mirror of
https://github.com/google-research/timesfm.git
synced 2026-10-02 03:24:55 +08:00
winklemad
af8f5952e2
The force_flip_invariance path reverses the quantile axis of the flipped forecast before averaging it against the original. The prefill outputs and the quantile spreads are flipped, but the autoregressive branch was appended without the flip. For horizons past the first output patch the flipped quantile channels ended up misaligned, so the quantile bands in that region were wrong. The mean and median are flip fixed-points, so the point forecast was unaffected. The Flax implementation flips flipped_ar_outputs before concatenating; this brings the PyTorch path in line. Adds a regression test asserting the flip-invariance identity holds across the full horizon.