In rstudio/keras: R Interface to 'Keras'

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 JAX backend.
os.environ["KERAS_BACKEND"] = "jax"

import jax
import keras
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 implement a fully-stateless compute_loss_and_updates() method to compute the loss as well as the updated values for the non-trainable variables of the model. Internally, it calls stateless_call() and the built-in compute_loss().
We implement a fully-stateless train_step() method to compute current metric values (including the loss) as well as updated values for the trainable variables, the optimizer variables, and the metric variables.

Note that you can also take into account the sample_weight argument by:

Unpacking the data as x, y, sample_weight = data
Passing sample_weight to compute_loss()
Passing sample_weight alongside y and y_pred to metrics in stateless_update_state()

class CustomModel(keras.Model):
    def compute_loss_and_updates(
        self,
        trainable_variables,
        non_trainable_variables,
        x,
        y,
        training=False,
    ):
        y_pred, non_trainable_variables = self.stateless_call(
            trainable_variables,
            non_trainable_variables,
            x,
            training=training,
        )
        loss = self.compute_loss(x, y, y_pred)
        return loss, (y_pred, non_trainable_variables)

    def train_step(self, state, data):
        (
            trainable_variables,
            non_trainable_variables,
            optimizer_variables,
            metrics_variables,
        ) = state
        x, y = data

        # Get the gradient function.
        grad_fn = jax.value_and_grad(self.compute_loss_and_updates, has_aux=True)

        # Compute the gradients.
        (loss, (y_pred, non_trainable_variables)), grads = grad_fn(
            trainable_variables,
            non_trainable_variables,
            x,
            y,
            training=True,
        )

        # Update trainable variables and optimizer variables.
        (
            trainable_variables,
            optimizer_variables,
        ) = self.optimizer.stateless_apply(
            optimizer_variables, grads, trainable_variables
        )

        # Update metrics.
        new_metrics_vars = []
        logs = {}
        for metric in self.metrics:
            this_metric_vars = metrics_variables[
                len(new_metrics_vars) : len(new_metrics_vars) + len(metric.variables)
            ]
            if metric.name == "loss":
                this_metric_vars = metric.stateless_update_state(this_metric_vars, loss)
            else:
                this_metric_vars = metric.stateless_update_state(
                    this_metric_vars, y, y_pred
                )
            logs[metric.name] = metric.stateless_result(this_metric_vars)
            new_metrics_vars += this_metric_vars

        # Return metric logs and updated state variables.
        state = (
            trainable_variables,
            non_trainable_variables,
            optimizer_variables,
            new_metrics_vars,
        )
        return logs, state

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:

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 compute_loss_and_updates(
        self,
        trainable_variables,
        non_trainable_variables,
        x,
        y,
        training=False,
    ):
        y_pred, non_trainable_variables = self.stateless_call(
            trainable_variables,
            non_trainable_variables,
            x,
            training=training,
        )
        loss = self.loss_fn(y, y_pred)
        return loss, (y_pred, non_trainable_variables)

    def train_step(self, state, data):
        (
            trainable_variables,
            non_trainable_variables,
            optimizer_variables,
            metrics_variables,
        ) = state
        x, y = data

        # Get the gradient function.
        grad_fn = jax.value_and_grad(self.compute_loss_and_updates, has_aux=True)

        # Compute the gradients.
        (loss, (y_pred, non_trainable_variables)), grads = grad_fn(
            trainable_variables,
            non_trainable_variables,
            x,
            y,
            training=True,
        )

        # Update trainable variables and optimizer variables.
        (
            trainable_variables,
            optimizer_variables,
        ) = self.optimizer.stateless_apply(
            optimizer_variables, grads, trainable_variables
        )

        # Update metrics.
        loss_tracker_vars = metrics_variables[: len(self.loss_tracker.variables)]
        mae_metric_vars = metrics_variables[len(self.loss_tracker.variables) :]

        loss_tracker_vars = self.loss_tracker.stateless_update_state(
            loss_tracker_vars, loss
        )
        mae_metric_vars = self.mae_metric.stateless_update_state(
            mae_metric_vars, y, y_pred
        )

        logs = {}
        logs[self.loss_tracker.name] = self.loss_tracker.stateless_result(
            loss_tracker_vars
        )
        logs[self.mae_metric.name] = self.mae_metric.stateless_result(mae_metric_vars)

        new_metrics_vars = loss_tracker_vars + mae_metric_vars

        # Return metric logs and updated state variables.
        state = (
            trainable_variables,
            non_trainable_variables,
            optimizer_variables,
            new_metrics_vars,
        )
        return logs, state

    @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 pass 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)

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, state, data):
        # Unpack the data.
        x, y = data
        (
            trainable_variables,
            non_trainable_variables,
            metrics_variables,
        ) = state

        # Compute predictions and loss.
        y_pred, non_trainable_variables = self.stateless_call(
            trainable_variables,
            non_trainable_variables,
            x,
            training=False,
        )
        loss = self.compute_loss(x, y, y_pred)

        # Update metrics.
        new_metrics_vars = []
        for metric in self.metrics:
            this_metric_vars = metrics_variables[
                len(new_metrics_vars) : len(new_metrics_vars) + len(metric.variables)
            ]
            if metric.name == "loss":
                this_metric_vars = metric.stateless_update_state(this_metric_vars, loss)
            else:
                this_metric_vars = metric.stateless_update_state(
                    this_metric_vars, y, y_pred
                )
            logs = metric.stateless_result(this_metric_vars)
            new_metrics_vars += this_metric_vars

        # Return metric logs and updated state variables.
        state = (
            trainable_variables,
            non_trainable_variables,
            new_metrics_vars,
        )
        return logs, state


# 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)