intro_to_keras_for_researchers.ipynb

Kernel: Python 3

Introduction to Keras for Researchers

Author: fchollet
Date created: 2020/04/01
Last modified: 2020/10/02
Description: Everything you need to know to use Keras & TensorFlow for deep learning research.

Setup

import tensorflow as tf
import keras

Introduction

Are you a machine learning researcher? Do you publish at NeurIPS and push the state-of-the-art in CV and NLP? This guide will serve as your first introduction to core Keras & TensorFlow API concepts.

In this guide, you will learn about:

• Tensors, variables, and gradients in TensorFlow

• Creating layers by subclassing the Layer class

• Writing low-level training loops

• Tracking losses created by layers via the add_loss() method

• Tracking metrics in a low-level training loop

• Speeding up execution with a compiled tf.function

• Executing layers in training or inference mode

• The Keras Functional API

You will also see the Keras API in action in two end-to-end research examples: a Variational Autoencoder, and a Hypernetwork.

Tensors

TensorFlow is an infrastructure layer for differentiable programming. At its heart, it's a framework for manipulating N-dimensional arrays (tensors), much like NumPy.

However, there are three key differences between NumPy and TensorFlow:

• TensorFlow can leverage hardware accelerators such as GPUs and TPUs.

• TensorFlow can automatically compute the gradient of arbitrary differentiable tensor expressions.

• TensorFlow computation can be distributed to large numbers of devices on a single machine, and large number of machines (potentially with multiple devices each).

Let's take a look at the object that is at the core of TensorFlow: the Tensor.

Here's a constant tensor:

x = tf.constant([[5, 2], [1, 3]])
print(x)

You can get its value as a NumPy array by calling .numpy():

x.numpy()

Much like a NumPy array, it features the attributes dtype and shape:

print("dtype:", x.dtype)
print("shape:", x.shape)

A common way to create constant tensors is via tf.ones and tf.zeros (just like np.ones and np.zeros):

print(tf.ones(shape=(2, 1)))
print(tf.zeros(shape=(2, 1)))

You can also create random constant tensors:

x = tf.random.normal(shape=(2, 2), mean=0.0, stddev=1.0)

x = tf.random.uniform(shape=(2, 2), minval=0, maxval=10, dtype="int32")

Variables

Variables are special tensors used to store mutable state (such as the weights of a neural network). You create a Variable using some initial value:

initial_value = tf.random.normal(shape=(2, 2))
a = tf.Variable(initial_value)
print(a)

You update the value of a Variable by using the methods .assign(value), .assign_add(increment), or .assign_sub(decrement):

new_value = tf.random.normal(shape=(2, 2))
a.assign(new_value)
for i in range(2):
    for j in range(2):
        assert a[i, j] == new_value[i, j]

added_value = tf.random.normal(shape=(2, 2))
a.assign_add(added_value)
for i in range(2):
    for j in range(2):
        assert a[i, j] == new_value[i, j] + added_value[i, j]

Doing math in TensorFlow

If you've used NumPy, doing math in TensorFlow will look very familiar. The main difference is that your TensorFlow code can run on GPU and TPU.

a = tf.random.normal(shape=(2, 2))
b = tf.random.normal(shape=(2, 2))

c = a + b
d = tf.square(c)
e = tf.exp(d)

Gradients

Here's another big difference with NumPy: you can automatically retrieve the gradient of any differentiable expression.

Just open a GradientTape, start "watching" a tensor via tape.watch(), and compose a differentiable expression using this tensor as input:

a = tf.random.normal(shape=(2, 2))
b = tf.random.normal(shape=(2, 2))

with tf.GradientTape() as tape:
    tape.watch(a)  # Start recording the history of operations applied to `a`
    c = tf.sqrt(tf.square(a) + tf.square(b))  # Do some math using `a`
    # What's the gradient of `c` with respect to `a`?
    dc_da = tape.gradient(c, a)
    print(dc_da)

By default, variables are watched automatically, so you don't need to manually watch them:

a = tf.Variable(a)

with tf.GradientTape() as tape:
    c = tf.sqrt(tf.square(a) + tf.square(b))
    dc_da = tape.gradient(c, a)
    print(dc_da)

Note that you can compute higher-order derivatives by nesting tapes:

with tf.GradientTape() as outer_tape:
    with tf.GradientTape() as tape:
        c = tf.sqrt(tf.square(a) + tf.square(b))
        dc_da = tape.gradient(c, a)
    d2c_da2 = outer_tape.gradient(dc_da, a)
    print(d2c_da2)

Keras layers

While TensorFlow is an infrastructure layer for differentiable programming, dealing with tensors, variables, and gradients, Keras is a user interface for deep learning, dealing with layers, models, optimizers, loss functions, metrics, and more.

Keras serves as the high-level API for TensorFlow: Keras is what makes TensorFlow simple and productive.

The Layer class is the fundamental abstraction in Keras. A Layer encapsulates a state (weights) and some computation (defined in the call method).

A simple layer looks like this. The self.add_weight() method gives you a shortcut for creating weights:

class Linear(keras.layers.Layer):
    """y = w.x + b"""

    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 instance much like a Python function:

# Instantiate our layer.
linear_layer = Linear(units=4, input_dim=2)

# The layer can be treated as a function.
# Here we call it on some data.
y = linear_layer(tf.ones((2, 2)))
assert y.shape == (2, 4)

The weight variables (created in __init__) are automatically tracked under the weights property:

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

You have many built-in layers available, from Dense to Conv2D to LSTM to fancier ones like Conv3DTranspose or ConvLSTM2D. Be smart about reusing built-in functionality.

Layer weight creation in build(input_shape)

It's often a good idea to defer weight creation to the build() method, so that you don't need to specify the input dim/shape at layer construction time:

class Linear(keras.layers.Layer):
    """y = w.x + b"""

    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


# Instantiate our layer.
linear_layer = Linear(4)

# This will also call `build(input_shape)` and create the weights.
y = linear_layer(tf.ones((2, 2)))

Layer gradients

You can automatically retrieve the gradients of the weights of a layer by calling it inside a GradientTape. Using these gradients, you can update the weights of the layer, either manually, or using an optimizer object. Of course, you can modify the gradients before using them, if you need to.

# Prepare a dataset.
(x_train, y_train), _ = keras.datasets.mnist.load_data()
dataset = tf.data.Dataset.from_tensor_slices(
    (x_train.reshape(60000, 784).astype("float32") / 255, y_train)
)
dataset = dataset.shuffle(buffer_size=1024).batch(64)

# Instantiate our linear layer (defined above) with 10 units.
linear_layer = Linear(10)

# Instantiate a logistic loss function that expects integer targets.
loss_fn = keras.losses.SparseCategoricalCrossentropy(from_logits=True)

# Instantiate an optimizer.
optimizer = keras.optimizers.SGD(learning_rate=1e-3)

# Iterate over the batches of the dataset.
for step, (x, y) in enumerate(dataset):
    # Open a GradientTape.
    with tf.GradientTape() as tape:
        # Forward pass.
        logits = linear_layer(x)

        # Loss value for this batch.
        loss = loss_fn(y, logits)

    # Get gradients of the loss wrt the weights.
    gradients = tape.gradient(loss, linear_layer.trainable_weights)

    # Update the weights of our linear layer.
    optimizer.apply_gradients(zip(gradients, linear_layer.trainable_weights))

    # Logging.
    if step % 100 == 0:
        print("Step:", step, "Loss:", float(loss))

Trainable and non-trainable weights

Weights created by layers can be either trainable or non-trainable. They're exposed in trainable_weights and non_trainable_weights respectively. Here's a layer with a non-trainable weight:

class ComputeSum(keras.layers.Layer):
    """Returns the sum of the inputs."""

    def __init__(self, input_dim):
        super().__init__()
        # Create a non-trainable weight.
        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


my_sum = ComputeSum(2)
x = tf.ones((2, 2))

y = my_sum(x)
print(y.numpy())  # [2. 2.]

y = my_sum(x)
print(y.numpy())  # [4. 4.]

assert my_sum.weights == [my_sum.total]
assert my_sum.non_trainable_weights == [my_sum.total]
assert my_sum.trainable_weights == []

Layers that own layers

Layers can be recursively nested to create bigger computation blocks. Each layer will track the weights of its sublayers (both trainable and non-trainable).

# Let's reuse the Linear class
# with a `build` method that we defined above.


class MLP(keras.layers.Layer):
    """Simple stack of Linear layers."""

    def __init__(self):
        super().__init__()
        self.linear_1 = Linear(32)
        self.linear_2 = Linear(32)
        self.linear_3 = Linear(10)

    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 = MLP()

# The first call to the `mlp` object will create the weights.
y = mlp(tf.ones(shape=(3, 64)))

# Weights are recursively tracked.
assert len(mlp.weights) == 6

Note that our manually-created MLP above is equivalent to the following built-in option:

mlp = keras.Sequential(
    [
        keras.layers.Dense(32, activation=tf.nn.relu),
        keras.layers.Dense(32, activation=tf.nn.relu),
        keras.layers.Dense(10),
    ]
)

Tracking losses created by layers

Layers can create losses during the forward pass via the add_loss() method. This is especially useful for regularization losses. The losses created by sublayers are recursively tracked by the parent layers.

Here's a layer that creates an activity regularization loss:

class ActivityRegularization(keras.layers.Layer):
    """Layer that creates an activity sparsity regularization loss."""

    def __init__(self, rate=1e-2):
        super().__init__()
        self.rate = rate

    def call(self, inputs):
        # We use `add_loss` to create a regularization loss
        # that depends on the inputs.
        self.add_loss(self.rate * tf.reduce_sum(inputs))
        return inputs

Any model incorporating this layer will track this regularization loss:

# Let's use the loss layer in a MLP block.


class SparseMLP(keras.layers.Layer):
    """Stack of Linear layers with a sparsity regularization loss."""

    def __init__(self):
        super().__init__()
        self.linear_1 = Linear(32)
        self.regularization = ActivityRegularization(1e-2)
        self.linear_3 = Linear(10)

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


mlp = SparseMLP()
y = mlp(tf.ones((10, 10)))

print(mlp.losses)  # List containing one float32 scalar

These losses are cleared by the top-level layer at the start of each forward pass -- they don't accumulate. layer.losses always contains only the losses created during the last forward pass. You would typically use these losses by summing them before computing your gradients when writing a training loop.

# Losses correspond to the *last* forward pass.
mlp = SparseMLP()
mlp(tf.ones((10, 10)))
assert len(mlp.losses) == 1
mlp(tf.ones((10, 10)))
assert len(mlp.losses) == 1  # No accumulation.

# Let's demonstrate how to use these losses in a training loop.

# Prepare a dataset.
(x_train, y_train), _ = keras.datasets.mnist.load_data()
dataset = tf.data.Dataset.from_tensor_slices(
    (x_train.reshape(60000, 784).astype("float32") / 255, y_train)
)
dataset = dataset.shuffle(buffer_size=1024).batch(64)

# A new MLP.
mlp = SparseMLP()

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

for step, (x, y) in enumerate(dataset):
    with tf.GradientTape() as tape:
        # Forward pass.
        logits = mlp(x)

        # External loss value for this batch.
        loss = loss_fn(y, logits)

        # Add the losses created during the forward pass.
        loss += sum(mlp.losses)

        # Get gradients of the loss wrt the weights.
        gradients = tape.gradient(loss, mlp.trainable_weights)

    # Update the weights of our linear layer.
    optimizer.apply_gradients(zip(gradients, mlp.trainable_weights))

    # Logging.
    if step % 100 == 0:
        print("Step:", step, "Loss:", float(loss))

Keeping track of training metrics

Keras offers a broad range of built-in metrics, like keras.metrics.AUC or keras.metrics.PrecisionAtRecall. It's also easy to create your own metrics in a few lines of code.

To use a metric in a custom training loop, you would:

• Instantiate the metric object, e.g. metric = keras.metrics.AUC()

• Call its metric.udpate_state(targets, predictions) method for each batch of data

• Query its result via metric.result()

• Reset the metric's state at the end of an epoch or at the start of an evaluation via metric.reset_state()

Here's a simple example:

# Instantiate a metric object
accuracy = keras.metrics.SparseCategoricalAccuracy()

# Prepare our layer, loss, and optimizer.
model = keras.Sequential(
    [
        keras.layers.Dense(32, activation="relu"),
        keras.layers.Dense(32, activation="relu"),
        keras.layers.Dense(10),
    ]
)
loss_fn = keras.losses.SparseCategoricalCrossentropy(from_logits=True)
optimizer = keras.optimizers.Adam(learning_rate=1e-3)

for epoch in range(2):
    # Iterate over the batches of a dataset.
    for step, (x, y) in enumerate(dataset):
        with tf.GradientTape() as tape:
            logits = model(x)
            # Compute the loss value for this batch.
            loss_value = loss_fn(y, logits)

        # Update the state of the `accuracy` metric.
        accuracy.update_state(y, logits)

        # Update the weights of the model to minimize the loss value.
        gradients = tape.gradient(loss_value, model.trainable_weights)
        optimizer.apply_gradients(zip(gradients, model.trainable_weights))

        # Logging the current accuracy value so far.
        if step % 200 == 0:
            print("Epoch:", epoch, "Step:", step)
            print("Total running accuracy so far: %.3f" % accuracy.result())

    # Reset the metric's state at the end of an epoch
    accuracy.reset_state()

You can also define your own metrics by subclassing keras.metrics.Metric. You need to override the three functions called above:

• Override update_state() to update the statistic values.

• Override result() to return the metric value.

• Override reset_state() to reset the metric to its initial state.

Here is an example where we implement the F1-score metric (with support for sample weighting).

class F1Score(keras.metrics.Metric):
    def __init__(self, name="f1_score", dtype="float32", threshold=0.5, **kwargs):
        super().__init__(name=name, dtype=dtype, **kwargs)
        self.threshold = 0.5
        self.true_positives = self.add_weight(
            name="tp", dtype=dtype, initializer="zeros"
        )
        self.false_positives = self.add_weight(
            name="fp", dtype=dtype, initializer="zeros"
        )
        self.false_negatives = self.add_weight(
            name="fn", dtype=dtype, initializer="zeros"
        )

    def update_state(self, y_true, y_pred, sample_weight=None):
        y_pred = tf.math.greater_equal(y_pred, self.threshold)
        y_true = tf.cast(y_true, tf.bool)
        y_pred = tf.cast(y_pred, tf.bool)

        true_positives = tf.cast(y_true & y_pred, self.dtype)
        false_positives = tf.cast(~y_true & y_pred, self.dtype)
        false_negatives = tf.cast(y_true & ~y_pred, self.dtype)

        if sample_weight is not None:
            sample_weight = tf.cast(sample_weight, self.dtype)
            true_positives *= sample_weight
            false_positives *= sample_weight
            false_negatives *= sample_weight

        self.true_positives.assign_add(tf.reduce_sum(true_positives))
        self.false_positives.assign_add(tf.reduce_sum(false_positives))
        self.false_negatives.assign_add(tf.reduce_sum(false_negatives))

    def result(self):
        precision = self.true_positives / (self.true_positives + self.false_positives)
        recall = self.true_positives / (self.true_positives + self.false_negatives)
        return precision * recall * 2.0 / (precision + recall)

    def reset_state(self):
        self.true_positives.assign(0)
        self.false_positives.assign(0)
        self.false_negatives.assign(0)

Let's test-drive it:

m = F1Score()
m.update_state([0, 1, 0, 0], [0.3, 0.5, 0.8, 0.9])
print("Intermediate result:", float(m.result()))

m.update_state([1, 1, 1, 1], [0.1, 0.7, 0.6, 0.0])
print("Final result:", float(m.result()))

Compiled functions

Running eagerly is great for debugging, but you will get better performance by compiling your computation into static graphs. Static graphs are a researcher's best friends. You can compile any function by wrapping it in a tf.function decorator.

# Prepare our layer, loss, and optimizer.
model = keras.Sequential(
    [
        keras.layers.Dense(32, activation="relu"),
        keras.layers.Dense(32, activation="relu"),
        keras.layers.Dense(10),
    ]
)
loss_fn = keras.losses.SparseCategoricalCrossentropy(from_logits=True)
optimizer = keras.optimizers.Adam(learning_rate=1e-3)

# Create a training step function.


@tf.function  # Make it fast.
def train_on_batch(x, y):
    with tf.GradientTape() as tape:
        logits = model(x)
        loss = loss_fn(y, logits)
        gradients = tape.gradient(loss, model.trainable_weights)
    optimizer.apply_gradients(zip(gradients, model.trainable_weights))
    return loss


# Prepare a dataset.
(x_train, y_train), _ = keras.datasets.mnist.load_data()
dataset = tf.data.Dataset.from_tensor_slices(
    (x_train.reshape(60000, 784).astype("float32") / 255, y_train)
)
dataset = dataset.shuffle(buffer_size=1024).batch(64)

for step, (x, y) in enumerate(dataset):
    loss = train_on_batch(x, y)
    if step % 100 == 0:
        print("Step:", step, "Loss:", float(loss))

Training mode & inference mode

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 modes.

class Dropout(keras.layers.Layer):
    def __init__(self, rate):
        super().__init__()
        self.rate = rate

    def call(self, inputs, training=None):
        if training:
            return tf.nn.dropout(inputs, rate=self.rate)
        return inputs


class MLPWithDropout(keras.layers.Layer):
    def __init__(self):
        super().__init__()
        self.linear_1 = Linear(32)
        self.dropout = Dropout(0.5)
        self.linear_3 = Linear(10)

    def call(self, inputs, training=None):
        x = self.linear_1(inputs)
        x = tf.nn.relu(x)
        x = self.dropout(x, training=training)
        return self.linear_3(x)


mlp = MLPWithDropout()
y_train = mlp(tf.ones((2, 2)), training=True)
y_test = mlp(tf.ones((2, 2)), training=False)

The Functional API for model-building

To build deep learning models, you don't have to use object-oriented programming all the time. All layers we've seen so far can also be composed functionally, like this (we call it the "Functional API"):

# We use an `Input` object to describe the shape and dtype of the inputs.
# This is the deep learning equivalent of *declaring a type*.
# The shape argument is per-sample; it does not include the batch size.
# The functional API focused on defining per-sample transformations.
# The model we create will automatically batch the per-sample transformations,
# so that it can be called on batches of data.
inputs = keras.Input(shape=(16,), dtype="float32")

# We call layers on these "type" objects
# and they return updated types (new shapes/dtypes).
x = Linear(32)(inputs)  # We are reusing the Linear layer we defined earlier.
x = Dropout(0.5)(x)  # We are reusing the Dropout layer we defined earlier.
outputs = Linear(10)(x)

# A functional `Model` can be defined by specifying inputs and outputs.
# A model is itself a layer like any other.
model = keras.Model(inputs, outputs)

# A functional model already has weights, before being called on any data.
# That's because we defined its input shape in advance (in `Input`).
assert len(model.weights) == 4

# Let's call our model on some data, for fun.
y = model(tf.ones((2, 16)))
assert y.shape == (2, 10)

# You can pass a `training` argument in `__call__`
# (it will get passed down to the Dropout layer).
y = model(tf.ones((2, 16)), training=True)

The Functional API tends to be more concise than subclassing, and provides a few other advantages (generally the same advantages that functional, typed languages provide over untyped OO development). However, it can only be used to define DAGs of layers -- recursive networks should be defined as Layer subclasses instead.

Learn more about the Functional API here.

In your research workflows, you may often find yourself mix-and-matching OO models and Functional models.

Note that the Model class also features built-in training & evaluation loops: fit(), predict() and evaluate() (configured via the compile() method). These built-in functions give you access to the following built-in training infrastructure features:

• Callbacks. You can leverage built-in callbacks for early-stopping, model checkpointing, and monitoring training with TensorBoard. You can also implement custom callbacks if needed.

• Distributed training. You can easily scale up your training to multiple GPUs, TPU, or even multiple machines with the tf.distribute API -- with no changes to your code.

• Step fusing. With the steps_per_execution argument in Model.compile(), you can process multiple batches in a single tf.function call, which greatly improves device utilization on TPUs.

We won't go into the details, but we provide a simple code example below. It leverages the built-in training infrastructure to implement the MNIST example above.

inputs = keras.Input(shape=(784,), dtype="float32")
x = keras.layers.Dense(32, activation="relu")(inputs)
x = keras.layers.Dense(32, activation="relu")(x)
outputs = keras.layers.Dense(10)(x)
model = keras.Model(inputs, outputs)

# Specify the loss, optimizer, and metrics with `compile()`.
model.compile(
    loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    optimizer=keras.optimizers.Adam(learning_rate=1e-3),
    metrics=[keras.metrics.SparseCategoricalAccuracy()],
)

# Train the model with the dataset for 2 epochs.
model.fit(dataset, epochs=2)
model.predict(dataset)
model.evaluate(dataset)

You can always subclass the Model class (it works exactly like subclassing Layer) if you want to leverage built-in training loops for your OO models. Just override the Model.train_step() to customize what happens in fit() while retaining support for the built-in infrastructure features outlined above -- callbacks, zero-code distribution support, and step fusing support. You may also override test_step() to customize what happens in evaluate(), and override predict_step() to customize what happens in predict(). For more information, please refer to this guide.

class CustomModel(keras.Model):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.loss_tracker = keras.metrics.Mean(name="loss")
        self.accuracy = keras.metrics.SparseCategoricalAccuracy()
        self.loss_fn = keras.losses.SparseCategoricalCrossentropy(from_logits=True)
        self.optimizer = keras.optimizers.Adam(learning_rate=1e-3)

    def train_step(self, data):
        # Unpack the data. Its structure depends on your model and
        # on what you pass to `fit()`.
        x, y = data
        with tf.GradientTape() as tape:
            y_pred = self(x, training=True)  # Forward pass
            loss = self.loss_fn(y, y_pred)
        gradients = tape.gradient(loss, self.trainable_weights)
        self.optimizer.apply_gradients(zip(gradients, self.trainable_weights))
        # Update metrics (includes the metric that tracks the loss)
        self.loss_tracker.update_state(loss)
        self.accuracy.update_state(y, y_pred)
        # Return a dict mapping metric names to current value
        return {"loss": self.loss_tracker.result(), "accuracy": self.accuracy.result()}

    @property
    def metrics(self):
        # We list our `Metric` objects here so that `reset_states()` can be
        # called automatically at the start of each epoch.
        return [self.loss_tracker, self.accuracy]


inputs = keras.Input(shape=(784,), dtype="float32")
x = keras.layers.Dense(32, activation="relu")(inputs)
x = keras.layers.Dense(32, activation="relu")(x)
outputs = keras.layers.Dense(10)(x)
model = CustomModel(inputs, outputs)
model.compile()
model.fit(dataset, epochs=2)

End-to-end experiment example 1: variational autoencoders.

Here are some of the things you've learned so far:

• A Layer encapsulates a state (created in __init__ or build) and some computation (defined in call).

• Layers can be recursively nested to create new, bigger computation blocks.

• You can easily write highly hackable training loops by opening a GradientTape, calling your model inside the tape's scope, then retrieving gradients and applying them via an optimizer.

• You can speed up your training loops using the @tf.function decorator.

• Layers can create and track losses (typically regularization losses) via self.add_loss().

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

Below is our model definition.

First, we have an Encoder class, which uses a Sampling layer to map a MNIST digit to a latent-space triplet (z_mean, z_log_var, z).

from tensorflow.keras import layers


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


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, **kwargs):
        super().__init__(**kwargs)
        self.dense_proj = layers.Dense(intermediate_dim, activation=tf.nn.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

Next, we have a Decoder class, which maps the probabilistic latent space coordinates back to a MNIST digit.

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

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

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

Finally, our VariationalAutoEncoder composes together an encoder and a decoder, and creates a KL divergence regularization loss via add_loss().

class VariationalAutoEncoder(layers.Layer):
    """Combines the encoder and decoder into an end-to-end model for training."""

    def __init__(self, original_dim, intermediate_dim=64, latent_dim=32, **kwargs):
        super().__init__(**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

Now, let's write a training loop. Our training step is decorated with a @tf.function to compile into a super fast graph function.

# Our model.
vae = VariationalAutoEncoder(original_dim=784, intermediate_dim=64, latent_dim=32)

# Loss and optimizer.
loss_fn = keras.losses.MeanSquaredError()
optimizer = keras.optimizers.Adam(learning_rate=1e-3)

# Prepare a dataset.
(x_train, _), _ = keras.datasets.mnist.load_data()
dataset = tf.data.Dataset.from_tensor_slices(
    x_train.reshape(60000, 784).astype("float32") / 255
)
dataset = dataset.shuffle(buffer_size=1024).batch(32)


@tf.function
def training_step(x):
    with tf.GradientTape() as tape:
        reconstructed = vae(x)  # Compute input reconstruction.
        # Compute loss.
        loss = loss_fn(x, reconstructed)
        loss += sum(vae.losses)  # Add KLD term.
    # Update the weights of the VAE.
    grads = tape.gradient(loss, vae.trainable_weights)
    optimizer.apply_gradients(zip(grads, vae.trainable_weights))
    return loss


losses = []  # Keep track of the losses over time.
for step, x in enumerate(dataset):
    loss = training_step(x)
    # Logging.
    losses.append(float(loss))
    if step % 100 == 0:
        print("Step:", step, "Loss:", sum(losses) / len(losses))

    # Stop after 1000 steps.
    # Training the model to convergence is left
    # as an exercise to the reader.
    if step >= 1000:
        break

As you can see, building and training this type of model in Keras is quick and painless.

End-to-end experiment example 2: hypernetworks.

Let's take a look at another kind of research experiment: hypernetworks.

The idea is to use a small deep neural network (the hypernetwork) to generate the weights for a larger network (the main network).

Let's implement a really trivial hypernetwork: we'll use a small 2-layer network to generate the weights of a larger 3-layer network.

import numpy as np

input_dim = 784
classes = 10

# This is the main network we'll actually use to predict labels.
main_network = keras.Sequential(
    [
        keras.layers.Dense(64, activation=tf.nn.relu),
        keras.layers.Dense(classes),
    ]
)

# It doesn't need to create its own weights, so let's mark its layers
# as already built. That way, calling `main_network` won't create new variables.
for layer in main_network.layers:
    layer.built = True

# This is the number of weight coefficients to generate. Each layer in the
# main network requires output_dim * input_dim + output_dim coefficients.
num_weights_to_generate = (classes * 64 + classes) + (64 * input_dim + 64)

# This is the hypernetwork that generates the weights of the `main_network` above.
hypernetwork = keras.Sequential(
    [
        keras.layers.Dense(16, activation=tf.nn.relu),
        keras.layers.Dense(num_weights_to_generate, activation=tf.nn.sigmoid),
    ]
)

This is our training loop. For each batch of data:

• We use hypernetwork to generate an array of weight coefficients, weights_pred

• We reshape these coefficients into kernel & bias tensors for the main_network

• We run the forward pass of the main_network to compute the actual MNIST predictions

• We run backprop through the weights of the hypernetwork to minimize the final classification loss

# Loss and optimizer.
loss_fn = keras.losses.SparseCategoricalCrossentropy(from_logits=True)
optimizer = keras.optimizers.Adam(learning_rate=1e-4)

# Prepare a dataset.
(x_train, y_train), _ = keras.datasets.mnist.load_data()
dataset = tf.data.Dataset.from_tensor_slices(
    (x_train.reshape(60000, 784).astype("float32") / 255, y_train)
)

# We'll use a batch size of 1 for this experiment.
dataset = dataset.shuffle(buffer_size=1024).batch(1)


@tf.function
def train_step(x, y):
    with tf.GradientTape() as tape:
        # Predict weights for the outer model.
        weights_pred = hypernetwork(x)

        # Reshape them to the expected shapes for w and b for the outer model.
        # Layer 0 kernel.
        start_index = 0
        w0_shape = (input_dim, 64)
        w0_coeffs = weights_pred[:, start_index : start_index + np.prod(w0_shape)]
        w0 = tf.reshape(w0_coeffs, w0_shape)
        start_index += np.prod(w0_shape)
        # Layer 0 bias.
        b0_shape = (64,)
        b0_coeffs = weights_pred[:, start_index : start_index + np.prod(b0_shape)]
        b0 = tf.reshape(b0_coeffs, b0_shape)
        start_index += np.prod(b0_shape)
        # Layer 1 kernel.
        w1_shape = (64, classes)
        w1_coeffs = weights_pred[:, start_index : start_index + np.prod(w1_shape)]
        w1 = tf.reshape(w1_coeffs, w1_shape)
        start_index += np.prod(w1_shape)
        # Layer 1 bias.
        b1_shape = (classes,)
        b1_coeffs = weights_pred[:, start_index : start_index + np.prod(b1_shape)]
        b1 = tf.reshape(b1_coeffs, b1_shape)
        start_index += np.prod(b1_shape)

        # Set the weight predictions as the weight variables on the outer model.
        main_network.layers[0].kernel = w0
        main_network.layers[0].bias = b0
        main_network.layers[1].kernel = w1
        main_network.layers[1].bias = b1

        # Inference on the outer model.
        preds = main_network(x)
        loss = loss_fn(y, preds)

    # Train only inner model.
    grads = tape.gradient(loss, hypernetwork.trainable_weights)
    optimizer.apply_gradients(zip(grads, hypernetwork.trainable_weights))
    return loss


losses = []  # Keep track of the losses over time.
for step, (x, y) in enumerate(dataset):
    loss = train_step(x, y)

    # Logging.
    losses.append(float(loss))
    if step % 100 == 0:
        print("Step:", step, "Loss:", sum(losses) / len(losses))

    # Stop after 1000 steps.
    # Training the model to convergence is left
    # as an exercise to the reader.
    if step >= 1000:
        break

Implementing arbitrary research ideas with Keras is straightforward and highly productive. Imagine trying out 25 ideas per day (20 minutes per experiment on average)!

Keras has been designed to go from idea to results as fast as possible, because we believe this is the key to doing great research.

We hope you enjoyed this quick introduction. Let us know what you build with Keras!