---
library_name: ilex
tags:
- jax
- equinox
- ilex
- neuroimaging
- fmri
license: apache-2.0
license_link: https://www.apache.org/licenses/LICENSE-2.0
---

# SwiFT (4D Swin fMRI Transformer, Mamba-free predecessor of NeuroSTORM) -- SwiFT contrastive-pretrained backbone

## Description

SwiFT (Kim, Kwon, Moon, Cha et al., arXiv:2307.05916) is a 4D
Swin Transformer for fMRI BOLD volumes. The architecture is the
Mamba-free predecessor of NeuroSTORM -- same 4-stage Swin
topology (depths [2, 2, 6, 2], channels [36, 72, 144, 288],
4D window [4, 4, 4, 4]) but using conventional WindowAttention4D
multi-head self-attention as the per-window mixer instead of
NeuroSTORM's Mamba selective-scan SSM.

v0 ships two variants:

* ``contrastive`` -- the contrastive-pretraining checkpoint
  (SimCLR-style backbone trained on a multi-cohort fMRI corpus).
  Use as a frozen feature extractor for downstream tasks.
* ``hcp-sex`` -- the supervised fine-tune on HCP-YA sex
  classification. Same architecture as `contrastive`; weights
  have been further fine-tuned end-to-end. Use as a starting
  point for additional fine-tuning or as an HCP-specific feature
  extractor.

The bundle stores only the backbone (``SwinTransformer4D``) --
the consumer-side SimCLR projection (``emb_mlp``) and the
downstream task heads (``clf.head``, ``reg.head``) are training-
time plumbing and are dropped at extract.


## Intended use

fMRI backbone for representation learning. Input: (1, 96, 96, 96, 20) MNI152 BOLD clip (consumer handles registration + normalisation). Output: deepest-stage backbone feature map (288, 2, 2, 2, 20) suitable for downstream linear probes or MLP heads. The bundle ships model.norm + model.head as static fields for state-dict round-trip but JAX __call__ does NOT invoke them (matching the upstream's encoder-only forward).

## Usage

```python
from ilex.models.swift import SwiFT
model = SwiFT.from_pretrained('ilex-hub/swift.contrastive.1')
```

## Authors

Kim P. Y., Kwon J., Moon T., Cha J. (Seoul National University M.IN.D Lab + Connectome Lab)

## Citation

Kim P. Y., Kwon J., Moon T., Cha J. et al. (2023). SwiFT -- Swin 4D fMRI Transformer. arXiv 2307.05916.

### References

- Kim P. Y., Kwon J., Moon T., Cha J. et al. (2023). SwiFT -- Swin 4D fMRI Transformer. arXiv:2307.05916.
- Liu Z. et al. (2021). Swin Transformer -- Hierarchical Vision Transformer using Shifted Windows. arXiv:2103.14030.
- Upstream code + weights -- github.com/Transconnectome/SwiFT (Apache-2.0).

## License

HF Hub license tag: `apache-2.0`

**Effective terms:** Apache-2.0 (both upstream code and the in-tree pretrained .ckpt checkpoints at github.com/Transconnectome/SwiFT). The ilex JAX / Equinox port code is separately Apache-2.0 / GPL-3.0.

Upstream license reference: https://www.apache.org/licenses/LICENSE-2.0

### Copyright

SwiFT is copyright (c) Transconnectome / Seoul National
University 2023, Apache-2.0-licensed on the upstream code +
the in-tree pretrained checkpoints
(github.com/Transconnectome/SwiFT). The ilex JAX / Equinox port
code is separately licensed under Apache-2.0 / GPL-3.0.


## Upstream source

Original weights / reference implementation: https://github.com/Transconnectome/SwiFT

## Provenance

This artefact was produced by [ilex](https://github.com/hypercoil/ilex)'s
save/load pipeline. The architecture is implemented in
`ilex.models.swift.SwiFT` and the weights have been converted
from their upstream format. See the upstream source above
for the canonical reference.