# Fake-Mamba — implementation notes

Arena wrapper for **Fake-Mamba** (PN-BiMamba, Xuan et al., IEEE ASRU 2025, arXiv 2508.09294),
built 2026-10-30 under the 6-model Mamba batch plan
(`speech-spoof-bench/docs/plans/2026-10-29-mamba-batch-arena-submission.md`).

## Source
- Code: https://github.com/xuanxixi/Fake-Mamba @ `e966d2f317b187f474d8b2d54b7e0f9d0a124ba0` — **no LICENSE**.
- Checkpoint: HF `xxuan-speech/fake-mamba-models` @ `cd99e0efb4d4f3f6dfe610ae8e27de9a73d6bf41`,
  `models/Exp65/best.pth` (1.29 GB). Local: `fake_mamba_exp65_best.pth` → symlink to
  `../_ckpt_dl/xxuan-speech__fake-mamba-models/models/Exp65/best.pth`. **Not re-hosted** (no license);
  `meta.yaml` `checkpoint` is the commit-pinned authors' URL.
- `best.pth` is byte-identical (md5 4601685f…) to `best/best_0.pth` (lowest LA-dev loss).
- 12 PN-BiMamba blocks (`encoder.layers.0-11`), paper/source default is 7 → built with `depth=12`.
- XLS-R arch: `xlsr2_300m_cfg.json` = the stored model cfg of fairseq `xlsr2_300m.pt`
  (symlink `xlsr2_300m.pt` → `../W2V2-AASIST/xlsr2_300m.pt`, used only to extract it), merged onto
  the installed `Wav2Vec2Config` (`merge_with_parent`, cross-version fix from W2V2-AASIST). The 3.6 GB
  base checkpoint is not loaded at runtime; the fine-tuned checkpoint overwrites every weight.

## Port (`_net.py`)
- Only `Fake_Mamba` + `PN_BiMambas(_Encoder)` ported; the missing modules imported by the source
  (`conformer.ConformerBlock/BiMambaEncoder`, `ConExBiMamba.ExBiMamba`) belong to other back-ends.
- Mamba = mamba_ssm 1.x `Mamba` (d_model 144, d_state 16, d_conv 4, expand 2; checkpoint keys match),
  non-fast-path branch (nn.Conv1d + selective scan). No RMSNorm anywhere (LayerNorm only).
- Selective scan: `selective_scan_parallel` — chunked (8) Hillis-Steele prefix scan + 26-step
  inter-chunk carry; equals the library's sequential `selective_scan_ref` (kept, `FAKE_MAMBA_SCAN=ref`)
  to ~4e-6 on encoder outputs, 2e-7 on scores. ~3.6× faster than the 208-step loop.
- Forward + time-flipped sequence go through the shared Mamba as one batch (same math).
- `extract_feat`'s `self.model.train()` on device mismatch removed; wrapper does `.to()` then `.eval()`.

## Wrapper (`fake_mamba.py:FakeMamba`)
- First 66,800 samples, repeat-tiled if shorter (`utils.pad`), no resampling/normalisation, FP32.
- Score = `logits[:, 1]` (source labels bonafide = 1; eval scripts write `out[:, 1]`).
- params: 321,290,117 (321.2901 M).

## Parity with the authors' In-the-Wild scores (`ref_scores/…/best/SITW/score.txt`, label 1 = spoof)
- Authors' file → EER 5.849% (paper 5.85%).
- Our `best.pth` on the full arena In-the-Wild parquet (31,779): **EER 5.97%**, per-clip Pearson
  r = 0.989 vs the authors' scores.
- The authors' `SITW-DF_test.py` defaults to `--average_model True --n_average_model 5`
  (weight-average of `best/best_0..4.pth`). Averaging those 5 files moves us to r = 0.999 vs their
  file, full-ITW EER 5.82% (median |Δ| 0.13 logit; not exact — their best/ set or environment differed slightly).
  Checked on 256/128 clips; the 16-bit FLAC arena audio vs the original float wavs makes no
  difference (r identical). CPU vs GPU, TF32 on/off, ref vs parallel scan: identical.
- Decision (plan): ship the single `best.pth` (Exp65) as approved; avg-5 noted as an option.

## Results (2026-10-30; both GPUs were shared with other unlocked jobs, so throughput is a lower bound)
| | RTX 4070 Ti SUPER (idx 2) | RTX 5060 Ti (idx 3) |
|---|---|---|
| PyTorch sweep (bs 4) | 34.6 utt/s | 17.2 utt/s |
| TRT FP32 (FP16 Spearman 0.998 < 0.9999 → FP32) | 60.2 utt/s @24 | 38.8 utt/s @16 |
| TRT build parity (64 ITW clips) | Spearman 0.99995 | 0.99995 |
| J-SPAW_LA EER, TRT vs PyTorch 15.000 | 15.000 (Δ0), Spearman 0.999996 | 14.966 (Δ0.034), Spearman 0.999996 |

PyTorch J-SPAW_LA run (results_pt) was on the RTX 5060 Ti. Both engines are accepted.
- 2026-10-30 (maintainer): weights switched to the authors' exact eval average (2*b1+b2+b3+b4)/5, averaged at load time from the pinned HF files; ITW r=0.99996 vs authors' scores. Old best.pth outputs in old_bestpth/.