TensorSpec(shape=(None, 4), dtype=tf.float32),
],
],
)
This also works with dictionary inputs:
model = keras.Model(inputs={"x1": x1, "x2": x2}, outputs=outputs)
export_archive = ExportArchive()
export_archive.track(model)
export_archive.add_endpoint(
name="serve",
fn=model.call,
input_signature=[
{
"x1": tf.TensorSpec(shape=(None, 3), dtype=tf.float32),
"x2": tf.TensorSpec(shape=(None, 4), dtype=tf.float32),
},
],
)
Adding an endpoint that is a tf.function
:
@tf.function()
def serving_fn(x):
return model(x)
# The function must be traced, i.e. it must be called at least once.
serving_fn(tf.random.normal(shape=(2, 3)))
export_archive = ExportArchive()
export_archive.track(model)
export_archive.add_endpoint(name="serve", fn=serving_fn)
add_variable_collection
add_variable_collection(
name, variables
)
Register a set of variables to be retrieved after reloading.
Arguments | |
---|---|
name
|
The string name for the collection. |
variables
|
A tuple/list/set of tf.Variable instances.
|
Example:
export_archive = ExportArchive()
export_archive.track(model)
# Register an endpoint
export_archive.add_endpoint(
name="serve",
fn=model.call,
input_signature=[tf.TensorSpec(shape=(None, 3), dtype=tf.float32)],
)
# Save a variable collection
export_archive.add_variable_collection(
name="optimizer_variables", variables=model.optimizer.variables)
export_archive.write_out("path/to/location")
# Reload the object
revived_object = tf.saved_model.load("path/to/location")
# Retrieve the variables
optimizer_variables = revived_object.optimizer_variables
track
track(
layer
)
Track the variables (and other resources) of a layer or model.
write_out
write_out(
filepath, options=None
)
Write the corresponding SavedModel to disk.
Arguments | |
---|---|
filepath
|
str or pathlib.Path object.
Path where to save the artifact.
|
options
|
tf.saved_model.SaveOptions object that specifies
SavedModel saving options.
|
Note on TF-Serving: all endpoints registered via add_endpoint()
are made visible for TF-Serving in the SavedModel artifact. In addition,
the first endpoint registered is made visible under the alias
"serving_default"
(unless an endpoint with the name
"serving_default"
was already registered manually),
since TF-Serving requires this endpoint to be set.