# GPU leftovers (do not run on the M3 8 GB Mac)

The MacBook Pro (2024, Apple M3, 8 GB) is for NumPy set algorithms only. Everything below needs CUDA (or at least the private JSON dump that the Mac cannot pull because of Hugging Face private-storage quota).

Use a Hugging Face **access token with read** on `jamie33` and write on `jamie33/924`. Do not paste the token into this file.

## G0 — dump real posteriors (blocker for `mode=real`)

**Why GPU/server:** the files are already on the private dataset / server cache. The laptop gets `403 Forbidden: Private repository storage limit reached for jamie33` on `p0_anchor_validation_safe/sub01.json`.

**Source (ours, not the public web):**

- Dataset: `jamie33/jmtseng-work-backup` (private)
- Files: `brain3d_refs/results/p0_anchor_validation_safe/sub{01..12}.json` (~400 KB each)
- Also: `.../p0_anchor_validation_safe/ensemble12.json`

If that tree is missing, re-infer with the validation-safe encoder (G1).

**Output to upload here:** `results/cpu/posteriors.npz` with `q` `(12,144,72)` and `y` `(144,)`. Then rerun:

```bash
python3 brainsets/cpu_run.py --mode real --npz results/cpu/posteriors.npz --out results/cpu/real_brainsets.json
```

## G1 — encoder re-inference (only if G0 JSON is gone)

**Download from the web / original release, not from us:**

- Neuro-3D / EEG-3D stimuli and EEG arrays: follow `jamie33/July` → `AAAI_SUPPLEMENT/BrainBits3D_code_and_data.zip` → `configs/DATA_LAYOUT.md`
- Typical local layout on the training box: EEG npy under the Neuro-3D data root; checkpoints under the private backup `brain3d_refs/`

**From us (token read):**

- `jamie33/jmtseng-work-backup` encoder checkpoints (`encoder_validation_safe` tree)
- Public overflow: `jamie33/jmtseng-work-backup-public/brain3d_refs/results/collective_set/*/model.pt`

This is **GPU**. Batch size 8, 12 subjects, 144 objects. Do not start this on the 8 GB Mac.

## G2 — 72×72 train-only Chamfer (BrainSets + geometry)

**From us (public):**

- `jamie33/jmtseng-work-backup-public` → `brain3d_refs/results/p0_category_geometry/category_00.pt` … `category_71.pt`

Export a `D72.npy` `(72,72)` on a machine with `torch`, upload to `jamie33/924/results/cpu/D72.npy`. Then CPU Chamfer-on-sets can run on the Mac.

## G3 — leftover 9.24 BrainBits GPU jobs (not BrainSets, still missing in tables)

Tracked in `PencilHu/JAMIETSENG` `9.24torun` and `jamie33/923`:

| ID | What | Data |
|---|---|---|
| T1 | Wrong-category F@0.1 / EMD | generated meshes + category means; **ours** under private backup `results/` |
| T2 | fMRI-Shape instance Chamfer | fMRI-Shape release **from the web** + our encoder |
| T3 | True Gaussian-splatting family (not Gaussian-noise proxy) | generator ckpt **ours** private backup |
| T8 | Wrong-category qualitative figures | same as T1 |

T4–T7 (encoder 12-sub, NLL/Brier, conformal multiseed, best-of-5) are CPU-or-light once G0 posteriors exist; run them with `cpu_run.py` / ToDo2 scripts, not on a fat GPU.

## G4 — if a file is not on our HF

Say so in the job log: **download from the original dataset page** (Neuro-3D / fMRI-Shape). Do not invent a `jamie33` path. The token only reads what we already uploaded.