Skip to content

Latest commit

 

History

History
155 lines (111 loc) · 6.96 KB

README.md

File metadata and controls

155 lines (111 loc) · 6.96 KB
Raw

Read our documentation!

dato is an open source library that provides a rapid, declarative ecosystem for reproducible data science within python. dato accomplishes this by (1) enabling piping with >> and (2) unifying common data science libraries under a common syntax.

df >> GroupBy('country') >> Sum >> Hist('revenue', col='age')

Dato has four major components:

  • dato.base.Pipeable Decorator that enables piping with >>.
  • dato.process Sub-module with pipe-compatible pandas operations.
  • dato.plot Sub-module with pipe-compatible plotting operations, following a consistent pandas-inspired syntax with seaborn-esque extended functionality.
  • dato.ml(in development) Simplifies and standardizes syntax across popular ML libraries.

Installation

pip install dato

Basic usage: the Pipeable class

dato is meant to be flexible, and therefore can accept (almost) anything as input. Creating custom functions compatible with the dato framework is therefore quite easy. The class dato.base.Pipeable can wrap or decorate any method to enable compatibility with the >> operator. For example:

from dato import Pipeable

@Pipeable
def Func(*args, **kwargs):
    return func(*args, **kwargs)

Or even more concisely, any existing function func that you'd like to use with dato can be trivially implemented as follows:

Func = Pipeable(func)

The entire piping framework is incredibly simple (it only takes up around 40 lines of code), and can be found in dato.base.Pipeable. If you write a custom function, please consider making a pull request. Happy piping!

Some illustrative examples

We used this framework to implement data science-specific methods to improve QOL when performing repetitive data-related tasks (and to illustrate the potential of dato). Our biggest pain points in this domain have been:

  • Remembering pandas syntax and defaults.
  • Styling matplotlib/pandas/seaborn visualizations.
  • Remembering scikit-learn model creation syntax, best practices, and evaluation metrics.

We have therefore focused on wrapping and consolidating these libraries. We provide a few examples for each of these use cases below, but for full functionality, see the documentation.

Full integration is currently still being built out, so contributions are very welcome. To contribute, please see the Contribution section of the docs.

More readable pandas

A common pattern in exploratory analyses is to aggregate one value with respect to another. In pandas, this is typically accomplished as follows:

df['date'] = pd.to_datetime(df.date)
gb = df.groupby('date').sum()['sale_value'].plot()

While pandas has already done an incredible amount of heavy lifting to make this aggregation syntactically quite simple, it still takes some thought, trial, and error to correctly write the above few commands. The same command in dato can be rewritten as follows:

from dato import *
df >> ToDatetime('date') >> GroupBy('date') >> Sum('sale_value') >> Plot()

Auto-styled matplotlib

matplotlib is a staple in data visualization, primarily for its flexibility and speed. However, generating a presentation-ready plot takes an extraordinarily long time with substantial cognitive load, owing to library-specific syntax and an immense styling dictionary (mpl.rcParams). Below is an example from ./examples/sample.ipynb here to illustrate how cumbersome this can be.

plt.figure(figsize=(8.5,5.2))
plt.scatter(a.lat, a.lng, alpha=0.5, s=100)
plt.scatter(b.lat, b.lng, alpha=0.1, s=100)
plt.scatter(c.lat, c.lng, alpha=0.1, s=100)
plt.grid('on', linestyle=':')
plt.rcParams.update({'font.size': 15})

While this script isn't particularly long, each argument (s for scatter, the 'on' arg for grid, the keys for rcParams), in our experience, warrants a stackoverflow crawl. Even with almost a decade of experience using matplotlib, it still takes about 5 minutes to write up that snippet.

We therefore implement some improved basic styling to reduce the overhead of using matplotlib (granted, style is incredibly subjective, and you may find our decisions horrendous). At the least, we hope that this will improve the readability of your code, and at best, reduce the need to use any matplotlib styling.

from dato import Scatter
(a.lat, a.lng) >> Scatter
(b.lat, b.lng) >> Scatter(alpha=0.1)
(c.lat, c.lng) >> Scatter(alpha=0.1)

Cleaner sklearn

We also provide limited, but ever-growing ML tooling, wrapping sklearn and xgboost. We do not intend this to replace existing libraries, but to more quickly test the feasibility of a model.

A disclaimer regarding ML: while, in general, dato does not modify outputs, because of the complex, branching nature of machine-learning workflows (creating and holding onto a validation set, for example), we created a hidden _ModelSpec method that holds model-related information (the train and test sets). A _ModelSpec class object (here represented as m) contains the following attributes:

  • m.train: the training data.
  • m.test: the test data.
  • m.estimator: the underlying scikit-learn estimator.

A typical full-on ML effort (without any hyperparameter tuning) can be condensed as follows:

import numpy as np
import sklearn
df = pd.merge(users, purchases, on='id_user')
le = sklearn.preprocessing.LabelEncoder()
df['city'] = le.fit_transform(df.city)
df = df[['population', 'density', 'sale_value', 'city']]
X = df[['population', 'density', 'city']]
y = df['sale_value']
X_train, X_test, y_train, y_test = sklearn.model_selection.train_test_split(X, y)
reg = sklearn.linear_model.LinearRegression()
reg.fit(X_train, y_train)
y_pred = reg.predict(X_test)
y_train_pred = reg.predict(X_train)
print('Mean squared error:', sklearn.metrics.mean_squared_error(y_train, y_train_pred))
print('Mean absolute error:', sklearn.metrics.mean_absolute_error(y_train, y_train_pred))
print('Root mean squared error:', np.sqrt(sklearn.metrics.mean_squared_error(y_train, y_train_pred)))
print('Mean squared error:', sklearn.metrics.mean_squared_error(y_test, y_pred))
print('Mean absolute error:', sklearn.metrics.mean_absolute_error(y_test, y_pred))
print('Root mean squared error:', np.sqrt(sklearn.metrics.mean_squared_error(y_test, y_pred)))

But it's still clearly quite cumbersome, even without the imports. With dato tooling, this entire process can be condensed as follows:

from dato import *
modelspec = (users, purchases) \
    >> Merge(on='id_user') \
    >> Select('population', 'density', 'sale_value', 'city') \
    >> InitModel(label='sale_value') \
    >> LabelEnc(columns=['city']) \
    >> TrainTestSplit \
    >> LinearReg