Skip to content

Latest commit

 

History

History
56 lines (42 loc) · 2.78 KB

v0.1.8.md

File metadata and controls

56 lines (42 loc) · 2.78 KB

Features

  • Granite PatchTST-FM r2 support: The shared PatchTSTFM wrapper now supports ibm-granite/granite-timeseries-patchtst-fm-r2 (~385M). r2 adds Conformer blocks, overlapping patches (patch_stride=8), and a 99-quantile head on top of r1. Requires granite-tsfm>=0.3.9. GIFT-Eval CI replicates Granite-PatchTST-FM-r2 on m4_weekly/short.

    import pandas as pd
    from foundationforecast.models.patchtst_fm import PatchTSTFM
    
    df = pd.read_csv(
        "https://timecopilot.s3.amazonaws.com/public/data/air_passengers.csv",
        parse_dates=["ds"],
    )
    model = PatchTSTFM(
        repo_id="ibm-granite/granite-timeseries-patchtst-fm-r2",
        alias="Granite-PatchTST-FM-r2",
        context_length=8192,
        batch_size=32,
    )
    fcst = model.forecast(df, h=12, freq="MS")
    assert "Granite-PatchTST-FM-r2" in fcst.columns
  • Shared panel processing: Forecasters reuse a common panel path for multi-series inputs (Chronos, TimesFM, Tafsut, and others), reducing redundant work and improving multi-model orchestration throughput.

  • T0 beta support: The shared T0 wrapper now works with theforecastingcompany/t0-beta (~256M). Register it in gift-eval models.yaml; GIFT-Eval CI still replicates t0-alpha until leaderboard results exist for beta.

    import pandas as pd
    from foundationforecast.models.t0 import T0
    
    df = pd.read_csv(
        "https://timecopilot.s3.amazonaws.com/public/data/air_passengers.csv",
        parse_dates=["ds"],
    )
    model = T0(
        repo_id="theforecastingcompany/t0-beta",
        alias="t0-beta",
    )
    fcst = model.forecast(df, h=12, freq="D", quantiles=[0.1, 0.5, 0.9])
    assert "t0-beta" in fcst.columns

Fixes

  • FlowState frequency aliases: Normalize pandas-inferred aliases (h, 6h, 15min, MS) before calling granite-tsfm get_fixed_factor, so hourly and month-start series work when freq is omitted.

  • T0 / gift-eval CI: Bump tfc-t0 to >=0.5.0 and call T0Forecaster.predict(quantile_levels=...) instead of the removed quantiles= kwarg. Fixes gift-eval Modal runs after PyPI resolved tfc-t0 0.5.0 alongside the t0-beta release.

Dependencies

  • Bumped tfc-t0 from >=0.2.3 to >=0.5.0 (required for t0-beta; still supports t0-alpha).
  • Replaced timecopilot-granite-tsfm with granite-tsfm>=0.3.9 (required for PatchTST-FM r2 overlapping-patch inference).

Full Changelog: https://github.com/TimeCopilot/foundationforecast/compare/v0.1.7...v0.1.8