Making new layers and models via subclassing

Author: fchollet

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

Setup

import tensorflow as tf
from tensorflow import keras

The Layer class: the combination of state (weights) and some computation

One of the central abstractions in Keras is the Layer class. A layer encapsulates both a state (the layer's "weights") and a transformation from inputs to outputs (a "call", the layer's forward pass).

Here's a densely-connected layer. It has a state: the variables w and b.

class Linear(keras.layers.Layer):
    def __init__(self, units=32, input_dim=32):
        super().__init__()
        self.w = self.add_weight(
            shape=(input_dim, units), initializer="random_normal", trainable=True
        )
        self.b = self.add_weight(shape=(units,), initializer="zeros", trainable=True)

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

You would use a layer by calling it on some tensor input(s), much like a Python function.

x = tf.ones((2, 2))
linear_layer = Linear(4, 2)
y = linear_layer(x)
print(y)

tf.Tensor(
[[-0.02419483 -0.06813122  0.00395634 -0.03124779]
 [-0.02419483 -0.06813122  0.00395634 -0.03124779]], shape=(2, 4), dtype=float32)

Note that the weights w and b are automatically tracked by the layer upon being set as layer attributes:

assert linear_layer.weights == [linear_layer.w, linear_layer.b]

Layers can have non-trainable weights

Besides trainable weights, you can add non-trainable weights to a layer as well. Such weights are meant not to be taken into account during backpropagation, when you are training the layer.

Here's how to add and use a non-trainable weight:

class ComputeSum(keras.layers.Layer):
    def __init__(self, input_dim):
        super().__init__()
        self.total = self.add_weight(
            initializer="zeros", shape=(input_dim,), trainable=False
        )

    def call(self, inputs):
        self.total.assign_add(tf.reduce_sum(inputs, axis=0))
        return self.total


x = tf.ones((2, 2))
my_sum = ComputeSum(2)
y = my_sum(x)
print(y.numpy())
y = my_sum(x)
print(y.numpy())

[2. 2.]
[4. 4.]

It's part of layer.weights, but it gets categorized as a non-trainable weight:

print("weights:", len(my_sum.weights))
print("non-trainable weights:", len(my_sum.non_trainable_weights))

# It's not included in the trainable weights:
print("trainable_weights:", my_sum.trainable_weights)

weights: 1
non-trainable weights: 1
trainable_weights: []

Best practice: deferring weight creation until the shape of the inputs is known

Our Linear layer above took an input_dim argument that was used to compute the shape of the weights w and b in __init__():

class Linear(keras.layers.Layer):
    def __init__(self, units=32, input_dim=32):
        super().__init__()
        self.w = self.add_weight(
            shape=(input_dim, units), initializer="random_normal", trainable=True
        )
        self.b = self.add_weight(shape=(units,), initializer="zeros", trainable=True)

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

In many cases, you may not know in advance the size of your inputs, and you would like to lazily create weights when that value becomes known, some time after instantiating the layer.

In the Keras API, we recommend creating layer weights in the build(self, inputs_shape) method of your layer. Like this:

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

The __call__() method of your layer will automatically run build the first time it is called. You now have a layer that's lazy and thus easier to use:

# At instantiation, we don't know on what inputs this is going to get called
linear_layer = Linear(32)

# The layer's weights are created dynamically the first time the layer is called
y = linear_layer(x)

Implementing build() separately as shown above nicely separates creating weights only once from using weights in every call. However, for some advanced custom layers, it can become impractical to separate the state creation and computation. Layer implementers are allowed to defer weight creation to the first __call__(), but need to take care that later calls use the same weights. In addition, since __call__() is likely to be executed for the first time inside a tf.function, any variable creation that takes place in __call__() should be wrapped in a tf.init_scope.

Layers are recursively composable

If you assign a Layer instance as an attribute of another Layer, the outer layer will start tracking the weights created by the inner layer.

We recommend creating such sublayers in the __init__() method and leave it to the first __call__() to trigger building their weights.

class MLPBlock(keras.layers.Layer):
    def __init__(self):
        super().__init__()
        self.linear_1 = Linear(32)
        self.linear_2 = Linear(32)
        self.linear_3 = Linear(1)

    def call(self, inputs):
        x = self.linear_1(inputs)
        x = tf.nn.relu(x)
        x = self.linear_2(x)
        x = tf.nn.relu(x)
        return self.linear_3(x)


mlp = MLPBlock()
y = mlp(tf.ones(shape=(3, 64)))  # The first call to the `mlp` will create the weights
print("weights:", len(mlp.weights))
print("trainable weights:", len(mlp.trainable_weights))

weights: 6
trainable weights: 6

The add_loss() method

When writing the call() method of a layer, you can create loss tensors that you will want to use later, when writing your training loop. This is doable by calling self.add_loss(value):

# A layer that creates an activity regularization loss
class ActivityRegularizationLayer(keras.layers.Layer):
    def __init__(self, rate=1e-2):
        super().__init__()
        self.rate = rate

    def call(self, inputs):
        self.add_loss(self.rate * tf.reduce_mean(inputs))
        return inputs

Notice that add_loss() can take the result of plain TensorFlow operations. There is no need to call a Loss object here.

These losses (including those created by any inner layer) can be retrieved via layer.losses. This property is reset at the start of every __call__() to the top-level layer, so that layer.losses always contains the loss values created during the last forward pass.

class OuterLayer(keras.layers.Layer):
    def __init__(self):
        super().__init__()
        self.activity_reg = ActivityRegularizationLayer(1e-2)

    def call(self, inputs):
        return self.activity_reg(inputs)


layer = OuterLayer()
assert len(layer.losses) == 0  # No losses yet since the layer has never been called

_ = layer(tf.zeros(1, 1))
assert len(layer.losses) == 1  # We created one loss value

# `layer.losses` gets reset at the start of each __call__
_ = layer(tf.zeros(1, 1))
assert len(layer.losses) == 1  # This is the loss created during the call above

In addition, the loss property also contains regularization losses created for the weights of any inner layer:

class OuterLayerWithKernelRegularizer(keras.layers.Layer):
    def __init__(self):
        super().__init__()
        self.dense = keras.layers.Dense(
            32, kernel_regularizer=keras.regularizers.l2(1e-3)
        )

    def call(self, inputs):
        return self.dense(inputs)


layer = OuterLayerWithKernelRegularizer()
_ = layer(tf.zeros((1, 1)))

# This is `1e-3 * sum(layer.dense.kernel ** 2)`,
# created by the `kernel_regularizer` above.
print(layer.losses)

[<tf.Tensor: shape=(), dtype=float32, numpy=0.0017542194>]

These losses are meant to be taken into account when writing training loops, like this:

# Instantiate an optimizer.
optimizer = keras.optimizers.SGD(learning_rate=1e-3)
loss_fn = keras.losses.SparseCategoricalCrossentropy(from_logits=True)

# Iterate over the batches of a dataset.
for x_batch_train, y_batch_train in train_dataset:
    with tf.GradientTape() as tape:
        logits = layer(x_batch_train)  # Logits for this minibatch
        # Loss value for this minibatch
        loss_value = loss_fn(y_batch_train, logits)
        # Add extra losses created during this forward pass:
        loss_value += sum(model.losses)

    grads = tape.gradient(loss_value, model.trainable_weights)
    optimizer.apply_gradients(zip(grads, model.trainable_weights))

For a detailed guide about writing training loops, see the guide to writing a training loop from scratch.

These losses also work seamlessly with fit() (they get automatically summed and added to the main loss, if any):

import numpy as np

inputs = keras.Input(shape=(3,))
outputs = ActivityRegularizationLayer()(inputs)
model = keras.Model(inputs, outputs)

# If there is a loss passed in `compile`, the regularization
# losses get added to it
model.compile(optimizer="adam", loss="mse")
model.fit(np.random.random((2, 3)), np.random.random((2, 3)))

# It's also possible not to pass any loss in `compile`,
# since the model already has a loss to minimize, via the `add_loss`
# call during the forward pass!
model.compile(optimizer="adam")
model.fit(np.random.random((2, 3)), np.random.random((2, 3)))

1/1 [==============================] - 0s 75ms/step - loss: 0.1081
1/1 [==============================] - 0s 31ms/step - loss: 0.0044
<keras.src.callbacks.History at 0x7fb23c0e3f40>

You can optionally enable serialization on your layers

If you need your custom layers to be serializable as part of a Functional model, you can optionally implement a get_config() method:

class Linear(keras.layers.Layer):
    def __init__(self, units=32):
        super().__init__()
        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):
        return {"units": self.units}


# Now you can recreate the layer from its config:
layer = Linear(64)
config = layer.get_config()
print(config)
new_layer = Linear.from_config(config)

{'units': 64}

Note that the __init__() method of the base Layer class takes some keyword arguments, in particular a name and a dtype. It's good practice to pass these arguments to the parent class in __init__() and to include them in the layer config:

class Linear(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


layer = Linear(64)
config = layer.get_config()
print(config)
new_layer = Linear.from_config(config)

{'name': 'linear_7', 'trainable': True, 'dtype': 'float32', 'units': 64}

If you need more flexibility when deserializing the layer from its config, you can also override the from_config() class method. This is the base implementation of from_config():

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

To learn more about serialization and saving, see the complete guide to saving and serializing models.

Privileged training argument in the call() method

Some layers, in particular the BatchNormalization layer and the Dropout layer, have different behaviors during training and inference. For such layers, it is standard practice to expose a training (boolean) argument in the call() method.

By exposing this argument in call(), you enable the built-in training and evaluation loops (e.g. fit()) to correctly use the layer in training and inference.

class CustomDropout(keras.layers.Layer):
    def __init__(self, rate, **kwargs):
        super().__init__(**kwargs)
        self.rate = rate

    def call(self, inputs, training=False):
        if training:
            return tf.nn.dropout(inputs, rate=self.rate)
        return inputs

Privileged mask argument in the call() method

The other privileged argument supported by call() is the mask argument.

You will find it in all Keras RNN layers. A mask is a boolean tensor (one boolean value per timestep in the input) used to skip certain input timesteps when processing timeseries data.

Keras will automatically pass the correct mask argument to __call__() for layers that support it, when a mask is generated by a prior layer. Mask-generating layers are the Embedding layer configured with mask_zero=True, and the Masking layer.

To learn more about masking and how to write masking-enabled layers, please check out the guide "understanding padding and masking".

The Model class

In general, you will use the Layer class to define inner computation blocks, and will use the Model class to define the outer model -- the object you will train.

For instance, in a ResNet50 model, you would have several ResNet blocks subclassing Layer, and a single Model encompassing the entire ResNet50 network.

The Model class has the same API as Layer, with the following differences:

  • It exposes built-in training, evaluation, and prediction loops (model.fit(), model.evaluate(), model.predict()).
  • It exposes the list of its inner layers, via the model.layers property.
  • It exposes saving and serialization APIs (save(), save_weights()...)

Effectively, the Layer class corresponds to what we refer to in the literature as a "layer" (as in "convolution layer" or "recurrent layer") or as a "block" (as in "ResNet block" or "Inception block").

Meanwhile, the Model class corresponds to what is referred to in the literature as a "model" (as in "deep learning model") or as a "network" (as in "deep neural network").

So if you're wondering, "should I use the Layer class or the Model class?", ask yourself: will I need to call fit() on it? Will I need to call save() on it? If so, go with Model. If not (either because your class is just a block in a bigger system, or because you are writing training & saving code yourself), use Layer.

For instance, we could take our mini-resnet example above, and use it to build a Model that we could train with fit(), and that we could save with save_weights():

class ResNet(keras.Model):

    def __init__(self, num_classes=1000):
        super().__init__()
        self.block_1 = ResNetBlock()
        self.block_2 = ResNetBlock()
        self.global_pool = layers.GlobalAveragePooling2D()
        self.classifier = Dense(num_classes)

    def call(self, inputs):
        x = self.block_1(inputs)
        x = self.block_2(x)
        x = self.global_pool(x)
        return self.classifier(x)


resnet = ResNet()
dataset = ...
resnet.fit(dataset, epochs=10)
resnet.save(filepath.keras)

Putting it all together: an end-to-end example

Here's what you've learned so far:

  • A Layer encapsulate a state (created in __init__() or build()) and some computation (defined in call()).
  • Layers can be recursively nested to create new, bigger computation blocks.
  • Layers can create and track losses (typically regularization losses) via add_loss().
  • The outer container, the thing you want to train, is a Model. A Model is just like a Layer, but with added training and serialization utilities.

Let's put all of these things together into an end-to-end example: we're going to implement a Variational AutoEncoder (VAE). We'll train it on MNIST digits.

Our VAE will be a subclass of Model, built as a nested composition of layers that subclass Layer. It will feature a regularization loss (KL divergence).

from keras import layers


@keras.saving.register_keras_serializable()
class Sampling(layers.Layer):
    """Uses (z_mean, z_log_var) to sample z, the vector encoding a digit."""

    def call(self, inputs):
        z_mean, z_log_var = inputs
        batch = tf.shape(z_mean)[0]
        dim = tf.shape(z_mean)[1]
        epsilon = keras.backend.random_normal(shape=(batch, dim))
        return z_mean + tf.exp(0.5 * z_log_var) * epsilon


@keras.saving.register_keras_serializable()
class Encoder(layers.Layer):
    """Maps MNIST digits to a triplet (z_mean, z_log_var, z)."""

    def __init__(self, latent_dim=32, intermediate_dim=64, name="encoder", **kwargs):
        super().__init__(name=name, **kwargs)
        self.dense_proj = layers.Dense(intermediate_dim, activation="relu")
        self.dense_mean = layers.Dense(latent_dim)
        self.dense_log_var = layers.Dense(latent_dim)
        self.sampling = Sampling()

    def call(self, inputs):
        x = self.dense_proj(inputs)
        z_mean = self.dense_mean(x)
        z_log_var = self.dense_log_var(x)
        z = self.sampling((z_mean, z_log_var))
        return z_mean, z_log_var, z


@keras.saving.register_keras_serializable()
class Decoder(layers.Layer):
    """Converts z, the encoded digit vector, back into a readable digit."""

    def __init__(self, original_dim, intermediate_dim=64, name="decoder", **kwargs):
        super().__init__(name=name, **kwargs)
        self.dense_proj = layers.Dense(intermediate_dim, activation="relu")
        self.dense_output = layers.Dense(original_dim, activation="sigmoid")

    def call(self, inputs):
        x = self.dense_proj(inputs)
        return self.dense_output(x)


@keras.saving.register_keras_serializable()
class VariationalAutoEncoder(keras.Model):
    """Combines the encoder and decoder into an end-to-end model for training."""

    def __init__(
        self,
        original_dim,
        intermediate_dim=64,
        latent_dim=32,
        name="autoencoder",
        **kwargs
    ):
        super().__init__(name=name, **kwargs)
        self.original_dim = original_dim
        self.encoder = Encoder(latent_dim=latent_dim, intermediate_dim=intermediate_dim)
        self.decoder = Decoder(original_dim, intermediate_dim=intermediate_dim)

    def call(self, inputs):
        z_mean, z_log_var, z = self.encoder(inputs)
        reconstructed = self.decoder(z)
        # Add KL divergence regularization loss.
        kl_loss = -0.5 * tf.reduce_mean(
            z_log_var - tf.square(z_mean) - tf.exp(z_log_var) + 1
        )
        self.add_loss(kl_loss)
        return reconstructed

Let's write a simple training loop on MNIST:

original_dim = 784
vae = VariationalAutoEncoder(original_dim, 64, 32)

optimizer = keras.optimizers.Adam(learning_rate=1e-3)
mse_loss_fn = keras.losses.MeanSquaredError()

loss_metric = keras.metrics.Mean()

(x_train, _), _ = keras.datasets.mnist.load_data()
x_train = x_train.reshape(60000, 784).astype("float32") / 255

train_dataset = tf.data.Dataset.from_tensor_slices(x_train)
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(64)

epochs = 2

# Iterate over epochs.
for epoch in range(epochs):
    print("Start of epoch %d" % (epoch,))

    # Iterate over the batches of the dataset.
    for step, x_batch_train in enumerate(train_dataset):
        with tf.GradientTape() as tape:
            reconstructed = vae(x_batch_train)
            # Compute reconstruction loss
            loss = mse_loss_fn(x_batch_train, reconstructed)
            loss += sum(vae.losses)  # Add KLD regularization loss

        grads = tape.gradient(loss, vae.trainable_weights)
        optimizer.apply_gradients(zip(grads, vae.trainable_weights))

        loss_metric(loss)

        if step % 100 == 0:
            print("step %d: mean loss = %.4f" % (step, loss_metric.result()))

Start of epoch 0
WARNING:tensorflow:5 out of the last 5 calls to <function _BaseOptimizer._update_step_xla at 0x7fb220066af0> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has reduce_retracing=True option that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for  more details.
WARNING:tensorflow:6 out of the last 6 calls to <function _BaseOptimizer._update_step_xla at 0x7fb220066af0> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has reduce_retracing=True option that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for  more details.
step 0: mean loss = 0.3433
step 100: mean loss = 0.1257
step 200: mean loss = 0.0994
step 300: mean loss = 0.0893
step 400: mean loss = 0.0844
step 500: mean loss = 0.0810
step 600: mean loss = 0.0788
step 700: mean loss = 0.0772
step 800: mean loss = 0.0760
step 900: mean loss = 0.0750
Start of epoch 1
step 0: mean loss = 0.0747
step 100: mean loss = 0.0741
step 200: mean loss = 0.0736
step 300: mean loss = 0.0731
step 400: mean loss = 0.0727
step 500: mean loss = 0.0723
step 600: mean loss = 0.0720
step 700: mean loss = 0.0717
step 800: mean loss = 0.0715
step 900: mean loss = 0.0712

Note that since the VAE is subclassing Model, it features built-in training loops. So you could also have trained it like this:

vae = VariationalAutoEncoder(784, 64, 32)

optimizer = keras.optimizers.Adam(learning_rate=1e-3)

vae.compile(optimizer, loss=keras.losses.MeanSquaredError())
vae.fit(x_train, x_train, epochs=2, batch_size=64)

Epoch 1/2
938/938 [==============================] - 4s 3ms/step - loss: 0.0746
Epoch 2/2
938/938 [==============================] - 3s 3ms/step - loss: 0.0676
<keras.src.callbacks.History at 0x7fb1e0533580>