tf.keras.Model

View source on GitHub

Model groups layers into an object with training and inference features.

Inherits From: Layer, Module

tf.keras.Model(
    *args, **kwargs
)

There are two ways to instantiate a Model:

1 - With the "Functional API", where you start from Input, you chain layer calls to specify the model's forward pass, and finally you create your model from inputs and outputs:

import tensorflow as tf

inputs = tf.keras.Input(shape=(3,))
x = tf.keras.layers.Dense(4, activation=tf.nn.relu)(inputs)
outputs = tf.keras.layers.Dense(5, activation=tf.nn.softmax)(x)
model = tf.keras.Model(inputs=inputs, outputs=outputs)

A new Functional API model can also be created by using the intermediate tensors. This enables you to quickly extract sub-components of the model.

Example:

inputs = keras.Input(shape=(None, None, 3))
processed = keras.layers.RandomCrop(width=32, height=32)(inputs)
conv = keras.layers.Conv2D(filters=2, kernel_size=3)(processed)
pooling = keras.layers.GlobalAveragePooling2D()(conv)
feature = keras.layers.Dense(10)(pooling)

full_model = keras.Model(inputs, feature)
backbone = keras.Model(processed, conv)
activations = keras.Model(conv, feature)

Note that the backbone and activations models are not created with keras.Input objects, but with the tensors that are originated from keras.Inputs objects. Under the hood, the layers and weights will be shared across these models, so that user can train the full_model, and use backbone or activations to do feature extraction. The inputs and outputs of the model can be nested structures of tensors as well, and the created models are standard Functional API models that support all the existing APIs.

2 - By subclassing the Model class: in that case, you should define your layers in __init__() and you should implement the model's forward pass in call().

import tensorflow as tf

class MyModel(tf.keras.Model):

  def __init__(self):
    super().__init__()
    self.dense1 = tf.keras.layers.Dense(4, activation=tf.nn.relu)
    self.dense2 = tf.keras.layers.Dense(5, activation=tf.nn.softmax)

  def call(self, inputs):
    x = self.dense1(inputs)
    return self.dense2(x)

model = MyModel()

If you subclass Model, you can optionally have a training argument (boolean) in call(), which you can use to specify a different behavior in training and inference:

import tensorflow as tf

class MyModel(tf.keras.Model):

  def __init__(self):
    super().__init__()
    self.dense1 = tf.keras.layers.Dense(4, activation=tf.nn.relu)
    self.dense2 = tf.keras.layers.Dense(5, activation=tf.nn.softmax)
    self.dropout = tf.keras.layers.Dropout(0.5)

  def call(self, inputs, training=False):
    x = self.dense1(inputs)
    if training:
      x = self.dropout(x, training=training)
    return self.dense2(x)

model = MyModel()

Once the model is created, you can config the model with losses and metrics with model.compile(), train the model with model.fit(), or use the model to do prediction with model.predict().

inputs = tf.keras.layers.Input(shape=(3,))
outputs = tf.keras.layers.Dense(2)(inputs)
model = tf.keras.models.Model(inputs=inputs, outputs=outputs)
model.compile(optimizer="Adam", loss="mse", metrics=["mae"])
model.metrics_names
[]

x = np.random.random((2, 3))
y = np.random.randint(0, 2, (2, 2))
model.fit(x, y)
model.metrics_names
['loss', 'mae']

inputs = tf.keras.layers.Input(shape=(3,))
d = tf.keras.layers.Dense(2, name='out')
output_1 = d(inputs)
output_2 = d(inputs)
model = tf.keras.models.Model(
   inputs=inputs, outputs=[output_1, output_2])
model.compile(optimizer="Adam", loss="mse", metrics=["mae", "acc"])
model.fit(x, (y, y))
model.metrics_names
['loss', 'out_loss', 'out_1_loss', 'out_mae', 'out_acc', 'out_1_mae',
'out_1_acc']

Methods

call

View source

call(
    inputs, training=None, mask=None
)

Calls the model on new inputs and returns the outputs as tensors.

In this case call() just reapplies all ops in the graph to the new inputs (e.g. build a new computational graph from the provided inputs).

compile

View source

compile(
    optimizer='rmsprop',
    loss=None,
    metrics=None,
    loss_weights=None,
    weighted_metrics=None,
    run_eagerly=None,
    steps_per_execution=None,
    jit_compile=None,
    **kwargs
)

Configures the model for training.

Example:

model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=1e-3),
              loss=tf.keras.losses.BinaryCrossentropy(),
              metrics=[tf.keras.metrics.BinaryAccuracy(),
                       tf.keras.metrics.FalseNegatives()])

compute_loss

View source

compute_loss(
    x=None, y=None, y_pred=None, sample_weight=None
)

Compute the total loss, validate it, and return it.

Subclasses can optionally override this method to provide custom loss computation logic.

Example:

class MyModel(tf.keras.Model):

  def __init__(self, *args, **kwargs):
    super(MyModel, self).__init__(*args, **kwargs)
    self.loss_tracker = tf.keras.metrics.Mean(name='loss')

  def compute_loss(self, x, y, y_pred, sample_weight):
    loss = tf.reduce_mean(tf.math.squared_difference(y_pred, y))
    loss += tf.add_n(self.losses)
    self.loss_tracker.update_state(loss)
    return loss

  def reset_metrics(self):
    self.loss_tracker.reset_states()

  @property
  def metrics(self):
    return [self.loss_tracker]

tensors = tf.random.uniform((10, 10)), tf.random.uniform((10,))
dataset = tf.data.Dataset.from_tensor_slices(tensors).repeat().batch(1)

inputs = tf.keras.layers.Input(shape=(10,), name='my_input')
outputs = tf.keras.layers.Dense(10)(inputs)
model = MyModel(inputs, outputs)
model.add_loss(tf.reduce_sum(outputs))

optimizer = tf.keras.optimizers.SGD()
model.compile(optimizer, loss='mse', steps_per_execution=10)
model.fit(dataset, epochs=2, steps_per_epoch=10)
print('My custom loss: ', model.loss_tracker.result().numpy())

compute_metrics

View source

compute_metrics(
    x, y, y_pred, sample_weight
)

Update metric states and collect all metrics to be returned.

Subclasses can optionally override this method to provide custom metric updating and collection logic.

Example:

class MyModel(tf.keras.Sequential):

  def compute_metrics(self, x, y, y_pred, sample_weight):

    # This super call updates `self.compiled_metrics` and returns results
    # for all metrics listed in `self.metrics`.
    metric_results = super(MyModel, self).compute_metrics(
        x, y, y_pred, sample_weight)

    # Note that `self.custom_metric` is not listed in `self.metrics`.
    self.custom_metric.update_state(x, y, y_pred, sample_weight)
    metric_results['custom_metric_name'] = self.custom_metric.result()
    return metric_results

evaluate

View source

evaluate(
    x=None,
    y=None,
    batch_size=None,
    verbose=1,
    sample_weight=None,
    steps=None,
    callbacks=None,
    max_queue_size=10,
    workers=1,
    use_multiprocessing=False,
    return_dict=False,
    **kwargs
)

Returns the loss value & metrics values for the model in test mode.

Computation is done in batches (see the batch_size arg.)

See the discussion of Unpacking behavior for iterator-like inputs for Model.fit.

fit

View source

fit(
    x=None,
    y=None,
    batch_size=None,
    epochs=1,
    verbose='auto',
    callbacks=None,
    validation_split=0.0,
    validation_data=None,
    shuffle=True,
    class_weight=None,
    sample_weight=None,
    initial_epoch=0,
    steps_per_epoch=None,
    validation_steps=None,
    validation_batch_size=None,
    validation_freq=1,
    max_queue_size=10,
    workers=1,
    use_multiprocessing=False
)

Trains the model for a fixed number of epochs (iterations on a dataset).

Unpacking behavior for iterator-like inputs: A common pattern is to pass a tf.data.Dataset, generator, or tf.keras.utils.Sequence to the x argument of fit, which will in fact yield not only features (x) but optionally targets (y) and sample weights. Keras requires that the output of such iterator-likes be unambiguous. The iterator should return a tuple of length 1, 2, or 3, where the optional second and third elements will be used for y and sample_weight respectively. Any other type provided will be wrapped in a length one tuple, effectively treating everything as 'x'. When yielding dicts, they should still adhere to the top-level tuple structure. e.g. ({"x0": x0, "x1": x1}, y). Keras will not attempt to separate features, targets, and weights from the keys of a single dict. A notable unsupported data type is the namedtuple. The reason is that it behaves like both an ordered datatype (tuple) and a mapping datatype (dict). So given a namedtuple of the form: namedtuple("example_tuple", ["y", "x"]) it is ambiguous whether to reverse the order of the elements when interpreting the value. Even worse is a tuple of the form: namedtuple("other_tuple", ["x", "y", "z"]) where it is unclear if the tuple was intended to be unpacked into x, y, and sample_weight or passed through as a single element to x. As a result the data processing code will simply raise a ValueError if it encounters a namedtuple. (Along with instructions to remedy the issue.)

get_layer

View source

get_layer(
    name=None, index=None
)

Retrieves a layer based on either its name (unique) or index.

If name and index are both provided, index will take precedence. Indices are based on order of horizontal graph traversal (bottom-up).

load_weights

View source

load_weights(
    filepath, by_name=False, skip_mismatch=False, options=None
)

Loads all layer weights, either from a TensorFlow or an HDF5 weight file.

If by_name is False weights are loaded based on the network's topology. This means the architecture should be the same as when the weights were saved. Note that layers that don't have weights are not taken into account in the topological ordering, so adding or removing layers is fine as long as they don't have weights.

If by_name is True, weights are loaded into layers only if they share the same name. This is useful for fine-tuning or transfer-learning models where some of the layers have changed.

Only topological loading (by_name=False) is supported when loading weights from the TensorFlow format. Note that topological loading differs slightly between TensorFlow and HDF5 formats for user-defined classes inheriting from tf.keras.Model: HDF5 loads based on a flattened list of weights, while the TensorFlow format loads based on the object-local names of attributes to which layers are assigned in the Model's constructor.

make_predict_function

View source

make_predict_function(
    force=False
)

Creates a function that executes one step of inference.

This method can be overridden to support custom inference logic. This method is called by Model.predict and Model.predict_on_batch.

Typically, this method directly controls tf.function and tf.distribute.Strategy settings, and delegates the actual evaluation logic to Model.predict_step.

This function is cached the first time Model.predict or Model.predict_on_batch is called. The cache is cleared whenever Model.compile is called. You can skip the cache and generate again the function with force=True.

make_test_function

View source

make_test_function(
    force=False
)

Creates a function that executes one step of evaluation.

This method can be overridden to support custom evaluation logic. This method is called by Model.evaluate and Model.test_on_batch.

Typically, this method directly controls tf.function and tf.distribute.Strategy settings, and delegates the actual evaluation logic to Model.test_step.

This function is cached the first time Model.evaluate or Model.test_on_batch is called. The cache is cleared whenever Model.compile is called. You can skip the cache and generate again the function with force=True.

make_train_function

View source

make_train_function(
    force=False
)

Creates a function that executes one step of training.

This method can be overridden to support custom training logic. This method is called by Model.fit and Model.train_on_batch.

Typically, this method directly controls tf.function and tf.distribute.Strategy settings, and delegates the actual training logic to Model.train_step.

This function is cached the first time Model.fit or Model.train_on_batch is called. The cache is cleared whenever Model.compile is called. You can skip the cache and generate again the function with force=True.

predict

View source

predict(
    x,
    batch_size=None,
    verbose=0,
    steps=None,
    callbacks=None,
    max_queue_size=10,
    workers=1,
    use_multiprocessing=False
)

Generates output predictions for the input samples.

Computation is done in batches. This method is designed for batch processing of large numbers of inputs. It is not intended for use inside of loops that iterate over your data and process small numbers of inputs at a time.

For small numbers of inputs that fit in one batch, directly use __call__() for faster execution, e.g., model(x), or model(x, training=False) if you have layers such as tf.keras.layers.BatchNormalization that behave differently during inference. You may pair the individual model call with a tf.function for additional performance inside your inner loop. If you need access to numpy array values instead of tensors after your model call, you can use tensor.numpy() to get the numpy array value of an eager tensor.

Also, note the fact that test loss is not affected by regularization layers like noise and dropout.

See the discussion of Unpacking behavior for iterator-like inputs for Model.fit. Note that Model.predict uses the same interpretation rules as Model.fit and Model.evaluate, so inputs must be unambiguous for all three methods.

predict_on_batch

View source

predict_on_batch(
    x
)

Returns predictions for a single batch of samples.

predict_step

View source

predict_step(
    data
)

The logic for one inference step.

This method can be overridden to support custom inference logic. This method is called by Model.make_predict_function.

This method should contain the mathematical logic for one step of inference. This typically includes the forward pass.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_predict_function, which can also be overridden.

reset_metrics

View source

reset_metrics()

Resets the state of all the metrics in the model.

Examples:

inputs = tf.keras.layers.Input(shape=(3,))
outputs = tf.keras.layers.Dense(2)(inputs)
model = tf.keras.models.Model(inputs=inputs, outputs=outputs)
model.compile(optimizer="Adam", loss="mse", metrics=["mae"])

x = np.random.random((2, 3))
y = np.random.randint(0, 2, (2, 2))
_ = model.fit(x, y, verbose=0)
assert all(float(m.result()) for m in model.metrics)

model.reset_metrics()
assert all(float(m.result()) == 0 for m in model.metrics)

reset_states

View source

reset_states()

save

View source

save(
    filepath,
    overwrite=True,
    include_optimizer=True,
    save_format=None,
    signatures=None,
    options=None,
    save_traces=True
)

Saves the model to Tensorflow SavedModel or a single HDF5 file.

Please see tf.keras.models.save_model or the Serialization and Saving guide for details.

Example:

from keras.models import load_model

model.save('my_model.h5')  # creates a HDF5 file 'my_model.h5'
del model  # deletes the existing model

# returns a compiled model
# identical to the previous one
model = load_model('my_model.h5')

save_spec

View source

save_spec(
    dynamic_batch=True
)

Returns the tf.TensorSpec of call inputs as a tuple (args, kwargs).

This value is automatically defined after calling the model for the first time. Afterwards, you can use it when exporting the model for serving:

model = tf.keras.Model(...)

@tf.function
def serve(*args, **kwargs):
  outputs = model(*args, **kwargs)
  # Apply postprocessing steps, or add additional outputs.
  ...
  return outputs

# arg_specs is `[tf.TensorSpec(...), ...]`. kwarg_specs, in this example, is
# an empty dict since functional models do not use keyword arguments.
arg_specs, kwarg_specs = model.save_spec()

model.save(path, signatures={
  'serving_default': serve.get_concrete_function(*arg_specs, **kwarg_specs)
})

save_weights

View source

save_weights(
    filepath, overwrite=True, save_format=None, options=None
)

Saves all layer weights.

Either saves in HDF5 or in TensorFlow format based on the save_format argument.

When saving in HDF5 format, the weight file has:

• layer_names (attribute), a list of strings (ordered names of model layers).
• For every layer, a group named layer.name
  • For every such layer group, a group attribute weight_names, a list of strings (ordered names of weights tensor of the layer).
  • For every weight in the layer, a dataset storing the weight value, named after the weight tensor.

When saving in TensorFlow format, all objects referenced by the network are saved in the same format as tf.train.Checkpoint, including any Layer instances or Optimizer instances assigned to object attributes. For networks constructed from inputs and outputs using tf.keras.Model(inputs, outputs), Layer instances used by the network are tracked/saved automatically. For user-defined classes which inherit from tf.keras.Model, Layer instances must be assigned to object attributes, typically in the constructor. See the documentation of tf.train.Checkpoint and tf.keras.Model for details.

While the formats are the same, do not mix save_weights and tf.train.Checkpoint. Checkpoints saved by Model.save_weights should be loaded using Model.load_weights. Checkpoints saved using tf.train.Checkpoint.save should be restored using the corresponding tf.train.Checkpoint.restore. Prefer tf.train.Checkpoint over save_weights for training checkpoints.

The TensorFlow format matches objects and variables by starting at a root object, self for save_weights, and greedily matching attribute names. For Model.save this is the Model, and for Checkpoint.save this is the Checkpoint even if the Checkpoint has a model attached. This means saving a tf.keras.Model using save_weights and loading into a tf.train.Checkpoint with a Model attached (or vice versa) will not match the Model's variables. See the guide to training checkpoints for details on the TensorFlow format.

summary

View source

summary(
    line_length=None,
    positions=None,
    print_fn=None,
    expand_nested=False,
    show_trainable=False
)

Prints a string summary of the network.

test_on_batch

View source

test_on_batch(
    x, y=None, sample_weight=None, reset_metrics=True, return_dict=False
)

Test the model on a single batch of samples.

test_step

View source

test_step(
    data
)

The logic for one evaluation step.

This method can be overridden to support custom evaluation logic. This method is called by Model.make_test_function.

This function should contain the mathematical logic for one step of evaluation. This typically includes the forward pass, loss calculation, and metrics updates.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_test_function, which can also be overridden.

to_json

View source

to_json(
    **kwargs
)

Returns a JSON string containing the network configuration.

To load a network from a JSON save file, use keras.models.model_from_json(json_string, custom_objects={}).

to_yaml

View source

to_yaml(
    **kwargs
)

Returns a yaml string containing the network configuration.

To load a network from a yaml save file, use keras.models.model_from_yaml(yaml_string, custom_objects={}).

custom_objects should be a dictionary mapping the names of custom losses / layers / etc to the corresponding functions / classes.

train_on_batch

View source

train_on_batch(
    x,
    y=None,
    sample_weight=None,
    class_weight=None,
    reset_metrics=True,
    return_dict=False
)

Runs a single gradient update on a single batch of data.

train_step

View source

train_step(
    data
)

The logic for one training step.

This method can be overridden to support custom training logic. For concrete examples of how to override this method see Customizing what happends in fit. This method is called by Model.make_train_function.

This method should contain the mathematical logic for one step of training. This typically includes the forward pass, loss calculation, backpropagation, and metric updates.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_train_function, which can also be overridden.