Skip to content

Latest commit

 

History

History
1055 lines (724 loc) · 40.8 KB

faq.md

File metadata and controls

1055 lines (724 loc) · 40.8 KB

Keras FAQ

A list of frequently Asked Keras Questions.

General questions

Training-related questions

Modeling-related questions

General questions

How can I train a Keras model on multiple GPUs (on a single machine)?

There are two ways to run a single model on multiple GPUs: data parallelism and device parallelism. In most cases, what you need is most likely data parallelism.

1) Data parallelism

Data parallelism consists in replicating the target model once on each device, and using each replica to process a different fraction of the input data.

The best way to do data parallelism with Keras models is to use the tf.distribute API. Make sure to read our guide about using tf.distribute with Keras.

The gist of it is the following:

a) instantiate a "distribution strategy" object, e.g. MirroredStrategy (which replicates your model on each available device and keeps the state of each model in sync):

strategy = tf.distribute.MirroredStrategy()

b) Create your model and compile it under the strategy's scope:

with strategy.scope():
    # This could be any kind of model -- Functional, subclass...
    model = tf.keras.Sequential([
        tf.keras.layers.Conv2D(32, 3, activation='relu', input_shape=(28, 28, 1)),
        tf.keras.layers.GlobalMaxPooling2D(),
        tf.keras.layers.Dense(10)
    ])
    model.compile(loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
                  optimizer=tf.keras.optimizers.Adam(),
                  metrics=[tf.keras.metrics.SparseCategoricalAccuracy()])

Note that it's important that all state variable creation should happen under the scope. So in case you create any additional variables, do that under the scope.

c) Call fit() with a tf.data.Dataset object as input. Distribution is broadly compatible with all callbacks, including custom callbacks. Note that this call does not need to be under the strategy scope, since it doesn't create new variables.

model.fit(train_dataset, epochs=12, callbacks=callbacks)

2) Model parallelism

Model parallelism consists in running different parts of a same model on different devices. It works best for models that have a parallel architecture, e.g. a model with two branches.

This can be achieved by using TensorFlow device scopes. Here is a quick example:

# Model where a shared LSTM is used to encode two different sequences in parallel
input_a = keras.Input(shape=(140, 256))
input_b = keras.Input(shape=(140, 256))

shared_lstm = keras.layers.LSTM(64)

# Process the first sequence on one GPU
with tf.device_scope('/gpu:0'):
    encoded_a = shared_lstm(input_a)
# Process the next sequence on another GPU
with tf.device_scope('/gpu:1'):
    encoded_b = shared_lstm(input_b)

# Concatenate results on CPU
with tf.device_scope('/cpu:0'):
    merged_vector = keras.layers.concatenate(
        [encoded_a, encoded_b], axis=-1)

How can I distribute training across multiple machines?

Like for single-machine parallelism, the best way to do distributed training with Keras is to use the tf.distribute API, in particular MultiWorkerMirroredStrategy. Make sure to read our guide about using tf.distribute with Keras.

Distributed training is somewhat more involved than single-machine multi-device training. Roughly, you will need to launch a remote cluster of machines, then run your code on a "chief" machine that holds a TF_CONFIG environment variable that specifies how to communicate with the other machines in the cluster. From there, the workflow is similar to using single-machine multi-GPU training, with the main difference being that you will use MultiWorkerMirroredStrategy as your distribution strategy.

Importantly, you should:

  • Make sure your dataset is so configured that all workers in the cluster are able to efficiently pull data from it (e.g. if your custer in on GCP, it's a good idea to host your data on GCS).
  • Make sure your training is fault-tolerant (e.g. by configuring a ModelCheckpoint callback).

How can I train a Keras model on TPU?

TPUs are a fast & efficient hardware accelerator for deep learning that is publicly available on Google Cloud. You can use TPUs via Colab, AI Platform (ML Engine), and Deep Learning VMs (provided the TPU_NAME environment variable is set on the VM).

Make sure to read the TPU usage guide first. Here's a quick summary:

After connecting to a TPU runtime (e.g. by selecting the TPU runtime in Colab), you will need to detect your TPU using a TPUClusterResolver, which automatically detects a linked TPU on all supported platforms:

tpu = tf.distribute.cluster_resolver.TPUClusterResolver()  # TPU detection
print('Running on TPU: ', tpu.cluster_spec().as_dict()['worker'])

tf.config.experimental_connect_to_cluster(tpu)
tf.tpu.experimental.initialize_tpu_system(tpu)
strategy = tf.distribute.experimental.TPUStrategy(tpu)
print('Replicas: ', strategy.num_replicas_in_sync)

with strategy.scope():
    # Create your model here.
    ...

After the initial setup, the workflow is similar to using single-machine multi-GPU training, with the main difference being that you will use TPUStrategy as your distribution strategy.

Importantly, you should:

  • Make sure your dataset yields batches with a fixed static shape. A TPU graph can only process inputs with a constant shape.
  • Make sure you are able to read your data fast enough to keep the TPU utilized. Using the TFRecord format to store your data may be a good idea.
  • Consider running multiple steps of gradient descent per graph execution in order to keep the TPU utilized. You can do this via the experimental_steps_per_execution argument compile(). It will yield a significant speed up for small models.

Where is the Keras configuration file stored?

The default directory where all Keras data is stored is:

$HOME/.keras/

For instance, for me, on a MacBook Pro, it's /Users/fchollet/.keras/.

Note that Windows users should replace $HOME with %USERPROFILE%.

In case Keras cannot create the above directory (e.g. due to permission issues), /tmp/.keras/ is used as a backup.

The Keras configuration file is a JSON file stored at $HOME/.keras/keras.json. The default configuration file looks like this:

{
    "image_data_format": "channels_last",
    "epsilon": 1e-07,
    "floatx": "float32",
    "backend": "tensorflow"
}

It contains the following fields:

  • The image data format to be used as default by image processing layers and utilities (either channels_last or channels_first).
  • The epsilon numerical fuzz factor to be used to prevent division by zero in some operations.
  • The default float data type.
  • The default backend. This is legacy; nowadays there is only TensorFlow.

Likewise, cached dataset files, such as those downloaded with get_file(), are stored by default in $HOME/.keras/datasets/, and cached model weights files from Keras Applications are stored by default in $HOME/.keras/models/.

How to do hyperparameter tuning with Keras?

We recommend using Keras Tuner.

How can I obtain reproducible results using Keras during development?

During development of a model, sometimes it is useful to be able to obtain reproducible results from run to run in order to determine if a change in performance is due to an actual model or data modification, or merely a result of a new random seed.

First, you need to set the PYTHONHASHSEED environment variable to 0 before the program starts (not within the program itself). This is necessary in Python 3.2.3 onwards to have reproducible behavior for certain hash-based operations (e.g., the item order in a set or a dict, see Python's documentation or issue #2280 for further details). One way to set the environment variable is when starting python like this:

$ cat test_hash.py
print(hash("keras"))
$ python3 test_hash.py                  # non-reproducible hash (Python 3.2.3+)
8127205062320133199
$ python3 test_hash.py                  # non-reproducible hash (Python 3.2.3+)
3204480642156461591
$ PYTHONHASHSEED=0 python3 test_hash.py # reproducible hash
4883664951434749476
$ PYTHONHASHSEED=0 python3 test_hash.py # reproducible hash
4883664951434749476
```endshell

Moreover, whenrunning on a GPU, some operations have non-deterministic outputs, in particular `tf.reduce_sum()`. This is due to the fact that GPUs run many operations in parallel, so the order of execution is not always guaranteed. Due to the limited precision of floats, even adding several numbers together may give slightly different results depending on the order in which you add them. You can try to avoid the non-deterministic operations, but some may be created automatically by TensorFlow to compute the gradients, so it is much simpler to just run the code on the CPU. For this, you can set the `CUDA_VISIBLE_DEVICES` environment variable to an empty string, for example:

```shell
$ CUDA_VISIBLE_DEVICES="" PYTHONHASHSEED=0 python your_program.py
```endshell

The below snippet of code provides an example of how to obtain reproducible results:

```python
import numpy as np
import tensorflow as tf
import random as python_random

# The below is necessary for starting Numpy generated random numbers
# in a well-defined initial state.
np.random.seed(123)

# The below is necessary for starting core Python generated random numbers
# in a well-defined state.
python_random.seed(123)

# The below set_seed() will make random number generation
# in the TensorFlow backend have a well-defined initial state.
# For further details, see:
# https://www.tensorflow.org/api_docs/python/tf/random/set_seed
tf.random.set_seed(1234)

# Rest of code follows ...

Note that you don't have to set seeds for individual initializers in your code if you do the steps above, because their seeds are determined by the combination of the seeds set above.

What are my options for saving models?

Note: it is not recommended to use pickle or cPickle to save a Keras model.

1) Whole-model saving (configuration + weights)

Whole-model saving means creating a file that will contain:

  • the architecture of the model, allowing to re-create the model
  • the weights of the model
  • the training configuration (loss, optimizer)
  • the state of the optimizer, allowing to resume training exactly where you left off.

The default and recommend format to use is the TensorFlow SavedModel format. In TensorFlow 2.0 and higher, you can just do: model.save(your_file_path).

For explicitness, you can also use model.save(your_file_path, save_format='tf').

Keras still supports its original HDF5-based saving format. To save a model in HDF5 format, use model.save(your_file_path, save_format='h5'). Note that this option is automatically used if your_file_path ends in .h5 or .keras. Please also see How can I install HDF5 or h5py to save my models? for instructions on how to install h5py.

After saving a model in either format, you can reinstantiate it via model = keras.models.load_model(your_file_path).

Example:

from tensorflow.keras.models import load_model

model.save('my_model')  # creates a HDF5 file 'my_model.h5'
del model  # deletes the existing model

# returns a compiled model
# identical to the previous one
model = load_model('my_model')

2) Weights-only saving

If you need to save the weights of a model, you can do so in HDF5 with the code below:

model.save_weights('my_model_weights.h5')

Assuming you have code for instantiating your model, you can then load the weights you saved into a model with the same architecture:

model.load_weights('my_model_weights.h5')

If you need to load the weights into a different architecture (with some layers in common), for instance for fine-tuning or transfer-learning, you can load them by layer name:

model.load_weights('my_model_weights.h5', by_name=True)

Example:

"""
Assuming the original model looks like this:

model = Sequential()
model.add(Dense(2, input_dim=3, name='dense_1'))
model.add(Dense(3, name='dense_2'))
...
model.save_weights(fname)
"""

# new model
model = Sequential()
model.add(Dense(2, input_dim=3, name='dense_1'))  # will be loaded
model.add(Dense(10, name='new_dense'))  # will not be loaded

# load weights from first model; will only affect the first layer, dense_1.
model.load_weights(fname, by_name=True)

Please also see How can I install HDF5 or h5py to save my models? for instructions on how to install h5py.

3) Configuration-only saving (serialization)

If you only need to save the architecture of a model, and not its weights or its training configuration, you can do:

# save as JSON
json_string = model.to_json()

The generated JSON file is human-readable and can be manually edited if needed.

You can then build a fresh model from this data:

# model reconstruction from JSON:
from tensorflow.keras.models import model_from_json
model = model_from_json(json_string)

4) Handling custom layers (or other custom objects) in saved models

If the model you want to load includes custom layers or other custom classes or functions, you can pass them to the loading mechanism via the custom_objects argument:

from tensorflow.keras.models import load_model
# Assuming your model includes instance of an "AttentionLayer" class
model = load_model('my_model.h5', custom_objects={'AttentionLayer': AttentionLayer})

Alternatively, you can use a custom object scope:

from tensorflow.keras.utils import CustomObjectScope

with CustomObjectScope({'AttentionLayer': AttentionLayer}):
    model = load_model('my_model.h5')

Custom objects handling works the same way for load_model & model_from_json:

from tensorflow.keras.models import model_from_json
model = model_from_json(json_string, custom_objects={'AttentionLayer': AttentionLayer})

How can I install HDF5 or h5py to save my models?

In order to save your Keras models as HDF5 files, Keras uses the h5py Python package. It is a dependency of Keras and should be installed by default. On Debian-based distributions, you will have to additionally install libhdf5:

``` sudo apt-get install libhdf5-serial-dev ```

If you are unsure if h5py is installed you can open a Python shell and load the module via

import h5py

If it imports without error it is installed, otherwise you can find detailed installation instructions here.

How should I cite Keras?

Please cite Keras in your publications if it helps your research. Here is an example BibTeX entry:

@misc{chollet2015keras,
  title={Keras},
  author={Chollet, Fran\c{c}ois and others},
  year={2015},
  howpublished={\url{https://keras.io}},
}

Training-related questions

What do "sample", "batch", and "epoch" mean?

Below are some common definitions that are necessary to know and understand to correctly utilize Keras fit():

  • Sample: one element of a dataset. For instance, one image is a sample in a convolutional network. One audio snippet is a sample for a speech recognition model.

  • Batch: a set of N samples. The samples in a batch are processed independently, in parallel. If training, a batch results in only one update to the model. A batch generally approximates the distribution of the input data better than a single input. The larger the batch, the better the approximation; however, it is also true that the batch will take longer to process and will still result in only one update. For inference (evaluate/predict), it is recommended to pick a batch size that is as large as you can afford without going out of memory (since larger batches will usually result in faster evaluation/prediction).

  • Epoch: an arbitrary cutoff, generally defined as "one pass over the entire dataset", used to separate training into distinct phases, which is useful for logging and periodic evaluation. When using validation_data or validation_split with the fit method of Keras models, evaluation will be run at the end of every epoch. Within Keras, there is the ability to add callbacks specifically designed to be run at the end of an epoch. Examples of these are learning rate changes and model checkpointing (saving).

Why is my training loss much higher than my testing loss?

A Keras model has two modes: training and testing. Regularization mechanisms, such as Dropout and L1/L2 weight regularization, are turned off at testing time. They are reflected in the training time loss but not in the test time loss.

Besides, the training loss that Keras displays is the average of the losses for each batch of training data, over the current epoch. Because your model is changing over time, the loss over the first batches of an epoch is generally higher than over the last batches. This can bring the epoch-wise average down. On the other hand, the testing loss for an epoch is computed using the model as it is at the end of the epoch, resulting in a lower loss.

How can I use Keras with datasets that don't fit in memory?

You should use the tf.data API to create tf.data.Dataset objects -- an abstraction over a data pipeline that can pull data from local disk, from a distribtued filesystem, from GCS, etc., as well as efficiently apply various data transformations.

For instance, the utility tf.keras.preprocessing.image_dataset_from_directory will create a dataset that reads image data from a local directory. Likewise, the utility tf.keras.preprocessing.text_dataset_from_directory will create a dataset that reads text files from a local directory.

Dataset objects can be directly passed to fit(), or can be iterated over in a custom low-level training loop.

model.fit(dataset, epochs=10, validation_data=val_dataset)

How can I regularly save Keras models during training?

To ensure the ability to recover from an interrupted training run at any time (fault tolerance), you should use a callback that regularly saves your model to disk. You should also set up your code to optionally reload that model at startup. Here's a simple example.

import os
from tensorflow import keras

# Prepare a directory to store all the checkpoints.
checkpoint_dir = './ckpt'
if not os.path.exists(checkpoint_dir):
    os.makedirs(checkpoint_dir)


def make_model():
    # Create a new linear regression model.
    model = keras.Sequential([keras.layers.Dense(1)])
    model.compile(optimizer='adam', loss='mse')
    return model


def make_or_restore_model():
    # Either restore the latest model, or create a fresh one
    # if there is no checkpoint available.
    checkpoints = [checkpoint_dir + '/' + name
                   for name in os.listdir(checkpoint_dir)]
    if checkpoints:
        latest_checkpoint = max(checkpoints, key=os.path.getctime)
        print('Restoring from', latest_checkpoint)
        return keras.models.load_model(latest_checkpoint)
    print('Creating a new model')
    return make_model()


model = make_or_restore_model()
callbacks = [
    # This callback saves a SavedModel every 100 batches.
    # We include the training loss in the folder name.
    keras.callbacks.ModelCheckpoint(
        filepath=checkpoint_dir + '/ckpt-loss={loss:.2f}',
        save_freq=100)
]
model.fit(train_data, epochs=10, callbacks=callbacks)

Find out more in the callbacks documentation.

How can I interrupt training when the validation loss isn't decreasing anymore?

You can use an EarlyStopping callback:

from tensorflow.keras.callbacks import EarlyStopping

early_stopping = EarlyStopping(monitor='val_loss', patience=2)
model.fit(x, y, validation_split=0.2, callbacks=[early_stopping])

Find out more in the callbacks documentation.

How can I freeze layers and do fine-tuning?

Setting the trainable attribute

All layers & models have a layer.trainable boolean attribute:

>>> layer = Dense(3)
>>> layer.trainable
True
```endshell

On all layers & models, the `trainable` attribute can be set (to True or False).
When set to `False`, the `layer.trainable_weights` attribute is empty:

```python
>>> layer = Dense(3)
>>> layer.build(input_shape=(3, 3)) # Create the weights of the layer
>>> layer.trainable
True
>>> layer.trainable_weights
[<tf.Variable 'kernel:0' shape=(3, 3) dtype=float32, numpy=
array([[...]], dtype=float32)>, <tf.Variable 'bias:0' shape=(3,) dtype=float32, numpy=array([...], dtype=float32)>]
>>> layer.trainable = False
>>> layer.trainable_weights
[]

Setting the trainable attribute on a layer recursively sets it on all children layers (contents of self.layers).

1) When training with fit():

To do fine-tuning with fit(), you would:

  • Instantiate a base model and load pre-trained weights
  • Freeze that base model
  • Add trainable layers on top
  • Call compile() and fit()

Like this:

model = Sequential([
    ResNet50Base(input_shape=(32, 32, 3), weights='pretrained'),
    Dense(10),
])
model.layers[0].trainable = False  # Freeze ResNet50Base.

assert model.layers[0].trainable_weights == []  # ResNet50Base has no trainable weights.
assert len(model.trainable_weights) == 2  # Just the bias & kernel of the Dense layer.

model.compile(...)
model.fit(...)  # Train Dense while excluding ResNet50Base.

You can follow a similar workflow with the Functional API or the model subclassing API. Make sure to call compile() after changing the value of trainable in order for your changes to be taken into account. Calling compile() will freeze the state of the training step of the model.

2) When using a custom training loop:

When writing a training loop, make sure to only update weights that are part of model.trainable_weights (and not all model.weights).

model = Sequential([
    ResNet50Base(input_shape=(32, 32, 3), weights='pretrained'),
    Dense(10),
])
model.layers[0].trainable = False  # Freeze ResNet50Base.

# Iterate over the batches of a dataset.
for inputs, targets in dataset:
    # Open a GradientTape.
    with tf.GradientTape() as tape:
        # Forward pass.
        predictions = model(inputs)
        # Compute the loss value for this batch.
        loss_value = loss_fn(targets, predictions)

    # Get gradients of loss wrt the *trainable* weights.
    gradients = tape.gradient(loss_value, model.trainable_weights)
    # Update the weights of the model.
    optimizer.apply_gradients(zip(gradients, model.trainable_weights))

Interaction between trainable and compile()

Calling compile() on a model is meant to "freeze" the behavior of that model. This implies that the trainable attribute values at the time the model is compiled should be preserved throughout the lifetime of that model, until compile is called again. Hence, if you change trainable, make sure to call compile() again on your model for your changes to be taken into account.

For instance, if two models A & B share some layers, and:

  • Model A gets compiled
  • The trainable attribute value on the shared layers is changed
  • Model B is compiled

Then model A and B are using different trainable values for the shared layers. This mechanism is critical for most existing GAN implementations, which do:

discriminator.compile(...)  # the weights of `discriminator` should be updated when `discriminator` is trained
discriminator.trainable = False
gan.compile(...)  # `discriminator` is a submodel of `gan`, which should not be updated when `gan` is trained

What's the difference between the training argument in call() and the trainable attribute?

training is a boolean argument in call that determines whether the call should be run in inference mode or training mode. For example, in training mode, a Dropout layer applies random dropout and rescales the output. In inference mode, the same layer does nothing. Example:

y = Dropout(0.5)(x, training=True)  # Applies dropout at training time *and* inference time

trainable is a boolean layer attribute that determines the trainable weights of the layer should be updated to minimize the loss during training. If layer.trainable is set to False, then layer.trainable_weights will always be an empty list. Example:

model = Sequential([
    ResNet50Base(input_shape=(32, 32, 3), weights='pretrained'),
    Dense(10),
])
model.layers[0].trainable = False  # Freeze ResNet50Base.

assert model.layers[0].trainable_weights == []  # ResNet50Base has no trainable weights.
assert len(model.trainable_weights) == 2  # Just the bias & kernel of the Dense layer.

model.compile(...)
model.fit(...)  # Train Dense while excluding ResNet50Base.

As you can see, "inference mode vs training mode" and "layer weight trainability" are two very different concepts.

You could imagine the following: a dropout layer where the scaling factor is learned during training, via backpropagation. Let's name it AutoScaleDropout. This layer would have simultaneously a trainable state, and a different behavior in inference and training. Because the trainable attribute and the training call argument are independent, you can do the following:

layer = AutoScaleDropout(0.5)

# Applies dropout at training time *and* inference time  
# *and* learns the scaling factor during training
y = layer(x, training=True)

assert len(layer.trainable_weights) == 1
# Applies dropout at training time *and* inference time  
# with a *frozen* scaling factor

layer = AutoScaleDropout(0.5)
layer.trainable = False
y = layer(x, training=True)

Special case of the BatchNormalization layer

Consider a BatchNormalization layer in the frozen part of a model that's used for fine-tuning.

It has long been debated whether the moving statistics of the BatchNormalization layer should stay frozen or adapt to the new data. Historically, bn.trainable = False would only stop backprop but would not prevent the training-time statistics update. After extensive testing, we have found that it is usually better to freeze the moving statistics in fine-tuning use cases. Starting in TensorFlow 2.0, setting bn.trainable = False will also force the layer to run in inference mode.

This behavior only applies for BatchNormalization. For every other layer, weight trainability and "inference vs training mode" remain independent.

In fit(), how is the validation split computed?

If you set the validation_split argument in model.fit to e.g. 0.1, then the validation data used will be the last 10% of the data. If you set it to 0.25, it will be the last 25% of the data, etc. Note that the data isn't shuffled before extracting the validation split, so the validation is literally just the last x% of samples in the input you passed.

The same validation set is used for all epochs (within a same call to fit).

Note that the validation_split option is only available if your data is passed as Numpy arrays (not tf.data.Datasets, which are not indexable).

In fit(), is the data shuffled during training?

If you pass your data as NumPy arrays and if the shuffle argument in model.fit() is set to True (which is the default), the training data will be globally randomly shuffled at each epoch.

If you pass your data as a tf.data.Dataset object and if the shuffle argument in model.fit() is set ot True, the dataset will be locally shuffled (buffered shuffling).

When using tf.data.Dataset objects, prefer shuffling your data beforehand (e.g. by calling dataset = dataset.shuffle(buffer_size)) so as to be in control of the buffer size.

Validation data is never shuffled.

What's the recommended way to monitor my metrics when training with fit()?

Loss values and metric values are reported via the default progress bar displayed by calls to fit(). However, staring at changing ascii numbers in a console ins't an optimal metric-monitoring experience. We recommend the use of TensorBoard, which will display nice-looking graphs of your training and validation metrics, regularly updated during training, which you can access from your browser.

You can use TensorBoard with fit() via the TensorBoard callback.

What if I need to customize what fit() does?

You have two options:

1) Write a low-level custom training looop

This is a good option if you want to be in control of every last little detail. But it can be somewhat verbose. Example:

# Prepare an optimizer.
optimizer = tf.keras.optimizers.Adam()
# Prepare a loss function.
loss_fn = tf.keras.losses.kl_divergence

# Iterate over the batches of a dataset.
for inputs, targets in dataset:
    # Open a GradientTape.
    with tf.GradientTape() as tape:
        # Forward pass.
        predictions = model(inputs)
        # Compute the loss value for this batch.
        loss_value = loss_fn(targets, predictions)

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

This examples does not include a lot of essential functionality like displaying a progress bar, calling callbacks, updating metrics, etc. You would have to do this yourself. It's not difficult at all, but it's a bit of work.

2) Subclass the Model class and override the train_step (and test_step) methods

This is a better option if you want to use custom update rules but still want to leverage the functionality provided by fit(), such as callbacks, efficient step fusing, etc.

Note that this pattern does not prevent you from building models with the Functional API (or even Sequential models).

The example below shows a Functional model with a custom train_step.

from tensorflow import keras
import tensorflow as tf
import numpy as np

class MyCustomModel(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

        with tf.GradientTape() as tape:
            y_pred = self(x, training=True)  # Forward pass
            # Compute the loss value
            # (the loss function is configured in `compile()`)
            loss = self.compiled_loss(y, y_pred,
                                      regularization_losses=self.losses)

        # Compute gradients
        trainable_vars = self.trainable_variables
        gradients = tape.gradient(loss, trainable_vars)
        # Update weights
        self.optimizer.apply_gradients(zip(gradients, trainable_vars))
        # Update metrics (includes the metric that tracks the loss)
        self.compiled_metrics.update_state(y, y_pred)
        # Return a dict mapping metric names to current value
        return {m.name: m.result() for m in self.metrics}


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

# Just use `fit` as usual
x = np.random.random((1000, 32))
y = np.random.random((1000, 1))
model.fit(x, y, epochs=10)

You can also easily add support for sample weighting:

class MyCustomModel(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:
            x, y = data

        with tf.GradientTape() as tape:
            y_pred = self(x, training=True)  # Forward pass
            # Compute the loss value.
            # The loss function is configured in `compile()`.
            loss = self.compiled_loss(y, y_pred,
                                      sample_weight=sample_weight,
                                      regularization_losses=self.losses)

        # Compute gradients
        trainable_vars = self.trainable_variables
        gradients = tape.gradient(loss, trainable_vars)

        # Update weights
        self.optimizer.apply_gradients(zip(gradients, trainable_vars))

        # Update the metrics.
        # Metrics are configured in `compile()`.
        self.compiled_metrics.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 MyCustomModel
inputs = keras.Input(shape=(32,))
outputs = keras.layers.Dense(1)(inputs)
model = MyCustomModel(inputs, outputs)
model.compile(optimizer='adam', loss='mse', metrics=['accuracy'])

# 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=10)

Similarly, you can also customize evaluation by overriding test_step:

class MyCustomModel(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
      self.compiled_loss(
          y, y_pred, regularization_losses=self.losses)
      # Update the metrics.
      self.compiled_metrics.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}

How can I train models in mixed precision?

Keras has built-in support for mixed precision training on GPU and TPU. See this extensive guide.

Modeling-related questions

How can I obtain the output of an intermediate layer (feature extraction)?

In the Functional API and Sequential API, if a layer has been called exactly once, you can retrieve its output via layer.output and its input via layer.input. This enables you do quickly instantiate feature-extraction models, like this one:

from tensorflow import keras
from tensorflow.keras import layers

model = Sequential([
    layers.Conv2D(32, 3, activation='relu'),
    layers.Conv2D(32, 3, activation='relu'),
    layers.MaxPooling2D(2),
    layers.Conv2D(32, 3, activation='relu'),
    layers.Conv2D(32, 3, activation='relu'),
    layers.GlobalMaxPooling2D(),
    layers.Dense(10),
])
extractor = keras.Model(inputs=model.inputs,
                        outputs=[layer.output for layer in model.layers])
features = extractor(data)

Naturally, this is not possible with models that are subclasses of Model that override call.

Here's another example: instantiating a Model that returns the output of a specific named layer:

model = ...  # create the original model

layer_name = 'my_layer'
intermediate_layer_model = keras.Model(inputs=model.input,
                                       outputs=model.get_layer(layer_name).output)
intermediate_output = intermediate_layer_model(data)

How can I use pre-trained models in Keras?

You could leverage the models available in keras.applications, or the models available on TensorFlow Hub. TensorFlow Hub is well-integrated with Keras.

How can I use stateful RNNs?

Making a RNN stateful means that the states for the samples of each batch will be reused as initial states for the samples in the next batch.

When using stateful RNNs, it is therefore assumed that:

  • all batches have the same number of samples
  • If x1 and x2 are successive batches of samples, then x2[i] is the follow-up sequence to x1[i], for every i.

To use statefulness in RNNs, you need to:

  • explicitly specify the batch size you are using, by passing a batch_size argument to the first layer in your model. E.g. batch_size=32 for a 32-samples batch of sequences of 10 timesteps with 16 features per timestep.
  • set stateful=True in your RNN layer(s).
  • specify shuffle=False when calling fit().

To reset the states accumulated:

  • use model.reset_states() to reset the states of all layers in the model
  • use layer.reset_states() to reset the states of a specific stateful RNN layer

Example:

from tensorflow import keras
from tensorflow.keras import layers
import numpy as np

x = np.random.random((32, 21, 16))  # this is our input data, of shape (32, 21, 16)
# we will feed it to our model in sequences of length 10

model = keras.Sequential()
model.add(layers.LSTM(32, input_shape=(10, 16), batch_size=32, stateful=True))
model.add(layers.Dense(16, activation='softmax'))

model.compile(optimizer='rmsprop', loss='categorical_crossentropy')

# we train the network to predict the 11th timestep given the first 10:
model.train_on_batch(x[:, :10, :], np.reshape(x[:, 10, :], (32, 16)))

# the state of the network has changed. We can feed the follow-up sequences:
model.train_on_batch(x[:, 10:20, :], np.reshape(x[:, 20, :], (32, 16)))

# let's reset the states of the LSTM layer:
model.reset_states()

# another way to do it in this case:
model.layers[0].reset_states()

Note that the methods predict, fit, train_on_batch, predict_classes, etc. will all update the states of the stateful layers in a model. This allows you to do not only stateful training, but also stateful prediction.