Fraud Detection With Categorical XGBoost

This example notebook shows how to train and deploy an XGBoost model with categorical features in Triton using the FIL backend. We begin by training two separate models on a fraud detection dataset with categorical variables: one small model designed to maximize runtime performance and one larger model designed to maximize accurate and precise detection of fraud. We then deploy both models on CPU and GPU and compare their performance using Triton’s perf_analyzer. Based on these results, we see that GPU deployment opens up the possibility of deploying a much larger and more accurate fraud model with higher throughput while also keeping to a tight latency budget.

Running the notebook

In order to launch the Triton server, you will need Docker installed on your system. The rest of the notebook also requires a few Python dependencies. To easily install these additional dependencies, you may make use of the provided conda environment file as follows:

conda env create -f environment.yml

You may then activate the conda environment and run the notebook as usual:

conda activate triton_example
jupyter notebook

The Jupyter interface should now be accessible from a browser, and you can follow the instructions within the notebook itself from there.