RAPIDS: GPU-Accelerated Data Analytics & Machine Learning

Key Takeaways

a data set that's very familiar to data scientists is a kaggle new york city taxicab data set we brought this data set into the omni-side gpu accelerated analytics platform where we can visualize and manipulate the tabular data in real time as well as quickly explore the data this data set contains records of taxi fares for a 12-month period and has information such as time passenger count as well as geographic locations the dot on the map are drop off locations where the size of the dot indicates how long the trip was and the color is the tip amount zero being blue and the max value being yellow you can see there's a lot of blue indicating zero tip which is a bit odd but as you explore the data you'll realize that the only way tips are being tracked is through credit card transactions which is not a true indicator of what trips truly are and is skewing the average tip amount i can show only credit card transactions by clicking on credit within this pie chart notice how the average tip changed from 11 to 20 percent the omni-side gpu accelerated analytics platform is based upon a gpu accelerated database that runs entirely in gpu memory it allows me as a business intelligence analyst to quickly gain insights and visualize large amounts of data i can build charts based upon this data so that i can further analyze maybe i wanted to show where taxi fares that are requiring larger vehicles here are the fares that had nine passengers for example i can add trips that had eight or seven passengers as well this allows me to gain an understanding where larger vehicles were requested it's very fast and a great way to allow data exploration across the enterprise this next part is where the data scientist comes into play what if you wanted to predict taxi fares maybe when somebody submitted a taxi fare request on an application like an app on their phone you wanted to give them an estimate and you wanted these estimates to be based upon real taxi fares that happened at the beginning of the month so that's where a data scientist comes in and you can use rapids to do those predictions rapids is an end-to-end open source ecosystem of gpu accelerated data science libraries that run on nvidia gpus a data scientist interacts with rapid libraries using a jupiter notebook and omnisci has jupiter notebook integration into their application the advantage to using rapids is that the data is coming directly from the omni-side gpu accelerated database no need to go back to csvs once the data is in the data frame we can use rapids and an etl workflow which imports inspects and cleans the data as well as selects a training set which is used for training a model to predict taxi fares this end-to-end workflow of data processing and model training is faster in using less hardware than cpu only so let's begin the rapids workflow here i go in and i import in libraries i connect to the data source and this brings in the data here is the data the next thing i'll do is inspect and clean up the data for example i have columns from 2014 that were named a certain way but they did not map the same in 2015. well through rapids you can concatenate fields and do data cleanup that's what we're doing here this is the data cleanup phase let's go on to the next part of the rapids workflow i can add new columns or new interesting features for example if i didn't have a way to understand if a taxi fare was on a weekend or maybe i wanted to know what day of the week it was without having to read it from a date field i can create that data and in this case i did i added it to the end of the data set i also used a distance calculator so i could see the distance of the trip itself as you can see this data has been added and exists as new columns within the data set the next part of the workflow is to pick a training set so here i will use 75 of the data to do training on and then i will use the remaining 25 to test my model against so i create that training set and we do the training it uses xg boost to do the training and lastly we test our model against that remaining 25 percent we then calculate a root mean square error which is basically a standard deviation to see how close we can get between the predictions and the ground truth the end result was that the taxi fare was within a two dollar range which i think is pretty fair estimate to give to taxi riders now that we have predictions we can bring this data back into the omni sci dashboard so that we can visualize and compare predictions against actual data the graph predicted amount versus actual fare provides this visual comparison blue are more accurate predictions whereas red are less accurate