Build Status

tl;dr: keras-pandas allows users to rapidly build and iterate on deep learning models.

Getting data formatted and into keras can be tedious, time consuming, and difficult, whether your a veteran or new to Keras. keras-pandas overcomes these issues by (automatically) providing:

  • A cleaned, transformed and correctly formatted X and y (good for keras, sklearn or any other ML platform)
  • An ‘input nub’, without the hassle of worrying about input shapes or data types
  • An output layer, correctly formatted for the kind of response variable provided

With these resources, it’s possible to rapidly build and iterate on deep learning models, and focus on the parts of modeling that you enjoy!

For more info, check out the:

Quick Start

Let’s build a model with the titanic data set. This data set is particularly fun because this data set contains a mix of categorical and numerical data types, and features a lot of null values.

We’ll keras-pandas

pip install -U keras-pandas

And then run the following snippet to create and train a model:

from keras import Model
from keras.layers import Dense

from keras_pandas.Automater import Automater
from keras_pandas.lib import load_titanic

observations = load_titanic()

# Transform the data set, using keras_pandas
categorical_vars = ['pclass', 'sex', 'survived']
numerical_vars = ['age', 'siblings_spouses_aboard', 'parents_children_aboard', 'fare']
text_vars = ['name']

auto = Automater(categorical_vars=categorical_vars, numerical_vars=numerical_vars, text_vars=text_vars,
X, y = auto.fit_transform(observations)

# Start model with provided input nub
x = auto.input_nub

# Fill in your own hidden layers
x = Dense(32)(x)
x = Dense(32, activation='relu')(x)
x = Dense(32)(x)

# End model with provided output nub
x = auto.output_nub(x)

model = Model(inputs=auto.input_layers, outputs=x)
model.compile(optimizer='Adam', loss=auto.loss, metrics=['accuracy'])

# Train model, y, epochs=4, validation_split=.2)



You can install keras-pandas with pip:

pip install -U keras-pandas

Creating an Automater

The core feature of keras-pandas is the Automater, which accepts lists of variable types (all optional), and a response variable (optional, for supervised problems). Together, all of these variables are the user_input_variables, which may be different than the variables fed into Keras.

As a side note, the response variable must be in one of the variable type lists (e.g. survived is in categorical_vars)

One variable type

If you only have one variable type, only use that variable type!

categorical_vars = ['pclass', 'sex', 'survived']
auto = Automater(categorical_vars=categorical_vars, response_var='survived')

Multiple variable types

If you have multiple variable types, throw them all in!

categorical_vars = ['pclass', 'sex', 'survived']
numerical_vars = ['age', 'siblings_spouses_aboard', 'parents_children_aboard', 'fare']

auto = Automater(categorical_vars=categorical_vars, numerical_vars=numerical_vars, response_var='survived')

No response_var

If all variables are always available, and / or your problems space doesn’t have a single response variable, you can omit the response variable.

categorical_vars = ['pclass', 'sex', 'survived']
numerical_vars = ['age', 'siblings_spouses_aboard', 'parents_children_aboard', 'fare']

auto = Automater(categorical_vars=categorical_vars, numerical_vars=numerical_vars)

In this case, an output nub will not be auto-generated

Fitting the Automater

Before use, the Automator must be fit. The fit() method accepts a pandas DataFrame, which must contain all of the columns listed during initialization.

Transforming data

Now, we can use our Automater to transform the dataset, from a pandas DataFrame to numpy objects properly formatted for Keras’s input and output layers.

X, y = auto.transform(observations, df_out=False)

This will return two objects:

  • X: An array, containing numpy object for each Keras input. This is generally one Keras input for each user input variable.
  • y: A numpy object, containing the response variable (if one was provided)

Using input / output nubs

Setting up correctly formatted, heuristically ‘good’ input and output layers is often

  • Tedious
  • Time consuming
  • Difficult for those new to Keras

With this in mind, keras-pandas provides correctly formatted input and output ‘nubs’.

The input nub is correctly formatted to accept the output from auto.transform(). It contains one Keras Input layer for each generated input, may contain addition layers, and has all input piplines joined with a Concatenate layer.

The output layer is correctly formatted to accept the response variable numpy object.


Hey, I’m Brendan Herger, avaiable at Please feel free to reach out to me at 13herger <at> gmail <dot> com

If you’d like to know a bit about me, I enjoy bridging the gap between data science and engineering, to build and deploy data products.

I’ve enjoyed building a unique combination of machine learning, deep learning, and software engineering skills. In my previous work at Capital One and startups, I’ve has built authorization fraud, insider threat, and legal discovery automation platforms. In each of these cases I’ve lead a team of data scientists and data engineers to enable and elevate our client’s business workflow (and capture some amazing data).

When I’m not knee deep in a code base, I can be found traveling, sharing my collection of Japanese teas, and playing board games with my partner in Seattle.


  • PR title (#PR number, or #Issue if no PR)


  • Nothing here yet


  • Remove a lot of the unnecessary dependencies (#75)
  • Update dependencies to contemporary versions (#74)


  • Fix issue w/ PyPi conflict


  • Adding CI/CD and PyPi links, and updating contact section w/ about the author (#70)
  • Major rewrite / update of examples (#72)
    • Fixes bug in embedding transformer. Embeddings will now be at least length 1.
    • Add functionality to check if resp_var is in the list of user provided variables
    • Added better null filling w/ CategoricalImputer
    • Added filling unseen values w/ CategoricalImputer
    • Converted default transformer pipeline to use copy.deepcopy instead of copy.copy. This was a hotfix for a previously unknown issue.
    • Standardizing setting logging level, only in test base class and examples (when __main__)


  • Adding regression example w/ inverse_transformation (#64)
  • Fixing issue where web socket connections were being opened needlessly (#65)


  • Adding, with including files references in (#54)


  • Fixed poorly written text embedding index unit test (#52)
  • Added license (#49)


  • Lots of things happened. Break things and move fast