Time Series Forecasting with the NVIDIA Time Series Prediction Platform and Triton Inference Server

Technical Blogs & Events Technical Blog
1

Originally published at: https://developer.nvidia.com/blog/time-series-forecasting-with-the-nvidia-time-series-prediction-platform-and-triton-inference-server/

Learn how the Time Series Prediction Platform provides an end-to-end framework that enables users to train, tune, and deploy time series models.

2

how to use TSF with C++? And how do I customize components? Such as the loss function.

3

Hey, thank you for you interest! At the moment, the NVIDIA TSPP is just for Python. In order to customize the loss function (or any component), there are 2 main things needed.

  1. One needs the, in the case of loss function, source code: DeepLearningExamples/criterion.py at master · NVIDIA/DeepLearningExamples · GitHub. We keep our quantile and GLL loss functions there.
  2. One also needs the configuration file present in this directory: DeepLearningExamples/Tools/PyTorch/TimeSeriesPredictionPlatform/conf/trainer/criterion at master · NVIDIA/DeepLearningExamples · GitHub. Taking a look at the GLL config file: DeepLearningExamples/GLL.yaml at master · NVIDIA/DeepLearningExamples · GitHub, we just have a pointer to the criterion file linked above and the name of the loss function built. In order to select the criterion at runtime, simply add “trainer/criterion=<config file name>” at train time.

Let us know if you have further issues/questions and don’t hesitate to open an issue on the github: Issues · NVIDIA/DeepLearningExamples · GitHub

4

Is there a way to do Classification only with the TFT? I have a time series dataset that I need to do classification on and I thought I could potentially do classification on it with the TFT. Thoughts?