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

All estimators in <code>Snap ML</code> support GPU acceleration, with one or multiple GPUs. For Single GPU training, simply set the parameter <code>use_gpu=True</code>. For multiple GPU training, in addition to setting <code>use_gpu</code>, pass a list containing the GPU IDs available to your job to <code>device_ids</code>. For example, inside a job that requested 2 GPUs, set <code>device_ids=[0,1]</code> to use both GPUs for training. The following example extends the performance comparison from the previous section to include training on GPU with <code>Snap ML</code>, this time training an SVM classifier with a non-linear kernel.

==Out-of-Memory Training== <!--T:58-->

All estimators in <code>Snap ML</code> use first-order iterative solvers, similar to SGD, by default. It is thus possible to perform training in batches and avoid loading entire datasets in memory. Unlike <code>scikit-learn</code> however, <code>Snap ML</code> accepts memory-mapped numpy arrays as inputs directly.