“Many-model” Time Series Forecasting: Scaling PyTorch Training Across 1000s...- V. Sridhar & S. Chen

“Many-model” Time Series Forecasting: Scaling PyTorch Training Across 1000s of Models - Vinay Sridhar & Shulin Chen, Snowflake Modern time series forecasting increasingly relies on transformer models, often deployed in scenarios where datasets are partitioned such as across product segments in retail, leading to the need to create a time-series model per segment, often 100s to 1000s at the same time. Training these sophisticated models involves large datasets and complex transforms, while dealing with resource limitations. This leads to training such 1000s of PyTorch models sequentially resulting in long training times and thereby much fewer iterations. This talk presents a solution that demonstrates how to train thousands of such PyTorch models in parallel leveraging Ray’s advanced scheduling and distributed processing while utilizing resources efficiently. By seamlessly integrating highly parallel data loading with distributed processing, we establish a pipeline that optimizes for cost-efficiency, resource utilization, and execution time. This approach empowers users to train larger, more complex transformer models on significantly bigger datasets and more often leading to a higher velocity of developing and deploying advanced, ML-powered forecasting applications with superior predictive quality.