Hugging Face's logo

SpeechAntiSpoofingBenchmarks
/
DPDA-GA-Mamba

PyTorch
audio
anti-spoofing
deepfake-detection
speech-anti-spoofing-arena
Model card Files
xet
 Community
DPDA-GA-Mamba / implementation-notes.md
korallll's picture
korallll
Wrapper code (speech-spoof-bench)
63ca006 verified
preview code
|
Raw History Blame Contribute Delete
5.89 kB

DPDA-GA-Mamba — implementation notes (2026-10-30)

Same network code is shipped in three self-contained dirs: XLSR-Mamba-LA/, XLSR-Mamba-DF/, DPDA-GA-Mamba/ (_net.py and _mamba_v1.py byte-identical; only the wrapper module / class / checkpoint differ). Not to be confused with XLSR-MamBo/ (arXiv 2601.02944, a different model).

Source

  • Code: github.com/swagshaw/XLSR-Mamba (MIT) model.py + mamba_blocks.py. DPDA-GA nets/mamba.py + nets/mamba_blocks.py are byte-identical (md5 checked).
  • Training recipe: DPDA-GA (Truong et al., ICASSP 2026, arXiv 2509.20682; github.com/ductuantruong/dpda_ga, MIT): dual-path data-augmented training (each batch = clean + RawBoost-augmented copies) with PCGrad gradient alignment between the two paths' losses, on the XLSR-Mamba architecture.
  • Checkpoints: HF tuantd/dpda_ga @ 531554e0a1c28d6a6ccf22a355281e3d3acefeb5 pcgrad/xlsr_mamba/best/best_{0..4}.pth (565 tensors each, same keys as XLSR-Mamba).
  • Weight average (make_avg5.py -> dpda_ga_mamba_avg5.pth) — maintainer decision 2026-10-30: reproduce main.py (--average_model True --n_average_model 5) EXACTLY, bug included. main.py does sd = model.state_dict() after loading best_0; state_dict() tensors share storage with the parameters, so the next load_state_dict(best_1) overwrites sd in place before the first sd[key] + sd2[key] (which then rebinds sd to fresh tensors). Result: (2·best_1 + best_2 + best_3 + best_4) / 5, best_0 dropped (verified on torch 2.7 with a toy Linear: weights 0..4 -> 2.2, not 2.0; and on the real file, max |diff| to that formula 1e-7). This is what the authors evaluated. Float tensors summed in float64, stored float32; first_bn.num_batches_tracked rounded (unused at eval). make_avg5.py --truemean gives the plain 5-way mean instead (not shipped).
  • Hyper-parameters (main.py defaults, confirmed by the checkpoint shapes): emb_size 144, num_encoders 12 -> 6 forward + 6 backward Mamba blocks; ssm_cfg {} -> d_state 16, d_conv 4, expand 2 (d_inner 288, dt_rank 9); rms_norm, fused_add_norm, residual_in_fp32 all True.

Parameter count

319,327,542 parameters + 3 BatchNorm buffer elements = 319,327,545 state_dict elements (the 319,327,545 in the brief / HF header counts the buffers). params_millions: 319.3275.

Edits to the source (all inference-only; listed in _net.py header)

  1. XLS-R built from xlsr2_300m.pt cfg via merge_with_parent (W2V2-AASIST pattern; symlink to ../W2V2-AASIST/xlsr2_300m.pt); the checkpoint overwrites every weight (strict load).
  2. extract_feat: removed self.model.train() (would re-enable dropout on a device/dtype mismatch) and the duplicate XLS-R forward whose layer_results was discarded.
  3. Model(emb_size, num_encoders) instead of argparse args.
  4. Kernels: mamba_ssm / causal_conv1d are not installed, and Triton is unreliable on the RTX 5060 Ti (sm_120; XLSR-MamBo hit "computeCapability not supported"). _mamba_v1.py vendors the needed mamba_ssm v1 subset (from benchmarks/RawBMamba/mamba_ssm 1.1.3.post1) as ONE pure-PyTorch file (not a mamba_ssm package -> no sys.modules clash with other dirs): Mamba (slow path: depthwise causal conv1d + SiLU, x_proj/dt_proj, selective scan, z-gating; == the fused mamba_inner_fn math), Block, RMSNorm, fused add+RMSNorm (rms_norm_fn, rms_norm_ref semantics incl. prenorm/residual_in_fp32). Pure torch is used for every score on every card.
  5. Selective scan = selective_scan_chunked: exact closed form of the recurrence in chunks of 8 (intra-chunk masked-exp sums + inter-chunk sums over chunk totals; every exp argument is a masked difference <= 0 -> no overflow). No Python loop -> fast eager and a compact ONNX graph. Checked against the sequential selective_scan_ref (kept in the file): random inputs max |diff| 2e-4 on |y|~128 (rel 2e-6); full-model scores within 1e-4 (test_chunked_scan_matches_reference).
  6. Only the released path of MixerModel is ported (bidirectional, fused_add_norm). Its quirk is preserved verbatim: the final fused add+norm of each direction adds the stack input hidden_states to that direction's residual (the last mixer output is not added), and the backward branch's residual stays in flipped time order. BiBlock/BiMixerModel (unused) dropped.

Inference conventions

  • First 66,800 samples; shorter clips repeat-tiled (source utils.pad). No resampling, no normalisation, no RawBoost. FP32. Score = logits[:, 1] (bona fide = class 1, source batch_score = batch_out[:, 1]). .to(device) then .eval().

Tests (test_dpda_ga_mamba.py)

pad truncate / repeat, strict load + eval mode + param/buffer count, batched == single (rtol=atol=1e-4), chunked scan == reference scan on the full net, determinism, finite.

Results (2026-10-30; GPUs shared with other agents' jobs, so throughputs are lower bounds)

  • PyTorch sweep (sweep.py): RTX 4070 Ti SUPER ~27 utt/s (plateau bs 8-32) -> batch_size = 16; RTX 5060 Ti ~15 utt/s at bs 2-4 -> 4 (set in load()).
  • TensorRT (trt_<module>.py, sidecar trt_<slug>.json): FP16 failed the ranking check on both cards -> FP32 engines (TF32 on). best_batch / utt/s: RTX 4070 Ti SUPER 32 / 33.9; RTX 5060 Ti 32 / 31.6. Build parity (64 InTheWild clips) Spearman >= 0.9999 on both cards.
  • Full J-SPAW_LA (2,397 trials): PyTorch EER 20.000% (results_pt/); TRT RTX 4070 Ti SUPER 20.038 (Spearman 1.0000) (results_trt/); TRT RTX 5060 Ti 20.100 (Spearman 1.0000) (results_trt3/). All |dEER| <= 0.10 -> both engines accepted.
  • Unit tests: 6/6 pass on both cards. Batched-vs-single uses rtol=atol=1e-4: cuDNN chooses batch-dependent conv algorithms (TF32 on by default), giving ~4e-5 relative jitter on logits ~4.
  • setup_lane.sh (in XLSR-Mamba-LA/) is the driver that produced all of the above.