Thanks for tuning in to Google I/O. View all sessions on demandWatch on demand

Serialization and saving

Authors: Kathy Wu, Francois Chollet

View on TensorFlow.org Run in Google Colab View source on GitHub View on keras.io

Introduction

A Keras model consists of multiple components:

  • The architecture, or configuration, which specifies what layers the model contain, and how they're connected.
  • A set of weights values (the "state of the model").
  • An optimizer (defined by compiling the model).
  • A set of losses and metrics (defined by compiling the model or calling add_loss() or add_metric()).

The Keras API makes it possible to save all of these pieces to disk at once, or to only selectively save some of them:

  • Saving everything into a single archive in the TensorFlow SavedModel format (or in the older Keras H5 format). This is the standard practice.
  • Saving the architecture / configuration only, typically as a JSON file.
  • Saving the weights values only. This is generally used when training the model.

Let's take a look at each of these options. When would you use one or the other, and how do they work?

How to save and load a model

If you only have 10 seconds to read this guide, here's what you need to know.

Saving a Keras model:

model = ...  # Get model (Sequential, Functional Model, or Model subclass)
model.save('path/to/location')

Loading the model back:

from tensorflow import keras
model = keras.models.load_model('path/to/location')

Now, let's look at the details.

Setup

import numpy as np
import tensorflow as tf
from tensorflow import keras

2023-01-14 00:02:24.697027: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer.so.7'; dlerror: libnvinfer.so.7: cannot open shared object file: No such file or directory
2023-01-14 00:02:24.697128: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer_plugin.so.7'; dlerror: libnvinfer_plugin.so.7: cannot open shared object file: No such file or directory
2023-01-14 00:02:24.697137: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.

Whole-model saving & loading

You can save an entire model to a single artifact. It will include:

  • The model's architecture/config
  • The model's weight values (which were learned during training)
  • The model's compilation information (if compile() was called)
  • The optimizer and its state, if any (this enables you to restart training where you left)

APIs

There are two formats you can use to save an entire model to disk: the TensorFlow SavedModel format, and the older Keras H5 format. The recommended format is SavedModel. It is the default when you use model.save().

You can switch to the H5 format by:

  • Passing save_format='h5' to save().
  • Passing a filename that ends in .h5 or .keras to save().

SavedModel format

SavedModel is the more comprehensive save format that saves the model architecture, weights, and the traced Tensorflow subgraphs of the call functions. This enables Keras to restore both built-in layers as well as custom objects.

Example:

def get_model():
    # Create a simple model.
    inputs = keras.Input(shape=(32,))
    outputs = keras.layers.Dense(1)(inputs)
    model = keras.Model(inputs, outputs)
    model.compile(optimizer="adam", loss="mean_squared_error")
    return model


model = get_model()

# Train the model.
test_input = np.random.random((128, 32))
test_target = np.random.random((128, 1))
model.fit(test_input, test_target)

# Calling `save('my_model')` creates a SavedModel folder `my_model`.
model.save("my_model")

# It can be used to reconstruct the model identically.
reconstructed_model = keras.models.load_model("my_model")

# Let's check:
np.testing.assert_allclose(
    model.predict(test_input), reconstructed_model.predict(test_input)
)

# The reconstructed model is already compiled and has retained the optimizer
# state, so training can resume:
reconstructed_model.fit(test_input, test_target)

4/4 [==============================] - 1s 3ms/step - loss: 0.3906
INFO:tensorflow:Assets written to: my_model/assets
4/4 [==============================] - 0s 1ms/step
4/4 [==============================] - 0s 1ms/step
4/4 [==============================] - 0s 3ms/step - loss: 0.3383
<keras.callbacks.History at 0x7f97a0073be0>

What the SavedModel contains

Calling model.save('my_model') creates a folder named my_model, containing the following:

ls my_model

assets  fingerprint.pb  keras_metadata.pb  saved_model.pb  variables

The model architecture, and training configuration (including the optimizer, losses, and metrics) are stored in saved_model.pb. The weights are saved in the variables/ directory.

For detailed information on the SavedModel format, see the SavedModel guide (The SavedModel format on disk).

How SavedModel handles custom objects

When saving the model and its layers, the SavedModel format stores the class name, call function, losses, and weights (and the config, if implemented). The call function defines the computation graph of the model/layer.

In the absence of the model/layer config, the call function is used to create a model that exists like the original model which can be trained, evaluated, and used for inference.

Nevertheless, it is always a good practice to define the get_config and from_config methods when writing a custom model or layer class. This allows you to easily update the computation later if needed. See the section about Custom objects for more information.

Example:

class CustomModel(keras.Model):
    def __init__(self, hidden_units):
        super().__init__()
        self.hidden_units = hidden_units
        self.dense_layers = [keras.layers.Dense(u) for u in hidden_units]

    def call(self, inputs):
        x = inputs
        for layer in self.dense_layers:
            x = layer(x)
        return x

    def get_config(self):
        return {"hidden_units": self.hidden_units}

    @classmethod
    def from_config(cls, config):
        return cls(**config)


model = CustomModel([16, 16, 10])
# Build the model by calling it
input_arr = tf.random.uniform((1, 5))
outputs = model(input_arr)
model.save("my_model")

# Option 1: Load with the custom_object argument.
loaded_1 = keras.models.load_model(
    "my_model", custom_objects={"CustomModel": CustomModel}
)

# Option 2: Load without the CustomModel class.

# Delete the custom-defined model class to ensure that the loader does not have
# access to it.
del CustomModel

loaded_2 = keras.models.load_model("my_model")
np.testing.assert_allclose(loaded_1(input_arr), outputs)
np.testing.assert_allclose(loaded_2(input_arr), outputs)

print("Original model:", model)
print("Model Loaded with custom objects:", loaded_1)
print("Model loaded without the custom object class:", loaded_2)

INFO:tensorflow:Assets written to: my_model/assets
WARNING:tensorflow:No training configuration found in save file, so the model was *not* compiled. Compile it manually.
WARNING:tensorflow:No training configuration found in save file, so the model was *not* compiled. Compile it manually.
Original model: <__main__.CustomModel object at 0x7f97b098a3a0>
Model Loaded with custom objects: <__main__.CustomModel object at 0x7f98ac87b3d0>
Model loaded without the custom object class: <keras.saving.legacy.saved_model.load.CustomModel object at 0x7f98ac8302e0>

The first loaded model is loaded using the config and CustomModel class. The second model is loaded by dynamically creating the model class that acts like the original model.

Configuring the SavedModel

New in TensorFlow 2.4 The argument save_traces has been added to model.save, which allows you to toggle SavedModel function tracing. Functions are saved to allow the Keras to re-load custom objects without the original class definitions, so when save_traces=False, all custom objects must have defined get_config/from_config methods. When loading, the custom objects must be passed to the custom_objects argument. save_traces=False reduces the disk space used by the SavedModel and saving time.

Keras H5 format

Keras also supports saving a single HDF5 file containing the model's architecture, weights values, and compile() information. It is a light-weight alternative to SavedModel.

Example:

model = get_model()

# Train the model.
test_input = np.random.random((128, 32))
test_target = np.random.random((128, 1))
model.fit(test_input, test_target)

# Calling `save('my_model.h5')` creates a h5 file `my_model.h5`.
model.save("my_h5_model.h5")

# It can be used to reconstruct the model identically.
reconstructed_model = keras.models.load_model("my_h5_model.h5")

# Let's check:
np.testing.assert_allclose(
    model.predict(test_input), reconstructed_model.predict(test_input)
)

# The reconstructed model is already compiled and has retained the optimizer
# state, so training can resume:
reconstructed_model.fit(test_input, test_target)

4/4 [==============================] - 0s 3ms/step - loss: 1.4363
4/4 [==============================] - 0s 2ms/step
4/4 [==============================] - 0s 1ms/step
4/4 [==============================] - 0s 2ms/step - loss: 1.2949
<keras.callbacks.History at 0x7f98ac611a30>

Format Limitations

Keras SavedModel format limitations:

The tracing done by SavedModel to produce the graphs of the layer call functions allows SavedModel be more portable than H5, but it comes with drawbacks.

  • Can be slower and bulkier than H5.
  • Cannot serialize the ops generated from the mask argument (i.e. if a layer is called with layer(..., mask=mask_value), the mask argument is not saved to SavedModel).
  • Does not save the overridden train_step() in subclassed models.

Custom objects that use masks or have a custom training loop can still be saved and loaded from SavedModel, except they must override get_config()/from_config(), and the classes must be passed to the custom_objects argument when loading.

H5 limitations:

  • External losses & metrics added via model.add_loss() & model.add_metric() are not saved (unlike SavedModel). If you have such losses & metrics on your model and you want to resume training, you need to add these losses back yourself after loading the model. Note that this does not apply to losses/metrics created inside layers via self.add_loss() & self.add_metric(). As long as the layer gets loaded, these losses & metrics are kept, since they are part of the call method of the layer.
  • The computation graph of custom objects such as custom layers is not included in the saved file. At loading time, Keras will need access to the Python classes/functions of these objects in order to reconstruct the model. See Custom objects.
  • Does not support preprocessing layers.

Saving the architecture

The model's configuration (or architecture) specifies what layers the model contains, and how these layers are connected*. If you have the configuration of a model, then the model can be created with a freshly initialized state for the weights and no compilation information.

*Note this only applies to models defined using the functional or Sequential apis not subclassed models.

Configuration of a Sequential model or Functional API model

These types of models are explicit graphs of layers: their configuration is always available in a structured form.

APIs

get_config() and from_config()

Calling config = model.get_config() will return a Python dict containing the configuration of the model. The same model can then be reconstructed via Sequential.from_config(config) (for a Sequential model) or Model.from_config(config) (for a Functional API model).

The same workflow also works for any serializable layer.

Layer example:

layer = keras.layers.Dense(3, activation="relu")
layer_config = layer.get_config()
new_layer = keras.layers.Dense.from_config(layer_config)

Sequential model example:

model = keras.Sequential([keras.Input((32,)), keras.layers.Dense(1)])
config = model.get_config()
new_model = keras.Sequential.from_config(config)

Functional model example:

inputs = keras.Input((32,))
outputs = keras.layers.Dense(1)(inputs)
model = keras.Model(inputs, outputs)
config = model.get_config()
new_model = keras.Model.from_config(config)

to_json() and tf.keras.models.model_from_json()

This is similar to get_config / from_config, except it turns the model into a JSON string, which can then be loaded without the original model class. It is also specific to models, it isn't meant for layers.

Example:

model = keras.Sequential([keras.Input((32,)), keras.layers.Dense(1)])
json_config = model.to_json()
new_model = keras.models.model_from_json(json_config)

Custom objects

Models and layers

The architecture of subclassed models and layers are defined in the methods __init__ and call. They are considered Python bytecode, which cannot be serialized into a JSON-compatible config -- you could try serializing the bytecode (e.g. via pickle), but it's completely unsafe and means your model cannot be loaded on a different system.

In order to save/load a model with custom-defined layers, or a subclassed model, you should overwrite the get_config and optionally from_config methods. Additionally, you should register the custom object so that Keras is aware of it.

Custom functions

Custom-defined functions (e.g. activation loss or initialization) do not need a get_config method. The function name is sufficient for loading as long as it is registered as a custom object.

Loading the TensorFlow graph only

It's possible to load the TensorFlow graph generated by the Keras. If you do so, you won't need to provide any custom_objects. You can do so like this:

model.save("my_model")
tensorflow_graph = tf.saved_model.load("my_model")
x = np.random.uniform(size=(4, 32)).astype(np.float32)
predicted = tensorflow_graph(x).numpy()

WARNING:tensorflow:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
INFO:tensorflow:Assets written to: my_model/assets

Note that this method has several drawbacks:

  • For traceability reasons, you should always have access to the custom objects that were used. You wouldn't want to put in production a model that you cannot re-create.
  • The object returned by tf.saved_model.load isn't a Keras model. So it's not as easy to use. For example, you won't have access to .predict() or .fit()

Even if its use is discouraged, it can help you if you're in a tight spot, for example, if you lost the code of your custom objects or have issues loading the model with tf.keras.models.load_model().

You can find out more in the page about tf.saved_model.load

Defining the config methods

Specifications:

  • get_config should return a JSON-serializable dictionary in order to be compatible with the Keras architecture- and model-saving APIs.
  • from_config(config) (classmethod) should return a new layer or model object that is created from the config. The default implementation returns cls(**config).

Example:

class CustomLayer(keras.layers.Layer):
    def __init__(self, a):
        self.var = tf.Variable(a, name="var_a")

    def call(self, inputs, training=False):
        if training:
            return inputs * self.var
        else:
            return inputs

    def get_config(self):
        return {"a": self.var.numpy()}

    # There's actually no need to define `from_config` here, since returning
    # `cls(**config)` is the default behavior.
    @classmethod
    def from_config(cls, config):
        return cls(**config)


layer = CustomLayer(5)
layer.var.assign(2)

serialized_layer = keras.layers.serialize(layer)
new_layer = keras.layers.deserialize(
    serialized_layer, custom_objects={"CustomLayer": CustomLayer}
)

Registering the custom object

Keras keeps a note of which class generated the config. From the example above, tf.keras.layers.serialize generates a serialized form of the custom layer:

{'class_name': 'CustomLayer', 'config': {'a': 2} }

Keras keeps a master list of all built-in layer, model, optimizer, and metric classes, which is used to find the correct class to call from_config. If the class can't be found, then an error is raised (Value Error: Unknown layer). There are a few ways to register custom classes to this list:

  1. Setting custom_objects argument in the loading function. (see the example in section above "Defining the config methods")
  2. tf.keras.utils.custom_object_scope or tf.keras.utils.CustomObjectScope
  3. tf.keras.utils.register_keras_serializable

Custom layer and function example

class CustomLayer(keras.layers.Layer):
    def __init__(self, units=32, **kwargs):
        super().__init__(**kwargs)
        self.units = units

    def build(self, input_shape):
        self.w = self.add_weight(
            shape=(input_shape[-1], self.units),
            initializer="random_normal",
            trainable=True,
        )
        self.b = self.add_weight(
            shape=(self.units,), initializer="random_normal", trainable=True
        )

    def call(self, inputs):
        return tf.matmul(inputs, self.w) + self.b

    def get_config(self):
        config = super().get_config()
        config.update({"units": self.units})
        return config


def custom_activation(x):
    return tf.nn.tanh(x) ** 2


# Make a model with the CustomLayer and custom_activation
inputs = keras.Input((32,))
x = CustomLayer(32)(inputs)
outputs = keras.layers.Activation(custom_activation)(x)
model = keras.Model(inputs, outputs)

# Retrieve the config
config = model.get_config()

# At loading time, register the custom objects with a `custom_object_scope`:
custom_objects = {"CustomLayer": CustomLayer, "custom_activation": custom_activation}
with keras.utils.custom_object_scope(custom_objects):
    new_model = keras.Model.from_config(config)

In-memory model cloning

You can also do in-memory cloning of a model via tf.keras.models.clone_model(). This is equivalent to getting the config then recreating the model from its config (so it does not preserve compilation information or layer weights values).

Example:

with keras.utils.custom_object_scope(custom_objects):
    new_model = keras.models.clone_model(model)

Saving & loading only the model's weights values

You can choose to only save & load a model's weights. This can be useful if:

  • You only need the model for inference: in this case you won't need to restart training, so you don't need the compilation information or optimizer state.
  • You are doing transfer learning: in this case you will be training a new model reusing the state of a prior model, so you don't need the compilation information of the prior model.

APIs for in-memory weight transfer

Weights can be copied between different objects by using get_weights and set_weights:

Examples below.

Transfering weights from one layer to another, in memory

def create_layer():
    layer = keras.layers.Dense(64, activation="relu", name="dense_2")
    layer.build((None, 784))
    return layer


layer_1 = create_layer()
layer_2 = create_layer()

# Copy weights from layer 1 to layer 2
layer_2.set_weights(layer_1.get_weights())

Transfering weights from one model to another model with a compatible architecture, in memory

# Create a simple functional model
inputs = keras.Input(shape=(784,), name="digits")
x = keras.layers.Dense(64, activation="relu", name="dense_1")(inputs)
x = keras.layers.Dense(64, activation="relu", name="dense_2")(x)
outputs = keras.layers.Dense(10, name="predictions")(x)
functional_model = keras.Model(inputs=inputs, outputs=outputs, name="3_layer_mlp")

# Define a subclassed model with the same architecture
class SubclassedModel(keras.Model):
    def __init__(self, output_dim, name=None):
        super().__init__(name=name)
        self.output_dim = output_dim
        self.dense_1 = keras.layers.Dense(64, activation="relu", name="dense_1")
        self.dense_2 = keras.layers.Dense(64, activation="relu", name="dense_2")
        self.dense_3 = keras.layers.Dense(output_dim, name="predictions")

    def call(self, inputs):
        x = self.dense_1(inputs)
        x = self.dense_2(x)
        x = self.dense_3(x)
        return x

    def get_config(self):
        return {"output_dim": self.output_dim, "name": self.name}


subclassed_model = SubclassedModel(10)
# Call the subclassed model once to create the weights.
subclassed_model(tf.ones((1, 784)))

# Copy weights from functional_model to subclassed_model.
subclassed_model.set_weights(functional_model.get_weights())

assert len(functional_model.weights) == len(subclassed_model.weights)
for a, b in zip(functional_model.weights, subclassed_model.weights):
    np.testing.assert_allclose(a.numpy(), b.numpy())

The case of stateless layers

Because stateless layers do not change the order or number of weights, models can have compatible architectures even if there are extra/missing stateless layers.

inputs = keras.Input(shape=(784,), name="digits")
x = keras.layers.Dense(64, activation="relu", name="dense_1")(inputs)
x = keras.layers.Dense(64, activation="relu", name="dense_2")(x)
outputs = keras.layers.Dense(10, name="predictions")(x)
functional_model = keras.Model(inputs=inputs, outputs=outputs, name="3_layer_mlp")

inputs = keras.Input(shape=(784,), name="digits")
x = keras.layers.Dense(64, activation="relu", name="dense_1")(inputs)
x = keras.layers.Dense(64, activation="relu", name="dense_2")(x)

# Add a dropout layer, which does not contain any weights.
x = keras.layers.Dropout(0.5)(x)
outputs = keras.layers.Dense(10, name="predictions")(x)
functional_model_with_dropout = keras.Model(
    inputs=inputs, outputs=outputs, name="3_layer_mlp"
)

functional_model_with_dropout.set_weights(functional_model.get_weights())

APIs for saving weights to disk & loading them back

Weights can be saved to disk by calling model.save_weights in the following formats:

  • TensorFlow Checkpoint
  • HDF5

The default format for model.save_weights is TensorFlow checkpoint. There are two ways to specify the save format:

  1. save_format argument: Set the value to save_format="tf" or save_format="h5".
  2. path argument: If the path ends with .h5 or .hdf5, then the HDF5 format is used. Other suffixes will result in a TensorFlow checkpoint unless save_format is set.

There is also an option of retrieving weights as in-memory numpy arrays. Each API has its pros and cons which are detailed below.

TF Checkpoint format

Example:

# Runnable example
sequential_model = keras.Sequential(
    [
        keras.Input(shape=(784,), name="digits"),
        keras.layers.Dense(64, activation="relu", name="dense_1"),
        keras.layers.Dense(64, activation="relu", name="dense_2"),
        keras.layers.Dense(10, name="predictions"),
    ]
)
sequential_model.save_weights("ckpt")
load_status = sequential_model.load_weights("ckpt")

# `assert_consumed` can be used as validation that all variable values have been
# restored from the checkpoint. See `tf.train.Checkpoint.restore` for other
# methods in the Status object.
load_status.assert_consumed()

<tensorflow.python.checkpoint.checkpoint.CheckpointLoadStatus at 0x7f98ac460d00>

Format details

The TensorFlow Checkpoint format saves and restores the weights using object attribute names. For instance, consider the tf.keras.layers.Dense layer. The layer contains two weights: dense.kernel and dense.bias. When the layer is saved to the tf format, the resulting checkpoint contains the keys "kernel" and "bias" and their corresponding weight values. For more information see "Loading mechanics" in the TF Checkpoint guide.

Note that attribute/graph edge is named after the name used in parent object, not the name of the variable. Consider the CustomLayer in the example below. The variable CustomLayer.var is saved with "var" as part of key, not "var_a".

class CustomLayer(keras.layers.Layer):
    def __init__(self, a):
        self.var = tf.Variable(a, name="var_a")


layer = CustomLayer(5)
layer_ckpt = tf.train.Checkpoint(layer=layer).save("custom_layer")

ckpt_reader = tf.train.load_checkpoint(layer_ckpt)

ckpt_reader.get_variable_to_dtype_map()

{'save_counter/.ATTRIBUTES/VARIABLE_VALUE': tf.int64,
 'layer/var/.ATTRIBUTES/VARIABLE_VALUE': tf.int32,
 '_CHECKPOINTABLE_OBJECT_GRAPH': tf.string}

Transfer learning example

Essentially, as long as two models have the same architecture, they are able to share the same checkpoint.

Example:

inputs = keras.Input(shape=(784,), name="digits")
x = keras.layers.Dense(64, activation="relu", name="dense_1")(inputs)
x = keras.layers.Dense(64, activation="relu", name="dense_2")(x)
outputs = keras.layers.Dense(10, name="predictions")(x)
functional_model = keras.Model(inputs=inputs, outputs=outputs, name="3_layer_mlp")

# Extract a portion of the functional model defined in the Setup section.
# The following lines produce a new model that excludes the final output
# layer of the functional model.
pretrained = keras.Model(
    functional_model.inputs, functional_model.layers[-1].input, name="pretrained_model"
)
# Randomly assign "trained" weights.
for w in pretrained.weights:
    w.assign(tf.random.normal(w.shape))
pretrained.save_weights("pretrained_ckpt")
pretrained.summary()

# Assume this is a separate program where only 'pretrained_ckpt' exists.
# Create a new functional model with a different output dimension.
inputs = keras.Input(shape=(784,), name="digits")
x = keras.layers.Dense(64, activation="relu", name="dense_1")(inputs)
x = keras.layers.Dense(64, activation="relu", name="dense_2")(x)
outputs = keras.layers.Dense(5, name="predictions")(x)
model = keras.Model(inputs=inputs, outputs=outputs, name="new_model")

# Load the weights from pretrained_ckpt into model.
model.load_weights("pretrained_ckpt")

# Check that all of the pretrained weights have been loaded.
for a, b in zip(pretrained.weights, model.weights):
    np.testing.assert_allclose(a.numpy(), b.numpy())

print("\n", "-" * 50)
model.summary()

# Example 2: Sequential model
# Recreate the pretrained model, and load the saved weights.
inputs = keras.Input(shape=(784,), name="digits")
x = keras.layers.Dense(64, activation="relu", name="dense_1")(inputs)
x = keras.layers.Dense(64, activation="relu", name="dense_2")(x)
pretrained_model = keras.Model(inputs=inputs, outputs=x, name="pretrained")

# Sequential example:
model = keras.Sequential([pretrained_model, keras.layers.Dense(5, name="predictions")])
model.summary()

pretrained_model.load_weights("pretrained_ckpt")

# Warning! Calling `model.load_weights('pretrained_ckpt')` won't throw an error,
# but will *not* work as expected. If you inspect the weights, you'll see that
# none of the weights will have loaded. `pretrained_model.load_weights()` is the
# correct method to call.

Model: "pretrained_model"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 digits (InputLayer)         [(None, 784)]             0         
                                                                 
 dense_1 (Dense)             (None, 64)                50240     
                                                                 
 dense_2 (Dense)             (None, 64)                4160      
                                                                 
=================================================================
Total params: 54,400
Trainable params: 54,400
Non-trainable params: 0
_________________________________________________________________

 --------------------------------------------------
Model: "new_model"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 digits (InputLayer)         [(None, 784)]             0         
                                                                 
 dense_1 (Dense)             (None, 64)                50240     
                                                                 
 dense_2 (Dense)             (None, 64)                4160      
                                                                 
 predictions (Dense)         (None, 5)                 325       
                                                                 
=================================================================
Total params: 54,725
Trainable params: 54,725
Non-trainable params: 0
_________________________________________________________________
Model: "sequential_3"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 pretrained (Functional)     (None, 64)                54400     
                                                                 
 predictions (Dense)         (None, 5)                 325       
                                                                 
=================================================================
Total params: 54,725
Trainable params: 54,725
Non-trainable params: 0
_________________________________________________________________
<tensorflow.python.checkpoint.checkpoint.CheckpointLoadStatus at 0x7f98ac671970>

It is generally recommended to stick to the same API for building models. If you switch between Sequential and Functional, or Functional and subclassed, etc., then always rebuild the pre-trained model and load the pre-trained weights to that model.

The next question is, how can weights be saved and loaded to different models if the model architectures are quite different? The solution is to use tf.train.Checkpoint to save and restore the exact layers/variables.

Example:

# Create a subclassed model that essentially uses functional_model's first
# and last layers.
# First, save the weights of functional_model's first and last dense layers.
first_dense = functional_model.layers[1]
last_dense = functional_model.layers[-1]
ckpt_path = tf.train.Checkpoint(
    dense=first_dense, kernel=last_dense.kernel, bias=last_dense.bias
).save("ckpt")

# Define the subclassed model.
class ContrivedModel(keras.Model):
    def __init__(self):
        super().__init__()
        self.first_dense = keras.layers.Dense(64)
        self.kernel = self.add_variable("kernel", shape=(64, 10))
        self.bias = self.add_variable("bias", shape=(10,))

    def call(self, inputs):
        x = self.first_dense(inputs)
        return tf.matmul(x, self.kernel) + self.bias


model = ContrivedModel()
# Call model on inputs to create the variables of the dense layer.
_ = model(tf.ones((1, 784)))

# Create a Checkpoint with the same structure as before, and load the weights.
tf.train.Checkpoint(
    dense=model.first_dense, kernel=model.kernel, bias=model.bias
).restore(ckpt_path).assert_consumed()

/tmpfs/tmp/ipykernel_16571/1419082508.py:15: UserWarning: `layer.add_variable` is deprecated and will be removed in a future version. Please use the `layer.add_weight()` method instead.
  self.kernel = self.add_variable("kernel", shape=(64, 10))
/tmpfs/tmp/ipykernel_16571/1419082508.py:16: UserWarning: `layer.add_variable` is deprecated and will be removed in a future version. Please use the `layer.add_weight()` method instead.
  self.bias = self.add_variable("bias", shape=(10,))
<tensorflow.python.checkpoint.checkpoint.CheckpointLoadStatus at 0x7f98ac4425e0>

HDF5 format

The HDF5 format contains weights grouped by layer names. The weights are lists ordered by concatenating the list of trainable weights to the list of non-trainable weights (same as layer.weights). Thus, a model can use a hdf5 checkpoint if it has the same layers and trainable statuses as saved in the checkpoint.

Example:

# Runnable example
sequential_model = keras.Sequential(
    [
        keras.Input(shape=(784,), name="digits"),
        keras.layers.Dense(64, activation="relu", name="dense_1"),
        keras.layers.Dense(64, activation="relu", name="dense_2"),
        keras.layers.Dense(10, name="predictions"),
    ]
)
sequential_model.save_weights("weights.h5")
sequential_model.load_weights("weights.h5")

Note that changing layer.trainable may result in a different layer.weights ordering when the model contains nested layers.

class NestedDenseLayer(keras.layers.Layer):
    def __init__(self, units, name=None):
        super().__init__(name=name)
        self.dense_1 = keras.layers.Dense(units, name="dense_1")
        self.dense_2 = keras.layers.Dense(units, name="dense_2")

    def call(self, inputs):
        return self.dense_2(self.dense_1(inputs))


nested_model = keras.Sequential([keras.Input((784,)), NestedDenseLayer(10, "nested")])
variable_names = [v.name for v in nested_model.weights]
print("variables: {}".format(variable_names))

print("\nChanging trainable status of one of the nested layers...")
nested_model.get_layer("nested").dense_1.trainable = False

variable_names_2 = [v.name for v in nested_model.weights]
print("\nvariables: {}".format(variable_names_2))
print("variable ordering changed:", variable_names != variable_names_2)

variables: ['nested/dense_1/kernel:0', 'nested/dense_1/bias:0', 'nested/dense_2/kernel:0', 'nested/dense_2/bias:0']

Changing trainable status of one of the nested layers...

variables: ['nested/dense_2/kernel:0', 'nested/dense_2/bias:0', 'nested/dense_1/kernel:0', 'nested/dense_1/bias:0']
variable ordering changed: True

Transfer learning example

When loading pretrained weights from HDF5, it is recommended to load the weights into the original checkpointed model, and then extract the desired weights/layers into a new model.

Example:

def create_functional_model():
    inputs = keras.Input(shape=(784,), name="digits")
    x = keras.layers.Dense(64, activation="relu", name="dense_1")(inputs)
    x = keras.layers.Dense(64, activation="relu", name="dense_2")(x)
    outputs = keras.layers.Dense(10, name="predictions")(x)
    return keras.Model(inputs=inputs, outputs=outputs, name="3_layer_mlp")


functional_model = create_functional_model()
functional_model.save_weights("pretrained_weights.h5")

# In a separate program:
pretrained_model = create_functional_model()
pretrained_model.load_weights("pretrained_weights.h5")

# Create a new model by extracting layers from the original model:
extracted_layers = pretrained_model.layers[:-1]
extracted_layers.append(keras.layers.Dense(5, name="dense_3"))
model = keras.Sequential(extracted_layers)
model.summary()

Model: "sequential_6"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense_1 (Dense)             (None, 64)                50240     
                                                                 
 dense_2 (Dense)             (None, 64)                4160      
                                                                 
 dense_3 (Dense)             (None, 5)                 325       
                                                                 
=================================================================
Total params: 54,725
Trainable params: 54,725
Non-trainable params: 0
_________________________________________________________________