Large Scale Machine Learning (Big Data): Difference between revisions

Another option that reduces memory usage even more, is to use [https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.SGDRegressor.html SGDRegressor] instead of <code>Ridge</code>. This class implements many types of Generalized Linear Models for regression, using vanilla Stochastic Gradient Descent as a solver. One caveat of using <code>SGDRegressor</code> is that it only works if the output is 1-dimensional (a scalar).

==Batch Learning== <!--T:21-->

In cases where your dataset is too large to fit in memory - or just large enough that it does not leave enough memory free for training - it is possible to leave your data on disk and load it in batches during training, similar to how Deep Learning packages work. <code>scikit-learn</code> refers to this as [https://scikit-learn.org/stable/computing/scaling_strategies.html out-of-core learning] and it is a viable option whenever an estimator has the <code>partial_fit</code> [https://scikit-learn.org/stable/computing/scaling_strategies.html?highlight=partial_fit#incremental-learning  method available]. In the examples below, we perform out-of-core learning by iterating over datasets stored on disk.