We know that the sklearn library is a collection of a variety of Machine learning algorithms that can help us to train our data. The problem arises when the training data is huge because training a huge amount of data takes immense time. To increase the speed of training our model by one hundred and fifty times faster than using sklearn data scientists have made it possible with cuML.

Before learning cuML, we should have acquaintance with RAPIDS. Let’s see what is RAPIDS. 

RAPIDS

RAPIDS is a set of open-source software libraries which help us to execute complete data science projects on a GPU (Graphics processing unit). It was developed by NVIDIA and was licensed under Apache 2.0. It was developed to execute projects that have a very large amount of data in optimal time. Data scientists put forth their thoughts and developed this based on extensive data science experience. 

RAPIDS uses NVIDIA CUDA to compute low-level optimization. And with the help of user-friendly python interfaces, it exposes high bandwidth memory speed and GPU parallelism. It also supports multi-node, multi-GPU deployments and enables accelerating processing and training of a huge amount of data sets. RAPIDS projects include cuDF it is a manipulation library much similar to panda’s data frame, cuML is a collection of machine learning libraries that provide GPU versions of algorithms that are available in scikit-learn. a graph is a collection of graph analytics libraries. RAPIDS get along very well with the python libraries. 

cuML

CuML is the collection of fast, GPU accelerated machine learning algorithms that are particularly designed for analytical tasks and data science. The application programming interface (API) of cuML is way similar to sklearns. It provides programmers with a fit predict transform paradigm without even writing a program for it on a GPU. 

When the amount of data increases or becomes huge then the algorithms that are running even on CPU become cumbersome and slow. To avoid this RAPIDS provides the users a well-planned approach in which we will initially load the in GPU and then perform the tasks directly on it. CuML is completely open-source and it welcomes new contributors, programmers, and users always.  

The image above depicts the speed at which RAPIDS cuML executes the machine learning algorithms. An algorithm runs six hundred times faster on RAPIDS GPU than compared to a normal CPU. To install cuML we need to have a Linux-like operating system. Installation in windows may be possible in near future.

The image above depicts the architecture of the cuML library along with GPU memory. 

In this article, we will compare the performance of two different libraries with a Support vector machine (SVM) and a random forest classifier.

Installation

This link will help you to install the cuML package. Rapids cuML installation. We can either install all packages or just cuML because cuML alone will suffice the need. If space permits, we can install cuDF along with cuML because cuDF compliments are cuML very well as it is a GPU data frame. Select the options that your system supports.

Data Preparation:

Here we are creating a data set on which we train our two models. Since cuML works better on large data sets here we will create data with 50000 rows of data using sklearn. datasets. 

We have to convert the datatypes to float 32 because cuML only accepts the input format in np.float32. 

Support Vector Machine: 

Here, we will write the support vector machine function to train our model. We are installing the SVM function from the cuML library that speeds up the training. The last two lines in the code represent the time taken by sklearn SVM and cuML SVM. 

CuML’s support vector machine model is 2.5 times faster than sklearn’s SVM model. 

Random Forest Classifier: 

Random forest classifier creates a group of decision trees out of randomly selected subset data from the training set. It then gathers the votes from each decision tree to declare the final class of the testing object. Now let’s see the code of the random forest classifier algorithm with the help of the cuML library. 

CuML’s Random forest classifier is much faster than sklearns random forest classifier. Approximately it is 60 times faster. If sklearn’s Random forest algorithm takes thirty seconds then cuML’s Random forest classifier algorithm takes just 0.5 seconds to train a model.  

Conclusion

In this article, I have compared two different models with the help of two different libraries. We can trust the cuML library for training a large amount of data within no time. 

Thanks for reading!