{
  "_ilex": {
    "architecture": "ilex.models.swift.model.SwiFT",
    "constructor_kwargs": {
      "c_multiplier": 2,
      "depths": [
        2,
        2,
        6,
        2
      ],
      "embed_dim": 36,
      "first_window_size": [
        4,
        4,
        4,
        4
      ],
      "img_size": [
        96,
        96,
        96,
        20
      ],
      "in_chans": 1,
      "last_layer_full_MSA": true,
      "num_heads": [
        3,
        6,
        12,
        24
      ],
      "patch_size": [
        6,
        6,
        6,
        1
      ],
      "window_size": [
        4,
        4,
        4,
        4
      ]
    },
    "format": "ilex",
    "framework_version": {
      "equinox": "0.13.8",
      "ilex": "0.0.0.dev0",
      "jax": "0.10.0",
      "jaxlib": "0.10.0",
      "numpy": "2.4.4",
      "safetensors": "0.7.0"
    },
    "has_state": false,
    "origin": "ilex-native",
    "weights_status": "bundled"
  },
  "authors": "Kim P. Y., Kwon J., Moon T., Cha J. (Seoul National University M.IN.D Lab + Connectome Lab)",
  "copyright": "SwiFT is copyright (c) Transconnectome / Seoul National\nUniversity 2023, Apache-2.0-licensed on the upstream code +\nthe in-tree pretrained checkpoints\n(github.com/Transconnectome/SwiFT). The ilex JAX / Equinox port\ncode is separately licensed under Apache-2.0 / GPL-3.0.\n",
  "data_type": "image",
  "description": "SwiFT (Kim, Kwon, Moon, Cha et al., arXiv:2307.05916) is a 4D\nSwin Transformer for fMRI BOLD volumes. The architecture is the\nMamba-free predecessor of NeuroSTORM -- same 4-stage Swin\ntopology (depths [2, 2, 6, 2], channels [36, 72, 144, 288],\n4D window [4, 4, 4, 4]) but using conventional WindowAttention4D\nmulti-head self-attention as the per-window mixer instead of\nNeuroSTORM's Mamba selective-scan SSM.\n\nv0 ships two variants:\n\n* ``contrastive`` -- the contrastive-pretraining checkpoint\n  (SimCLR-style backbone trained on a multi-cohort fMRI corpus).\n  Use as a frozen feature extractor for downstream tasks.\n* ``hcp-sex`` -- the supervised fine-tune on HCP-YA sex\n  classification. Same architecture as `contrastive`; weights\n  have been further fine-tuned end-to-end. Use as a starting\n  point for additional fine-tuning or as an HCP-specific feature\n  extractor.\n\nThe bundle stores only the backbone (``SwinTransformer4D``) --\nthe consumer-side SimCLR projection (``emb_mlp``) and the\ndownstream task heads (``clf.head``, ``reg.head``) are training-\ntime plumbing and are dropped at extract.\n",
  "equinox_version": "0.13.8",
  "ilex_version": "0.0.0.dev0",
  "image_classes": "Inputs are 4D fMRI BOLD volumes (channels-first per-sample\n``(1, 96, 96, 96, 20)`` -- 1 channel, 96^3 MNI152 brain voxels,\n20 time frames per clip). Preprocessing is consumer-side per\nthe upstream's pipeline.\n",
  "intended_use": "Research use for fMRI BOLD analysis where the consumer wants a\nlearned representation of a 4D volume rather than hand-engineered\nfeatures. Typical downstream consumers:\n\n* Phenotype regression (sex, age, IQ, cognitive measures) via a\n  linear probe on the backbone embedding.\n* Disease classification (HCP-YA, ABCD-derived tasks) via the\n  `hcp-sex` variant as initialisation.\n* Self-supervised representation comparisons against NeuroSTORM\n  (the Mamba-augmented successor) for ablation studies.\n\nThe ilex port wraps the JAX backbone forward only; consumers\nown the upstream's preprocessing pipeline (MNI152 registration,\n``(96, 96, 96)`` resampling, normalisation, 20-frame clips).\nThe bundle ships ``model.norm`` and ``model.head`` Linear\nparameters as static fields for state-dict round-trip integrity\nbut the JAX ``__call__`` does NOT invoke them -- matching the\nupstream ``SwinTransformer4D.forward`` which returns the deepest-\nstage feature map directly.\n\nNot a clinical diagnostic tool.\n",
  "jax_version": "0.10.0",
  "network_data_format": {
    "inputs": {},
    "outputs": {}
  },
  "numpy_version": "2.4.4",
  "pred_classes": "Output is the deepest-stage backbone feature map\n``(288, 2, 2, 2, 20)`` -- a per-clip, channels-first 4D\nrepresentation suitable for downstream task heads (linear probe,\nattention pool, MLP regression / classification). Consumers\ncan chain the bundle's stored ``model.norm`` + a custom head\nfor inference per the upstream's ``clf_mlp`` / ``reg_mlp``\npattern.\n",
  "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)."
  ],
  "schema": "https://github.com/Project-MONAI/MONAI-extra-test-data/releases/download/0.8.1/meta_schema_20240725.json",
  "task": "fMRI representation learning + downstream prediction via a 4D Swin Transformer backbone (the Mamba-free predecessor of NeuroSTORM)",
  "version": "0.0.0"
}