Training machine learning models on big datasets

Parallel Programming with Dask in Python

James Fulton

Climate Informatics Researcher

Dask-ML

import dask_ml

Speeds up machine learning tasks

Linear regression

An example of some data with a linear relationship between x and y.

Linear regression

A straight line has been fit to the data.

Linear regression

The distance between the straight line and the actual data points is highlighted.

Fitting a linear regression model

# Import regression model
from sklearn.linear_model import SGDRegressor

# Create instance of model
model = SGDRegressor()

# Fit model to data
model.fit(X, y)

# Make predictions
y_pred = model.predict(X)

Using a scikit-learn model with Dask

# Import regression model
from sklearn.linear_model import SGDRegressor

# Create instance of model
model = SGDRegressor()


# Import Dask-ML wrapper for model
from dask_ml.wrappers import Incremental


# Wrap model
dask_model = Incremental(model, scoring='neg_mean_squared_error')


# Fit on Dask DataFrames or arrays
dask_model.fit(dask_X, dask_y)  # not lazy

Fitting takes multiple iterations

The animation shows that the straight line fits more accurately after multiple iterations of fitting.

Training an Incremental model

# Loop through data multiple times
for i in range(10):
    dask_model.partial_fit(dask_X, dask_y)  # not lazy

Generating predictions

y_pred = dask_model.predict(dask_X)

print(y_pred)

dask.array<_predict, shape=(nan,), dtype=int64, chunksize=(nan,), chunktype=...>

print(y_pred.compute())

array([0.465557, 0.905675, 0.285214, ..., 0.249454, 0.559624, 0.823475])

Let's practice!

Parallel Programming with Dask in Python