README.md

# I-JEPA deepfake detectors (per generator)

Fine-tune **`facebook/ijepa_vith14_1k`** (ViT-H/14) as a binary **fake vs real** detector, **one model per generator**.

## Protocol

For each trainable generator `G` ∈  
`ADM, BigGAN, Midjourney, VQDM, glide, wukong, stable_diffusion_v_1_4, stable_diffusion_v_1_5, imagenetv3`:

| Split | Fakes | Reals |
| --- | --- | --- |
| Train | `dataset/G/train/ai` | `dataset/train_nature` (shared) |
| Eval | every `dataset/*/val/ai` | matching `val/nature` or `val/real` |

- **Never train** on `flux2` or `drifting` (test-only / OOD).
- Labels: `0 = real`, `1 = fake`.
- A 10% train holdout is used only for in-run monitoring (not a GenImage holdout).

## Dual-GPU launch (not DDP)

`run_all.py` starts **one process per GPU**. Each process:

1. sets `CUDA_VISIBLE_DEVICES=<gpu>`
2. trains + evaluates **one** detector
3. shuts down DataLoader workers and frees CUDA
4. pulls the next unfinished generator until the queue is empty

Training uses a **5% balanced holdout** from `train/ai` + `train_nature` (metric=`acc`, patience=4, checks every 1000 steps, max 1 epoch). Classes are balanced 50/50.

**`imagenetv3`-only extras** (other generators unchanged):
- no patience (full epoch)
- LR hypertune on a small split (`2k/class`, 800 steps, `1e-5…1e-4`)
- after training, pick **last-epoch vs best-holdout** on its own `val` split


```bash
cd /home/themis/workspace_firas/appli/i-jepa
conda activate test   # or any env with torch + transformers
pip install -r requirements.txt

# full sweep on GPU 0 and GPU 1
python run_all.py --gpus 0,1

# retrain ADM only (overwrite broken checkpoint)
python train.py --generator ADM --device cuda:0 --force

# smoke test
python run_all.py --gpus 0,1 --limit-per-class 200 --epochs 1

# subset
python run_all.py --gpus 0,1 --generators ADM,BigGAN
```

Single-generator (one GPU):

```bash
python train.py --generator ADM --device cuda:0
```

## Metrics

- **Accuracy**: argmax over 2-class logits
- **AUROC**: ranking score = **MSP** (`max` softmax probability), against labels `0=real`, `1=fake`

To recompute AUROC with MSP for already-trained detectors (rewrites `metrics.json` / CSVs):

```bash
python recompute_auroc_msp.py --gpus 0,1
# or single GPU:
python recompute_auroc_msp.py --device cuda:0
```

## Outputs

```text
runs/<generator>/
  detector.pt
  train_log.json
  metrics.json
  metrics.csv
runs/summary_auroc.csv
runs/summary_acc.csv
```

Resume-safe:
- If `runs/<gen>/detector.pt` exists → **training is skipped**
- If `runs/<gen>/metrics.json` also exists → **eval is skipped** (job fully done)
- If only the detector exists → eval-only resume
- Use `--force` to retrain + re-eval from scratch

## Memory / workers

- Default `--batch-size 8` + `--amp bf16` for ViT-H. Raise batch if VRAM allows.
- `--num-workers 2` with `persistent_workers=False`; workers are shut down after every job.
- Optional `--freeze-backbone` for a linear probe (much lighter).

## Layout

| File | Role |
| --- | --- |
| `config.py` | paths + hyperparameters |
| `data.py` | train/eval loaders |
| `model.py` | I-JEPA binary detector |
| `train.py` | train + cross-generator eval |
| `cleanup.py` | DataLoader / CUDA teardown |
| `run_all.py` | multi-GPU job queue |
| `upload_ijepa_hf.py` | upload this folder (code + `runs/`) to Hugging Face |

### Upload to Hugging Face

```bash
cd /home/themis/workspace_firas/appli/i-jepa
export HF_TOKEN=hf_...
python upload_ijepa_hf.py --repo-id fira7s/ijepa
# private / dataset / resumable workers:
python upload_ijepa_hf.py --repo-id fira7s/ijepa --private --num-workers 4
```