This repository hosts the development of the Keras library. Read the documentation at keras.io.
Keras is a deep learning API written in Python, running on top of the machine learning platform TensorFlow. It was developed with a focus on enabling fast experimentation. Being able to go from idea to result as fast as possible is key to doing good research.
Keras is:
- Simple -- but not simplistic. Keras reduces developer cognitive load to free you to focus on the parts of the problem that really matter.
- Flexible -- Keras adopts the principle of progressive disclosure of complexity: simple workflows should be quick and easy, while arbitrarily advanced workflows should be possible via a clear path that builds upon what you've already learned.
- Powerful -- Keras provides industry-strength performance and scalability: it is used by organizations and companies including NASA, YouTube, and Waymo.
TensorFlow 2 is an end-to-end, open-source machine learning platform. You can think of it as an infrastructure layer for differentiable programming. It combines four key abilities:
- Efficiently executing low-level tensor operations on CPU, GPU, or TPU.
- Computing the gradient of arbitrary differentiable expressions.
- Scaling computation to many devices, such as clusters of hundreds of GPUs.
- Exporting programs ("graphs") to external runtimes such as servers, browsers, mobile and embedded devices.
Keras is the high-level API of TensorFlow 2: an approachable, highly-productive interface for solving machine learning problems, with a focus on modern deep learning. It provides essential abstractions and building blocks for developing and shipping machine learning solutions with high iteration velocity.
Keras empowers engineers and researchers to take full advantage of the scalability and cross-platform capabilities of TensorFlow 2: you can run Keras on TPU or on large clusters of GPUs, and you can export your Keras models to run in the browser or on a mobile device.
The core data structures of Keras are layers and models.
The simplest type of model is the Sequential
model, a linear stack of layers.
For more complex architectures, you should use the Keras functional API,
which allows you to build arbitrary graphs of layers or write models entirely from scratch via subclassing.
Here is the Sequential
model:
from tensorflow.keras.models import Sequential
model = Sequential()
Stacking layers is as easy as .add()
:
from tensorflow.keras.layers import Dense
model.add(Dense(units=64, activation='relu'))
model.add(Dense(units=10, activation='softmax'))
Once your model looks good, configure its learning process with .compile()
:
model.compile(loss='categorical_crossentropy',
optimizer='sgd',
metrics=['accuracy'])
If you need to, you can further configure your optimizer. The Keras philosophy is to keep simple things simple, while allowing the user to be fully in control when they need to (the ultimate control being the easy extensibility of the source code via subclassing).
model.compile(loss=tf.keras.losses.categorical_crossentropy,
optimizer=tf.keras.optimizers.SGD(
learning_rate=0.01, momentum=0.9, nesterov=True))
You can now iterate on your training data in batches:
# x_train and y_train are Numpy arrays.
model.fit(x_train, y_train, epochs=5, batch_size=32)
Evaluate your test loss and metrics in one line:
loss_and_metrics = model.evaluate(x_test, y_test, batch_size=128)
Or generate predictions on new data:
classes = model.predict(x_test, batch_size=128)
What you just saw is the most elementary way to use Keras.
Here is a very minimal use of Keras to train, compile and score a model:
from tensorflow.keras.models import Sequential
model = Sequential()
model.compile(loss='categorical_crossentropy')
model.fit(x_train, y_train)
loss_and_metrics = model.evaluate(x_test, y_test)
classes = model.predict(x_test)
Note: x_train, x_test, y_train and y_test need to be created and set up in addition to the code above.
However, Keras is also a highly-flexible framework suitable to iterate on state-of-the-art research ideas. Keras follows the principle of progressive disclosure of complexity: it makes it easy to get started, yet it makes it possible to handle arbitrarily advanced use cases, only requiring incremental learning at each step.
In much the same way that you were able to train & evaluate a simple neural network above in a few lines,
you can use Keras to quickly develop new training procedures or exotic model architectures.
Here's a low-level training loop example, combining Keras functionality with the TensorFlow GradientTape
:
import tensorflow as tf
# 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))
For more in-depth tutorials about Keras, you can check out:
- Introduction to Keras for engineers
- Introduction to Keras for researchers
- Developer guides
- Other learning resources
Keras comes packaged with TensorFlow 2 as tensorflow.keras
.
To start using Keras, simply install TensorFlow 2 through pip or conda.
To install pip, which is needed to install TensorFlow 2, download through ensure.pip (with python) or get-pip.py
If using the ensure.pip method, run this command to install pip:
Windows:
C:> py -m ensurepip --upgrade
MacOS:
$ python -m ensurepip --upgrade
Linux:
$ python -m ensurepip --upgrade
If using the get-pip.py method, download get-pip.py with the link and cd to the directory containing the file, then run this command:
Windows:
C:> py get-pip.py
MacOS:
$ python get-pip.py
Linux:
$ python get-pip.py
To upgrade pip:
Windows:
C:> py -m pip install --upgrade pip
MacOS:
$ python -m pip install --upgrade pip
Linux:
$ python -m pip install --upgrade pip
Keras has nightly releases (keras-nightly
on PyPI)
and stable releases (keras
on PyPI).
The nightly Keras releases are usually compatible with the corresponding version
of the tf-nightly
releases
(e.g. keras-nightly==2.7.0.dev2021100607
should be
used with tf-nightly==2.7.0.dev2021100607
).
We don't maintain backward compatibility for nightly releases.
For stable releases, each Keras
version maps to a specific stable version of TensorFlow.
The table below shows the compatibility version mapping between TensorFlow versions and Keras versions.
All the release branches can be found on GitHub.
All the release binaries can be found on Pypi.
Keras release | Note | Compatible Tensorflow version |
---|---|---|
2.4 | Last stable release of multi-backend Keras | < 2.5 |
2.5-pre | Pre-release (not formal) for standalone Keras repo | >= 2.5 < 2.6 |
2.6 | First formal release of standalone Keras. | >= 2.6 < 2.7 |
2.7 | (Upcoming release) | >= 2.7 < 2.8 |
nightly | tf-nightly |
You can ask questions and join the development discussion:
- In the TensorFlow forum.
- On the Keras Google group.
You can also post bug reports and feature requests (only) in GitHub issues.
We welcome contributions! Before opening a PR, please read our contributor guide, and the API design guideline.