Training Example: Keras – Review the Data, Give Your Score & Compare to the Real AI Evaluation

KERAS 3.0 RELEASED
[H1] A superpower for ML developers
Keras is a deep learning API designed for human beings, not
machines. Keras focuses on debugging speed, code elegance &
conciseness, maintainability, and deployability. When you choose
Keras, your codebase is smaller, more readable, easier to iterate
on.
API DOCS
GUIDES
EXAMPLES

[IMG: K graphic]

Copied
inputs = keras.Input(shape=(32, 32, 3))
x = layers.Conv2D(32, 3, activation="relu")(inputs)
x = layers.Conv2D(64, 3, activation="relu")(x)
residual = x = layers.MaxPooling2D(3)(x)
x = layers.Conv2D(64, 3, padding="same")(x)
x = layers.Activation("relu")(x)
x = layers.Conv2D(64, 3, padding="same")(x)
x = layers.Activation("relu")(x)
x = x + residual
x = layers.Conv2D(64, 3, activation="relu")(x)
x = layers.GlobalAveragePooling2D()(x)
outputs = layers.Dense(10, activation="softmax")(x)
model = keras.Model(inputs, outputs, name="mini_resnet")
keras.utils.plot_model(model, "mini_resnet.png")
model.fit(dataset, epochs=10)
Run quickstart

Copied
causal_lm = keras_hub.models.CausalLM.from_preset(
"gemma2_instruct_2b_en",
dtype="float16",
)
prompt = """<start_of_turn>user
Write python code to print the first 100 primes.
<end_of_turn>
<start_of_turn>model
"""
text_output = causal_lm.generate(prompt, max_length=512)
text_to_image = keras_hub.models.TextToImage.from_preset(
"stable_diffusion_3_medium",
dtype="float16",
)
prompt = "Astronaut in a jungle, detailed"
image_output = text_to_image.generate(prompt)
Run quickstart

[IMG: Backend logos]

[H2]
Welcome to multi-framework machine learning
With its multi-backend approach, Keras gives you the freedom to
work with JAX, TensorFlow, and PyTorch. Build models that can move
seamlessly across these frameworks and leverage the strengths of
each ecosystem.

GET STARTED

[H2] Developer Guides
VIEW ALL

Copied
inputs = keras.Input(shape=(28, 28, 1))
x = inputs
x = layers.Conv2D(16, 3, activation="relu")(x)
x = layers.Conv2D(32, 3, activation="relu")(x)
x = layers.MaxPooling2D(3)(x)
x = layers.Conv2D(32, 3, activation="relu")(x)
x = layers.Conv2D(16, 3, activation="relu")(x)
x = layers.GlobalMaxPooling2D()(x)
x = layers.Dropout(0.5)
outputs = layers.Dense(10)
model = keras.Model(inputs, outputs)
model.summary()

[H3] The Functional API

Starting from the beginning and learn how to build models using the functional building pattern.

VIEW GUIDE

Copied
model.compile(
optimizer="rmsprop",
loss="categorical_crossentropy",
metrics=["accuracy"],
)
history = model.fit(
x_train,
y_train,
batch_size=64,
epochs=2,
validation_data=(x_val, y_val),
)

[H3] Training & evaluation with the built-in methods

Train and evaluate your model using model.fit(...).

VIEW GUIDE

Copied
class MLPBlock(keras.layers.Layer):
def __init__(self):
super().__init__()
self.dense_1 = layers.Dense(32)
self.dense_2 = layers.Dense(32)
self.dense_3 = layers.Dense(1)
def call(self, inputs):
x = self.dense_1(inputs)
x = keras.activations.relu(x)
x = self.dense_2(x)
x = keras.activations.relu(x)
return self.dense_3(x)

[H3] Making new layers and models via subclassing

Learn how to customize your model via subclassing Keras layers.

VIEW GUIDE

VIEW ALL

[H2] KerasHub
The KerasHub library provides Keras 3 implementations of popular model architectures, paired with a collection of pretrained checkpoints available on Kaggle Models. Models can be used for both training and inference, on any of the TensorFlow, JAX, and PyTorch backends.

SEE ALL

[H4] GEMMA

Google’s family of lightweight language models built from the same research and technology used to create Gemini.
VIEW DOCUMENTATION
KAGGLE DETAILS

[H4] LLAMA

Meta’s flagship open text generation models available in a wide range of sizes and precisions.
VIEW DOCUMENTATION
KAGGLE DETAILS

[H4] STABLE DIFFUSION

Generate image content with this state of the art diffusion model from Stability AI.
VIEW DOCUMENTATION
KAGGLE DETAILS

[H4] MISTRAL

A generative language from the French company Mistral AI, making frontier models accessible to all.
VIEW DOCUMENTATION
KAGGLE DETAILS

SEE ALL

[H2] Code examples
VIEW ALL

[IMG: eye]

[H3] Computer vision

Take a look at our examples for doing image classification, object detection, video processing, and more.

SEE EXAMPLE

[IMG: text]

[H3] Natural Language Processing

We also have many guides for doing NLP including text classification, machine translation, and language modeling.

SEE EXAMPLE

[IMG: flower]

[H3] Generative Deep Learning

Get started with generative deep learning with our wealth of guides involving state-of-the-art diffusion models, GANs, and transformer models.

SEE EXAMPLE

VIEW ALL

[H2] Trusted for research and production

Keras is used by CERN, NASA, NIH, and many more scientific
organizations around the world (and yes, Keras is used at the Large
Hadron Collider). Keras is used by Waymo to power self-driving
vehicles. Keras partners with Kaggle and HuggingFace to meet ML
developers in the tools they use daily.

[IMG: youtube logo]

[IMG: google logo]

[IMG: waymo logo]

[IMG: amazon logo]

[IMG: spotify logo]

[IMG: uber logo]

[IMG: netflix logo]

[H2] Stay in touch
Sign up to our mailing list for regular updates and discussions about the Keras ecosystem. Listen in at our community meetings, and follow us on social media!

JOIN GOOGLE GROUP

JOIN COMMUNITY MEETING
DISCORD
GOOGLE AI FORUM

[H2] Contributions welcome!
We welcome your code, ideas, and feedback as we continue to grow. Visit our roadmap, contribution guide or GitHub for more information.
VIEW ROADMAP
CONTRIBUTION GUIDE
GITHUB

None

► Code examples

[H1] Code examples
Our code examples are short (less than 300 lines of code), focused demonstrations of vertical deep learning workflows.
All of our examples are written as Jupyter notebooks and can be run in one click in Google Colab,
a hosted notebook environment that requires no setup and runs in the cloud. Google Colab includes GPU and TPU runtimes.
★ = Good starter example
V3 = Keras 3 example

[H2] Computer Vision

[H3] Image classification

★

V3

Image classification from scratch

★

V3

Simple MNIST convnet

★

V3

Image classification via fine-tuning with EfficientNet

V3

Image classification with Vision Transformer

V3

Classification using Attention-based Deep Multiple Instance Learning

V3

Image classification with modern MLP models

V3

A mobile-friendly Transformer-based model for image classification

V3

Pneumonia Classification on TPU

V3

Compact Convolutional Transformers

V3

Image classification with ConvMixer

V3

Image classification with EANet (External Attention Transformer)

V3

Involutional neural networks

V3

Image classification with Perceiver

V3

Few-Shot learning with Reptile

V3

Semi-supervised image classification using contrastive pretraining with SimCLR

V3

Image classification with Swin Transformers

V3

Train a Vision Transformer on small datasets

V3

A Vision Transformer without Attention

V3

Image Classification using Global Context Vision Transformer

V3

When Recurrence meets Transformers

V3

Using the Forward-Forward Algorithm for Image Classification

V3

Image Classification using BigTransfer (BiT)

V3

Focal Modulation: A replacement for Self-Attention

[H3] Image segmentation

★

V3

Image segmentation with a U-Net-like architecture

V3

Multiclass semantic segmentation using DeepLabV3+

V3

Highly accurate boundaries segmentation using BASNet

V3

Image Segmentation using Composable Fully-Convolutional Networks

[H3] Object detection

V3

Keypoint Detection with Transfer Learning

V3

Object detection with Vision Transformers

[H3] 3D

V3

3D Multimodal Brain Tumor Segmentation

V3

3D image classification from CT scans

V3

Monocular depth estimation

★

V3

3D volumetric rendering with NeRF

V3

Point cloud segmentation with PointNet

V3

Point cloud classification

[H3] OCR

V3

OCR model for reading Captchas

V3

Handwriting recognition

[H3] Image enhancement

V3

Convolutional autoencoder for image denoising

V3

Low-light image enhancement using MIRNet

V3

Image Super-Resolution using an Efficient Sub-Pixel CNN

V3

Enhanced Deep Residual Networks for single-image super-resolution

V3

Zero-DCE for low-light image enhancement

[H3] Data augmentation

V3

CutMix data augmentation for image classification

V3

MixUp augmentation for image classification

V3

RandAugment for Image Classification for Improved Robustness

[H3] Image & Text

★

V3

Image captioning

[H3] Vision models interpretability

V3

Visualizing what convnets learn

V3

Model interpretability with Integrated Gradients

V3

Investigating Vision Transformer representations

V3

Grad-CAM class activation visualization

[H3] Image similarity search

V3

Semantic Image Clustering

V3

Image similarity estimation using a Siamese Network with a contrastive loss

V3

Image similarity estimation using a Siamese Network with a triplet loss

V3

Metric learning for image similarity search

V3

Self-supervised contrastive learning with NNCLR

V3

Self-supervised contrastive learning with SimSiam

[H3] Video

V3

Video Classification with a CNN-RNN Architecture

V3

Next-Frame Video Prediction with Convolutional LSTMs

V3

Video Classification with Transformers

V3

Video Vision Transformer

[H3] Performance recipes

V3

Gradient Centralization for Better Training Performance

V3

Learning to tokenize in Vision Transformers

V3

Knowledge Distillation

V3

FixRes: Fixing train-test resolution discrepancy

V3

Class Attention Image Transformers with LayerScale

V3

Augmenting convnets with aggregated attention

V3

Learning to Resize

V3

Semi-supervision and domain adaptation with AdaMatch

V3

Consistency training with supervision

V3

Distilling Vision Transformers

V3

Masked image modeling with Autoencoders

[H2] Natural Language Processing

[H3] Text classification

★

V3

Text classification from scratch

V3

Review Classification using Active Learning

V3

Text Classification using FNet

V3

Large-scale multi-label text classification

V3

Text classification with Transformer

V3

Text classification with Switch Transformer

V3

Using pre-trained word embeddings

V3

Bidirectional LSTM on IMDB

V3

Data Parallel Training with KerasHub and tf.distribute

V3

MultipleChoice Task with Transfer Learning

[H3] Machine translation

V3

English-to-Spanish translation with KerasHub

★

V3

English-to-Spanish translation with a sequence-to-sequence Transformer

V3

Character-level recurrent sequence-to-sequence model

[H3] Entailment prediction

V3

Multimodal entailment

[H3] Named entity recognition

V3

Named Entity Recognition using Transformers

[H3] Sequence-to-sequence

V3

Sequence to sequence learning for performing number addition

[H3] Text similarity search

V3

Semantic Similarity with KerasHub

V3

Semantic Similarity with BERT

V3

Sentence embeddings using Siamese RoBERTa-networks

[H3] Language modeling

V3

End-to-end Masked Language Modeling with BERT

V3

Abstractive Text Summarization with BART

[H3] Parameter efficient fine-tuning

V3

Parameter-efficient fine-tuning of GPT-2 with LoRA

[H2] Structured Data

[H3] Structured data classification

★

V3

Structured data classification with FeatureSpace

★

V3

FeatureSpace advanced use cases

★

V3

Imbalanced classification: credit card fraud detection

V3

Structured data classification from scratch

V3

Structured data learning with Wide, Deep, and Cross networks

V3

Classification with Gated Residual and Variable Selection Networks

V3

Classification with Neural Decision Forests

V3

Structured data learning with TabTransformer

V3

Classification with Gated Residual and Variable Selection Networks with HyperParameters tuning

[H3] Structured data regression

V3

Deep Learning for Customer Lifetime Value

[H3] Recommendation

V3

Collaborative Filtering for Movie Recommendations

V3

A Transformer-based recommendation system

[H2] Timeseries

[H3] Timeseries classification

★

V3

Timeseries classification from scratch

V3

Timeseries classification with a Transformer model

V3

Electroencephalogram Signal Classification for action identification

V3

Event classification for payment card fraud detection

V3

Electroencephalogram Signal Classification for Brain-Computer Interface

[H3] Anomaly detection

V3

Timeseries anomaly detection using an Autoencoder

[H3] Timeseries forecasting

V3

Traffic forecasting using graph neural networks and LSTM

V3

Timeseries forecasting for weather prediction

[H2] Generative Deep Learning

[H3] Image generation

★

V3

Denoising Diffusion Implicit Models

★

V3

A walk through latent space with Stable Diffusion 3

V3

DreamBooth

V3

Variational AutoEncoder

V3

GAN overriding Model.train_step

V3

WGAN-GP overriding Model.train_step

V3

Conditional GAN

V3

CycleGAN

V3

Data-efficient GANs with Adaptive Discriminator Augmentation

V3

Deep Dream

V3

GauGAN for conditional image generation

V3

PixelCNN

V3

Vector-Quantized Variational Autoencoders

V3

A walk through latent space with Stable Diffusion

[H3] Style transfer

V3

Neural style transfer

[H3] Text generation

★

V3

GPT2 Text Generation with KerasHub

V3

GPT text generation from scratch with KerasHub

V3

Text generation with a miniature GPT

V3

Character-level text generation with LSTM

V3

Text Generation using FNet

[H3] Audio generation

V3

Music Generation with Transformer Models

[H3] Graph generation

V3

Drug Molecule Generation with VAE

[H2] Audio Data

[H3] Vocal track separation

V3

Vocal Track Separation with Encoder-Decoder Architecture

[H3] Speech recognition

V3

Automatic Speech Recognition with Transformer

V3

Automatic Speech Recognition using CTC

[H3] Audio classification

V3

Audio Classification with the STFTSpectrogram layer

V3

Speaker Recognition

[H2] Reinforcement Learning

Actor Critic Method

Proximal Policy Optimization

Deep Q-Learning for Atari Breakout

Deep Deterministic Policy Gradient (DDPG)

[H2] Graph Data

[H3] Node classification

V3

Graph attention network (GAT) for node classification

V3

Node Classification with Graph Neural Networks

[H3] Graph representation learning

V3

Graph representation learning with node2vec

[H2] Quick Keras Recipes

[H3] Keras usage tips

V3

Parameter-efficient fine-tuning of Gemma with LoRA and QLoRA

V3

Float8 training and inference with a simple Transformer model

V3

Keras debugging tips

V3

Customizing the convolution operation of a Conv2D layer

V3

Trainer pattern

V3

Endpoint layer pattern

V3

Reproducibility in Keras Models

V3

Writing Keras Models With TensorFlow NumPy

V3

Simple custom layer example: Antirectifier

V3

Packaging Keras models for wide distribution using Functional Subclassing

V3

Approximating non-Function Mappings with Mixture Density Networks

V3

Evaluating and exporting scikit-learn metrics in a Keras callback

[H3] Serving

V3

Serving TensorFlow models with TFServing

[H3] ML best practices

V3

Estimating required sample size for model training

V3

Memory-efficient embeddings for recommendation systems

V3

Creating TFRecords

V3

Knowledge distillation recipes

[H2] Adding a new code example
We welcome new code examples! Here are our rules:
They should be shorter than 300 lines of code (comments may be as long as you want).
They should demonstrate modern Keras best practices.
They should be substantially different in topic from all examples listed above.
They should be extensively documented & commented.
New examples are added via Pull Requests to the keras.io repository.
They must be submitted as a .py file that follows a specific format. They are usually generated from Jupyter notebooks.
See the tutobooks documentation for more details.
If you would like to convert a Keras 2 example to Keras 3, please open a Pull Request to the keras.io repository.

Code examples

Adding a new code example

None

► Developer guides

[H1] Developer guides
Our developer guides are deep-dives into specific topics such as layer subclassing, fine-tuning, or model saving.
They're one of the best ways to become a Keras expert.
Most of our guides are written as Jupyter notebooks and can be run in one click in Google Colab,
a hosted notebook environment that requires no setup and runs in the cloud. Google Colab includes GPU and TPU runtimes.
[H2] Available guides
The Functional API
The Sequential model
Making new layers & models via subclassing
Training & evaluation with the built-in methods
Customizing fit() with JAX
Customizing fit() with TensorFlow
Customizing fit() with PyTorch
Writing a custom training loop in JAX
Writing a custom training loop in TensorFlow
Writing a custom training loop in PyTorch
Serialization & saving
Customizing saving & serialization
Writing your own callbacks
Transfer learning & fine-tuning
Distributed training with JAX
Distributed training with TensorFlow
Distributed training with PyTorch
Distributed training with Keras 3
Migrating Keras 2 code to Keras 3
How to use Keras with NNX backend
Orbax Checkpointing in Keras
Quantization in Keras
8-bit integer quantization in Keras
4-bit integer quantization in Keras
GPTQ quantization in Keras
AWQ quantization in Keras
Writing quantization-compatible layers in Keras
Customizing quantization in Keras
Define a Custom TPU/GPU Kernel

Developer guides

Available guides

None

► KerasHub

[H1] KerasHub
Star
KerasHub is a pretrained modeling library that aims to be simple, flexible,
and fast. The library provides Keras 3
implementations of popular model architectures, paired with a collection of
pretrained checkpoints available on Kaggle Models.
Models can be used for both training and inference, on any of the TensorFlow,
Jax, and Torch backends.
KerasHub is an extension of the core Keras API; KerasHub components are provided
as keras.layers.Layer and keras.Model implementations. If you are familiar
with Keras, congratulations! You already understand most of KerasHub.
[H2] Quick links
Getting started with KerasHub
Developer guides
API documentation
KerasHub on GitHub
KerasHub models on Kaggle
Pretrained model list
[H2] Installation
To install the latest KerasHub release with Keras 3, simply run:
pip install --upgrade keras-hub
To install the latest nightly changes for both KerasHub and Keras, you can use
our nightly package.
pip install --upgrade keras-hub-nightly
Currently, installing KerasHub will always pull in TensorFlow for use of the
tf.data API for preprocessing. When pre-processing with tf.data, training
can still happen on any backend.
Visit the core Keras getting started page
for more information on installing Keras 3, accelerator support, and
compatibility with different frameworks.
[H2] Quickstart
Choose a backend:
import os
os.environ["KERAS_BACKEND"] = "jax" # Or "tensorflow" or "torch"!
Import KerasHub and other libraries:
import keras
import keras_hub
import numpy as np
import tensorflow_datasets as tfds
Load a resnet model and use it to predict a label for an image:
classifier = keras_hub.models.ImageClassifier.from_preset(
"resnet_50_imagenet",
activation="softmax",
)
url = "https://upload.wikimedia.org/wikipedia/commons/a/aa/California_quail.jpg"
path = keras.utils.get_file(origin=url)
image = keras.utils.load_img(path)
preds = classifier.predict(np.array([image]))
print(keras_hub.utils.decode_imagenet_predictions(preds))
Load a Bert model and fine-tune it on IMDb movie reviews:
classifier = keras_hub.models.BertClassifier.from_preset(
"bert_base_en_uncased",
activation="softmax",
num_classes=2,
)
imdb_train, imdb_test = tfds.load(
"imdb_reviews",
split=["train", "test"],
as_supervised=True,
batch_size=16,
)
classifier.fit(imdb_train, validation_data=imdb_test)
preds = classifier.predict(["What an amazing movie!", "A total waste of time."])
print(preds)
[H2] Compatibility
We follow Semantic Versioning, and plan to
provide backwards compatibility guarantees both for code and saved models built
with our components. While we continue with pre-release 0.y.z development, we
may break compatibility at any time and APIs should not be consider stable.
[H2] Disclaimer
KerasHub provides access to pre-trained models via the keras_hub.models API.
These pre-trained models are provided on an "as is" basis, without warranties
or conditions of any kind.
[H2] Citing KerasHub
If KerasHub helps your research, we appreciate your citations.
Here is the BibTeX entry:
@misc{kerashub2024,
title={KerasHub},
author={Watson, Matthew, and Chollet, Fran\c{c}ois and Sreepathihalli,
Divyashree, and Saadat, Samaneh and Sampath, Ramesh, and Rasskin, Gabriel and
and Zhu, Scott and Singh, Varun and Wood, Luke and Tan, Zhenyu and Stenbit,
Ian and Qian, Chen, and Bischof, Jonathan and others},
year={2024},
howpublished={\url{https://github.com/keras-team/keras-hub}},
}

KerasHub

Quick links

Installation

Quickstart

Compatibility

Disclaimer

Citing KerasHub