timesfm 2.0 500m pytorch

Introduction

TimesFM (Time Series Foundation Model) is a pretrained model developed by Google Research for time-series forecasting. It supports univariate time series forecasting with a focus on point forecasts and offers some experimental features like quantile heads.

Architecture

TimesFM can handle context lengths up to 2048 time points and various horizon lengths. It uses a decoder-only architecture, allowing the model to manage univariate time series forecasting by filling in missing values through linear interpolation. It operates optimally with contiguous context and frequency-consistent horizon data.

Training

The TimesFM 2.0 model was pretrained using a subset of datasets from the LOTSA pretraining data, which includes a variety of data sources such as residential power loads, traffic data, and more. The training process aimed to equip the model with robust capabilities to handle various time-series forecasting tasks.

Guide: Running Locally

1. Installation

   • Install the timesfm library from the GitHub repository: TimesFM GitHub.

2. Initialize the Model

   import timesfm
   
   tfm = timesfm.TimesFm(
       hparams=timesfm.TimesFmHparams(
           backend=<backend>,
           per_core_batch_size=32,
           horizon_len=128,
           input_patch_len=32,
           output_patch_len=128,
           num_layers=50,
           model_dims=1280,
           use_positional_embedding=False,
       ),
       checkpoint=timesfm.TimesFmCheckpoint(
           huggingface_repo_id="google/timesfm-2.0-500m-pytorch"),
   )
   

3. Perform Inference

   • Utilize the tfm.forecast() function for array inputs or tfm.forecast_on_df() for pandas dataframe inputs.
   • Ensure the frequency input is consistent with the model’s categorical expectations (e.g., 0 for high frequency).

4. Compute Requirements

   • Given the model's complexity, using a cloud GPU is recommended for efficient performance.

License

The TimesFM model and its associated code are distributed under the Apache-2.0 license.