Patch Time Series Transformer in Hugging Face
Hugging Face Blog
PatchTST, a time series forecasting model based on Transformers, was added to Hugging Face with documentation showing how to train and apply it to electricity data using transfer learning. The model divides time series into patches of configurable length (the example uses patch_length=16 with context_length=512) and reduces computational complexity quadratically compared to processing full sequences. Users can now train PatchTST directly on datasets, perform zero-shot forecasting on new domains, and fine-tune pretrained models using the Hugging Face Trainer API.