We can train models on large datasets one batch at a time. Many Scikit-Learn estimators implement a partial_fit
method to enable incremental learning in batches.
est = SGDClassifier(...)
est.partial_fit(X_train_1, y_train_1)
est.partial_fit(X_train_2, y_train_2)
...
The Scikit-Learn documentation discusses this approach in more depth in their user guide.
This notebook demonstrates the use of Dask-ML's Incremental
meta-estimator, which automates the use of Scikit-Learn's partial_fit
over Dask arrays and dataframes. Scikit-Learn handles all of the computation while Dask handles the data management, loading and moving batches of data as necessary. This allows scaling to large datasets distributed across many machines, or to datasets that do not fit in memory, all with a familiar workflow.
This example shows ...
partial_fit
with the Dask-ML Incremental meta-estimatorAlthough this example uses Scikit-Learn's SGDClassifer, the Incremental
meta-estimator will work for any class that implements partial_fit
and the scikit-learn base estimator API.
We first start a Dask client in order to get access to the Dask dashboard, which will provide progress and performance metrics.
You can view the dashboard by clicking on the dashboard link after you run the cell
from dask.distributed import Client
client = Client(n_workers=4, threads_per_worker=1)
client
We create a synthetic dataset that is large enough to be interesting, but small enough to run quickly.
Our dataset has 1,000,000 examples and 100 features.
import dask
import dask.array as da
from dask_ml.datasets import make_classification
n, d = 100000, 100
X, y = make_classification(n_samples=n, n_features=d,
chunks=n // 10, flip_y=0.2)
X
For more information on creating dask arrays and dataframes from real data, see documentation on Dask arrays and Dask dataframes.
We split our dataset into training and testing data to aid evaluation by making sure we have a fair test:
from dask_ml.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y)
X_train
This dataset is small enough to fit in distributed memory, so we call dask.persist
to ask Dask to execute the computations above and keep the results in memory.
X_train, X_test, y_train, y_test = dask.persist(X_train, X_test, y_train, y_test)
If you are working in a situation where your dataset does not fit in memory then you should skip this step. Everything will still work, but will be slower and use less memory.
Calling dask.persist
will preserve our data in memory, so no computation will be needed as we pass over our data many times. For example if our data came from CSV files and was not persisted, then the CSV files would have to be re-read on each pass. This is desirable if the data does not fit in RAM, but not slows down our computation otherwise.
We pre-compute the classes from our training data, which is required for this classification example:
classes = da.unique(y_train).compute()
classes
We make the underlying Scikit-Learn estimator, an SGDClassifier
:
from sklearn.linear_model import SGDClassifier
est = SGDClassifier(loss='log', penalty='l2', tol=1e-3)
Here we use SGDClassifier
, but any estimator that implements the partial_fit
method will work. A list of Scikit-Learn models that implement this API is available here.
We now wrap our SGDClassifer
with the dask_ml.wrappers.Incremental
meta-estimator.
from dask_ml.wrappers import Incremental
inc = Incremental(est, scoring='accuracy')
Recall that Incremental
only does data management while leaving the actual algorithm to the underlying Scikit-Learn estimator.
Note: We set the scoring parameter above in the Dask estimator to tell it to handle scoring. This works better when using Dask arrays for test data.
Incremental
implements a fit
method, which will perform one loop over the dataset, calling partial_fit
over each chunk in the Dask array.
You may want to watch the dashboard during this fit process to see the sequential fitting of many batches.
inc.fit(X_train, y_train, classes=classes)
inc.score(X_test, y_test)
Calling .fit
passes over all chunks our data once. However, in many cases we may want to pass over the training data many times. To do this we can use the Incremental.partial_fit
method and a for loop.
est = SGDClassifier(loss='log', penalty='l2', tol=0e-3)
inc = Incremental(est, scoring='accuracy')
for _ in range(10):
inc.partial_fit(X_train, y_train, classes=classes)
print('Score:', inc.score(X_test, y_test))
Finally we can also call Incremental.predict
and Incremental.score
on our testing data
inc.predict(X_test) # Predict produces lazy dask arrays
inc.predict(X_test)[:100].compute() # call compute to get results
inc.score(X_test, y_test)
In this notebook we went over using Dask-ML's Incremental
meta-estimator to automate the process of incremental training with Scikit-Learn estimators that implement the partial_fit
method. If you want to learn more about this process you might want to investigate the following documentation: