Train and evaluate with Keras

View on TensorFlow.org Run in Google Colab View source on GitHub Download notebook

This guide covers training, evaluation, and prediction (inference) models in TensorFlow 2.0 in two broad situations:

  • When using built-in APIs for training and validation (such as model.fit(), model.evaluate(), model.predict()). This is covered in the section "Using built-in training and evaluation loops".
  • When writing custom loops from scratch using eager execution and the GradientTape object. This is covered in the section "Writing your own training and evaluation loops from scratch".

In general, whether you are using built-in loops or writing your own, model training and evaluation works strictly in the same way across every kind of Keras model -- Sequential models, models built with the Functional API, and models written from scratch via model subclassing.

This guide doesn't cover distributed training.

Setup

import tensorflow as tf

import numpy as np

Part I: Using built-in training and evaluation loops

When passing data to the built-in training loops of a model, you should either use Numpy arrays (if your data is small and fits in memory) or tf.data Dataset objects. In the next few paragraphs, we'll use the MNIST dataset as Numpy arrays, in order to demonstrate how to use optimizers, losses, and metrics.

API overview: a first end-to-end example

Let's consider the following model (here, we build in with the Functional API, but it could be a Sequential model or a subclassed model as well):

from tensorflow import keras
from tensorflow.keras import layers

inputs = keras.Input(shape=(784,), name='digits')
x = layers.Dense(64, activation='relu', name='dense_1')(inputs)
x = layers.Dense(64, activation='relu', name='dense_2')(x)
outputs = layers.Dense(10, name='predictions')(x)

model = keras.Model(inputs=inputs, outputs=outputs)

Here's what the typical end-to-end workflow looks like, consisting of training, validation on a holdout set generated from the original training data, and finally evaluation on the test data:

Load a toy dataset for the sake of this example

(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()

# Preprocess the data (these are Numpy arrays)
x_train = x_train.reshape(60000, 784).astype('float32') / 255
x_test = x_test.reshape(10000, 784).astype('float32') / 255

y_train = y_train.astype('float32')
y_test = y_test.astype('float32')

# Reserve 10,000 samples for validation
x_val = x_train[-10000:]
y_val = y_train[-10000:]
x_train = x_train[:-10000]
y_train = y_train[:-10000]
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz
11493376/11490434 [==============================] - 0s 0us/step

Specify the training configuration (optimizer, loss, metrics)

model.compile(optimizer=keras.optimizers.RMSprop(),  # Optimizer
              # Loss function to minimize
              loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              # List of metrics to monitor
              metrics=['sparse_categorical_accuracy'])

Train the model by slicing the data into "batches" of size "batch_size", and repeatedly iterating over the entire dataset for a given number of "epochs"

print('# Fit model on training data')
history = model.fit(x_train, y_train,
                    batch_size=64,
                    epochs=3,
                    # We pass some validation for
                    # monitoring validation loss and metrics
                    # at the end of each epoch
                    validation_data=(x_val, y_val))

print('\nhistory dict:', history.history)
# Fit model on training data
Epoch 1/3
782/782 [==============================] - 2s 3ms/step - loss: 0.3299 - sparse_categorical_accuracy: 0.9061 - val_loss: 0.1687 - val_sparse_categorical_accuracy: 0.9521
Epoch 2/3
782/782 [==============================] - 2s 2ms/step - loss: 0.1547 - sparse_categorical_accuracy: 0.9539 - val_loss: 0.1300 - val_sparse_categorical_accuracy: 0.9625
Epoch 3/3
782/782 [==============================] - 2s 2ms/step - loss: 0.1141 - sparse_categorical_accuracy: 0.9654 - val_loss: 0.1371 - val_sparse_categorical_accuracy: 0.9587

history dict: {'loss': [0.32989633083343506, 0.15469889342784882, 0.11413515359163284], 'sparse_categorical_accuracy': [0.9060800075531006, 0.9538800120353699, 0.9654399752616882], 'val_loss': [0.168654665350914, 0.1300402134656906, 0.1370943784713745], 'val_sparse_categorical_accuracy': [0.9520999789237976, 0.9624999761581421, 0.9587000012397766]}

The returned "history" object holds a record of the loss values and metric values during training

# Evaluate the model on the test data using `evaluate`
print('\n# Evaluate on test data')
results = model.evaluate(x_test, y_test, batch_size=128)
print('test loss, test acc:', results)

# Generate predictions (probabilities -- the output of the last layer)
# on new data using `predict`
print('\n# Generate predictions for 3 samples')
predictions = model.predict(x_test[:3])
print('predictions shape:', predictions.shape)

# Evaluate on test data
79/79 [==============================] - 0s 2ms/step - loss: 0.1298 - sparse_categorical_accuracy: 0.9612
test loss, test acc: [0.12980282306671143, 0.9611999988555908]

# Generate predictions for 3 samples
predictions shape: (3, 10)

Specifying a loss, metrics, and an optimizer

To train a model with fit, you need to specify a loss function, an optimizer, and optionally, some metrics to monitor.

You pass these to the model as arguments to the compile() method:

model.compile(optimizer=keras.optimizers.RMSprop(learning_rate=1e-3),
              loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=[keras.metrics.sparse_categorical_accuracy])

The metrics argument should be a list -- you model can have any number of metrics.

If your model has multiple outputs, your can specify different losses and metrics for each output, and you can modulate the contribution of each output to the total loss of the model. You will find more details about this in the section "Passing data to multi-input, multi-output models".

Note that if you're satisfied with the default settings, in many cases the optimizer, loss, and metrics can be specified via string identifiers as a shortcut:

model.compile(optimizer='rmsprop',
              loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['sparse_categorical_accuracy'])

For later reuse, let's put our model definition and compile step in functions; we will call them several times across different examples in this guide.

def get_uncompiled_model():
  inputs = keras.Input(shape=(784,), name='digits')
  x = layers.Dense(64, activation='relu', name='dense_1')(inputs)
  x = layers.Dense(64, activation='relu', name='dense_2')(x)
  outputs = layers.Dense(10, name='predictions')(x)
  model = keras.Model(inputs=inputs, outputs=outputs)
  return model

def get_compiled_model():
  model = get_uncompiled_model()
  model.compile(optimizer=keras.optimizers.RMSprop(learning_rate=1e-3),
                loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
                metrics=['sparse_categorical_accuracy'])
  return model

Many built-in optimizers, losses, and metrics are available

In general, you won't have to create from scratch your own losses, metrics, or optimizers, because what you need is likely already part of the Keras API:

Optimizers:

  • SGD() (with or without momentum)
  • RMSprop()
  • Adam()
  • etc.

Losses:

  • MeanSquaredError()
  • KLDivergence()
  • CosineSimilarity()
  • etc.

Metrics:

  • AUC()
  • Precision()
  • Recall()
  • etc.

Custom losses

There are two ways to provide custom losses with Keras. The first example creates a function that accepts inputs y_true and y_pred. The following example shows a loss function that computes the average absolute error between the real data and the predictions:

def basic_loss_function(y_true, y_pred):
    return tf.math.reduce_mean(tf.abs(y_true - y_pred))

model.compile(optimizer=keras.optimizers.Adam(),
              loss=basic_loss_function)

model.fit(x_train, y_train, batch_size=64, epochs=3)
Epoch 1/3
782/782 [==============================] - 1s 2ms/step - loss: 1.3280
Epoch 2/3
782/782 [==============================] - 1s 2ms/step - loss: 0.7154
Epoch 3/3
782/782 [==============================] - 1s 2ms/step - loss: 0.5950

<tensorflow.python.keras.callbacks.History at 0x7fdd480de438>

If you need a loss function that takes in parameters beside y_true and y_pred, you can subclass the tf.keras.losses.Loss class and implement the following two methods:

  • __init__(self) —Accept parameters to pass during the call of your loss function
  • call(self, y_true, y_pred) —Use the targets (y_true) and the model predictions (y_pred) to compute the model's loss

Parameters passed into __init__() can be used during call() when calculating loss.

The following example shows how to implement a WeightedCrossEntropy loss function that calculates a BinaryCrossEntropy loss, where the loss of a certain class or the whole function can be modified by a scalar.

class WeightedBinaryCrossEntropy(keras.losses.Loss):
    """
    Args:
      pos_weight: Scalar to affect the positive labels of the loss function.
      weight: Scalar to affect the entirety of the loss function.
      from_logits: Whether to compute loss from logits or the probability.
      reduction: Type of tf.keras.losses.Reduction to apply to loss.
      name: Name of the loss function.
    """
    def __init__(self, pos_weight, weight, from_logits=False,
                 reduction=keras.losses.Reduction.AUTO,
                 name='weighted_binary_crossentropy'):
        super().__init__(reduction=reduction, name=name)
        self.pos_weight = pos_weight
        self.weight = weight
        self.from_logits = from_logits

    def call(self, y_true, y_pred):
        ce = tf.losses.binary_crossentropy(
            y_true, y_pred, from_logits=self.from_logits)[:,None]
        ce = self.weight * (ce*(1-y_true) + self.pos_weight*ce*(y_true))
        return ce

This is a binary loss but the dataset has 10 classes, so apply the loss as if the model were making an independent binary prediction for each class. To do that, start by creating one-hot vectors from the class indices:

one_hot_y_train = tf.one_hot(y_train.astype(np.int32), depth=10)

Now use those one-hots, and the custom loss to train a model:

model = get_uncompiled_model()

model.compile(
    optimizer=keras.optimizers.Adam(),
    loss=WeightedBinaryCrossEntropy(
        pos_weight=0.5, weight = 2, from_logits=True)
)

model.fit(x_train, one_hot_y_train, batch_size=64, epochs=5)
Epoch 1/5
782/782 [==============================] - 2s 2ms/step - loss: 0.1672
Epoch 2/5
782/782 [==============================] - 2s 2ms/step - loss: 0.0661
Epoch 3/5
782/782 [==============================] - 2s 2ms/step - loss: 0.0473
Epoch 4/5
782/782 [==============================] - 2s 2ms/step - loss: 0.0368
Epoch 5/5
782/782 [==============================] - 2s 2ms/step - loss: 0.0302

<tensorflow.python.keras.callbacks.History at 0x7fdd48075630>

Custom metrics

If you need a metric that isn't part of the API, you can easily create custom metrics by subclassing the Metric class. You will need to implement 4 methods:

  • __init__(self), in which you will create state variables for your metric.
  • update_state(self, y_true, y_pred, sample_weight=None), which uses the targets y_true and the model predictions y_pred to update the state variables.
  • result(self), which uses the state variables to compute the final results.
  • reset_states(self), which reinitializes the state of the metric.

State update and results computation are kept separate (in update_state() and result(), respectively) because in some cases, results computation might be very expensive, and would only be done periodically.

Here's a simple example showing how to implement a CategoricalTruePositives metric, that counts how many samples where correctly classified as belonging to a given class:

class CategoricalTruePositives(keras.metrics.Metric):

    def __init__(self, name='categorical_true_positives', **kwargs):
      super(CategoricalTruePositives, self).__init__(name=name, **kwargs)
      self.true_positives = self.add_weight(name='tp', initializer='zeros')

    def update_state(self, y_true, y_pred, sample_weight=None):
      y_pred = tf.reshape(tf.argmax(y_pred, axis=1), shape=(-1, 1))
      values = tf.cast(y_true, 'int32') == tf.cast(y_pred, 'int32')
      values = tf.cast(values, 'float32')
      if sample_weight is not None:
        sample_weight = tf.cast(sample_weight, 'float32')
        values = tf.multiply(values, sample_weight)
      self.true_positives.assign_add(tf.reduce_sum(values))

    def result(self):
      return self.true_positives

    def reset_states(self):
      # The state of the metric will be reset at the start of each epoch.
      self.true_positives.assign(0.)
model.compile(optimizer=keras.optimizers.RMSprop(learning_rate=1e-3),
              loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=[CategoricalTruePositives()])
model.fit(x_train, y_train,
          batch_size=64,
          epochs=3)
Epoch 1/3
782/782 [==============================] - 2s 2ms/step - loss: 0.0631 - categorical_true_positives: 49056.0000
Epoch 2/3
782/782 [==============================] - 2s 2ms/step - loss: 0.0516 - categorical_true_positives: 49206.0000
Epoch 3/3
782/782 [==============================] - 2s 2ms/step - loss: 0.0443 - categorical_true_positives: 49336.0000

<tensorflow.python.keras.callbacks.History at 0x7fdd48118e10>

Handling losses and metrics that don't fit the standard signature

The overwhelming majority of losses and metrics can be computed from y_true and y_pred, where y_pred is an output of your model. But not all of them. For instance, a regularization loss may only require the activation of a layer (there are no targets in this case), and this activation may not be a model output.

In such cases, you can call self.add_loss(loss_value) from inside the call method of a custom layer. Here's a simple example that adds activity regularization (note that activity regularization is built-in in all Keras layers -- this layer is just for the sake of providing a concrete example):

class ActivityRegularizationLayer(layers.Layer):

  def call(self, inputs):
    self.add_loss(tf.reduce_sum(inputs) * 0.1)
    return inputs  # Pass-through layer.

inputs = keras.Input(shape=(784,), name='digits')
x = layers.Dense(64, activation='relu', name='dense_1')(inputs)

# Insert activity regularization as a layer
x = ActivityRegularizationLayer()(x)

x = layers.Dense(64, activation='relu', name='dense_2')(x)
outputs = layers.Dense(10, name='predictions')(x)

model = keras.Model(inputs=inputs, outputs=outputs)
model.compile(optimizer=keras.optimizers.RMSprop(learning_rate=1e-3),
              loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True))

# The displayed loss will be much higher than before
# due to the regularization component.
model.fit(x_train, y_train,
          batch_size=64,
          epochs=1)
782/782 [==============================] - 2s 2ms/step - loss: 2.4970

<tensorflow.python.keras.callbacks.History at 0x7fdcf03cb5c0>

You can do the same for logging metric values:

class MetricLoggingLayer(layers.Layer):

  def call(self, inputs):
    # The `aggregation` argument defines
    # how to aggregate the per-batch values
    # over each epoch:
    # in this case we simply average them.
    self.add_metric(keras.backend.std(inputs),
                    name='std_of_activation',
                    aggregation='mean')
    return inputs  # Pass-through layer.


inputs = keras.Input(shape=(784,), name='digits')
x = layers.Dense(64, activation='relu', name='dense_1')(inputs)

# Insert std logging as a layer.
x = MetricLoggingLayer()(x)

x = layers.Dense(64, activation='relu', name='dense_2')(x)
outputs = layers.Dense(10, name='predictions')(x)

model = keras.Model(inputs=inputs, outputs=outputs)
model.compile(optimizer=keras.optimizers.RMSprop(learning_rate=1e-3),
              loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True))
model.fit(x_train, y_train,
          batch_size=64,
          epochs=1)
782/782 [==============================] - 2s 2ms/step - loss: 0.3486 - std_of_activation: 0.9018

<tensorflow.python.keras.callbacks.History at 0x7fdcf0260518>

In the Functional API, you can also call model.add_loss(loss_tensor), or model.add_metric(metric_tensor, name, aggregation).

Here's a simple example:

inputs = keras.Input(shape=(784,), name='digits')
x1 = layers.Dense(64, activation='relu', name='dense_1')(inputs)
x2 = layers.Dense(64, activation='relu', name='dense_2')(x1)
outputs = layers.Dense(10, name='predictions')(x2)
model = keras.Model(inputs=inputs, outputs=outputs)

model.add_loss(tf.reduce_sum(x1) * 0.1)

model.add_metric(keras.backend.std(x1),
                 name='std_of_activation',
                 aggregation='mean')

model.compile(optimizer=keras.optimizers.RMSprop(1e-3),
              loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True))
model.fit(x_train, y_train,
          batch_size=64,
          epochs=1)
782/782 [==============================] - 2s 2ms/step - loss: 2.4707 - std_of_activation: 0.0019

<tensorflow.python.keras.callbacks.History at 0x7fdcf012c240>

Automatically setting apart a validation holdout set

In the first end-to-end example you saw, we used the validation_data argument to pass a tuple of Numpy arrays (x_val, y_val) to the model for evaluating a validation loss and validation metrics at the end of each epoch.

Here's another option: the argument validation_split allows you to automatically reserve part of your training data for validation. The argument value represents the fraction of the data to be reserved for validation, so it should be set to a number higher than 0 and lower than 1. For instance, validation_split=0.2 means "use 20% of the data for validation", and validation_split=0.6 means "use 60% of the data for validation".

The way the validation is computed is by taking the last x% samples of the arrays received by the fit call, before any shuffling.

You can only use validation_split when training with Numpy data.

model = get_compiled_model()
model.fit(x_train, y_train, batch_size=64, validation_split=0.2, epochs=1, steps_per_epoch=1)
1/1 [==============================] - 0s 449ms/step - loss: 2.3406 - sparse_categorical_accuracy: 0.0312 - val_loss: 2.2199 - val_sparse_categorical_accuracy: 0.2010

<tensorflow.python.keras.callbacks.History at 0x7fdcf00360f0>

Training and evaluation from tf.data Datasets

In the past few paragraphs, you've seen how to handle losses, metrics, and optimizers, and you've seen how to use the validation_data and validation_split arguments in fit, when your data is passed as Numpy arrays.

Let's now take a look at the case where your data comes in the form of a tf.data Dataset.

The tf.data API is a set of utilities in TensorFlow 2.0 for loading and preprocessing data in a way that's fast and scalable.

For a complete guide about creating Datasets, see the tf.data documentation.

You can pass a Dataset instance directly to the methods fit(), evaluate(), and predict():

model = get_compiled_model()

# First, let's create a training Dataset instance.
# For the sake of our example, we'll use the same MNIST data as before.
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
# Shuffle and slice the dataset.
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(64)

# Now we get a test dataset.
test_dataset = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_dataset = test_dataset.batch(64)

# Since the dataset already takes care of batching,
# we don't pass a `batch_size` argument.
model.fit(train_dataset, epochs=3)

# You can also evaluate or predict on a dataset.
print('\n# Evaluate')
result = model.evaluate(test_dataset)
dict(zip(model.metrics_names, result))
Epoch 1/3
782/782 [==============================] - 2s 2ms/step - loss: 0.3360 - sparse_categorical_accuracy: 0.9056
Epoch 2/3
782/782 [==============================] - 2s 2ms/step - loss: 0.1613 - sparse_categorical_accuracy: 0.9522
Epoch 3/3
782/782 [==============================] - 2s 2ms/step - loss: 0.1203 - sparse_categorical_accuracy: 0.9651

# Evaluate
157/157 [==============================] - 0s 2ms/step - loss: 0.1157 - sparse_categorical_accuracy: 0.9669

{'loss': 0.11568339914083481,
 'sparse_categorical_accuracy': 0.9668999910354614}

Note that the Dataset is reset at the end of each epoch, so it can be reused of the next epoch.

If you want to run training only on a specific number of batches from this Dataset, you can pass the steps_per_epoch argument, which specifies how many training steps the model should run using this Dataset before moving on to the next epoch.

If you do this, the dataset is not reset at the end of each epoch, instead we just keep drawing the next batches. The dataset will eventually run out of data (unless it is an infinitely-looping dataset).

model = get_compiled_model()

# Prepare the training dataset
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(64).repeat()

# Only use the 100 batches per epoch (that's 64 * 100 samples)
model.fit(train_dataset, steps_per_epoch=100, epochs=3)
Epoch 1/3
100/100 [==============================] - 0s 2ms/step - loss: 0.7803 - sparse_categorical_accuracy: 0.7991
Epoch 2/3
100/100 [==============================] - 0s 2ms/step - loss: 0.3638 - sparse_categorical_accuracy: 0.8969
Epoch 3/3
100/100 [==============================] - 0s 2ms/step - loss: 0.3205 - sparse_categorical_accuracy: 0.9052

<tensorflow.python.keras.callbacks.History at 0x7fdce0678710>

Using a validation dataset

You can pass a Dataset instance as the validation_data argument in fit:

model = get_compiled_model()

# Prepare the training dataset
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(64)

# Prepare the validation dataset
val_dataset = tf.data.Dataset.from_tensor_slices((x_val, y_val))
val_dataset = val_dataset.batch(64)

model.fit(train_dataset, epochs=3, validation_data=val_dataset)
Epoch 1/3
782/782 [==============================] - 2s 3ms/step - loss: 0.3233 - sparse_categorical_accuracy: 0.9089 - val_loss: 0.1881 - val_sparse_categorical_accuracy: 0.9426
Epoch 2/3
782/782 [==============================] - 2s 2ms/step - loss: 0.1556 - sparse_categorical_accuracy: 0.9535 - val_loss: 0.1320 - val_sparse_categorical_accuracy: 0.9607
Epoch 3/3
782/782 [==============================] - 2s 2ms/step - loss: 0.1135 - sparse_categorical_accuracy: 0.9654 - val_loss: 0.1266 - val_sparse_categorical_accuracy: 0.9615

<tensorflow.python.keras.callbacks.History at 0x7fdce04fe2e8>

At the end of each epoch, the model will iterate over the validation Dataset and compute the validation loss and validation metrics.

If you want to run validation only on a specific number of batches from this Dataset, you can pass the validation_steps argument, which specifies how many validation steps the model should run with the validation Dataset before interrupting validation and moving on to the next epoch:

model = get_compiled_model()

# Prepare the training dataset
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(64)

# Prepare the validation dataset
val_dataset = tf.data.Dataset.from_tensor_slices((x_val, y_val))
val_dataset = val_dataset.batch(64)

model.fit(train_dataset, epochs=3,
          # Only run validation using the first 10 batches of the dataset
          # using the `validation_steps` argument
          validation_data=val_dataset, validation_steps=10)
Epoch 1/3
782/782 [==============================] - 2s 2ms/step - loss: 0.3448 - sparse_categorical_accuracy: 0.9031 - val_loss: 0.2954 - val_sparse_categorical_accuracy: 0.9234
Epoch 2/3
782/782 [==============================] - 2s 2ms/step - loss: 0.1584 - sparse_categorical_accuracy: 0.9527 - val_loss: 0.2611 - val_sparse_categorical_accuracy: 0.9219
Epoch 3/3
782/782 [==============================] - 2s 2ms/step - loss: 0.1148 - sparse_categorical_accuracy: 0.9658 - val_loss: 0.2105 - val_sparse_categorical_accuracy: 0.9422

<tensorflow.python.keras.callbacks.History at 0x7fdce042b748>

Note that the validation Dataset will be reset after each use (so that you will always be evaluating on the same samples from epoch to epoch).

The argument validation_split (generating a holdout set from the training data) is not supported when training from Dataset objects, since this features requires the ability to index the samples of the datasets, which is not possible in general with the Dataset API.

Other input formats supported

Besides Numpy arrays and TensorFlow Datasets, it's possible to train a Keras model using Pandas dataframes, or from Python generators that yield batches.

In general, we recommend that you use Numpy input data if your data is small and fits in memory, and Datasets otherwise.

Using sample weighting and class weighting

Besides input data and target data, it is possible to pass sample weights or class weights to a model when using fit:

  • When training from Numpy data: via the sample_weight and class_weight arguments.
  • When training from Datasets: by having the Dataset return a tuple (input_batch, target_batch, sample_weight_batch) .

A "sample weights" array is an array of numbers that specify how much weight each sample in a batch should have in computing the total loss. It is commonly used in imbalanced classification problems (the idea being to give more weight to rarely-seen classes). When the weights used are ones and zeros, the array can be used as a mask for the loss function (entirely discarding the contribution of certain samples to the total loss).

A "class weights" dict is a more specific instance of the same concept: it maps class indices to the sample weight that should be used for samples belonging to this class. For instance, if class "0" is twice less represented than class "1" in your data, you could use class_weight={0: 1., 1: 0.5}.

Here's a Numpy example where we use class weights or sample weights to give more importance to the correct classification of class #5 (which is the digit "5" in the MNIST dataset).

import numpy as np

class_weight = {0: 1., 1: 1., 2: 1., 3: 1., 4: 1.,
                # Set weight "2" for class "5",
                # making this class 2x more important
                5: 2.,
                6: 1., 7: 1., 8: 1., 9: 1.}
print('Fit with class weight')
model.fit(x_train, y_train,
          class_weight=class_weight,
          batch_size=64,
          epochs=4)
Fit with class weight
Epoch 1/4
782/782 [==============================] - 2s 2ms/step - loss: 0.1020 - sparse_categorical_accuracy: 0.9714
Epoch 2/4
782/782 [==============================] - 2s 2ms/step - loss: 0.0839 - sparse_categorical_accuracy: 0.9760
Epoch 3/4
782/782 [==============================] - 2s 2ms/step - loss: 0.0739 - sparse_categorical_accuracy: 0.9793
Epoch 4/4
782/782 [==============================] - 2s 2ms/step - loss: 0.0633 - sparse_categorical_accuracy: 0.9822

<tensorflow.python.keras.callbacks.History at 0x7fdce02dc390>
# Here's the same example using `sample_weight` instead:
sample_weight = np.ones(shape=(len(y_train),))
sample_weight[y_train == 5] = 2.
print('\nFit with sample weight')

model = get_compiled_model()
model.fit(x_train, y_train,
          sample_weight=sample_weight,
          batch_size=64,
          epochs=4)

Fit with sample weight
Epoch 1/4
782/782 [==============================] - 2s 2ms/step - loss: 0.3736 - sparse_categorical_accuracy: 0.9021
Epoch 2/4
782/782 [==============================] - 2s 2ms/step - loss: 0.1751 - sparse_categorical_accuracy: 0.9513
Epoch 3/4
782/782 [==============================] - 2s 2ms/step - loss: 0.1293 - sparse_categorical_accuracy: 0.9643
Epoch 4/4
782/782 [==============================] - 2s 2ms/step - loss: 0.1020 - sparse_categorical_accuracy: 0.9706

<tensorflow.python.keras.callbacks.History at 0x7fdce0238dd8>

Here's a matching Dataset example:

sample_weight = np.ones(shape=(len(y_train),))
sample_weight[y_train == 5] = 2.

# Create a Dataset that includes sample weights
# (3rd element in the return tuple).
train_dataset = tf.data.Dataset.from_tensor_slices(
    (x_train, y_train, sample_weight))

# Shuffle and slice the dataset.
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(64)

model = get_compiled_model()
model.fit(train_dataset, epochs=3)
Epoch 1/3
782/782 [==============================] - 2s 2ms/step - loss: 0.3552 - sparse_categorical_accuracy: 0.9079
Epoch 2/3
782/782 [==============================] - 2s 2ms/step - loss: 0.1673 - sparse_categorical_accuracy: 0.9544
Epoch 3/3
782/782 [==============================] - 2s 2ms/step - loss: 0.1241 - sparse_categorical_accuracy: 0.9653

<tensorflow.python.keras.callbacks.History at 0x7fdce0230550>

Passing data to multi-input, multi-output models

In the previous examples, we were considering a model with a single input (a tensor of shape (764,)) and a single output (a prediction tensor of shape (10,)). But what about models that have multiple inputs or outputs?

Consider the following model, which has an image input of shape (32, 32, 3) (that's (height, width, channels)) and a timeseries input of shape (None, 10) (that's (timesteps, features)). Our model will have two outputs computed from the combination of these inputs: a "score" (of shape (1,)) and a probability distribution over five classes (of shape (5,)).

from tensorflow import keras
from tensorflow.keras import layers

image_input = keras.Input(shape=(32, 32, 3), name='img_input')
timeseries_input = keras.Input(shape=(None, 10), name='ts_input')

x1 = layers.Conv2D(3, 3)(image_input)
x1 = layers.GlobalMaxPooling2D()(x1)

x2 = layers.Conv1D(3, 3)(timeseries_input)
x2 = layers.GlobalMaxPooling1D()(x2)

x = layers.concatenate([x1, x2])

score_output = layers.Dense(1, name='score_output')(x)
class_output = layers.Dense(5, name='class_output')(x)

model = keras.Model(inputs=[image_input, timeseries_input],
                    outputs=[score_output, class_output])

Let's plot this model, so you can clearly see what we're doing here (note that the shapes shown in the plot are batch shapes, rather than per-sample shapes).

keras.utils.plot_model(model, 'multi_input_and_output_model.png', show_shapes=True)

png

At compilation time, we can specify different losses to different outputs, by passing the loss functions as a list:

model.compile(
    optimizer=keras.optimizers.RMSprop(1e-3),
    loss=[keras.losses.MeanSquaredError(),
          keras.losses.CategoricalCrossentropy(from_logits=True)])

If we only passed a single loss function to the model, the same loss function would be applied to every output, which is not appropriate here.

Likewise for metrics:

model.compile(
    optimizer=keras.optimizers.RMSprop(1e-3),
    loss=[keras.losses.MeanSquaredError(),
          keras.losses.CategoricalCrossentropy(from_logits=True)],
    metrics=[[keras.metrics.MeanAbsolutePercentageError(),
              keras.metrics.MeanAbsoluteError()],
             [keras.metrics.CategoricalAccuracy()]])

Since we gave names to our output layers, we could also specify per-output losses and metrics via a dict:

model.compile(
    optimizer=keras.optimizers.RMSprop(1e-3),
    loss={'score_output': keras.losses.MeanSquaredError(),
          'class_output': keras.losses.CategoricalCrossentropy(from_logits=True)},
    metrics={'score_output': [keras.metrics.MeanAbsolutePercentageError(),
                              keras.metrics.MeanAbsoluteError()],
             'class_output': [keras.metrics.CategoricalAccuracy()]})

We recommend the use of explicit names and dicts if you have more than 2 outputs.

It's possible to give different weights to different output-specific losses (for instance, one might wish to privilege the "score" loss in our example, by giving to 2x the importance of the class loss), using the loss_weights argument:

model.compile(
    optimizer=keras.optimizers.RMSprop(1e-3),
    loss={'score_output': keras.losses.MeanSquaredError(),
          'class_output': keras.losses.CategoricalCrossentropy(from_logits=True)},
    metrics={'score_output': [keras.metrics.MeanAbsolutePercentageError(),
                              keras.metrics.MeanAbsoluteError()],
             'class_output': [keras.metrics.CategoricalAccuracy()]},
    loss_weights={'score_output': 2., 'class_output': 1.})

You could also chose not to compute a loss for certain outputs, if these outputs meant for prediction but not for training:

# List loss version
model.compile(
    optimizer=keras.optimizers.RMSprop(1e-3),
    loss=[None, keras.losses.CategoricalCrossentropy(from_logits=True)])

# Or dict loss version
model.compile(
    optimizer=keras.optimizers.RMSprop(1e-3),
    loss={'class_output':keras.losses.CategoricalCrossentropy(from_logits=True)})

Passing data to a multi-input or multi-output model in fit works in a similar way as specifying a loss function in compile: you can pass lists of Numpy arrays (with 1:1 mapping to the outputs that received a loss function) or dicts mapping output names to Numpy arrays of training data.

model.compile(
    optimizer=keras.optimizers.RMSprop(1e-3),
    loss=[keras.losses.MeanSquaredError(),
          keras.losses.CategoricalCrossentropy(from_logits=True)])

# Generate dummy Numpy data
img_data = np.random.random_sample(size=(100, 32, 32, 3))
ts_data = np.random.random_sample(size=(100, 20, 10))
score_targets = np.random.random_sample(size=(100, 1))
class_targets = np.random.random_sample(size=(100, 5))

# Fit on lists
model.fit([img_data, ts_data], [score_targets, class_targets],
          batch_size=32,
          epochs=3)

# Alternatively, fit on dicts
model.fit({'img_input': img_data, 'ts_input': ts_data},
          {'score_output': score_targets, 'class_output': class_targets},
          batch_size=32,
          epochs=3)
Epoch 1/3
4/4 [==============================] - 0s 9ms/step - loss: 5.7132 - score_output_loss: 1.4049 - class_output_loss: 4.3082
Epoch 2/3
4/4 [==============================] - 0s 3ms/step - loss: 5.1200 - score_output_loss: 0.8224 - class_output_loss: 4.2976
Epoch 3/3
4/4 [==============================] - 0s 3ms/step - loss: 4.8205 - score_output_loss: 0.5262 - class_output_loss: 4.2943
Epoch 1/3
4/4 [==============================] - 0s 3ms/step - loss: 4.6439 - score_output_loss: 0.3457 - class_output_loss: 4.2982
Epoch 2/3
4/4 [==============================] - 0s 3ms/step - loss: 4.5539 - score_output_loss: 0.2441 - class_output_loss: 4.3098
Epoch 3/3
4/4 [==============================] - 0s 3ms/step - loss: 4.4978 - score_output_loss: 0.1737 - class_output_loss: 4.3241

<tensorflow.python.keras.callbacks.History at 0x7fdcb0571c88>

Here's the Dataset use case: similarly as what we did for Numpy arrays, the Dataset should return a tuple of dicts.

train_dataset = tf.data.Dataset.from_tensor_slices(
    ({'img_input': img_data, 'ts_input': ts_data},
     {'score_output': score_targets, 'class_output': class_targets}))
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(64)

model.fit(train_dataset, epochs=3)
Epoch 1/3
2/2 [==============================] - 0s 13ms/step - loss: 4.4827 - score_output_loss: 0.1392 - class_output_loss: 4.3435
Epoch 2/3
2/2 [==============================] - 0s 2ms/step - loss: 4.4755 - score_output_loss: 0.1241 - class_output_loss: 4.3513
Epoch 3/3
2/2 [==============================] - 0s 3ms/step - loss: 4.4767 - score_output_loss: 0.1138 - class_output_loss: 4.3629

<tensorflow.python.keras.callbacks.History at 0x7fdcb06bdb38>

Using callbacks

Callbacks in Keras are objects that are called at different point during training (at the start of an epoch, at the end of a batch, at the end of an epoch, etc.) and which can be used to implement behaviors such as:

  • Doing validation at different points during training (beyond the built-in per-epoch validation)
  • Checkpointing the model at regular intervals or when it exceeds a certain accuracy threshold
  • Changing the learning rate of the model when training seems to be plateauing
  • Doing fine-tuning of the top layers when training seems to be plateauing
  • Sending email or instant message notifications when training ends or where a certain performance threshold is exceeded
  • Etc.

Callbacks can be passed as a list to your call to fit:

model = get_compiled_model()

callbacks = [
    keras.callbacks.EarlyStopping(
        # Stop training when `val_loss` is no longer improving
        monitor='val_loss',
        # "no longer improving" being defined as "no better than 1e-2 less"
        min_delta=1e-2,
        # "no longer improving" being further defined as "for at least 2 epochs"
        patience=2,
        verbose=1)
]
model.fit(x_train, y_train,
          epochs=20,
          batch_size=64,
          callbacks=callbacks,
          validation_split=0.2)
Epoch 1/20
625/625 [==============================] - 1s 2ms/step - loss: 0.3679 - sparse_categorical_accuracy: 0.8956 - val_loss: 0.2356 - val_sparse_categorical_accuracy: 0.9323
Epoch 2/20
625/625 [==============================] - 1s 2ms/step - loss: 0.1738 - sparse_categorical_accuracy: 0.9492 - val_loss: 0.1805 - val_sparse_categorical_accuracy: 0.9451
Epoch 3/20
625/625 [==============================] - 1s 2ms/step - loss: 0.1274 - sparse_categorical_accuracy: 0.9624 - val_loss: 0.1596 - val_sparse_categorical_accuracy: 0.9517
Epoch 4/20
625/625 [==============================] - 1s 2ms/step - loss: 0.1018 - sparse_categorical_accuracy: 0.9691 - val_loss: 0.1614 - val_sparse_categorical_accuracy: 0.9524
Epoch 5/20
625/625 [==============================] - 1s 2ms/step - loss: 0.0817 - sparse_categorical_accuracy: 0.9742 - val_loss: 0.1406 - val_sparse_categorical_accuracy: 0.9597
Epoch 6/20
625/625 [==============================] - 1s 2ms/step - loss: 0.0691 - sparse_categorical_accuracy: 0.9789 - val_loss: 0.1359 - val_sparse_categorical_accuracy: 0.9615
Epoch 7/20
625/625 [==============================] - 1s 2ms/step - loss: 0.0584 - sparse_categorical_accuracy: 0.9834 - val_loss: 0.1490 - val_sparse_categorical_accuracy: 0.9581
Epoch 00007: early stopping

<tensorflow.python.keras.callbacks.History at 0x7fdd557112b0>

Many built-in callbacks are available

  • ModelCheckpoint: Periodically save the model.
  • EarlyStopping: Stop training when training is no longer improving the validation metrics.
  • TensorBoard: periodically write model logs that can be visualized in TensorBoard (more details in the section "Visualization").
  • CSVLogger: streams loss and metrics data to a CSV file.
  • etc.

Writing your own callback

You can create a custom callback by extending the base class keras.callbacks.Callback. A callback has access to its associated model through the class property self.model.

Here's a simple example saving a list of per-batch loss values during training:

class LossHistory(keras.callbacks.Callback):

    def on_train_begin(self, logs):
        self.losses = []

    def on_batch_end(self, batch, logs):
        self.losses.append(logs.get('loss'))

Checkpointing models

When you're training model on relatively large datasets, it's crucial to save checkpoints of your model at frequent intervals.

The easiest way to achieve this is with the ModelCheckpoint callback:

model = get_compiled_model()

callbacks = [
    keras.callbacks.ModelCheckpoint(
        filepath='mymodel_{epoch}',
        # Path where to save the model
        # The two parameters below mean that we will overwrite
        # the current checkpoint if and only if
        # the `val_loss` score has improved.
        save_best_only=True,
        monitor='val_loss',
        verbose=1)
]
model.fit(x_train, y_train,
          epochs=3,
          batch_size=64,
          callbacks=callbacks,
          validation_split=0.2)
Epoch 1/3
598/625 [===========================>..] - ETA: 0s - loss: 0.3812 - sparse_categorical_accuracy: 0.8938
Epoch 00001: val_loss improved from inf to 0.25517, saving model to mymodel_1
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/ops/resource_variable_ops.py:1817: calling BaseResourceVariable.__init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version.
Instructions for updating:
If using Keras pass *_constraint arguments to layers.
INFO:tensorflow:Assets written to: mymodel_1/assets
625/625 [==============================] - 2s 3ms/step - loss: 0.3754 - sparse_categorical_accuracy: 0.8958 - val_loss: 0.2552 - val_sparse_categorical_accuracy: 0.9234
Epoch 2/3
624/625 [============================>.] - ETA: 0s - loss: 0.1679 - sparse_categorical_accuracy: 0.9505
Epoch 00002: val_loss improved from 0.25517 to 0.17379, saving model to mymodel_2
INFO:tensorflow:Assets written to: mymodel_2/assets
625/625 [==============================] - 2s 3ms/step - loss: 0.1680 - sparse_categorical_accuracy: 0.9505 - val_loss: 0.1738 - val_sparse_categorical_accuracy: 0.9463
Epoch 3/3
624/625 [============================>.] - ETA: 0s - loss: 0.1213 - sparse_categorical_accuracy: 0.9628
Epoch 00003: val_loss improved from 0.17379 to 0.15227, saving model to mymodel_3
INFO:tensorflow:Assets written to: mymodel_3/assets
625/625 [==============================] - 2s 3ms/step - loss: 0.1212 - sparse_categorical_accuracy: 0.9628 - val_loss: 0.1523 - val_sparse_categorical_accuracy: 0.9546

<tensorflow.python.keras.callbacks.History at 0x7fdd48145630>

You call also write your own callback for saving and restoring models.

For a complete guide on serialization and saving, see Guide to Saving and Serializing Models.

Using learning rate schedules

A common pattern when training deep learning models is to gradually reduce the learning as training progresses. This is generally known as "learning rate decay".

The learning decay schedule could be static (fixed in advance, as a function of the current epoch or the current batch index), or dynamic (responding to the current behavior of the model, in particular the validation loss).

Passing a schedule to an optimizer

You can easily use a static learning rate decay schedule by passing a schedule object as the learning_rate argument in your optimizer:

initial_learning_rate = 0.1
lr_schedule = keras.optimizers.schedules.ExponentialDecay(
    initial_learning_rate,
    decay_steps=100000,
    decay_rate=0.96,
    staircase=True)

optimizer = keras.optimizers.RMSprop(learning_rate=lr_schedule)

Several built-in schedules are available: ExponentialDecay, PiecewiseConstantDecay, PolynomialDecay, and InverseTimeDecay.

Using callbacks to implement a dynamic learning rate schedule

A dynamic learning rate schedule (for instance, decreasing the learning rate when the validation loss is no longer improving) cannot be achieved with these schedule objects since the optimizer does not have access to validation metrics.

However, callbacks do have access to all metrics, including validation metrics! You can thus achieve this pattern by using a callback that modifies the current learning rate on the optimizer. In fact, this is even built-in as the ReduceLROnPlateau callback.

Visualizing loss and metrics during training

The best way to keep an eye on your model during training is to use TensorBoard, a browser-based application that you can run locally that provides you with:

  • Live plots of the loss and metrics for training and evaluation
  • (optionally) Visualizations of the histograms of your layer activations
  • (optionally) 3D visualizations of the embedding spaces learned by your Embedding layers

If you have installed TensorFlow with pip, you should be able to launch TensorBoard from the command line:

tensorboard --logdir=/full_path_to_your_logs

Using the TensorBoard callback

The easiest way to use TensorBoard with a Keras model and the fit method is the TensorBoard callback.

In the simplest case, just specify where you want the callback to write logs, and you're good to go:

tensorboard_cbk = keras.callbacks.TensorBoard(log_dir='/full_path_to_your_logs')
model.fit(dataset, epochs=10, callbacks=[tensorboard_cbk])

The TensorBoard callback has many useful options, including whether to log embeddings, histograms, and how often to write logs:

keras.callbacks.TensorBoard(
  log_dir='/full_path_to_your_logs',
  histogram_freq=0,  # How often to log histogram visualizations
  embeddings_freq=0,  # How often to log embedding visualizations
  update_freq='epoch')  # How often to write logs (default: once per epoch)

Part II: Writing your own training and evaluation loops from scratch

If you want lower-level over your training and evaluation loops than what fit() and evaluate() provide, you should write your own. It's actually pretty simple! But you should be ready to have a lot more debugging to do on your own.

Using the GradientTape: a first end-to-end example

Calling a model inside a GradientTape scope enables you to retrieve the gradients of the trainable weights of the layer with respect to a loss value. Using an optimizer instance, you can use these gradients to update these variables (which you can retrieve using model.trainable_weights).

Let's reuse our initial MNIST model from Part I, and let's train it using mini-batch gradient with a training loop.

# Get the model.
inputs = keras.Input(shape=(784,), name='digits')
x = layers.Dense(64, activation='relu', name='dense_1')(inputs)
x = layers.Dense(64, activation='relu', name='dense_2')(x)
outputs = layers.Dense(10, name='predictions')(x)
model = keras.Model(inputs=inputs, outputs=outputs)

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

# Prepare the training dataset.
batch_size = 64
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(batch_size)

Run a training loop for a few epochs:

epochs = 3
for epoch in range(epochs):
  print('Start of epoch %d' % (epoch,))

  # Iterate over the batches of the dataset.
  for step, (x_batch_train, y_batch_train) in enumerate(train_dataset):

    # Open a GradientTape to record the operations run
    # during the forward pass, which enables autodifferentiation.
    with tf.GradientTape() as tape:

      # Run the forward pass of the layer.
      # The operations that the layer applies
      # to its inputs are going to be recorded
      # on the GradientTape.
      logits = model(x_batch_train, training=True)  # Logits for this minibatch

      # Compute the loss value for this minibatch.
      loss_value = loss_fn(y_batch_train, logits)

    # Use the gradient tape to automatically retrieve
    # the gradients of the trainable variables with respect to the loss.
    grads = tape.gradient(loss_value, model.trainable_weights)

    # Run one step of gradient descent by updating
    # the value of the variables to minimize the loss.
    optimizer.apply_gradients(zip(grads, model.trainable_weights))

    # Log every 200 batches.
    if step % 200 == 0:
        print('Training loss (for one batch) at step %s: %s' % (step, float(loss_value)))
        print('Seen so far: %s samples' % ((step + 1) * 64))
Start of epoch 0
Training loss (for one batch) at step 0: 2.3342301845550537
Seen so far: 64 samples
Training loss (for one batch) at step 200: 2.2239229679107666
Seen so far: 12864 samples
Training loss (for one batch) at step 400: 2.1380720138549805
Seen so far: 25664 samples
Training loss (for one batch) at step 600: 2.064563751220703
Seen so far: 38464 samples
Start of epoch 1
Training loss (for one batch) at step 0: 2.061167001724243
Seen so far: 64 samples
Training loss (for one batch) at step 200: 1.9205878973007202
Seen so far: 12864 samples
Training loss (for one batch) at step 400: 1.8126640319824219
Seen so far: 25664 samples
Training loss (for one batch) at step 600: 1.7128626108169556
Seen so far: 38464 samples
Start of epoch 2
Training loss (for one batch) at step 0: 1.4824482202529907
Seen so far: 64 samples
Training loss (for one batch) at step 200: 1.4608227014541626
Seen so far: 12864 samples
Training loss (for one batch) at step 400: 1.4236290454864502
Seen so far: 25664 samples
Training loss (for one batch) at step 600: 1.22068190574646
Seen so far: 38464 samples

Low-level handling of metrics

Let's add metrics to the mix. You can readily reuse the built-in metrics (or custom ones you wrote) in such training loops written from scratch. Here's the flow:

  • Instantiate the metric at the start of the loop
  • Call metric.update_state() after each batch
  • Call metric.result() when you need to display the current value of the metric
  • Call metric.reset_states() when you need to clear the state of the metric (typically at the end of an epoch)

Let's use this knowledge to compute SparseCategoricalAccuracy on validation data at the end of each epoch:

# Get model
inputs = keras.Input(shape=(784,), name='digits')
x = layers.Dense(64, activation='relu', name='dense_1')(inputs)
x = layers.Dense(64, activation='relu', name='dense_2')(x)
outputs = layers.Dense(10, name='predictions')(x)
model = keras.Model(inputs=inputs, outputs=outputs)

# Instantiate an optimizer to train the model.
optimizer = keras.optimizers.SGD(learning_rate=1e-3)
# Instantiate a loss function.
loss_fn = keras.losses.SparseCategoricalCrossentropy(from_logits=True)

# Prepare the metrics.
train_acc_metric = keras.metrics.SparseCategoricalAccuracy()
val_acc_metric = keras.metrics.SparseCategoricalAccuracy()

# Prepare the training dataset.
batch_size = 64
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(batch_size)

# Prepare the validation dataset.
val_dataset = tf.data.Dataset.from_tensor_slices((x_val, y_val))
val_dataset = val_dataset.batch(64)

Run a training loop for a few epochs:

epochs = 3
for epoch in range(epochs):
  print('Start of epoch %d' % (epoch,))

  # Iterate over the batches of the dataset.
  for step, (x_batch_train, y_batch_train) in enumerate(train_dataset):
    with tf.GradientTape() as tape:
      logits = model(x_batch_train)
      loss_value = loss_fn(y_batch_train, logits)
    grads = tape.gradient(loss_value, model.trainable_weights)
    optimizer.apply_gradients(zip(grads, model.trainable_weights))

    # Update training metric.
    train_acc_metric(y_batch_train, logits)

    # Log every 200 batches.
    if step % 200 == 0:
        print('Training loss (for one batch) at step %s: %s' % (step, float(loss_value)))
        print('Seen so far: %s samples' % ((step + 1) * 64))

  # Display metrics at the end of each epoch.
  train_acc = train_acc_metric.result()
  print('Training acc over epoch: %s' % (float(train_acc),))
  # Reset training metrics at the end of each epoch
  train_acc_metric.reset_states()

  # Run a validation loop at the end of each epoch.
  for x_batch_val, y_batch_val in val_dataset:
    val_logits = model(x_batch_val)
    # Update val metrics
    val_acc_metric(y_batch_val, val_logits)
  val_acc = val_acc_metric.result()
  val_acc_metric.reset_states()
  print('Validation acc: %s' % (float(val_acc),))
Start of epoch 0
Training loss (for one batch) at step 0: 2.288959503173828
Seen so far: 64 samples
Training loss (for one batch) at step 200: 2.2390151023864746
Seen so far: 12864 samples
Training loss (for one batch) at step 400: 2.1368484497070312
Seen so far: 25664 samples
Training loss (for one batch) at step 600: 2.0690793991088867
Seen so far: 38464 samples
Training acc over epoch: 0.29109999537467957
Validation acc: 0.46389999985694885
Start of epoch 1
Training loss (for one batch) at step 0: 1.9826419353485107
Seen so far: 64 samples
Training loss (for one batch) at step 200: 1.9439365863800049
Seen so far: 12864 samples
Training loss (for one batch) at step 400: 1.8446139097213745
Seen so far: 25664 samples
Training loss (for one batch) at step 600: 1.7223378419876099
Seen so far: 38464 samples
Training acc over epoch: 0.5405600070953369
Validation acc: 0.6363000273704529
Start of epoch 2
Training loss (for one batch) at step 0: 1.6395390033721924
Seen so far: 64 samples
Training loss (for one batch) at step 200: 1.5451477766036987
Seen so far: 12864 samples
Training loss (for one batch) at step 400: 1.443529725074768
Seen so far: 25664 samples
Training loss (for one batch) at step 600: 1.3105210065841675
Seen so far: 38464 samples
Training acc over epoch: 0.6864799857139587
Validation acc: 0.7537000179290771

Low-level handling of extra losses

You saw in the previous section that it is possible for regularization losses to be added by a layer by calling self.add_loss(value) in the call method.

In the general case, you will want to take these losses into account in your training loops (unless you've written the model yourself and you already know that it creates no such losses).

Recall this example from the previous section, featuring a layer that creates a regularization loss:

class ActivityRegularizationLayer(layers.Layer):

  def call(self, inputs):
    self.add_loss(1e-2 * tf.reduce_sum(inputs))
    return inputs

inputs = keras.Input(shape=(784,), name='digits')
x = layers.Dense(64, activation='relu', name='dense_1')(inputs)
# Insert activity regularization as a layer
x = ActivityRegularizationLayer()(x)
x = layers.Dense(64, activation='relu', name='dense_2')(x)
outputs = layers.Dense(10, name='predictions')(x)

model = keras.Model(inputs=inputs, outputs=outputs)

When you call a model, like this:

logits = model(x_train)

the losses it creates during the forward pass are added to the model.losses attribute:

logits = model(x_train[:64])
print(model.losses)
[<tf.Tensor: shape=(), dtype=float32, numpy=6.3990765>]

The tracked losses are first cleared at the start of the model __call__, so you will only see the losses created during this one forward pass. For instance, calling the model repeatedly and then querying losses only displays the latest losses, created during the last call:

logits = model(x_train[:64])
logits = model(x_train[64: 128])
logits = model(x_train[128: 192])
print(model.losses)
[<tf.Tensor: shape=(), dtype=float32, numpy=6.4714365>]

To take these losses into account during training, all you have to do is to modify your training loop to add sum(model.losses) to your total loss:

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

epochs = 3
for epoch in range(epochs):
  print('Start of epoch %d' % (epoch,))

  for step, (x_batch_train, y_batch_train) in enumerate(train_dataset):
    with tf.GradientTape() as tape:
      logits = model(x_batch_train)
      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))

    # Log every 200 batches.
    if step % 200 == 0:
        print('Training loss (for one batch) at step %s: %s' % (step, float(loss_value)))
        print('Seen so far: %s samples' % ((step + 1) * 64))
Start of epoch 0
Training loss (for one batch) at step 0: 8.943479537963867
Seen so far: 64 samples
Training loss (for one batch) at step 200: 2.503769636154175
Seen so far: 12864 samples
Training loss (for one batch) at step 400: 2.3875625133514404
Seen so far: 25664 samples
Training loss (for one batch) at step 600: 2.344862937927246
Seen so far: 38464 samples
Start of epoch 1
Training loss (for one batch) at step 0: 2.3388476371765137
Seen so far: 64 samples
Training loss (for one batch) at step 200: 2.33182430267334
Seen so far: 12864 samples
Training loss (for one batch) at step 400: 2.3141746520996094
Seen so far: 25664 samples
Training loss (for one batch) at step 600: 2.36799693107605
Seen so far: 38464 samples
Start of epoch 2
Training loss (for one batch) at step 0: 2.3305697441101074
Seen so far: 64 samples
Training loss (for one batch) at step 200: 2.3220674991607666
Seen so far: 12864 samples
Training loss (for one batch) at step 400: 2.319568634033203
Seen so far: 25664 samples
Training loss (for one batch) at step 600: 2.3073623180389404
Seen so far: 38464 samples

That was the last piece of the puzzle! You've reached the end of this guide.

Now you know everything there is to know about using built-in training loops and writing your own from scratch.