Customizing what happens in `fit()` with PyTorch

Introduction

When you’re doing supervised learning, you can use fit() and everything works smoothly.

When you need to take control of every little detail, you can write your own training loop entirely from scratch.

But what if you need a custom training algorithm, but you still want to benefit from the convenient features of fit(), such as callbacks, built-in distribution support, or step fusing?

A core principle of Keras is progressive disclosure of complexity. You should always be able to get into lower-level workflows in a gradual way. You shouldn’t fall off a cliff if the high-level functionality doesn’t exactly match your use case. You should be able to gain more control over the small details while retaining a commensurate amount of high-level convenience.

When you need to customize what fit() does, you should override the training step function of the Model class. This is the function that is called by fit() for every batch of data. You will then be able to call fit() as usual – and it will be running your own learning algorithm.

Note that this pattern does not prevent you from building models with the Functional API. You can do this whether you’re building Sequential models, Functional API models, or subclassed models.

Let’s see how that works.

Setup

import os

# This guide can only be run with the torch backend.
os.environ["KERAS_BACKEND"] = "torch"

import torch
import keras
from keras import layers
import numpy as np

A first simple example

Let’s start from a simple example:

• We create a new class that subclasses keras.Model.
• We just override the method train_step(self, data).
• We return a dictionary mapping metric names (including the loss) to their current value.

The input argument data is what gets passed to fit as training data:

• If you pass NumPy arrays, by calling fit(x, y, ...), then data will be the tuple (x, y)
• If you pass a torch.utils.data.DataLoader or a tf.data.Dataset, by calling fit(dataset, ...), then data will be what gets yielded by dataset at each batch.

In the body of the train_step() method, we implement a regular training update, similar to what you are already familiar with. Importantly, we compute the loss via self.compute_loss(), which wraps the loss(es) function(s) that were passed to compile().

Similarly, we call metric.update_state(y, y_pred) on metrics from self.metrics, to update the state of the metrics that were passed in compile(), and we query results from self.metrics at the end to retrieve their current value.

class CustomModel(keras.Model):
    def train_step(self, data):
        # Unpack the data. Its structure depends on your model and
        # on what you pass to `fit()`.
        x, y = data

        # Call torch.nn.Module.zero_grad() to clear the leftover gradients
        # for the weights from the previous train step.
        self.zero_grad()

        # Compute loss
        y_pred = self(x, training=True)  # Forward pass
        loss = self.compute_loss(y=y, y_pred=y_pred)

        # Call torch.Tensor.backward() on the loss to compute gradients
        # for the weights.
        loss.backward()

        trainable_weights = [v for v in self.trainable_weights]
        gradients = [v.value.grad for v in trainable_weights]

        # Update weights
        with torch.no_grad():
            self.optimizer.apply(gradients, trainable_weights)

        # Update metrics (includes the metric that tracks the loss)
        for metric in self.metrics:
            if metric.name == "loss":
                metric.update_state(loss)
            else:
                metric.update_state(y, y_pred)

        # Return a dict mapping metric names to current value
        # Note that it will include the loss (tracked in self.metrics).
        return {m.name: m.result() for m in self.metrics}

Let’s try this out:

# Construct and compile an instance of CustomModel
inputs = keras.Input(shape=(32,))
outputs = keras.layers.Dense(1)(inputs)
model = CustomModel(inputs, outputs)
model.compile(optimizer="adam", loss="mse", metrics=["mae"])

# Just use `fit` as usual
x = np.random.random((1000, 32))
y = np.random.random((1000, 1))
model.fit(x, y, epochs=3)

Going lower-level

Naturally, you could just skip passing a loss function in compile(), and instead do everything manually in train_step. Likewise for metrics.

Here’s a lower-level example, that only uses compile() to configure the optimizer:

• We start by creating Metric instances to track our loss and a MAE score (in __init__()).
• We implement a custom train_step() that updates the state of these metrics (by calling update_state() on them), then query them (via result()) to return their current average value, to be displayed by the progress bar and to be pass to any callback.
• Note that we would need to call reset_states() on our metrics between each epoch! Otherwise calling result() would return an average since the start of training, whereas we usually work with per-epoch averages. Thankfully, the framework can do that for us: just list any metric you want to reset in the metrics property of the model. The model will call reset_states() on any object listed here at the beginning of each fit() epoch or at the beginning of a call to evaluate().

class CustomModel(keras.Model):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.loss_tracker = keras.metrics.Mean(name="loss")
        self.mae_metric = keras.metrics.MeanAbsoluteError(name="mae")
        self.loss_fn = keras.losses.MeanSquaredError()

    def train_step(self, data):
        x, y = data

        # Call torch.nn.Module.zero_grad() to clear the leftover gradients
        # for the weights from the previous train step.
        self.zero_grad()

        # Compute loss
        y_pred = self(x, training=True)  # Forward pass
        loss = self.loss_fn(y, y_pred)

        # Call torch.Tensor.backward() on the loss to compute gradients
        # for the weights.
        loss.backward()

        trainable_weights = [v for v in self.trainable_weights]
        gradients = [v.value.grad for v in trainable_weights]

        # Update weights
        with torch.no_grad():
            self.optimizer.apply(gradients, trainable_weights)

        # Compute our own metrics
        self.loss_tracker.update_state(loss)
        self.mae_metric.update_state(y, y_pred)
        return {
            "loss": self.loss_tracker.result(),
            "mae": self.mae_metric.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
        # or at the start of `evaluate()`.
        return [self.loss_tracker, self.mae_metric]


# Construct an instance of CustomModel
inputs = keras.Input(shape=(32,))
outputs = keras.layers.Dense(1)(inputs)
model = CustomModel(inputs, outputs)

# We don't passs a loss or metrics here.
model.compile(optimizer="adam")

# Just use `fit` as usual -- you can use callbacks, etc.
x = np.random.random((1000, 32))
y = np.random.random((1000, 1))
model.fit(x, y, epochs=5)

Supporting sample_weight & class_weight

You may have noticed that our first basic example didn’t make any mention of sample weighting. If you want to support the fit() arguments sample_weight and class_weight, you’d simply do the following:

• Unpack sample_weight from the data argument
• Pass it to compute_loss & update_state (of course, you could also just apply it manually if you don’t rely on compile() for losses & metrics)
• That’s it.

class CustomModel(keras.Model):
    def train_step(self, data):
        # Unpack the data. Its structure depends on your model and
        # on what you pass to `fit()`.
        if len(data) == 3:
            x, y, sample_weight = data
        else:
            sample_weight = None
            x, y = data

        # Call torch.nn.Module.zero_grad() to clear the leftover gradients
        # for the weights from the previous train step.
        self.zero_grad()

        # Compute loss
        y_pred = self(x, training=True)  # Forward pass
        loss = self.compute_loss(
            y=y,
            y_pred=y_pred,
            sample_weight=sample_weight,
        )

        # Call torch.Tensor.backward() on the loss to compute gradients
        # for the weights.
        loss.backward()

        trainable_weights = [v for v in self.trainable_weights]
        gradients = [v.value.grad for v in trainable_weights]

        # Update weights
        with torch.no_grad():
            self.optimizer.apply(gradients, trainable_weights)

        # Update metrics (includes the metric that tracks the loss)
        for metric in self.metrics:
            if metric.name == "loss":
                metric.update_state(loss)
            else:
                metric.update_state(y, y_pred, sample_weight=sample_weight)

        # Return a dict mapping metric names to current value
        # Note that it will include the loss (tracked in self.metrics).
        return {m.name: m.result() for m in self.metrics}


# Construct and compile an instance of CustomModel
inputs = keras.Input(shape=(32,))
outputs = keras.layers.Dense(1)(inputs)
model = CustomModel(inputs, outputs)
model.compile(optimizer="adam", loss="mse", metrics=["mae"])

# You can now use sample_weight argument
x = np.random.random((1000, 32))
y = np.random.random((1000, 1))
sw = np.random.random((1000, 1))
model.fit(x, y, sample_weight=sw, epochs=3)

Providing your own evaluation step

What if you want to do the same for calls to model.evaluate()? Then you would override test_step in exactly the same way. Here’s what it looks like:

class CustomModel(keras.Model):
    def test_step(self, data):
        # Unpack the data
        x, y = data
        # Compute predictions
        y_pred = self(x, training=False)
        # Updates the metrics tracking the loss
        loss = self.compute_loss(y=y, y_pred=y_pred)
        # Update the metrics.
        for metric in self.metrics:
            if metric.name == "loss":
                metric.update_state(loss)
            else:
                metric.update_state(y, y_pred)
        # Return a dict mapping metric names to current value.
        # Note that it will include the loss (tracked in self.metrics).
        return {m.name: m.result() for m in self.metrics}


# Construct an instance of CustomModel
inputs = keras.Input(shape=(32,))
outputs = keras.layers.Dense(1)(inputs)
model = CustomModel(inputs, outputs)
model.compile(loss="mse", metrics=["mae"])

# Evaluate with our custom test_step
x = np.random.random((1000, 32))
y = np.random.random((1000, 1))
model.evaluate(x, y)

Wrapping up: an end-to-end GAN example

Let’s walk through an end-to-end example that leverages everything you just learned.

Let’s consider:

• A generator network meant to generate 28x28x1 images.
• A discriminator network meant to classify 28x28x1 images into two classes (“fake” and “real”).
• One optimizer for each.
• A loss function to train the discriminator.

# Create the discriminator
discriminator = keras.Sequential(
    [
        keras.Input(shape=(28, 28, 1)),
        layers.Conv2D(64, (3, 3), strides=(2, 2), padding="same"),
        layers.LeakyReLU(negative_slope=0.2),
        layers.Conv2D(128, (3, 3), strides=(2, 2), padding="same"),
        layers.LeakyReLU(negative_slope=0.2),
        layers.GlobalMaxPooling2D(),
        layers.Dense(1),
    ],
    name="discriminator",
)

# Create the generator
latent_dim = 128
generator = keras.Sequential(
    [
        keras.Input(shape=(latent_dim,)),
        # We want to generate 128 coefficients to reshape into a 7x7x128 map
        layers.Dense(7 * 7 * 128),
        layers.LeakyReLU(negative_slope=0.2),
        layers.Reshape((7, 7, 128)),
        layers.Conv2DTranspose(128, (4, 4), strides=(2, 2), padding="same"),
        layers.LeakyReLU(negative_slope=0.2),
        layers.Conv2DTranspose(128, (4, 4), strides=(2, 2), padding="same"),
        layers.LeakyReLU(negative_slope=0.2),
        layers.Conv2D(1, (7, 7), padding="same", activation="sigmoid"),
    ],
    name="generator",
)

Here’s a feature-complete GAN class, overriding compile() to use its own signature, and implementing the entire GAN algorithm in 17 lines in train_step:

class GAN(keras.Model):
    def __init__(self, discriminator, generator, latent_dim):
        super().__init__()
        self.discriminator = discriminator
        self.generator = generator
        self.latent_dim = latent_dim
        self.d_loss_tracker = keras.metrics.Mean(name="d_loss")
        self.g_loss_tracker = keras.metrics.Mean(name="g_loss")
        self.seed_generator = keras.random.SeedGenerator(1337)
        self.built = True

    @property
    def metrics(self):
        return [self.d_loss_tracker, self.g_loss_tracker]

    def compile(self, d_optimizer, g_optimizer, loss_fn):
        super().compile()
        self.d_optimizer = d_optimizer
        self.g_optimizer = g_optimizer
        self.loss_fn = loss_fn

    def train_step(self, real_images):
        if isinstance(real_images, tuple):
            real_images = real_images[0]
        # Sample random points in the latent space
        batch_size = real_images.shape[0]
        random_latent_vectors = keras.random.normal(
            shape=(batch_size, self.latent_dim), seed=self.seed_generator
        )

        # Decode them to fake images
        generated_images = self.generator(random_latent_vectors)

        # Combine them with real images
        real_images = torch.tensor(real_images)
        combined_images = torch.concat([generated_images, real_images], axis=0)

        # Assemble labels discriminating real from fake images
        labels = torch.concat(
            [torch.ones((batch_size, 1)), torch.zeros((batch_size, 1))], axis=0
        )
        # Add random noise to the labels - important trick!
        labels += 0.05 * keras.random.uniform(
            labels.shape, seed=self.seed_generator
        )

        # Train the discriminator
        self.zero_grad()
        predictions = self.discriminator(combined_images)
        d_loss = self.loss_fn(labels, predictions)
        d_loss.backward()
        grads = [v.value.grad for v in self.discriminator.trainable_weights]
        with torch.no_grad():
            self.d_optimizer.apply(grads, self.discriminator.trainable_weights)

        # Sample random points in the latent space
        random_latent_vectors = keras.random.normal(
            shape=(batch_size, self.latent_dim), seed=self.seed_generator
        )

        # Assemble labels that say "all real images"
        misleading_labels = torch.zeros((batch_size, 1))

        # Train the generator (note that we should *not* update the weights
        # of the discriminator)!
        self.zero_grad()
        predictions = self.discriminator(self.generator(random_latent_vectors))
        g_loss = self.loss_fn(misleading_labels, predictions)
        grads = g_loss.backward()
        grads = [v.value.grad for v in self.generator.trainable_weights]
        with torch.no_grad():
            self.g_optimizer.apply(grads, self.generator.trainable_weights)

        # Update metrics and return their value.
        self.d_loss_tracker.update_state(d_loss)
        self.g_loss_tracker.update_state(g_loss)
        return {
            "d_loss": self.d_loss_tracker.result(),
            "g_loss": self.g_loss_tracker.result(),
        }

Let’s test-drive it:

# Prepare the dataset. We use both the training & test MNIST digits.
batch_size = 64
(x_train, _), (x_test, _) = keras.datasets.mnist.load_data()
all_digits = np.concatenate([x_train, x_test])
all_digits = all_digits.astype("float32") / 255.0
all_digits = np.reshape(all_digits, (-1, 28, 28, 1))

# Create a TensorDataset
dataset = torch.utils.data.TensorDataset(
    torch.from_numpy(all_digits), torch.from_numpy(all_digits)
)
# Create a DataLoader
dataloader = torch.utils.data.DataLoader(
    dataset, batch_size=batch_size, shuffle=True
)

gan = GAN(
    discriminator=discriminator, generator=generator, latent_dim=latent_dim
)
gan.compile(
    d_optimizer=keras.optimizers.Adam(learning_rate=0.0003),
    g_optimizer=keras.optimizers.Adam(learning_rate=0.0003),
    loss_fn=keras.losses.BinaryCrossentropy(from_logits=True),
)

gan.fit(dataloader, epochs=1)

The ideas behind deep learning are simple, so why should their implementation be painful?