Models and Checkpoints¶

Download a Checkpoint¶

Checkpoints are distributed separately from the Python package. Pass a name from the first column to FOXESModel and it’s downloaded the first time and cached in ~/.cache/helio_foxes (or HELIO_FOXES_CACHE). You can also download a file yourself, or run git lfs pull from a clone of the repository, and pass its path instead.

Name	Repository path	Model	Download
`"uncertainty"` (default)	`model_assets/UncertaintyModel.ckpt`	Current model: local + global attention, with Gaussian uncertainty	Weights
`"inverted-attention-mask"`	`model_assets/LegacyModels/inverted-attention-mask.ckpt`	Legacy deterministic, distant-patch attention (paper model)	Weights
`"localized-attention-mask"`	`model_assets/LegacyModels/localized-attention-mask.ckpt`	Legacy deterministic, local attention	Weights
`"no-attention-mask"`	`model_assets/LegacyModels/no-attention-mask.ckpt`	Legacy deterministic, full attention	Weights

Current Uncertainty Model¶

UncertaintyModel.ckpt consists of a local transformer backbone (3 x 3 attention) + a global attention transformer with an uncertainty MLP head.

No legacy correction is applied to it. The SXR normalization is stored inside the checkpoint, so you don’t need a separate normalization file.

Legacy Models¶

The three deterministic checkpoints were trained on an older version of the SXR targets and don’t produce uncertainty estimates. Loading one by name turns on the legacy correction automatically:

import helio_foxes

model = helio_foxes.FOXESModel("inverted-attention-mask")

If you load a legacy checkpoint from a file path instead, pass legacy_correction_enabled=True yourself, since FOXES can’t tell from the path. Pass legacy_correction_enabled=False to see the uncorrected output.

With the correction on, any prediction before 2016-11-19 00:00 UTC is multiplied by 0.7 (to remove a SWPC scaling bias), and anything after is left alone. The factor applies to both the full-disk flux and the patch map, so they still add up. Each prediction needs a timestamp for this to work; an Observation already carries one.