You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This bundle contains data derived from the Natural Scenes Dataset (NSD), which is released under a data-use agreement, and stimulus images from MS-COCO. Access is granted for research use within or in collaboration with GISense Lab. By requesting access you agree not to redistribute these files.
Log in or Sign Up to review the conditions and access this dataset content.
Mind2Place — fMRI onboarding bundle
A small, self-contained slice of the GISense Lab Mind2Place fMRI experiment, sized so that a new lab member can reproduce the brain-to-image pipeline in a free Google Colab session.
Pair this with Mind2Place_fMRI_Onboarding.ipynb from the (private) GISense/Mind2Place repo.
Contents
| Path | What it is |
|---|---|
mind2place_fmri_attnfix_ema.pth |
Trained stage-4 diffusion model, EMA weights in fp16 (1.2 GB), stored under the state_dict key so the repo's 5inference_sub.py loads it directly. Slimmed from the 9.6 GB training checkpoint of the repaired retrain (see below), not from the checkpoint originally released with the repo. Architecture SounDiT_MoE_builder: depth 28, hidden 1152, patch 2, 16 heads, ~606 M parameters. |
data.tar |
Everything below, as one archive. The notebook extracts it after download; shipping ~1460 loose files instead causes rate-limited stalls. |
data/meta.csv |
Trial table: pair_id, split, subject, nsd_id, coco_id, roi, and paths |
data/fmri-1024/<pair_id>.npy |
1024-D fMRI embedding — the model's only conditioning input |
data/fmri-roi/<pair_id>.npy |
Raw ROI voxels, (3 repeats, ~2077 voxels), z-scored |
data/img/<nsd_id>.jpg |
The stimulus photograph (shared across the 4 subjects) |
data/img-vae/<pair_id>.npy |
SD-VAE latent of the stimulus, (8, 32, 32) = 4 mean + 4 std |
data/captions/<nsd_id>.json |
The five COCO captions for that photograph |
data/fmri_mean_1024.npy |
Mean training embedding — the "average brain" null used for guidance |
data/train/ |
200 training trials, for demonstrating the training loop |
data/eval/ |
ResNet-50 Places365 scene classifier, used for the identification metric |
data/demo/ |
8 trials, for the 5inference_sub.py demo (that script has no --limit) |
60 photographs × 4 subjects (subj01, 02, 05, 07) = 240 test trials. Every photograph is
present for all four subjects, so you can decode the same stimulus from four different brains. The
full experiment has 541 test photographs and ~17 900 training trials.
Provenance
- fMRI: Natural Scenes Dataset, 7T, ROI
nsdgeneral ∩ floc-places - Stimuli and captions: MS-COCO
- Scene classifier: Places365 ResNet-50 (MIT CSAIL, CC BY)
- Image decoder used by the notebook:
stabilityai/sd-vae-ft-ema
Why this is a retrain
The stage-4 checkpoint originally released with the repo is near-unconditional: swapping the fMRI vector for an unrelated subject's leaves the output almost unchanged (latent correlation 1.00), while changing only the random seed changes it completely. Its Places365 identification accuracy over the 60 scenes here is 0.017 at the repo's default settings — exactly chance.
The cause is in MultiHeadCrossAttention.forward, which flattens the batch into a single sequence
for the fast attention path and only restores per-sample boundaries when a padding mask is supplied.
The fMRI path supplies none, so at --batch-size 32 every sample attended to every other sample's
conditioning vector. The operation is a no-op at batch size 1, so inference was never affected and
only training was corrupted — and the failure is invisible in the loss curve (0.1749 repaired vs
0.1752 broken over a full 100-epoch run).
This bundle ships a retrain with that fixed:
| guidance | this checkpoint | originally released | chance |
|---|---|---|---|
| 1.0 (repo default) | 0.183 | 0.017 | 0.017 |
| 2.0 | 0.250 | 0.033 | 0.017 |
| 4.0 | 0.350 | 0.083 | 0.017 |
Identification accuracy, 60 scenes, subj01. The repaired model recovers scene identity — roughly what kind of place was viewed — not the photograph. It does not recover perceptual qualities. Quote it as scene decoding, not image reconstruction.
Citation
Somang Kim, Junbo Wang, Yuhao Kang. Mind2Place: Reconstructing Viewed Places from Brain Activity. GISense Lab, The University of Texas at Austin. Paper in preparation.
- Downloads last month
- 15