CuPy Functionality#

NNSOM leverages CuPy, a library for GPU-accelerated computing with a NumPy-compatible interface, to enhance the performance of SOM computations. By utilizing the parallel processing power of GPUs, NNSOM can process large datasets significantly faster than using CPU-based implementations. This document outlines how to use NNSOM with CuPy and the benefits it offers.

Requirements#

Before using CuPy with NNSOM, ensure you have the following:

A compatible GPU (NVIDIA CUDA-capable GPU)
CUDA Toolkit (check compatibility with your CuPy version)
CuPy library installed in your Python environment

To install CuPy, you can use pip:

pip install cupy

Usage#

To utilize CuPy with NNSOM, import the CuPy-enabled version of the SOM:

from NNSOM.som_gpu import SOMGpu

Here is an example of initializing and training a SOM on GPU:

dimensions = (10, 10)  # Dimensions of the SOM grid
som = SOMGpu(dimensions)  # Initialize SOM
data = np.random.rand(1000, 10)  # Generate some random data
som.init_w(data)  # Initialize weights
som.train(data)  # Train SOM

Benefits#

Using CuPy with NNSOM provides the following benefits:

Faster computation: Leveraging GPU for the training and simulation processes of SOMs speeds up computations, especially beneficial for large datasets.
Efficient memory usage: CuPy efficiently manages memory on GPUs, which can be more restrictive than CPU memory.
Seamless integration: Works with existing NumPy arrays and functions, minimizing the need to modify your existing codebase.

Limitations#

While CuPy accelerates computations, there are a few limitations to consider:

Hardware dependency: Requires an NVIDIA GPU; not compatible with other GPU brands.
Software compatibility: Ensure compatibility with CUDA versions and the operating system.
Debugging and profiling: GPU debugging can be more complex compared to CPU.

Automatic CuPy Detection#

The SOMPlots class in NNSOM is designed to automatically detect if CuPy is available in the user’s environment. If CuPy is detected, SOMPlots will utilize GPU acceleration for all applicable operations, optimizing performance without requiring manual configuration from the user.

from NNSOM.som_plots import SOMPlots

# SOMPlots instantiation automatically checks for CuPy
som_plots = SOMPlots(dimensions=(10, 10))
data = np.random.rand(1000, 10)  # Example data
som_plots.init_w(data)  # Initialize weights with potential GPU acceleration
som_plots.train(data)  # Train SOM with potential GPU acceleration

If CuPy is not available, SOMPlots will automatically fall back to a CPU-based implementation, ensuring that SOM computations can still be performed, albeit at a slower speed compared to GPU acceleration. This feature provides a seamless user experience by adapting to the available system resources.

Conclusion#

Integrating CuPy with NNSOM allows users to exploit the computational power of GPUs, making it a suitable choice for large-scale and real-time data processing scenarios in machine learning and